ASPECTS OF THETHEORY OF SYNTA ASPECTS OF THE THEORY OF SYNTAX Noam Chomsky 1111111 THE M.LT. PRESS Massachusetts Institute of Technolog Cambridge, Massachusetts ACKNOWLEDGMENT This is Specal Tecnical Report Number II of the Researc Labra tory of Electrnics of the Masschusetts Institute of Tecnology. The Researc Lbratory of Electonics is an interdepartment labrator in which facult members and graduate student from numer ous acdemic departments conduct researh. The research reported in ts document was made possible in part by support extended the Massachustts Institute of Technology. Resarch Laboratory of Electronics. by te JOINT SVICES ELCONICS PRORMS (U.S. Army. U.S. Navy. and U.S. Air Force) under Contract No. DAS6 o39-AMC-03l(E); additional supprt was received from the U.S. Air Force (Electronic Systems Division under Contract AFI9(628)-2487). te National Science Foundation (Grant GP-2495). the National Insti tutes of Health (Grant MH-D4737-D4). and The National Aeronautics and Space Administration (Grant NsG-496). Reproduction in whole or in part is peritted for any purpose of the United States Gverment. Copright @ I965 by The Masachwetts Intitute Of Technolog All Right Resered Librar of Congress Catalog Card Number: 65-I9080 Printed in the United States of America Preace The idea that a language is based on a system of rules deter mining the interpretation of its infnitely many sentences is by no means novel. Well over a century ago. it was expressed with reasonable clarity by Wilhelm von Humbold� in his famous but rarely studied introduction to general linguistics (Humboldt. 18 3 6). His view that a language "makes infnite use of fnite means" and that its grammar must describe te processes that make tis possible is. furtherore. an outgrowth of a persistent concer. within rationalistic philosophy of language and mind, with this "creative" aspct of languag use (for discussion, see Chomsky. 196 4 . forthcoming). What is more, it seems tat even Panini's grammar can be interpreted as a fragment of such a "gnerative grammar," in esentially the contemporary sense of this term. Nevertheless, within modem linguistics. it is ciefy within the last few years that fairly substantial attempts have been made to construct explicit generative grammars for particular lan guages and to explore their consequences. No great surprse should be occasioned by te extensive discussion and debate concering the proper forulation of the teory of generative grammar and the correct desciption of the languages that have been most intensively studied. The tentative charac�er of any conclusions that can now be advanced concerning lingistic theor, or, for that matter, English gammar, should certainy be obvious to anyone working in tis area. (I� is sufcient to v vi PREFACE consider the vast range of linguistic phenomena that have re sisted insightful forulation in any ters.) Still, it seems that certain fairly substantial conclusions are emerging and receiving continually increased support. In particular, the central role of grammatical transformations in any empirically adequate gen erative grammar seems to me to be established quite frly, though there remain many questions as to the proper form of the theory of transformational grammar . This monograph is an exploratory study of various problems that have arisen in the course of work on transformational gram mar, which is presupposed throughout as a general framework for the discussion. What is at issue here is precisely how this theory should be formulated. This study deals, then, wi th ques tions that are at the border of research in transforational gram mar. For some, defnite answers will be proposed; but more often the discussion will merely raise issues and consider pos sible approaches to them without reaching any defnite conclu sion. In Chapter 3, I shall sketch briefy what seems to me, in the light of this discussion, the most promising direction for the theory of generative grammar to take. But I should like to riter ate that this can be only a highly tentative proposal. The monograph is oranized in the following way . Chapter 1 sketches backgound assumptions. It contains little that is new, but aims only to summarize and to clarify certain points that are essential and that in some instances have been repeatedly misunderstood. Chapter 1 and 3 deal with a variety of defects in earlier versions of the theory of transformational grammar . The position discussed is that of Chomsky ( 1 957), Lees ( 1 960), and many others. These witers take the syntactic component of a transformational grammar to consist of a phrase structure grammar as its base, and a system of transforations that map structures generated by the base into actual sentences . This posi tion is restated briefy at the beginning of Chapter 3. Chapter 1 is concerned with the base of the syntactic component, and with difculties that aise from the assumption that it is, strictly speaking, a phrase structure grammar. Chapter 3 suggests a revi sion of the transformational component and its relation to base PRFAC vii structures. The notion of "grammatical transforation" itself is taken over without change (though with some simplifcations). In Chapter 4, various residual problems are raised, and discussed briefy and quite inconclusively. I should like to acknowledge wit. h gratitude the very helpful comments of many friends and colleagues who have taken the truble to read earlier versions of this manuscript. In particular, I am indebted to Morris Halle and Paul Postal, who have sug gested many valuable improvements, as well as to lerrold Katz, lames Mc C awley, George Miller, and G. H. Matthews; and to many students whose reactions and ideas when this material has been presented have led to quite substantial modifcations. The writing of this book was completed while I was at Harvard University, Center for Cognitive Studies. supported in part by Grant No. MH 05120-04 and -05 from the National Institutes of Health to Harard University. and in part by a fellowship of the American Council of Leared Societies. Cambridge, Massachuetts October Ig64 NOAM CHOMSKY Contents Preface v 1 Methodological Preliminaries 3 § 1. GENERATIVE GRAMMARS AS THEORIES OF LIN- GUISTIC COMPETENCE 3 § 2. TOWARD A THEORY OF PERFORMANCE 10 § 3. THE ORGANIZATION OF A GENERTIVE GRAMMAR 15 § 4. JUSTIFICATION OF GRAMMARS 18 § 5. FORMA AND SUBSTANTIVE UNIVERSALS 27 § 6. FURTHER REMARKS ON DESCRIPTIVE AND EX- PLANATORY THEORIES 30 § 7. ON EVALUATION PROCEDURES 37 § 8. LNGUISIC THEORY AND LNGUAGE LEARNING 4 7 § 9. GENERATIVE CAPACIT AND ITS LINGUISTIC REL- EVANCE 60 2 Categories and Relations in Syntactic Theory 6 3 § 1. THE SCOPE OF THE BASE 63 § 2. ASPEcS OF DEEP STRUcURE 64 §��.�� � § 2.2. Functional notions 68 ix x CONTNT § 2.J. Syntactic features 7 5 § 2.J.I. The problem 75 § 2.J.2. Some formal similarities between sntax and phonolog 79 § 2. J .J. General structure of the base component 84 § 2.J+ Context-sensitive subcategori%tion rules 9 § 3. AN ILLUSTRATIVE FRAGMENT OF THE BAE COM- PONENT 106 § 4. TYPES OF BASE RULES I I I § 4.1. Summar HI § 4.2. Selectional rules and grammatical relations HJ § 4.J. Further remarks on subcategoriztion rules 120 § 4+ The role of categorial rules I2J 3 Deep Structures and Grammatical Transformations I28 4 Some Residual Problems I48 § l. THE BOUNDARIES OF SYNTAX AND SEMANTICS 148 § 1.1. Degrees of grammaticalness 148 § 1.2. Further remarks on selectional rules I5J § I.J. Some additional problems of semantic M� � § 2. THE STRUCTURE OF THE LEXICON 164 § 2.1. Redundancy 164 § 2.2. Infectional processes 170 § 2.J. Derivational processes 184 &� � Notes to Chapter 1 1 9 3 Notes to Chapter. 208 Notes to Chapter 3 222 Notes to Chapter 4 227 Bibliography 237 Index 2 4 7 ASPECTS OF THE THEORY OF SYNTAX I Methodological Preliminaries § 1. GENERATIVE GRAMMARS AS THEORIES OF LINGUISTIC COMPETENCE THIS study will touch on a variety of topics in syntactic theory and English syntax, a few in some detail, several quite suprfcially, and none exhaustively. It will b concered with the syntactic component of a gnerative gammar, that is, with te rules that specify the well-fored string of minimal sy tactically functioning units (omat ives) and assign structural inforation of various kinds both to these strings and to string that deviate from well-formedness in certain respects. The gneral framework within which this investigation will proceed has been presented in many places, and some familiarity with the theoretical and descriptive studies listed in the bibliog raphy is prsupposed. In this chapter, I shall survey briefy some of the main background assumptions, making no serious attempt her to justify them but only to sketch them clearly. Linguistic theory is concerned primarily with an ideal speaker listener, in a completely homogeneous speech-community, who knows it languag perfectly and is unafected by such grammati cally irrelevant conditions as memory limitations, distractions, shift of attention and interest, and errors (random or character istic) in applying his knowledge of the language in actual per formance. This seems to me to have been the position of the founder of modem gneral linguistics, and no cogent reason for § METHODOLOGICAL PRELIMINARIE modifying it has been ofered. To study actual linguistic per formance, we must consider the interaction of a variety of factors, of which the underlying competence of the speaker-hearer is only one. In this respect, study of language is no diferent from empirical investigation of other complex phenomena. We thus make a fundamental distinction between competence (the speaker-hearer's knowledg of his language) and performance (the actual use of language in concrete situations) . Only under the idealization set forth in the preceding paragraph is per formance a direct refection of competence. In actual fact, it obviously could not directly refect competence. A record of natural speech will show numerous false starts, deviations from rules, changes of plan in mid·course, and so on. The problem for the linguist, as well as for the child learing the language, is to determine from the data of performance the underlying system of rules that has been mastered by the speaker-hearer and that he puts to use in actual performance. Hence, in the technical sense, linguistic theory is mentalistic, since it is concerned wi th discovering a mental reality underlying actual behavior. l Ob served use of languag or hypothesized dispositions to respond, habits, and so on, may provide evidence as to the nature of this mental reali ty, but surely cannot constitute the actual subj ect matter of linguistics, if this is to be a serious discipline. The distinction I am noting here is related to the langue-parole distinction of Saussure; but it is necessary to reject his concept of langue as merely a systematic inventory of items and to return rather to the Humboldtian conception of underlying competence as a system of generative processes. For discussion, see Chomsky ( 1 964). A grammar of a languag purports to be a description of the ideal speaker-hearer's intrinsic competence. If the grammar is, furthermore, perfectly explicit -in other words, if it does not rely on the intellignce of the understanding reader but rather provides an explicit analysis of his contribution -we may (somewhat redundantly) call it a generative grammar. A fully adequate grammar must assign to each of an infnite range of sentences a structural description indicating how this § Ï+ GENERATIVE GRAMMARS AND LINGUISTIC COMPETENCE 5 sentence is understood by the ideal speaker-hearer. This is the traditional problem of descriptive linguistics, and tradi tional grammars give a wealth of information concerning structural descriptions of sentences. However, valuable as they obviously are, traditional grammars are defcient in that they leave un expressed many of the basic regularities of the language with which they are concerned. This fact is particularly clear on the level of syntax, where no traditional or structuralist grammar goes beyond classifcation of particular examples to the stage of formulation of gnerative rules on any signifcant scale. An analysis of the best existing grammars will quickly reveal that this is a defect of principle, not just a matter of empirical detail or logical preciseness. Nevertheless, it seems obvious that the attempt to explore this largely uncharted territory can most proftably begin with a study of the kind of structural information presented by traditional grammars and the kind of linguistic processes that have been exhibited, however informally, in these grammars. 2 The limitations of traditional and structuralist grammars should be clearly appreciated. Although such grammars may contain full and explicit list of exceptions and irregularities, they provide only examples and hints concering the regular and productive syntactic processes. Traditional linguistic theory was not unaware of this fact. For example, James Beattie ( 1 788) remarks that Languages, therefore, resemble men in this respect, that, though each has peculiarities, whereby it is distinguished from every other, yet all have certain qualities in common. The peculiarities of individual tongues are explained in teir respective grammars and dictionaries. Those things, that all languages have in common, or that are necessary to every language, are treated of in a science, which some have called Universal or Philosophical grammar. Somewhat earlier, Du Marsais defnes universal and particular grammar in the following way ( 1 729; quoted in Sahlin, 1 928, pp. 29-30) : 11 y a dans la grammaire des observations qui conviennent a toutes les langues; ces obserations forent ce qU'on appelle la gammaire 6 METHODOLOGICAL PRUMINARE generale: telles sont les remarques que ron a faites sur les sons articles, sur les lettres qui sont les signes de ces sons; sur la nature des mots, et sur les diferentes manieres dont ils doivent �tre ou arranges ou termines pour faire un sens. Outre ces obserations generales, il y en a qui ne sont propres qu'a une langue particuliere; et c'est ce qui forme les gam maires particulieres de chaque langue. Within traditional linguistic theory, furthermore, it was clearly understood that one of the qualities that all languages have in common is their "creative" aspect. Thus an essential property of language is that it provides the means for expressing indefnitely many thoughts and for reacting appropriately in an indefnite range of new situations (for references, cf. Chomsky, 1 964, forth coming). The grammar of a particular language, then, is to be supplemented by a universal grammar that accommodates the creative aspect of language use and expresses the deep-seated regularities which, being universal, are omitted from the grammar itself. Therefore it is quite proper for a grammar to discuss only exceptions and irregularities in any detail. It is only when supplemented by a universal grammar that the grammar of a language provides a full account of the speaker-hearer's competence. Modern linguistics, however, has not explicitly recogized the necessity for supplementing a "particular gammar" of a lan guage by a universal grammar if it is to achieve descriptive adequacy. It has, in fact, characteristically rejected the study of universal grammar as misguided; and, as noted before, it has not attempted to deal with the creative aspect of language use. It thus suggst no way to overcome the fundamental descriptive inadequacy of structuralist grammars. Another reason for the failure of traditional grammars, particular or universal, to attempt a precise statement of regular processes of sentence formation and sentence interpretation lay in the widely held belief that there is a "natural order of thoughts" that is mirrored by the order of words. Hence, the rules of sentence formation do not really belong to gammar but to some other subject in which the "order of thoughts" is studied. Thus in the Grammaire genera le et raisonnee (Lancelot § 1 . GENERATIVE GRAMMARS AND LINGUISTIC COMPETNCE 7 et al., 1 660) it is asserted that, aside from fgurative speech, the sequence of words follows an "ordre naturel," which conforms "a l' expression naturelle de nos pensees. " Consequently, few gram matical rules need be formulated beyond the rules of ellipsis, inversion, and so on, which determine the fgurative use of lan guage. The same view appears in many forms and variants. To mention j ust one additional example, in an interesting essay devoted largely to the question of how the simul taneous and sequential array of ideas is refected in the order of words, Diderot concludes that French is unique among languages in the degree to which the order of words corresponds to the natural order of thoughts and ideas (Diderot, 1 75 1 ). Thus "quel que soit l'ordre des termes dans une langue ancienne ou moderne, l' esprit de l'ecrivain a suivi 1'0rdre didactique de la syntaxe fran�aise" (p. 390) ; "Nous disons les choses en fran�ais, comme l' esprit est force de les considerer en quelque langue qu'on ecrive" (p. 37 1). With admirable consistency he goes on to conclude that "notre langue pedestre a sur les autres l' avantage de l' utile sur l' agreable" (p. 372); thus French is appropriate for the sciences, whereas Greek, Latin, Italian, and English "sont plus avanta geuses pour les lettres." Moreover, le bons sens choisirait la langue fran�aise; mais . . . l'imagination et les passions donneront la preference aux langues anciennes et a celIes d nos voisins . . . il faut parler fran�ais dans la societe et dans les «oles de philosophie; et grec, latin, anglais, dans les chaires et sur les teA tres; . . . notre langue sera ceIle de la verite, si jamais elle revient sur la tere; et . . . la grecque, la latine et les autres seront les langes de la fable et du mensonge. Le fran�ais est fait pour instruire, eclairer et con vaincre; le grec, le latin, l'italien, l'anglais, pour persuader, emouvoir et tromper: parlez grec, latin, italien au peuple; mais parlez fran�is ÜW sage. (pp. 37 1 -372) In any event, insofar as the order of words is determined by factors independent of language, it is not necessary to describe it in a particular or universal gammar, and we therefore have principled grounds for excluding an explicit formulation of syntactic processes from grammar. It is worth noting that this naive view of language structure persists to modem times in 8 METHODOLOGICAL PRELIMINARIES various forms, for example, in Saussure's image of a sequence of expressions corresponding to an amorphous sequence of concepts or in the common characterization of language use as merely a matter of use of words and phrases (for example, Ryle, 1 9 53 ) . But the fundamental reason for this inadequacy of traditional grammars is a more technical one. Although it was well under stood that linguistic processes are in some sense "creative," the technical devices for expressing a system of recursive processes were simply not available until much more recently. In fact, a real understanding of how a language can (in Humboldt's words) "make infni te use of fnite means" has developed only within the last thirty years, in the course of studies in the foundations of mathematics. Now that these insights are readily available it is possible to return to the problems that were raised, but not solved, in traditional linguistic theory, and to attempt an explicit formulation of the "creative" processes of language. There is, in short, no longer a technical barrier to the full-scale study of generative grammars. Returning to the main theme, by a generative grammar I mean simply a system of rules that in some explicit and well defned way assigns structural descriptions to sentences. Obviously, every speaker of a language has mastered and internalized a gen erative grammar that expresses his knowledge of his language. This is not to say that he is aware of the rules of the grammar or even that he can become aware of them, or that his statements about his intuitive knowledge of the language are necessarily accurate. Any interesting generative grammar will be dealing, for the most part, with mental processes that are far beyond the level of actual or even potenti a l consciousness; furthermore, it is quite apparent that a speaker's reports and viewpoints about his behavior and his competence may be in error. Thus a generative grammar attempts to specify what the speaker actually knows, not what he may report about his knowledge. Similarly, a theory of visual perception would attempt to account for what a person actually sees and the mechanisms that determine this rather than his statements about what he sees and why, though these state- § 1 . GENERATIVE GRAMMARS AND LINGUISTIC COMPETENCE 9 ments may provide useful, in fact, compelling evidence for such a theory. To avoid what has been a continuing misunderstanding, it is perhaps worth while to reiterate that a generative grammar is not a model for a speaker or a hearer. It attempts to characterize in the most neutral possible terms the knowledge of the language that provides the basis for actual use of language by a speaker hearer. When we speak of a grammar as generating a sentence wi th a certain structural description, we mean simply that the grammar assigns this structural description to the sentence. When we say that a sentence has a certain derivation with respect to a particular generative grammar, we say nothing about how the speaker or hearer might proceed, in some practical or efcient way, to construct such a derivation. These questions belong to the theory of language use -the theory of per formance. No doubt, a reasonable model of language use will incorporate, as a basic component, the generative grammar that expresses the speaker-hearer's knowledge of the language; but this generative grammar does not, in itself, prescribe the char acter or functioning of a perceptual model or a model of speech production. For various attempts to clarify this point, see Chomsky ( 1957), Gleason ( 1961), Miller and Chomsky ( 1 9 63), and many other publications. Confusion over this matter has been sufciently persistent to suggest that a terminological change might be in order. Never theless, I think that the term "generative grammar" is completely appropriate, and have therefore continued to use it. The term "generate" is familiar in the sense intended here in logic, particularly in Post's theory of combinatorial systems. Further more, "generate" seems to be the most appropriate translation for Humboldt's term erzeugen, which he frequently uses, it seems, in essentially the sense here intended. Since this use of the term "generate" is well established both in logic and in the tradition of linguistic theory, I can see no reason for a revision of terminology. 10 METHODOLOGICAL PREUMINARIES § 2. TOWARD A THEORY OF PERFORMANCE There seems to be li ttle reason to question the traditional view that investigation of performance will proceed only so far as understanding of underlying competence permits. Further more, recent work on performance seems to give new support to this assumption. To my knowledge, the only concrete results that have been achieved and the only clear suggestions that have been put forth concerning the theory of performance, outside of phonetics, have come from studies of performance models that incorporate generative grammars of specifc kinds -that is, from studies that have been based on assumptions about underlying competence. 3 In particular, there are some suggestive observations concerning limitations on performance imposed by organization of memory and bounds on memory, and concerning the ex ploitation of grammatical devices to form deviant sentences of various types. The latter question is one to which we shall return in Chapters 2 and 4. To clarify further the distinction between competence and performance, it may be useful to summarize briefy some of the suggestions and results that have appeared in the last few years in the study of performance models with limita tions of memory, time, and access. For the purposes of this discussion, let us use the term "ac ceptable" to refer to utterances that are perfectly natural and immediately comprehensible without paper-and-pencil analysis, and in no way bizarre or outlandish. Obviously, acceptability will be a matter of degree, along various dimensions. One could go on to propose various operational tests to specify the notion more precisely (for example, rapidi ty, correctness, and uniformity of recall and recognition, normalcy of intonation). " For present purposes, it is unnecessary to delimi t it more carefully. To illus trate, the sentences of ( Ï ) are somewhat more acceptable, in the intended sense, than those of (2) : ( Ï ) (i) I called up the man who wrote the book that you told me about (ii) quite a few of the students who you met who come from New York are friends of mine § l. TOWARD A THEORY OF PERFORMANCE 11 (iii) John, Bill, Tom, and several of their friends visited u last night (l) (i) I called the man who wrote the book that you told me about up (ii) the man who the boy who the students recognized pointed out is a friend of mine The more acceptable sentences are those that are more likely to be produced, more easily understood, less clumsy, and in some sense more natural.1 The unacceptable sentences one would tend to avoid and replace by more acceptable variants, wherever possible, in actual discourse. The notion "acceptable" is not to be confused with "gram matical." Acceptability is a concept that belongs to the study of performance, whereas grammaticalness · belongs to the study of competence. The sentences of (l) are low on the scale of ac ceptability but high on the scale of grammaticalness, in the technical sense of this term. That is, the generative rules of the language assign an interpretation to them in exactly the way in which they assign an interpretation to the somewhat more ac ceptable sentences of ( 1 ) . Like acceptability, grammaticalness is, no doubt, a matter of degree (cf. Chomsky, 1 955, 1 957, 1 961 ), but the scales of grammaticalness and acceptability do not coincide. Grammaticalness is only one of many factors that interact to determine acceptability. Correspondingly, although one might propose various operational tests for acceptability, it is unlikely that a necessary and sufcient operational criterion might be invented for the much more abstract and· far more important notion of grammaticalness. The unacceptable grammatical sen tences often cannot be used, for reasons having to do, not with grammar, but rather with memory limitations, intonational and stylistic factors, "iconic" elements of discourse (for example, a tendency to place logical subject and object early rather than late; cf. note 31, Chapter l, and note 9, Chapter 3 ) , and so on. Note that it would be quite impossible to characterize the un acceptable sentences in grammatical terms. For example, we can not forulate particular rules of the gammar in such a way a METHODOLOGICAL PRELIMINARIE to exclude them. Nor, obviously, can we exclude them by limiting the number of reapplications of grammatical rules in the gen eration of a sentence, since unacceptability can j ust as well arise from application of distinct rules, each being applied only once. In fact, it is clear that we can characterize unacceptable sentences only in terms of some "global" property of derivations and te structures they defne -a property that is attributable, not to a particular rule, but rather to the way in which the rules inter relate in a derivation. This observation suggests that the study of performance could proftably begin with an investigation of the acceptability of the simplest formal structures in grammatical sentences. The most obvious formal property of utterances is their bracketing into constituents of various types, that is, the "tree structure" as sociated with them. Among such structures we can distinguish various kinds -for example, those to which we give the follow ing conventional technical names, for the purposes of this discussion: (3) (i) nested constructions (ii) self-embedded constructions (iii) multiple-branching constructions (iv) left-branching constructions (v) right-branching constructions The phrases A and B form a nested construction if A falls totally within B, wi th some non null element to its left within B and some nonnull element to its right within B . Thus the phrase "the man who wrote the book that you told me about" is nested in the phrase "called the man who wrote the book that you told me about up, " in (2i). The phrase A is self-embedded in B if A is nested in B and, furthermore, A is a phrase of the same type as B . Thus "who the students recognized" is self-embedded in "who the boy who the students recognized pointed out," in (2ii), since both are relative clauses. Thus nesting has to do with bracketing, and self-embedding with labeling of brackets as well. A multiple-branching construction is one with no interal structure. In (liii), the Subject Noun Phrase is multiple-branch- § 2. TOWARD A THEORY OF PERFORMANCE ing, since "John," "Bill," "Tom," and "several of their friends" are its immediate constituents, and have no further association among themselves. In terms of bracketing, a multiple-branching construction has the form [[A][B]·· . [M]] . A left-branching struc ture is of the form [[[ . . . ] . . . ] . . . ] -for example, in English, such indefnitely i terable structures as [[[[ ohn]' s brother], s father], s uncle] or [[[the man who you met] from Boston] who was on the train], or (Iii), which combines several kinds of left-branching. Right-branching structures are those with the opposite prop erty -for example, the Direct-Object of (Ii) or [this is [the cat that caught [the rat that stole the cheese]]] . The efect of these superfcial aspects of sentence structure on performance has been a topic of study since almost the very inception of recent work on generative grammar, and there are some suggestive observations concerning their role in determin ing acceptability (that is, their role in limiting performance). Summarizing this work briefy, the following observations seem plausible: (4) (i) repeated nesting contributes to unacceptability (ii) self-embedding contributes still more radically to unac ceptabili ty (iii) multiple-branching constructions are optimal in accepta bility (iv) nesting of a long and complex element reduces accepta bility (v) there are no clear examples of unacceptability involving only left-branching or only right-branching, although these constructions are unnatural in other ways -thus, for example, in reading the right-branching construction "this is the cat that caught the rat that stole the cheese," the intonation breaks are ordinarily inserted in the wrong places (that is, after "cat" and "rat, " instead of where the main brackets appear) In some measure, these phenomena are easily explained . Thus it is known (cf . Chomsky, 1 959a; and for discussion, Chomsky, 1 96 1 , and Miller and Chomsky, 1 96 3 ) that an optimal perceptual METHODLOGICAL PRELIMINARIE device. even with a bounded memory. can accept unbounded left-brnching and right-branching structures. though nested (enc ultimately self-embedded) structures g beyond it memory capacity. Thus case (4i) is simply a consequence of fniteness of memory. and te unacceptability of such examples as (aii) raises no problem. If (4ii) is correct.a then we have evidence for a conclusion about oranization of memory that goes beyond the trviality that it must be fnite i size. An optimal fnite perceptual device of the ty discussed in Chomsky (19594) need have no mor difculty with self-embedding than with other kind of nesting (see Bar HilIel. Kaher. and Shamir. 1963. for a discussion of this point). To account for the greater unacceptability of self-embeddng (assuming this to be a fact), we must add other conditions on the perceptual device beyond mere limitation of memory. We might aSsume, for example, that the perceptual device has a stock of analytic procedures available to it, one corsponding to each kind of phrase, and that it is orgnized in such a way that it is unable (or fnds it difcult) to utilize a procedure r while it is in the coure of executing r. This is not a necessary feature of a prceptual model. but it is Ü rather plausible one. and it would account for (4ii). See, in this connection. Miller and Isard (1964). The high acceptability of multiple-branching, as in cae (4iii), is easily explained on the rather plausible assumption tat the ratio of number of phrases to number of formatives (the node-to terminal node ratio, in a tree-diagram of a sentence) is a rough measure of the amount of computation that has to be performed in analyis. Thu multiple coordination would be the simplest kind of construction for an analytic device -it would impose the least strain on memory.7 For discusion. see Miller and Chomsky (1963). Cae (4iv) sugests decay of memory, -perhaps, but raises un solved problems (see ChoIky, 1961, note 19). Case (4v) follows fom the rsult about optimal perceptual models mentioned earlier. But it is unclear why left- and right braching structures should become unnatural after a certain pint, if they actually do.8 § 3. TE ORGAIZTON OF A GENETVE GRMAR One mÌght a8K whether attention to les suprcial aspct of grammaticl structur tan to of (3) could lead to somewhat deepr conclusions aDUt perforance D00CÌ8. TDi seem entirely pssible. For example, in Miller ad Chomsky ( 1 9 63) 8Dme sytactic and prceptual considerations are adduced in support of a 8UC8lÍDD [WDÍO Ì8, to be sure, highly spculative) a8 to the somewhat more detailed organization of a prceptual device. In gneral, Ít sems that the study of prforance models incDrpDIating gnerative grammar may b a fruitful study; furthermore, it i difcult to iDagine any other basis on whic a theory of prfDrmance might develop. There Da8 been a fair amout of citic8D of work in generatve gamar on the _IDUDO8 that it sligDts study of perforaDce in favor of study of underlying comptence. The fact, howevery seem to be that the only studies of prforance, outide of phDnetic (ut see note 3 ) , are those carried out as a by-prouct of work in generative grammar. I particular, the study of memor limitations just 8ummaized ad te study of deviation from IU¡C8, a8 a 8t y li8tic device, to WUÍO we retur in Chapter 2 and 4, have developed in this way. Furtherore, it seems that these, lines of investigtiDn can provide some insight into per forance. CDn8equent¡y, tDÍ8 citicism i unwarantedj ad, furtherore¿ completely misdirected. It is te desciptivst limitation-in-principle to classifcation and organization of data, to "extacting patterD8 from a corpus of observed speech, tD descbing speeO habits" or "habit structures," insofar a8 these may exist, etc., that precludes the development of a theory of actual perforance. § 3. THE ORGANIZATION OF A GENERATIVE GRAMMAR Returing now to te question of competence and the gen erative gmmars that purprt to describe it, we stss again that knowledg of a langag involves the implicit ability to ud stand idefnitely many sentences.8 Hence, a generative gramma Dut b a system of 1UÌC8 that can iterte to gnerate an in- METHODOLOGICAL PRELIMINARIES defnitely large number of structures. This system of rules can be analyzed into the three major components of a generative grammar: the syntactic, phonological, and semantic com ponents.10 The syntactic component specifes an infnite set of abstract formal objects, each of which incorporates all information relevant to a single interpretation of a particular sentence. ll Since I shall be concerned here only with the syntactic com ponent, I shall use the term "sentence" to refer to strings of formatives rather than to strings of phones. It will be recalled that a string of formatives specifes a string of phones uniquely (up to free variation), but not conversely. The phonological component of a grammar determines the phonetic form of a sentence generated by the syntactic rules. That is, it relates a structure generated by the syntactic com ponent to a phonetically represented signal. The semantic com ponent determines the semantic interpretation of a sentence. That is, it relates a structure generated by the syntactic com ponent to a certain semantic representation. Both the phono logical and semantic components are therefore purely inter pretive. Each utilizes information provided by the syntactic component concering formatives, their inherent properties, and their interrelations in a given sentence. Consequently, the syn tactic component of a grammar must specify, for each sentence, a deep structure that determines its semantic interpretation and a surface structure that determines its phonetic interpretation. The frst of these is interpreted by the semantic component; the second, by the phonological component. 1 2 It might be supposed that surface structure and deep structure will always be identical. In fact, one might briefy characterize the syntactic theories that have arisen in modern structural (taxonomic) linguistics as based on the assumption that deep and surface structures are actually the same (cf. Postal, Ig64a, Chomsky, Ig64) . The central idea of transformational grammar is that they are, in general, distinct and that the surface structure is deter mined by repeated application of certain formal operations called "grammatical transformations" to objects of a more § 3. THE ORGANIZATION OF A GENERATIVE GRAMMAR elementary sort. If this is true (as I assume, henceforth), then the syntactic component must generate deep and surface structures, for each sentence, and must interrelate them. This idea has been clarifed substantially in recent work, in ways that will be described later. In Chapter 3, I shall present a specifc and, in part, new proposal as to precisely how it should be formulated. For the moment, it is sufcient to observe that although the Immediate Constituent analysis (labeled bracketing) of an actual string of formatives may be adequate as an account of surface structure, it is certainly not adequate as an account of deep structure. My concern in this book is primarily with deep struc ture and, in particular, with the elementary objects of which deep structure is constituted. To clarify exposition, I shall use the following terminology, with occasional revisions as the discussion proceeds. The base of the syntactic component is a system of rules that generate a highly restricted (perhaps fnite) set of basic strings, each with an associated structural description called a base Phrase-marker . These base Phrase-markers are the elementary units of which deep structures are constituted. I shall assume that no ambiguity is introduced by rules of the base. This assumption seems to me correct, but has no important conse quences for what follows here, though it simplifes exposition. Underlying each sentence of the language there is a sequence of base Phrase-markers, each generated by the base of the syntactic component. I shall refer to this sequence as the basis of the sentence that it underlies. In addition to its base, the syntactic component of a generative grammar contains a transformational sub component. This is concerned with generating a sentence, with i ts surface structure, from its basis. Some familiarity with the operation and efects of transformational rules is henceforth presupposed. Since the base generates only a restricted set of base Phrase markers, most sentences will have a sequence of such objects as an underlying basis. Among the sentences with a single base Phrase-marker as basis, we can delimit a proper subset called "kernel sentences. " These are sentences of a particularly simple METODOLICL PRUMINA sort that involve a minimum of transformational apparatus in their generation. The notion "kerel sentence" has, I think, an important intuitive signifcance, but since kerel sentences play no distinctive role in generation or interpretation of sentences, I shall say nothing more about them here. One must be careful not to confuse kernel sentences with the basic string that under lie them. The basic strings and base Phrase-marker do, it seems, play a distinctive and crucial role in language use. Since transformations will not be considered her in detail, no careful distinction will be made, in the case of a sentence with a single element in it basis, between the basic string underlying this sentence and the sentence itself. In other words, at many points in the exposition I shall make the tacit simplifying (and contrary-to-fact) assumption that the underlying basic string is the sentence, in this case, and that the base Phrase-marker is the surface structure as well as the deep structure. I shall try to select examples in such a way as to minimize possible confusion, but the simplifying assumption should be bore in mind through out. § 4. JUSTIFICATION OF GRAMMARS Before entering directly into an investigation of the syntactic component of a generative grammar, it is important to give some thought to several methodological questions of justifcation and adequacy. There is, frst of all, the question of how one is to obtain information about the speaker-hearer's competence, about his knowledge of the language. Like most facts of interest and importance, this is neither presented for direct obseration nor extractable from data by inductive procedures of any known sort. Clearly, the actual data of linguistic performance will provide much evidence for determining the correctness of hypotheses about underlying linguistic structure, along with introspective reports (by the native speaker, or the linguist who has learned the language) . This is the position that is universally adopted in practice, although there are methodological discus- § 4 . JUSTIFICATION OF GRAMMARS sions that seem to imply a reluctance to use obsered perfor ance or introspective reports a evidence for some underlying reality. In brief, it is unfortunately the case that no adequate for malizable techniques are known for obtaining reliable informa tion concerning the facts of linguistic structure (nor is this particularly surprising). There are, in other words, very few reliable experimental or data-processing procedures for obtaining signifcant information concerning the linguistic intuition of the native speaker. It is important to bear in mind that when an operational procedure is pl'Oposed, it must be tested for adequacy (exactly as a theory of linguistic intui tion -a grammar -must be tested for adequacy) by measuring it against the standard provided by the tacit knowledge that it attempts to specify and describe_ Thus a proposed operational test for, say, segmenta tion into words, must meet the empirical condition of conform ing, in a mass of crucial and clear cases, to the linguistic intuition of the native speaker concerning such elements. Otherwise, it is without value. The same, obviously, is true in the case of any proposed operational procedure or any proposed grammatical description. If operational procedures were available that met this test, we might be justifed in relying on their results in unclear and difcult cases. This remains a hope for the future rather than a present reality, however. This is the objective situa tion of present-day linguistic work; allusions to presumably well known "procedures of elicitation" or "objective methods" simply obscure the actual situation in which linguistic work must, for the present, proceed. Furthermore, there is no reason to expect that reliable operational criteria for the deeper and more important theoretical notions of linguistics (such as "gram maticalness" and "paraphrase") will ever be forthcoming. Even though few reliable operational procedures have been developed, the theoretical (that is, grammatical) investigation of the knowledge of the native speaker can proceed perfectly well. The critical problem for grammatical theory today is not a paucity of evidence but rather the inadequacy of present theories of languag to account for masses of evidence that are hardly 20 METHODOLOGICAL PRELIMINARIES open to serious question. The problem for te grammarian is to construct a description and, where possible, an explanation for the enormous mas of unquestionable data concerning the linguistic intuition of the native speaker (often, himself); the problem for one concerned with operational procedures is to develop tests that give the correct results and make relevant distinctions. Neither"the study of grammar nor the attempt to develop useful test is hampred by lack of evidence with which to check result, for the present . We may hLpe that these eforts will converge, but they must obviously converge on the tacit knowledge of the native speaker if they are to be of any signifcance. One may ask whether the necessity for present·day linguistics to give such priority to introspective evidence and to the linguistic intuition of the native spaker excludes it from the domain of science. The answer to this essentially terminological question seems to have no bearing at all on any serious issue. At most, it determines how we shall denote the kind of research that can be efectively carried out in the present state of our tecnique and understanding. However, this terminological question actually does relate to a diferent issue of some interest, namely the question whether the important feature of the scs ful sciences has been their search for insigt or tei cncer for objectivity. The social and behavioral sciences provide ample evidence that objectivity can be pursued with little consequent gain in insight and understanding . On the other hand, a good case can be made for the view that the natural sciences have, by and large, sought objectivity primarily insofar as it is a tool for gaining insight (for providing phenomena that can suggest or test deeper explanatory hypotheses). In any event, at a given stage of investigation, one whose con cern is for insight and understanding (rather than for objectivity as a goal in itself must ask whether or to what extent a wider range and more exact description of phenomena is relevant to solving the problems that he faces. In linguistics, it seems to me that sharpning of the data by more objective tests i a matter of small importance for te" problems at hand. One who disagrees with this estimate of the present situation in linguistics can § 4 . JUSTIFICATION OF GRMMA u jutify his belief in the current importance of mor objective operational tests by showing how they can lead to new and deeper understanding of linguistic structure. Perhaps the day will come when the kinds of data that we now can obtain in abundance will be insufcient to resolve deeper questions con cering the structure of language. However, many questions that can realistically and sigifcantly be formulated today do not demand evidence of a kind that is unavailable or unattainable without signifcant improvements in obj ectivity of experimental technique. Although there is no way to avoid the traditional assumption that the speaker-hearer's linguistic intuition is the ultimate standard that deterines the accuracy of any proposed grammar, linguistic theory. or operational test, it must be emphasized, once again, that this tacit knowledge may very well not be immediately available to the user of the language . To eliminate what has seemed to some an air of paradox in this remark, let me illustrate with a few examples. If a sentence such as "Hying planes can b dangerous" is presented in an appropriately constructed context, the listener will interpret it immediately in a unique way, and will fail to detect the ambiguity. In fact, he may reject the second inter pretation, when this is pointed out to him, as forced or un natural (independently of which interpretation he originally selected under contextual pressure). Nevertheless, his intuitive knowledg of the language is clearly such that both of the inter pretations (coresponding to "Hying planes are dangrous" and "Hying planes is dangerous") are assigned to the sentence by the grammar he has interalized in some form . In the case just mentioned, the ambiguity may be fairly trans parent. But consider such a setence as (5) I had a book stolen Few hearrs may be aware of the fact that their interalized grammar in fact provides at least three structural descriptions for this sentence. Nevertheless, this fact can be brought to consciousness by consideration of slight elaborations of sentenc MODOLOICL PRUMNAIE ( 5 ), for example: (i) "I had a book stolen from my LÎ when I stupidly left the window open," tat is, "someone stole a book from my car"; (ii) "I had a bok stolen from his librar by a professional thief who I hired to do te job," that is, "I had some one steal a bok"; (iii ) "I almot had a book stolen, but tey caught me leaving the library with it," that is, "I had almot succeeded in stealing a bok." In bringing to consciousness te triple ambiguity of ( 5 ) in this way, we present no new informa tion to te hear and teach him noting new about his languag but simply arrang matters in such a way tat his linguistic intuition, previously obscured, becomes evident to him. A a fnal illusttion, consider the sentence (6) I peruaded John to leave (7) I expected John to leave The frst imprsion of the hearer may be tt tese sentences rceive the same structural analysis. Even fairly careful thought may fil to show him that his interalizd gramar assigns very diferent sytactic desciptions to these sentencs. In fact, so far as I have been able to discover, no English grammar has pointed out the fundamental distinction between tese to constructions (in particular, my ow sketche of English grammar in Chomky, I g 55 , 1 962a, failed to note this). However, it is clear tat te sentencs (6) and (7) are not parallel in structure. The difernc cn be brought out by consideration of the sentences (S) (i) I peruaded a specialist to examine John (ii) I peruaded Joh to be examined by a spcialist (9) (i) I expcted a spcialist to examine Joh (ii) I expected Joh to b examined by a spcialist The sentences ( gi) and (gii) are "cogitively synonymous" : one i true if and only if te other is te. But no variety of even weak paraphrae hold between (Si) and (8ii). Thus (Si) c b true or fl quite independendy of te truth or falsity of (Sii) . Wat ever difrenc of connotation or "topic" or emphasis one may fnd beteen ( gi) and (gii) is just te diference that exit b- § 4 . JUSTIFICATION OF GRMMAR tween the active sentence "a specialist will examine John" and it passive counterpart "John will be examined by a specialist." This is not at all the case with respect to (8), however. In fact, the underlying deep structure for (6) and (8ii) must show that "John" is the Direct-Object of the Verb Phrase a well as the grammatical Subject of the embedded sentence. Furtherore, in (8ii) "John" is the logical Direct-Object of the embedded sentence, whereas in (8i) the phrase "a specialist" is the Direct-Object of the Verb Phrase and the logical Subject of the embedded sentence. In ( 7 ) , (gi), and (gii), however, the Noun Phrases "John," "a specialist," and "John, " respectively, have no grammatical functions other than those that are i nternal to the embedded sentence; i n pa ticular, "John" is the logical Direct-Object and "a specialist" te loical Subject in the embedded sentences of (g). Thus the under lying deep structures for (8i), (8ii), (gi), and (gii) are, respectively, the following: 48 (1 0) (i) Noun Phrae - Verb - Noun Phrase - Sentence (I - persuaded - a specialist - a specialist will examine John) (ii) Noun Phrase - Verb - Noun Phrase - Sentence (I - persuaded - John - a specialit will examine John) ( 1 1 ) (i) Noun Phrae - Verb - Sentence (I - expected - a specialist will examine John) (ii) Noun Phrase - Verb - Sentence (I - expected - a specialist will examine John) In the case of ( I oii) and ( I l ii), the pasive transformation will apply to the embedded sentence, and in all four cases other operations will gve the fnal surface fors of (8) and (g) . The important point in the present connection is that (8i) difers from (8ii) in underlying structure, although (9i) and (gii) are essentially the same in underlying structure. This accounts for the diference in meaning. Notice, in support of this diference in analysis, that we can have "I persuaded John that (of the fact that) Sentence, " but not "I expected John that (of the fact that) Sentence. " METHODOLOGICAL PRELIMINARIE The example (6)-(7) serves to illustrate two important points. First, it shows how unrevealing surface structure may be as to underlying deep structure. Thus (6) and (7) are the same in surface structure, but very diferent in the deep structure that underlies them and determines their semantic interpretations. Second, it illustrates the elusiveness of the speaker's tacit knowl edge. Until such examples as (8) and (9) are adduced, it may not be in the least clear to a speaker of English that the grammar that he has internalized in fact assigns very diferent syntactic analyses to the superfcially analogous sentences (6) and (7). In short, we must be careful not to overlook the fact that surface similarities may hide underlying distinctions of a funda mental nature, and that it may be necessary to guide and draw out the speaker's intuition in perhaps fairly subtle ways before we can determine what is the actual character of his knowledge of his language or of anything else. Neither point is new (the former is a commonplace of traditional linguistic theory and analytic philosophy; the latter is as old as Plato's Meno) ; both are too often overlooked. A grammar can be regarded as a theory of a language; it is descriptively adequate to the extent that it correctly describes the intrinsic competence of the idealized native speaker. The struc tural descriptions assigned to sentences by the grammar, the distinctions that it makes between well-formed and deviant, and so on, must, for descriptive adequacy, correspond to the linguistic intuition of the native speaker (whether or not he may be immediately aware of this) in a substantial and signifcant class of crucial cases. A linguistic theory must contain a defnition of "grammar," that is, a specifcation of the class of potential grammars. We may, correspondingly, say that a linguistic theory is descriptively adequate if it makes a descriptively adequate grammar available for each natural language. Although even descriptive adequacy on a large scale is by no means easy to approach, it is crucial for the productive develop ment of linguistic theory that much higher goals than this be pursued. To facilitate the clear formulation of deeper questions, § 4. JUSTIFICATION OF GRAMMARS 25 it is useful to consider the abstract problem of constructing an "acquisition model" for language, that is, a theory of language learning or grammar construction. Clearly, a child who has learned a language has developed an internal representation of a system of rules that determine how sentences are to be formed, used, and understood. Using the term "grammar" with a sys tematic ambiguity (to refer, frst, to the native speaker's internally represented "theory of his language" and, second, to the linguist's account of this), we can say that the child has developed and internally represented a generative grammar, in the sense de scribed. He has done this on the basis of observation of what we may call primary linguistic data. This must include examples of linguistic performance that are taken to be well-formed sen tences, and may include also examples designated as non sentences, and no doubt much other information of the sort that is required for language learning, whatever this may be (see pp. 3 1 -32). On the basis of such data, the child constructs a grammar -that is, a theory of the language of which the well-formed sentences of the primary linguistic data constitute a small sample. 14 To learn a language, then, the child must have a method for devising an appropriate grammar, given primary linguistic data. As a precondition for language learning, he must possess, frst, a linguistic theory that specifes the form of the grammar of a possible human language, and, second, a strategy for selecting a grammar of the appropriate form that is com patible with the primary linguistic data. As a long-range task for general linguistics, we might set the problem of developing an account of this innate linguistic theory that provides the basis for language learing. (Note that we are again using the term "theory" -in this case "theory of language" rather than "theory of a particular language" -with a systematic ambiguity, to refer both to the child's innate predisposition to learn a language of a certain type and to the linguist'S account of this.) To the extent that a linguistic theory succeeds in selecting a descriptively adequate grammar on the basis of primary linguistic data, we can say that it meets the condition of explanatory ade quacy. That is, to this extent, it ofers an explanation for the METHODOLOGICAL PREUMINARIE intuition of te native speaker on the basis of an empirical hypothesis concering the innate predisposition of the child to develop a certain kind of theory to deal with the evidence presented to him. Any such hypothesis can be falsifed (all too easily, in actual fact) by showing that it fails to provide a descriptively adequate grammar for primary linguistic data from SOme oter language -evidently the child .is not pre disposed to learn one language rather than another. It is sup ported when it does provide an adequate explanation for some aspect of linguistic structure, an account of the way in which such knowledge might have been obtained. Clearly, it would b utopian to expect to achieve explanatory adequacy on a larg scale in the present state of linguistics. Nevertheless, considerations of explanator adequacy are often critical for advancing linguistic theory. Gross coverag of a larg mass of data can often be attained by conficting theories; for precisely this reason it is not, in itself, an achievement of any particular theoretical interest or importance. A in any other feld, the important problem in li nguistics is to discover a complex of data that diferentiates between conficting concep tions of linguistic structure in that one of these conficting theories can describe these data only by ad hoc means whereas the other can explain it on the basis of some empirical assump tion about the form of languag. Such small-scale studies of explanatory adequacy have, in fact, provided most of the evi dence tat has any serious bearing on the nature of linguistic structure. Thus whether we are comparing radically diferent theories of gammar or trying to determine the correctness of some particular aspect of one such theory, it is questions of explanatory adequacy that must, quite often, bear the burden of justifcation. This remark is in no way inconsistent with the fact that explanatory adequacy on a large scale is out of reach, for the present. It simply brings out te highly tentative character of any attempt to j ustify a empirical claim about linguistic structure. To summarize briefy, there are two respects in which one can speak of "jutifying a gnerative gammar. " On one level (that § 5. FORMAL AND SUBSTANTIVE UNIVERSAL of descriptive adequacy), the grammar is justifed to the extent that it correctly describes its object, namely the linguistic intui tion -the tacit comptence -of the native speaker. In this sense, the grammar is justifed on external gounds, on gounds of correspondence to linguistic fact. On a much deeper and hence much more rarely attainable level (that of explanatory adequacy), a gammar is justifed to the extent that it is a pincipled descrip tively adequate system, in that the linguistic theory with which it is associated selects this grammar over others, given primary linguistic data with which all are compatible. In this sense, the grammar is justifed on internal gounds, on grounds of its rela tion to a linguistic theory that constitutes an explanator hypoth esis about the form of language as such. The problem of internal justifcation -of explanatory adequacy -is essentially the problem of constructing a theory of languag acquisition, an account of the specifc innate abilities that make this achieve ment possible. § 5. FORMAL AND SUBSTANTIVE UNIVERSALS A theory of lingistic structure that aims for explanatory adequacy incorporates an account of linguistic universals, and it attributes tacit knowledg of these universals to the child. It proposes, then, that the child approaches the data with the presumption that they are drawn from a language of a certain antecedently well-defned type, his problem being to deterine which of the (humanly) possible languages is that of the com munity in which he is placed. Langage learing would be impossible unless this were the case. The important question is: What are the initial assumptions concering the nature of languag that the child bring to languag learning, and how detailed and specifc is the innate schema (the gneral defnition of "grammar") that gradually becomes more explicit and difer entiated a the child lea the langage? For the present we cannot come at all dose to making a hypothesis about innate schemata that is rich, detailed, and specifc enough to account for the fact of languag acquisition. Consequently, the main METHODOLOGICAL PRELIMINARIE task of linguistic theory must be to develop an account of linguistic universals that, on the one hand, will not be falsifed by the actual diversity of languages and, on the other, will be sufciently rich and explicit to account for the rapidity and uniformity of language learning, and the remarkable com plexity and range of the generative grammars that are the product of language learning. The study of linguistic universals is the study of the prop erties of any generative grammar for a natural language. Partic ular assumptions about linguistic universals may pertain to either the syntactic, semantic, or phonological component, or to interrelations among the three components. It is useful to classify linguistic universals as formal or sub stantive . A theory of substantive universals claims that items of a particular kind in any language must be drawn from a fxed class of items. For example, Jakobson's theory of distinctive features can be interpreted as making an assertion about substantive universals with respect to the phonological component of a generative grammar. It asserts that each output of this component consists of elements that are characterized in terms of some small number of fxed, universal, phonetic features (perhaps on the order of ffteen or twenty), each of which has a substantive acoustic-articulatory characterization independent of any partic ular language. Traditional universal grammar was also a theory of substantive universals, in this sense. It not only put forth interesting views as to the nature of universal phonetics, but also advanced the posi tion that certain fxed syntactic categories (Noun, Verb, etc.) can be found in the syntactic representations of the sentences of any language, and that these provide the general underlying syntactic structure of each language. A theory of substantive semantic universals might hold for ex ample, that certain designative functions must be carried out in a specifed way in each language. Thus i t might assert that each language will contain terms that desigate persons or lexical items referring to certain specifc kinds of objects, feelings, be havior, and so on. It is also possible, however, to search for universal properties § 5. FORMAL AND SUBSTANTIVE UNIVERSALS 29 of a more abstract sort. Consider a claim that the grammar of every language meets certain specifed formal condi tions. The truth of this hypothesis would not in itself imply that any particular rule must appear in all or even in any two grammars. The property of having a grammar meeting a certain abstract condition might be called a formal linguistic universal, if shown to be a general property of natural languages. Recent attempts to specify the abstract conditions that Ü generative grammar must meet have produced a variety of proposals concerning formal uni versals, in this sense. For example, consider the proposal that the syntactic component of a grammar must contain transforational rules (these being operations of a highly special kind) mapping semantically interpreted deep structures into phonetically inter preted surface structures, or the proposal that the phonological component of a grammar consists of a sequence of rules, a subset of which may apply cycl ically to successively more dominant con stituents of the surface structure (a transformational cycle, in the sense of much recent work on phonology) . Such proposals make claims of a quite diferent sort from the claim that certai n sub stantive phonetic elements are available for phonetic representa tion in all languages, or that certain specifc categories must be central to the syntax of all languages, or that certain semantic features or categories provide a universal framework for semantic description. Substantive universals such as these concern the vocabulary for the description of language; formal universals involve rather the character of the rules that appear in grammars and the ways in which they can be interconnected. On the semantic level, too, it is possible to search for what might be called formal universals, in essentially the sense j ust described. Consider, for example, the assumption that proper names, in any language, must designate obj ects meeting a condi tion of spatiotemporal contigui ty, 15 and that the same is true of other terms designating objects; or the condi tion that the color words of any language must subdivide the col or spectrum into continuous segments; or the condi tion that artifacts are defned in terms of certain human goals, needs, and functions instead of solely in terms of physical qualities. lo Formal con- METHODOLOGICAL PRELIMINARIES straints of this sort on a system of concepts may severely limit the choice (by the child, or the linguist) of a descriptive grammar, given primary linguistic data. The existence of deep-seated formal universals, in the sense suggested by such examples as these, implies that all languages are cut to the same pattern, but does not imply that there is any point by point correspondence between particular languages. It does not, for example, imply that there must be some reason able procedure for translating between languages. 1 7 In general, there is no doubt that a theory of language, re garded as a hypothesis about the innate "language-forming capacity" of humans, should concern itself with both substantive and formal universals. But whereas substantive universals have been the traditional concer of general linguistic theory, investi gations of the abstract conditions that must be satisfed by any generative grammar have been undertaken only quite recently. They seem to ofer extremely rich and varied possibilities for study in all aspects of grammar. § 6. FURTHER REMARKS ON DESCRIPTIVE AND EXPLANATORY THEORIES Let us consider with somewhat greater care j ust what is involved in the construction of an "acquisition model" for language. A child who is capable of language learning must have ( 1 2) (i) a technique for representing input signals (ii) a way of representing structural information about these sigals (iii) some initial delimitation of a class of possible hypotheses about language structure (iv) a method for determining what each such hypothesis im plies with respect to each sentence (v). a method for selecting one of the (presumably, infnitely many) hypotheses that are allowe by (iii) and are com patible with the given primary linguistic data § 6. DESCRIPTIVE AND EXPLANATORY THEORIES Correspondingly, a theory of linguistic structure that aims for explanatory adequacy must contain ( 1 3) (i) a universal phonetic theory that defnes the notion "possi ble sentence" (ii) a defnition of "structural description" (iii) a defnition of "generative grammar" (iv) a method for determi ning the structural description of a sentence, given a grammar (v) a way of evaluating alternative proposed grammars Putting the same requirements in somewhat diferent terms, we must require of such a linguistic theory that i t provide for ( 1 4) (i) an enumeration of the class S 1 ' S2' • • • of possible sentences (ii) an enumeration of the class SD 1 , SD2, • • • of possible structural descriptions (iii) an enumeration of the class G 1 , G2, • • • of possible genera tive grammars (iv) specifcation of a function f such that SD,«(, J ) is the struc tural description assigned to sentence ð_ by grammar Gi, for arbitrary i,j 18 (N) specifcation of a function M such that M(Î) is an integr associated with the grammar G, as its value (with, let us say, lower value indicated by higher number) Conditions of at least this strength are entailed by the decision to aim for explanatory adequacy. A theory meeting these conditions would attempt to account for language learning in the following way. Consider frst the nature of primary linguistic data. This consists of a fnite amount of information about sentences, which, furthermore, must be rather restricted in scope, considering the time limitations that are in efect, and fairly degenerate in quality (cf. note 1 4). For example, certain signals might be accepted as properly formed sentences, while others are classed as nonsentences, as a result of correction of the learner's attempts on the part of the linguistic community. Furthermore, the conditions of use might be such METHODOLOICAL PRELIMINARIES as to require that structural descriptions be assigned to these object in certain ways. That the latter is a prerequisite for language acquisition seems to follow from the widely accepted (but. for the moment. quite unsupported) view that there must be a partially semantic basis for the acquisition of syntax or for the justifcation of hypotheses about the syntactic component of a grammar. Incidentally. it is often not realized how strong a claim this is about the innate concept-forming abilities of the child and the system of linguistic universals that these abilities imply. Thus what is maintained. presumably. is that the child has an innate theory of potential structural descriptions that is sufciently rich and fully developed so that he is able to deter mine. from a real situation in which a sigal occurs, which struc tural descriptions may be appropriate to this sigal, and also that he is able to do this in part in advance of any assumption as to the linguistic structure of this signal. To say that the assumption about innate capacity is extremely strong is, of course. not to say that it is incorrect. Let us, in any event, assume tentatively that the primar linguistic data consist of signals classifed as sentences and nonsentences. and a partial and tenta tive pairing of signals with structural descriptions. A language-acquisition device that meets conditions (i)-(iv) is capable of utilizing such primary linguistic data as the empirical basis for language learing. This device must search through the set of possible hypotheses GI• G2• • • • • which are available to it by virtue of condition (Hi), and must select grammars that are compatible with the primary linguistic data, represented in ters of (i) and (ii) . It is possible to test compatibility by virtue of the fact that the device meets condition (iv) . The device would then select one of these potential grammars by the evaluation measure guaranteed by (V).19 The selected grammar now pro vides the device with a method for interpreting an arbitrary sentence, by virtue of (H) and (iv) . That is to say, the device has now constructed a theory of the language of which the primary lin g uistic data are a sample. The theory that the device has now selected and internally represented specifes its tacit comptence, its knowledge of the language. The child who acquires a language § 6. DESCRIPTIVE AND EXPLANATORY THEORIES 33 in this way of course knows a great deal more than he has "learned. " His knowledge of the language, as this is determined by his internalized grammar, goes far beyond the presented primary linguistic data and is in no sense an "inductive gen eralization" from these data. This account of language learning can, obviously, b para phrased directly as a description of how the linguist whose work is guided by a linguistic theory meeting conditions (i)-(v) would justify a grammar that he constructs for a language on the basis of given primary linguistic data.2 Notice, incidentally, that care must be taken to distinguish several diferent ways in which primary linguistic data may be necessary for language learing. In part, such data determine to which of the possible languages (that is, the languages pro vided with grammars in accordance with the a priori constraint (iii» the language learner is being exposed, and it is this function of the primary linguistic data that we are considering here. But such data may play an entirely diferent role as well; namely, certain kinds of data and experience may be required in order to set the language-acquisition device into operation, although they may not afect the manner of i ts functioning in the least. Thus it has been found that semantic reference may greatly facilitate perforance in a syntax-learning experiment, even though it does not, apparently, afect the manner in which acquisition of syntax proceeds; that is, it plays no role in deter mining which hypotheses are selected by the learer (Miller and Norman, 1 964) . Similarly, it would not be at all surprising to fnd that normal language learning requires use of language in real-life situations, in some way. But this, if true, would not be sufcient to show that information regarding situational context (in particular, a pairing of sigals wi th structural descrip tions that is at least in part prior to assumptions about syntactic structure) plays any role in determining how language is acquired, once the mechanism is put to work and the task of language learning is undertaken by the child. This distinction is quite familiar outside of the domain of language acquisition. For example, Richard Held has shown in numerous experiments 34 METHODOLOGICAL PRELIMINARIES that under certain circumstances reaferent stimulation (that is, stimulation resulting from voluntary activity) is a prerequisite to the development of a concept of visual space, although it may not detenine the character of this concept (cf. Held and Hein, 1 963; Held and Freedman, 1 963, and references cited there) . Or, to take one of innumerable examples from studies of animal learing, it has been observed (Lemmon and Patterson, 1 964) that depth prception in lambs is considerably facilitated by mother-neonate contact, although again there is no reason to suppose that the nature of the lamb's "theory of visual space" depends on this contact. In studying the actual character of learning, linguistic or oterwise, it is of course necessary to distinguish carefully be tween these two functions of exteral data -the function of initiating or facilitating the operation of innate mechanisms and the function of detennining in part te direction that learning will take.21 Returning now to the main theme, we shall call a theory of linguistic structure that meets conditions (i)-(v) an explanator theor, and a theory that meets conditions (i)-(iv) a descriptive theor. In fact, a linguistic theory that is concerned only with descriptive adequacy will limit it attentlon to topics (i)-(iv). Such a theory must, in other word, make available a class of generative grammars containing. for each language. a descrip tively adequate grammar of this language -a grammar that (by means of (iv» assigs structural descriptions to sentences in accordance with the linguistic competence of the native speaker. A theory of language is empirically signifcant only to the extent tat it meets conditions (i)-(iv). The further question of explana tory adequacy arises only in connection with a theory that also meets condition (v) (but see p. 36). In other words, it Üribes only to the extent that the theory provides a principled basis for selecting a descriptively adequate gammar on the basis of primary linguistic data by the use of a well·defned evaluation measure. This account is misleading in one important respct. It sug gests that to raise a descriptively adequate theory to the level § 6. DESCRIPTIVE AND EXPLANATORY THEORIES 3 5 of explanatory adequacy one needs only to defne an appropriate evaluation measure. This is incorrect, however. A theory may be descriptively adequate, in the sense just defned, and yet provide such a wide rang of potential grammars that there is no possi bility of discovering a foral property distinguishing the de scriptively adequate grammars, in general. from among the mass of gammars compatible with whatever data are available. In fact, the real problem is almost always to restrict the range of possible hypotheses by adding additional structure to the notion "generative grammar." For the construction of a reasonable acquisition model, it is necessary to reduce the class of attain able22 gammars compatible with given primary linguistic data to the point where selection among them can be made by a foral evaluation measure. This requires a precise and narrow delimitation of the notion "generative gammar" ¯a restrictive and rich hypothesis concering the universal properties that determine the for of language, in the traditional sense of this ter. The same point can be put in a somewhat diferent way. Given a variety of descriptively adequate grammar for natural languages, we are interested in determining to what extent they are unique and to what extent there are deep underlying similari ties among them that are attributable to the form of language as such. Real progress in linguistics consists in the discovery that certain features of given languages can be reduced to universal properties of language, and explained in terms of these deeper aspects of linguistic form. Thus the major endeavor of the linguist must be to enrich the theory of linguistic form by for mulating more specifc constraints and conditions on the notion "generative grammar." Where this can be done. particular gram mars can be simplifed by eliminating from them descriptive statement that are attributable to the general theory of grammar (cf. § 5). For example, if we conclude that the transformational cycle23 is a universal feature of the phonological component, it is unnecessary, in the gammar of English, to describe the man ner of functioning of those phonological rules that involve syntactic structur. This description will now have been ab- METHODOLOGICAL PRLIMINARIES stracted from the grammar of English and stated as a formal linguistic universal, as part of the theory of generative grammar. Obviously, this conclusion, if j ustifed, woul d represent an im portant advance in the theory of language, since it would then have been shown that what appears to be a peculiarity of English is actually explicable in terms of a general and deep empirical assumption about the nature of languag, an assumption that can be refuted, if false, by study of descriptively adequate gammars of other languages. In short, the most serious problem that arises in the attempt to achieve explanatory adequacy is that of characterizing the notion "gnerative gammar" in a sufciently rich, detailed, and highly structured way. A theory of gammar may be descriptively adequate and yet leave unexpressed major features that are defning properties of natural language and that distinguish natural languages from arbitrary symbolic systems. It is for just this rason that the attempt to achieve explanatory adequacy -the attempt to discover linguistic universals -is so crucial at every stage of undertanding of linguistic structure, despi te the fact that even descriptive adequacy on a broad scale may be an unrealized goal. It is not necessary to achieve descriptive ade quacy before raising questions of explanatory adequacy. On the contrary, the crucial questions, the quetions that have the geatest bearing on our concept of language and on descriptive practice as well, are almost always those involving explanatory adequacy with respect to particular aspects of languag structure. To acquire language, a child must devise a hypothesis compa tible with presented data -he must select from the store of potential grammars a specifc one that is appropriate to the data available to him. It is logically possible that the data might be sufciently rich and the class of ptential gammars sufciently limited so that no more than a single prmitted grammar will be compatible with the available data at the moment of success ful language acquisition, in our idealized "instantaneous" model (cf. notes 1 9 and 22) . In this case, no evaluation procedure will be necessar as a part of linguistic theory -that is, as an innate property of an organism or a device capable of language acquisi- § 7. ON EVALUATION PROCEDURES 37 tion. It is rather difcult to imagine how in detail this logical possibility might be realized, and all concrete attempts to formulate an empirically adequate linguistic theory certainly leave ample room for mutually inconsistent gammars, all com patible with primary data of any conceivable sort. All such theories therefore require supplementation by an evaluation measure if language acquisition is to be accounted for and selec tion of specifc gammars is to be justifed; and I shall continue to assume tentatively, as heretofore, that this is an empirical fact about the innate human faculte de langage and consequently about general linguistic theory as well. § 7. ON EVALUA TION PROCEDURES The status of an evaluation procedure for gammars (see condi tion (v) of ( 1 2)-( 1 4» has often been misconstrued. It must frt of all be kept clearly in mind that such a measure is not given a priori, in some manner. Rather, any proposal concering such Ü meÜsuÏe is an empiÏical hypothesis about the nature of language. This is evident from the preceding discussion. Suppose that we have a descriptive theor, meeting conditions (i)-(iv) of ( 1 2)-( 1 4) in some fxed way. Given primarily linguistic data D, diferent choices of an evaluation measure will assign quite diferent ranks to alternative hypotheses (alternative grammars) as to the lan guage of which D is a sample, and will therefore lead to entirely diferent predictions as to how a person who learns a language on the basis of D will interpret new sentences not in D . Con sequently, choice of an evaluation measure is an empirical matter, and particular proposals are correct or incorrect. Perhaps confusion about this matter can be traced to the use of the term "simplicity measure" for particular proposed evalua tion measures, it being assumed that "simplicity" is a gneral notion somehow understood in advance outside of linguistic theory. This is a misconception, however. In the context of this discussion, "simplicity" (that is, the evaluation measure M of (v» is a notion to be defned within linguistic theor along with "gammar," "phoneme, " etc. Choic of a simplici ty measur is METHODOLOGICAL PRUMINAIES rather like determination of the value of a physical constant. We are given, in part, an empirical pairing of certain kinds of primary linguistic data with certain gammars that are in fact constructed by people presented with such data. A pro posed simplicity measure constitutes part of the attempt to deter mine precisely the nature of this association. If a particular formulation of (i)-(iv) is assumed, and if pairs (D1, G 1 ) , (D 2 , G 2 ), • • • of primary linguistic data and descriptively adequate grammars are gven, the problem of defning "simplicity" is just the problem of discovering how G. is determined by D., for each i. Suppose, in other words, that we regard an acquisition model for language as an input-output device that determines a particular generative grammar as "output," given certain primary linguistic data as input. A proposed simplicity measure, taken together with a specifcation of (i)-(iv), constitutes a hypothesis concering the nature of such a device. Choice of a simplicity measure is there fore an empirical matter with empirical consequences. All of this has been said before. I repeat it at such length be cause it has been so grossly misunderstood. It is also apparent that evaluation measures of the kinds that have been discussed in the literature on generative grammar cannot be used to compare diferent theories of grammar; comparison of a grammar from one class of proposed grammars with a grammar from another clas, by such a MÜÜMÏÜ¿ is utterly without sense. Rather, an evaluation measure of this kind is an essential part of a particular theory of grammar that aims at explanatory adequacy. It is true that there is a sense in which alternative theories of language (or alterative theories in other domains) can be compared as to simplicity and elegance. Wat we have been discussing here, however, is not this general ques tion but rather the problem of comparing two theories of a language -two grammars of this language -in terms of a particular gneral linguistic theory. This is, then, a matter of formulating an explanatory theory of language; it is not to be confused with the problem of choosing among competing theories of language. Choice among competing theories of language is of course a fundamental question and should also be § 7 . ON EVALUATON PROCEDURES 39 settled, insofar as possible, on empirical grounds of descriptive and explanatory adequacy. But it is not the question involved in the use of an evaluation measure in the attempt to achieve explanatory adequacy. As a concrete illustration, consider the question of whether the rules of a grammar should be unordered (let us call this the linguistic theory Tu) or ordered in some specifc way (the theory To). A priori, there is no way to decide which of the to is correct. There is no known absolute sense of "simplicity" or "elegance. " developed within linguistic theory or gneral epis temology, in accordance wi th which Tu and To can be compared. It is quite meaningless, therefore, to maintain that in some absolute sense Tu is "simpler" than To or conversely. One can easily invent a gneral concept of "simplicity" that will prefer Tu to To, or To to Tu; in neither case W i ll this concept have any known justifcation. Certain measures of evaluation have been proposed and in part empirically justifed within linguistics -for example, minimization of feature specifcation (as discussed in Halle, 1 959a, 1 9 61 , 1 962a, 1 964) or the measure based on abbreviatory notations (discussed on pp. 42f.) . These measures do not apply, because they are internal to a specifc linguistic theory and their empirical justifcation relies essentially on this fact. To choose between Tu and To, we must proceed in an entirely diferent way. We must ask whether Tu or To provides descrip tively adequate gammars for natural languages, or leads to explanatory adequacy. This is a perfectly meaningful empirical question if the theories in question are stated with sufcient care. For example, if TuB is the familiar theory of phrase struc ture grammar and T OB is the same theory, with the further condi tion that the rules are linearly ordered and apply cyclically, with at least one rule A � X being obligatory for each category A, so as to guarantee that each cycle is nonvacuous, then it can be shown that TuB and T OB are incomparable in descriptive power (in "strong generative capacity" -see § 9 ; see Chomsky, 1 955, Chapters 6 and 7, and Chomsky, 1 956, for some discus sion of such systems). Consequently, we might ask whether natural languags in fact fall under T Jl or T OB, these being non- METHODOLOGICAL PRELIMINARIES equivalent and empirically distinguishable theories. Or, sup posing T @" and ToP to be theories of the phonological component (where T uP holds phonological rules to be unordered and ToP holds them to be partially ordered), it is easy to invent hypo thetical "languages" for which signifcant generalizations are expressible in terms of ToP but not T uP, or conversely . We can therefore try to determine whether there are signifcant gen eralizations that are expressible in terms of one but not the other theory in the case of empirically given languags. In principle, either result is possible; it is an entirely factual question, having to do with the properties of natural languages. We shall see later that T 08 is rather well motivated as a theory of the base, and strong arguments have been ofered to show that ToP is conect and T � is wrong, as a theory of phonological processes (cf. Chomsky, 1 95 1 , 1 964; HaIIe, 1 959a, 1 959b, 1 962a, 1 964). In both cases, the argument turs on the factual question of expressibiIity of linguistically signifcant generalizations in terms of one or the other theory, not on any presumed absolute sense of "simplicity" that might rank Tu and To rlative to one another. Failure to appreciate this fact has led to a great deal of vacuous and pointles discussion. Confusion about these questions may also have been engen dered by the fact that there are several diferent senses in which one can talk of "justifying" a grammar, as noted on pp. 26-27. To repeat the major point: on the one hand, the grammar can b justifed on exteral gounds of descriptive adequacy -we may ask whether i t states the facts about the language cor rectly, whether it predicts correctly how the idealized native speaker would understand arbitrary sentences and gives a conect account of the basis for this achievement; on the other hand, a grammar can be justifed on interal grounds if, given an ex planatory linguistic theory, it can be shown that this grammar is the highest-valued gammar permitted by the theory and com � patible with given primary linguistic data. In the latter case, a principled basis is presented for the construction of this grammar, and it is therefore justifed on much deeper empirical gounds. Both kind of justifcation are of course necessary; it is im- § 7. ON EVALUATION PROCEDURES portant. however. not to confuse them. In the case of a linguistic theor that is merely descriptive. only one kind of j ustifcation can be given -namely. we can show that it permit grammars that meet the exteral condition of descriptive adequacy. 2 4 It is only when all of the conditions (i)-(v) of ( 1 2)-( 1 4) are met that the deeper question of interal j ustifcation can be raised. It is also apparent that the discussion as to whether an evaluation measure is a "necessary" part of linguistic theory is quite without substance (see. however. pp. 36-37). If the linguist is content to formulate descriptions one way or another with little concern for j ustifcation. and if he does not intend to proceed from the study of facts about particular languages to an investigation of the characteristic properties of natural language as such. then construction of an evaluation procedure and the associated concers that relate to explanatory adequacy need not concern him. In this case. since interest in j ustifcation has been abandoned. neither evidence nor argument (beyond minimal requirements of consistency) has any bearing on what the linguist presents as a linguistic description. On the other hand. if he wishes to achieve descriptive adequacy in his account of language structure. he must concern himself with the problem of develop ing an explanatory theory of the form of grammar. since this provides one of the main tools for arriving at a descriptively adequate grammar in any particular case. In other words. choice of a grammar for a particular language L will always be much underdetermined by the data drawn from L alone. Moreover. other relevant data (namely, successful gammars for other languages or successful fragments for other subparts of L) will be available to the linguist only if he possesses an explanatory theory. Such a theory limits the choice of grammar by the dual method of imposing formal conditions on grammar and providing an evaluation procedure to be applied for the language L with which he is now concerned. Both te formal conditions and the evaluation procedure can be empirically justifed by their success in other cases. Hence. any far-reaching concern for descriptive adequacy must lead to an attempt to develop an explanatory theory that fulflls these dual functions. and concer with ex- METHODOLOGICAL PRELIMINARIE planatory adequacy surely requires an investigation of evaluation procedurs. The major problem in constructing an evaluation measure for grammar is that of determining which generalizations about a language are signifcant ones; an evaluation measure must be selected in such a way as to favor these. We have a generalization when a set of rules about distinct i tems can be replaced by a single rule (or. more gnerally. partially identical rules) about the whole set. or when it can be shown that a "natural class" of items undergoes a certain process or set of similar processes. Thus. choice of an evaluation measure constitutes a decision as to what are "similar processes" and "natural classes" - in short. what are signifcant generalizations. The problem is to devise a procedure that will assign a numerical measure of valuation to a grammar in ters of the degree of linguistically sigifcant generalization that this grammar achieves. The obvious numerical measure to be applied to a grammar is length. in terms of number of symbols. But if this is to be a meaningful measure. it is necessary to devise notations and to restrict the form of rules in such a way that signifcant considerations of complexity and gen erality are converted into considerations of length. so that real generalizations shorten the grammar and spurious ones do not. Thu it is the notational conventions used in presenting a grammar that defne "sigifcant generalization." if the evalua tion measure i taken as length. This is. in fact. the rationale behind the conventions for use of parentheses. brackets, etc ¤ • that have been adopted in explicit (that is. generative) grammars. For a detailed discussion of these, see Chomsky ( 1 95 1 , 1 955), Postal ( 1 962a), and Matthews ( 1 964). To take just one example. consider the analysis of the English Verbal Auxili a . The facts are that such a phrase must contain Tense (which is, furthermore, Past or Present). and then may or may not contain a Modal and either the Perfect or Progressive Aspect (or both), where the elements must appear in the order just given. Using familiar notational conventions. we can state this rule in the following form: § ÿ. ON EVALUATON PROCEDUR [ I _j PuX � ¯CD8C [NDU8¡j [Ï6t]66lj [Ït0_t688ÌU6j 43 [DuÍtlÍD_ UCl8Í¡8 lD8l 8IC DDl 1CÌCVaDl DCIcj. ÛDÌC [ I _j Í8 aD 8DDICVÍ8lIDD ÍDI CÍ_Dl IuÌC8 lDal aH8Ìy2C lDC CÌCDCDl PuX ÍDtD Ít8 CÍ_Dl gDä8ID¡C ÍDQä. bl8lCU ÍD ÍD¡Ì, lDC8C cÍ_Dl ID¡C8 WDuÌU ÍDVDÌVC lWCDLy 8yDDD¡8, WDCIC88 IDÌC [ I gj ÍDVDÏVC8 ÍDuI [DDl CDuDtÍD_ 1uX, ÍD DDtD C38C8j. ¯DC _8tCDlDC8Í8 DDl8lÍDD, ÍD lDÍ8 C88C, D88 lDC ÍDÌ¡DWÍD_ mCaDÍD_. Ïl 888CIl8 lD8t tDC UÍhCICDCC DClWCCD ÍDDI 8Dd tWCDly 8yDDD¡8 Í8 a mC8DIc DÍ lDC UC@CC DÍ ¡ÎD_uÎ8lÎC8Ì¡y äÎ_DÍhC8Dl _CDCI8ÌÍ28lÍDD 8CDÍCVcU ÍD a Ì8D_u8_C tD8l D88 IDC ÎDD8 _ÌVCD ÍD ÌÍ8l (16), ÎDI tDC JUÍ¡Íæ ÏDt88C, 88 CDmg8ICU WÍlD 8 ¡8D_D8_C lD8l Da8, ÍDI CXðDgÌC, lDC ÍDO8 _ÎVCD ÍD ¡Í8t [ I ÿj 88 tDC IcgIC8CDl8tÍVC8 DÍtDÍ8 C8lC_Iy¦ [ I Oj ±CD8C, ¯CDäC´NDC8Ì, ¯CD8C¯Ï6t]66l, ¯CD8C¯Ït0_r6ããÌU6, ¯CD8C¯NDda¡¯Ï6t]6cl, ¯CD8C¯NDU8̯Ït0_t688ÌU6, ¯CD8C ¯Ï6t]6cf´ Ït0_688ÌU6, ¯CD8C¯NDU8̯ ÏCt]66l¯ Ït0_t6ã8ÌU6 [ I ÿj ¯CD8C¯NDd8̯ Ï6t]6cl´ Ït0_t688ÌU6, NDU8Ì´ Ï6t]66l´Ït0- _t6MÌU6´¯CD8C, Ï6t]6cl´ Ït0_6MÌU6¯CD8C¯NDU8Î, Ïf0- _t688lU6¯±CD8C´NDU8̯ Ï6t]6cl, ¯CD8C¯ Ï6t]66l, mDO8¡´ Ït0_t688ÌU6 ÏD lDC C88C of DDlD ¡Î8l [ I Uj 8DU ÌÍ8l [ I ÿj, lWCDly 8yDDDÌ8 aD ÍDVD¡VCO. ÏÍ8l (1 6) aDDICVÍatC8 lD IuÍC [I _j Dy tDC DDt8tÍDD8Í CDDVCDlÍDD, ÌÍ8l [ I ÿj C8DDDl DC 8DDICVÍalCU Dy lDÍ8 CDDVCDlÍDD. ÏCDCC, adDglÍDD DÍ lDC ÍauÍÌÍ8I DDlalÍDD8Ì CDDVCDlÎDD8 ÍDVD¡VÍD_ tDC u8C DÍ Q8ICDlD6ä68 8mDUDt8 tD a C¡8Ím lD8l lDCIC Í8 8 ÌÍD_uÍ8tÍC8ÌÌy 8Í_DÌhCaDl _CDCI8ÌÍZ8tÍDD uDUCIÍyÍD_ tDC 8CI DÍ ÍDIm8 ÍD ¡Î8l (1 6) Dut DDl tDC 8Cl DÍ tDIm8 ÍD ¡Í8t [ I ÿj . Ïl 8mDuDt8 lD lDC Cu_IIÍC8¡ DygDtDC8Í8 lD3l tC_u¡8IÍlÍC8 DÍ lDC tygC Cx- Cmg¡ÍhCU ÍD [ I Oj aIC lDD8C ÍDDDU ÍD D8lDIaÌ ÌHD_D8@8, aDU 8IC of tDC lygC UHl CDI¡OtCD ÌCaMÍD_ a Í8D@a@ WÍÍÍ CXgCC| , WDCtC88 CyC¡ÍC IC_uÌ8IÍtÍC8 DÍlDc lygC CXCmg¡ÍhCU Í D [ I ÿj, lDDu_D QtÍCClÌy @DDÎDC, 8D8lt8ClÌy, 8Ic DDl CDaIaCLCIÎ8lÍC DÍ H8lUIaÌ ¡8D_u8@, 8IC DDl DÍ tDC ly_C ÍDI WDÍCD CDÍ¡dICD WÍ¡¡ ÍDlDÍlÍVCÌy 8C8tCD in 18H_u8_C D8I61aÎä, 8Dd 8tC DuCD DOtC dÎmCu1I ÎDt lDC ¡8D_u8_C·ÍC8MCI tD CDD8ltuCt DD lDC D8äÍä DÍ äC8tICtCO d3t8 44 METHODOLOGICAL PRELIMINARIE or to use. What is claimed, then, is that when given scattered examples fom ( 1 6) , the language learer will construct the rule ( 1 5) generating the full set with their semantic interpretations, whereas when given scattered examples that could be subsumed under a cyclic rule, he will not incorporate this "generalization" in his grammar -he will not, for example, conclude from the existence of "yesterday John arrived" and "John arrived yesterday" that there is a third form "arrived yesterday John, " or from the existence of "is John here" and "here is John" that there is a third form "John here is," etc. One might easily propose a diferent notational convention that would abbreviate list ( 1 7) to a shorter rule than list ( 1 6), thus making a diferent empirical assumption about what constitutes a linguistically sigifcant generalization. There is no a priori reason for preferring the usual convention; it ðimply embodies a factual claim about the structure of natural language and the predisposition of the child to search for certain types of regularity in natural language. The illustrative examples of the preceding paragraph must be regarded with some caution. It is the full set of notational con ventions that constitute an evaluation procedure, in the manner outlined earlier. The factual coÎtent of an explanatory theory lies in its claim that the most highly valued grammar of the permitted for will be selected, on the basis of given data. Hence, descriptions of particular subsystems of the grammar must be evaluated in terms of their efect on the entire system of rules. The extent to which particular parts of the grammar can be selected independently of others is an empirical matter about which very li ttle is known, at present. Although alternatives can be clearly formulated, deeper studies of particular languages than are presently available are needed to settle the questions that immediately arise when these extremely important issues are raised. To my knowledge, the only attempt to evaluate a fairly full and complex subsystem of a grammar is in Chomsky ( 1 95 1 ), but even here all that is shown is that the value of the system i s a "local maximum" in the sense that interchange of adjacent rules decreases value. The efect of modifcations on a largr § 7. ON EVALUATION PROCEDURES 4 5 scale is not investigated. Certain aspects of the general question, relating to lexical and phonological structure, are discussed in Halle and Chomsky (forthcoming). One special case of this general approach to evaluation that ha been worked out in a particularly convincing way is the condition of minimization of distinctive feature specifcations in the phonological component of the grammar . A very plausible argument can be given to the efect that this convention defnes the notions of "natural class" and "signifcant generalization" that have been relied on implicitly in descriptive and com parative-historical phonological investigations, and that determine the intuitively given distinction between "phonologically p sible" and "phonologically impossible" nonsense forms. For discussion, see Halle ( 1 959a, 1 959b, 1 961 , 1 962a, 1 964), Halle and Chomsky (forthcoming). It is important to observe that the efectiveness of this partiLular evaluation measure i completely dependent on a strong assumption about the form of grammar, namely, the assumption that only feature notation is per mitted . If phonemic notation is allowed in addition to feature notation, the measure gives absurd consequences, as Halle shows. It is clear, then, that choice of notations and other conventions is not an arbi trary or "merely technical" matter, if length is to be taken as the measure of valuation for a grammar. It is, rather, a matter that has immediate and perhaps quite drastic empirical consequences. When particular notational devices are in corporated into a linguistic theory of the sort we are discussing, a certain empirical claim is made, implicitly, concerning natural language . It is implied that a person learing a language will attempt to formulate gneralizations that can easily be expressed (that is, with few symbols) in terms of the notations available in this theory, and that he will select grammars containing these generalizations over other grammars that are also compatible with the given data but that contain diferent sorts of generaliza tion, diferent concepts of "natural class," and so on. These may be very strong claims, and need by no means be true on any a priori grounds. To avoid any possible lingring confusion on this matter, METHODOLOGICAL PRELIMINARIES let me repeat once more that this discussion of language learning in terms of formulation of rules, hypotheses, etc., does not refer to conscious forulation and expression of these but rather to the process of arriving at an interal representation of a gen erative system, which can be appropriately described in these terms. In brief, it is clear that no prsent-day theory of languag can hope to attain explanatory adequacy beyond very restricted domains. In other words, we are very far from being able to present a system of formal and substantive linguistic universals that will be sufciently rich and detailed to account for the facts of language learning. To advance linguistic theory in the direction of explanatory adequacy, we can attempt to refne the evaluation measure for grammars or to tighten the formal con straints on gammars so that it becomes more difcult to fnd a highly valued hypothesis compatible with primary linguistic data. There can be no doubt that present theories of grammar require modifcation in both of these ways, the latter, in general, being the more promisingy · Thus the most crucial problem for linguistic theory seems to be to abstract statements and gn eralizations from particular descriptively adequate gammars and, wherYer possible, to attribute them to the general theory of linguistic structure, thus enriching this theory and imposing more structure on the schema for gammatical description. Wenever this is done, an assertion about a particular language is replaced by a corresponding assertion, from which the frst follows, about language in general. If this formulation of a deeper hypothesis is incorrect, this fact should become evident when it efect on the desLÎiption of other aspects of the language or the description of other languags is ascertained. In short, I am making the obvious comment that, wherever possible, general assumptions about the nature of language should be formulated from which particular features of the gammars of individual languages can b deduced. In this way, linguistic theory may move toward explanatory adequacy and contribute to the study of human mental processes and intellectual capacity ¯more specifcally, to the determination of the abilities that make § 8. UNGUISTC TEORY AND LANGUAGE LEARNING 47 languag learing possible under the empirically given limita tions of time and data. § 8. LINGUISTIC THEORY AND LANGUAGE LEARNING In the preceding discussion, certain problems of linguistic theory have been formulated as questions about the construction of a hypothetical language-acquisition device. This seems a use ful and suggestive framework within which to pose and consider these problems. We may think of the theorist as given an empirical pairing of collections of primary linguistic data as sociated with gamars that a constucted by the device on the bais of such data. Much information can be obtained about both the primary data that constitute the input and the grammar that is the "output" of such a device, and the theorist has the problem of deterining the intrinsic properties of a device capable of mediating this input-output relation. It may be of some interest to set this discusion in a somewhat mor general and traditional framework. Historically, we can distinguish two general lines of approach to the problem of acquisition of knowledge, of whic the problem of acquisition of languag is a special and particularly informative case. The empiricist approach has assumed that the structure of the acquisi tion device is limited to certain elementary "peripheral proc essing mechanisms" -for example, in recent versions, an innate "quality space" with a innate "distance" defned on it (Quine, 1 960, pp. 83f.),25 a set of primitive uncondi tioned refexes (Hull, 1 943), or, in the case of language, the set of all "aurally distin guishable component" of the full "audi tory impression" (Bloch, 1 950). Beyond this, it assumes that the device has certain analytical data-processing mechanisms or inductive principles of a very elementary sort, for example, certain principles of association, weak principles of "generalization" involving gradi ent along the dimensions of the given quality space, or, in our case, taxonomic principles of segmentation and clasifcation suc a toe tat have been developed with some car i moem linguistic, in accordance with te Saussurian emphasis METHODOLOGICAL PRELIMINARIES on the fundamental character of such principles. It is then assumed that a preliminary analysis of experience is provided by the peripheral processing mechanisms, and that one's concepts and knowledge, beyond this, are acquired by application of the available inductive principles to this initially analyzed ex perience . 26 Such views can be formulated dearly in one way or another a empirical hypotheses about the nature of mind. A rather diferent approach to the problem of acquisi tion of knowledg has been characteristic of rationalist speculation about mental processes. The rationalist approach holds that beyond the peripheral processing mechanisms,27 there are innate ideas and principles of various kinds that determine the form of the acquired knowledge in what may be a rather restricted and highly organized way. A condition for innate mechanisms to become activated is that apprQpriate stimulation be presented. Thus for Descartes ( 1 647), the innate ideas are those arising from the faculty of thinking rather than from external objects : . . . noting reaches our mind from external objects through the organs of sense beyond certain corporeal movements . . . but even these move ments, and the fgures which arise from them. are not conceived by us in the shape they assume in the organs of sense . . . . Hence it follows that the ideas of the movements and fgures are themselves innate in us. So much the more must the ideas of pain, colour, sound and the like be innate, that our mind may, on occasion of certain corporeal move ments, envisage these ideas, for tey have no likeness to the corporeal movements . . . [po 443]. Similarly, such notions as that things equal to the same thing are equal to each other are innate, since they cannot arise as necessary principles from "particular movements: ' In general, sight . . . presents nothing beyond pictures, and hearing nothing be yond voices or sounds, so that all tese tings tat we think of, beyond these voices or pictures, as being symbolized by them. are presented to us by means of ideas which come from no other source than our faculty of thinking. and are accordingly together wit that faculty innate in us, that is, always existing in us potentially; for existence in any faculty is not actual but merely potential existence, since the very word "fac ulty" designates nothing more or less than a potentiality. . . . [Thus § 8. LINGUISTIC THEORY AND LANGUAGE LEARNING 49 ideas are innate in the sense that] in some families generosity is innate, i n other certain diseases like gout or gravel, not that on this account te babes of tese families sufer from these diseases in teir moter's womb, but because they are born with a certain disposi tion or propen sity for contracting them . . . [po 442]. Still earlier, Lord Herbert ( 1 624) maintains that innate ideas and principles "remain latent when their corresponding objects are not present, and even disappear and give no sign of their existence"; they "must be deemed not b much the outcome of experience as principles without which we should have no ex perience at all . . . [o 1 32]. " Without these principles, "we could have no experience at all nor be capable of obserations"; "we should never come to distinguish between things, or to grasp any general nature . . . [o 1 05] ." These notions are extensively developed throughout seventeenth-century rationalist philosophy . To mention j ust one example, Cudworth ( 1 73 1 ) gives an extensive argument in support of his view that "there are many ideas of the mind, which though the cogitations of them be often oc casionally invited from the motion or appulse of sensible objects without made upon our bodies; yet notwithstanding the ideas themselves could not possibly be stamped or impressed upon the soul from them, because sense takes no cognizance at all of any such things in those corporeal objects, and therefore they must needs arise from the innate vigour and activity of the mind itself . . . [Book IV]. " Even in Locke one fnds essentially the same conception, as was pointed out by Leibniz and many com mentators since. ÏD the Port-Royal Logic (Arnauld, 1 662), the same point of view is expressed in the following way: It is false, terefore, tat all our ideas come trough sense. On the con trary, it may be afrmed tat no idea which we have in our minds has taken its rise from sense, except on occasion of tose movements which are made in the brain trough sense, te impulse from sense giving oc casion to te mind to form diferent ideas which it would not have formed witout it, tough these ideas have very rarely any resemblance to what takes place in the sense and in the brain; and there are at least a very great number of ideas which, having no connection with any METODOLOGICAL PRELIMINARIE bodily image, cannot, without manifest absurdity, be refered t sense • . . [Chapter 1]. In the same vein. Leibniz refuses to accept a sharp distinction between innate and learned: I agee that we learn ideas and innate truths eiter in considering teir source or in verifying them through experience. . . . And I cannot admit this proposition: all that one learns is not innate. The truths of numb en are in us, yet nonetheless one learns them,28 eiter by drawing tem from their source when we learn them through demonstrative proof (which shows tat they are innate), or by testing tem in exam ples, a do ordinary arithmeticians . . . [New Essays, p. 7 5 ] . [Thus ] all arithmetic and all geometry are in us virtually, so tat we can fnd them there if we consider attentively and set in order what we already have in the mind • • • [po 78] . [In general,] we have an infnite amount of knowledge of which we are not always conscious, not even when we need it [po 77 ] . The senses, although necessary for all our actual knowl edge, a not sufcient to give it all to us, since the senses never give us anything but examples, i.e., particular or individual truts. Now all the examples whic conf a general tt, whatever their number, do not sufce to establish the univenal necessity of that same trut . ø • [pp. 4 2 -4S ] . Necessar truths . . . must have principles whose proof does not depend on examples, nor consequently upon te testimony of te senses, although witout the senses it would never have occured to us t think of them. . . . It is true tat we must not imagine that these eternal laws of t e reason can be read in the soul a in an open book • ø ¤ but it is sufcient that tey can be discovered in us by dint of at tention, for which te snses furnish occasions, and successful experience seres to confr reason ø • ø [po 44] . [here are innate general princi· pIes that] enter into our tought, of which they form the soul and the connection. They a as necessary thereto as the muscles and sinews are for walking, although we do not at all think of them. The mind leans upon tese principles every moment, but it does not come so easily to distinguish them and to represent them distinctly and separately, be- caue that demands grat attention to it acts . . . . Thus it is that one posseses many thing witout kowing it . . . [po 74] . (a, for example, the Chinese possess articulate sounds, and therefore the basis for alphabetic writing, although they have not invented tis). § 8. LINGUISTC THEORY AND LGUAGE LENING Notice, incidentally, that throughout these classical discussions of the interplay between sense and mind in the formation of ideas, no sharp distinction is made between perception and acquisition, although there would be no inconsistency in the assumption that latent innate mental structures, once "activated," are then available for interpretation of the data of sense in a way in which they were not previously. Applyng this rationalist view to the spcial cae of language learning, Humboldt ( 1 836) concludes that one cannot really teach language but can only prsent the conditions under which it will develop spontaneously in the mind in its own way. Thus the form of a language, the schema for its grammar, is to a large ex tent given, though it will not be available for use without ap propriate experience to set the language-foring processes into operation. Like Leibniz, he reiterates the Platonistic view that, for the individual, learning is largely a matter of Wiederer zeugung, that is, of drawing out what is innate in the mind.29 This view contrasts sharply with the empiricist notion (the prevailing modern view) that language is essentially an ad ventitious construct, taught by "conditioning" (as would be maintained, for example, by Skinner or Quine) or by drill and explicit explanation (as was claimed by Wittgenstein), or built up by elementary "data-processing" procedures (as moder linguistic typically maintains), but, in any event, relatively independent in its structure of any innate mental faculties. In short, empiricist speculation has characteristically assumed that only the procedures and mechanism for the acquisition of knowledge constitute an innate property of the mind. Thus for Hume, the method of "experimental reasoning" is a basic instinct in animals and humans, on a par with the instinct "which teaches a bird, with such exactness, the art of incubation, and the whole economy and order of it nursery" -it is derived "from the original hand of nature" (Hume, 1 748. § IX). The for of knowledge, however, is otherwise quite free. On the other hand. rationalist speculation has assumed that the general form of a system of knowledg is fed in advance as a disposition of the mind. and the function of experience is to cause this gneral METHODOLOGICAL PRELIMINARIE schematic structure to be realized and more fully diferentiated. To follow Leibniz's enlightening analogy, we may make . . . the comparison of a block of marble which has veins, rather than a block of marble wholly even, or of blank tablets, i.e. , of what is called among philosophers a tabula rasa. For if the soul resembled these blank tablets, trths would be in us as the fgure of Hercules is in the marble, when the marble is wholly indiferent to the reception of this fgure or some other. But if tere were veins in the block which should i ndicate the fgure of Hercules rather than oter fgures, this block would be more determined thereto, and Hercules would be in it ð in some sense innate, although it would be needful to labor to discover these veins, to clear them by polishing, and by cutting away what prevents them frm appearing. Thus it is tat ideas and truths are for us innate, as inclinations, dispositions, habits, or natural potentialities, and not as actions; although these potentialities are always accompanied by some actions, often insensible, which corespnd to them [Leibniz, New Es says, pp. 45-46J. It is not, of course, necessary to assume that empiricist and rationalist views can always be sharply distinguished and that these currents cannot cross. Nevertheless, it is historically ac curate a well as heuristically valuable to distinguish these two very diferent approaches to the problem of acquisition of knowledge. Particular empiricist and rationalist views can be made qui te precise and can then b presented as explicit hypotheses about acquisition of knowledg, in particular, about the innate structure of a language-acquisition device. In fact, it would not be inaccurate to describe the taxonomic, data prcessing apprach of modem lingistic as ÜÏ empiricist view that contrasts with the essentially rationalist alterative proposed in recent theories of transformational grammar. Taxonomic linguistic is empiricist in its assumption that general linguistic theory consists only of a body of procedures for determining the grammar of a language fom a corpus of data, the form of language being unspecifed except insofar as restrictions on possible grammars are determined by this set of procedures. If we interpret taxonomic linguistic as making an empirical claim, Bo § 8. UNGUISTIC THEORY AND LANGUAGE LEARNING 5 3 this claim must be that the grammars that result from application of the postulated procedures to a sufciently rich selection of data will be descriptively adequate -in other words, that the set of procedures can be regarded as constituting a hypothesis about the innate language-acquisition system. In contrast, the discussion of language acquisition in preceding sections was rationalistic in its assumption that various formal and sub stantive universals are intrinsic properties of the language-acqui sition system, these proNiding a schema that is applied to data and that determines in a highly restricted way the general form and, in part, even the substantive features of the grammar that may emer upon presentation of appropriate data. A general linguistic theory of the sort roughly described earlier, and elaborated in more detail in the following chapters and in other studies of transformational grammar, must therefore be regarded as a specifc hypothesis, of an essentially rationalist cast, as to the nature of mental structures and processes. See Chomsky ( 1 959b, 1 962b, 1 964) and Katz (forthcoming) for some further discussion of this point. Wen such constrasting views are clearly formulated, we may ask, as an empirical question, which (if either) is correct. There is no a priori way to settle this issue. Where empiricist and rationalist views have been presented with sufcient care so that the question of correctness can be seriously raised, it cannot, for example, be maintained that in any clear sense one is "simpler" than the other in terms of its potential physical realization,3 1 and even if this could be shown, one way or the other, it would have no bearing on what is completely a factual issue. This factual question can be approached in several ways. In particular, restricting ourselves now to the question of language acquisition, we must bear in mind that any concrete empiricist proposal does impose certain condi tions on the f orm of the grammars that can result from application of its inductive principles to primary data. We may therefore ak whether the grammars that these principles can provide, in principle, are at all close to those which we in fact discover when we investigate 54 METODOLOGICAL PRLIMINARIES real languages. The same question can be asked about a concrete rationalist proposal. This has. in the past, proved to be a useful way to subject such hypotheses to one sort of empirical test. 1£ the answer to this question of adequacy-in-principle is poitive. in either case, we c then tur to the question of feasibility: can the inductive procedures (in the empiricist case) or the mechanisms of elaboration and realization of innate schemata (in the rationalist case) succeed in producing grammars within the given constraints of time and access, and within the range of observed unifority of output? In fact, the second question has rarely been raised in any serious way in connection with empiricist views (but cf. Miller, Galanter, and Pribram, 1 960, pp. 145-1 48, and Miller and Chomsky, 1 963. p. 430, for some comment). since study of the frst question has been sufcient to rule out whatever explicit proposals of an essentially empiricist character have emerged in modem discussions of language acquisition. The only proposals that are explicit enough to support serious study are those that have been developed within taxonomic linguistics. It seems to have been demonstrated beyond any reasonable doubt that, quite apart from any question of feasibility, methods of the sort that have been studied in taxonomic linguistics are intrinsically incapable of yielding the systems of grammatical knowledge that must be attributed to the speaker of a language (cf. Chomsky, 1 956, 1 957, 1 964; Postal, I g62 b, I g64a, I g64c; Katz and Postal, I g64' § 5. 5, and many other publications for discussion of these questions that seems un answerable and is, for the moment, not challenged). In general, then, it seems to me correct to say that empiricist theories about language acquisition are refutable wherever they are clear, and that further empiricist speculations have been quite empty and uninformative. On the other hand, the rationalist approach ex emplifed by recent work in the theory of transformational grammar seems to have proved fairly productive, to be fully in accord with what is known about language, and to ofer at least some hope of providing a hypothesis about the intrinsic structure of a language-acquisition system that will meet the condition of adequacy-in-principle and do so in a sufciently § 8. LINGUISTIC THEORY AND LANGUAGE LEARNING 5 5 narrow and interesting way so that the question of feasibility can, for the frst time, be seriously raised. One might seek other ways of testing particular hypotheses about a language-acquisition device. A theory that attributes possession of certain linguistic universals to a language-acquisition system, as a property to be realized under appropriate exteral conditions, implies that only certain kind of symbolic system can be acquired and used as languags by this device. Others should be beyond its language-acquisition capacity. Systems can certainly be invented that fail the conditions, formal and sub stantive, that have been prposed as tentative linguistic uni versals in, for example, Jakobsonian distinctive-feature theory or the theory of transformational grammar. In principle, one might try to determine whether invented systems that fail these condi tions do pose inordinately difcult problems for language lear ing, and do fall beyond the domain for which the language acquisition system i designed. A a concrete example, consider the fact that, according to the theory of transformational grammar, only certain kinds of formal operations on string can appear in grammars -operations that, furthermore, have no a priori justifcation. For example, the peritted operations cannot be shown in any sense to be the most "simple" or "elementar" ones that might be invented. In fact, what might in gneral be considered "elementary operations" on strings do not qualify as gammatical transformations at all, while many of the operations that do qualify are far from elementary, in any general sense. Specifcally, grammatical transformations are necessarily "struc ture-dependent" in that they manipulate substrings only in terms of their assignment to categories. Thus it is possible to formulate a transformation that can insert all or part of the Auxiliary Verb to the left of a Noun Phrase that precedes it, independently of what te length or interal complexity of the strings belonging to these categries may be. It is impossible, however, to formulate as a transformation such a simple operation as refection of an arbitrary string (that is, replacement of any string a l " - af, where each Üg is a single symbol, by a f" ' al), or interchange of the (2n - l)th word with the 2n th word throughout a string of METODOLOGICAL PREUMINARIE arbitrary length, or insertion of a symbol in the middle of a string of even length. Similarly, if the structural analyses that defne transformations are restricted to Boolean conditions on Ana lyzability, as suggested later, it will be impossible to formulate many "structure-dependent" operations as transformations -for example, an operation that will iterate a symbol that is the left mOst member of a cateory (impossible, short of listing all categories of the grammar in the structural analysis), or an operation that will iterate a symbol that belongs to a many rightmost as leftmost categories). Hence, one who proposes this theory would have to predict that although a language might form interrogatives, for example, by interchanging the order of certain categories (as in English), it could not form inter rogatives by refection, or interchange of odd and even words, or insertion of a marker in the middle of the sentence. Many other such predictions, none of them at all obvious in any a priori sense, can be deduced from any sufciently explicit theory of linguistic universals that is attributed to a language-acquisition device as an intrinsic property. For some initial approaches to the very difcult but tantalizing problem of investigating ques tions of this sort, see Miller and Stein ( 1 963), Miller and Norman ( 1 964). Notice that when we maintain that a system is not learnable by a language-acquisition device that mirrors human capacities, we do not imply that this system cannot be mastered by a human in some other way, if treated as a puzzle or intellectual exercise of some sort. The language-acquisition device is only one component of the total system of intellectual structures that can be applied to problem solving and concept foration; in other words, the faculte de langage is only one of the faculties of the mind. What one would expect, however, is that there should be a qualitative diference in the way in which Ü organism with a functional language-acquisition system8 2 will approach and deal with systems that are languagelike ad others that are not. The problem of mapping the intrinsic cognitive capacities of § 8. LINGUISTIC TEORY AND LANGUAGE LEARNING 5 7 an organism and identifying the systems of belief and the organization of behavior that it can readily attain should be central to experimental psychology. However, the feld has not developed in this way. Learning theory has, for the most part, concentrated on what seems a much more marginal topic, namely the question of species-independent regularities in acquisition of items of a "behavioral repertoire" under exprimentally manipulable conditions. Consequently, it has necessarily directed i ts attention to tasks that are extrinsic to an organism's cogitive capacities -tasks that must be approached in a devious, indirect, and piecemeal fashion. In the course of this work, some incidental information has been obtained about the efect of intrinsic cogitive structure and intrinsic organization of behavior on what is learned, but this has rarely been the focus of serious attention (outside of ethology). The sporadic exceptions to this observation (see, for example, the discussion of "instinctual drift" in Brland and Breland, 1 961 ) are quite sugstive, a are many etological studies of lower organisms. The gneral question and its many ramifcations, however, remain in a primitive state. In brief, it seems clear that the present situation with regard to the study of language learing is essentially as follows . We have a certain amount of evidence about the character of the generative grammars that must be the "output" of an acquisition model for languag. This evidence shows clearly that taxonomic views of linguistic structure are inadequate and that knowledge of grammatical structure cannot arise by application of step-by step inductive operations (segmentation, classifcation, substitu tion procedures, flling of slots in frames, association, etc.) of any sort that have yet been developed within linguistic, psy chology, or philosophy. Further empiricist speculations contribute noting that even faintly sugst a way of overcoming the intrinsic limitations of the methods that have so far been proposed and elaborated. In particular, such speculations have not provided any way to account for or even to express the fundamental fact about the normal ue of language, namely the speaker's ability to produce and undertand instantly new METHODOLOGICAL PRELIMINARIE sentences that are not similar to those previously heard in any physically defned sense or in ters of any notion of frames or clases of elements, nor asociated with those previously heard by conditioning, nor obtainable from them by any sort of "gen eralization" known to psychology or philosophy. It seems plain that languag acquisition is based on the child's discovery of what fom a formal point of view is a deep and abstract theory a gnerative gammar of his languag -many of the concepts and principles of which are only remotely related to experience by long and intricate cains of unconscious quasi-inferential steps. P consideration of the character of the grammar that is acquired, the degnerate quality and narrowly limited extent of the available data, the striking unifority of the resulting gamars. and their independence of intelligence, motivation, and emotional state, over wide rangs of variation, leave little hop that much of the structure of the languag can be leared by an organism initially uninformed as to its general character. It is. for the present, impossible to forulate an assumption about initial. innate structure rich enough to account for the fact that grammatical knowledge is attained on the basis of the evidence available to the learner. Consequently, the empiricist efort to show how the assumptions. about a language·acquisition device can be reduced to a conceptual minimum38 is quite mis placed. The real problem is that of developing a hypothesis about initial structure that is sufciently rich to account for acquisition of languag, yet not so rich as to be inconsistent with the known diversity of languag. It is a matter of no concern and of only historical interest tat such a hypothesis will evidently not satisfy the preconceptions about learing that derive from centuries of empiricist doctrine. These preconceptions are not only quite implausible. to begin with, but are without factual support and are hardly consistent with what little is known about how animals or humans construct a "theory of the external world." It is clear why the view that all knowledge derives solely from the senses by elementary operations of association and "gn- § 8. LINGUISTIC THEORY AND LANGUAGE LEARNING 59 eralization" should have had much appeal in the context of eighteenth-century struggles for scientifc naturalism. However, there is surely no reason today for taking seriously a position that attributes a complex human achievement entirely to months (or at most years) of experience, rather than to millions of years of evolution or to principles of neural organization that may be even more deeply gounded in physical law -a position that would, furthermore, yield the conclusion that man is, apparently, unique among animals in the way in which he acquires knowl edge. Such a position is particularly implausible with regard to language, an aspect of the child's world that is a human creation and would naturally be expected to refect intrinsic human capacity in its interal organization. In short, the structure of particular languages may very well be largely determined by factors over which the individual has no conscious control and concering which society may have little choice or freedom. On the bais of the best information now available, it seem reasonable to suppose that a child cannot help constructing a particular sort of transforational grammar to account for the data presented to him, any more than he can control his perception of solid objects or his attention to line and angle. Thus it may well be that the general features of language structure refect, not so much the course of one's experience, but rather the gneral character of one's capacity to acquire knowl edge-in the tradit.ional sense, one's innate ideas and innate principles. It seems to me that the problem of clarifying this issue and sharpening our understanding of its many facets provides the most interesting and important reason for the study of descriptively adequate grammars and, beyond this, the forulation and j ustifcation of a general linguistic theory that meets the condition of explanatory adequacy. By pursuing this investigation, one may hope to give some real substance to the traditional belief that "the principles of grammar form an im portant, and very curious, part of the philosophy of the human mind" (Beattie, 1 788). 60 METODOLOICAL PRELIMINARIES § 9. GENERATIVE CAPACITY AND ITS LINGUISTIC RELEVANCE It may be useful to make one additional methodological obser vation in connection with the topiL discussed in the last few sec tions. Given a descriptive theory of language structure,84 we can distinguish its weak generative capacity from its strong genera tive capacity in the following way. Let u say that a grammar weakly generates a set of sentences and that it strongly generates a set of structural descriptions (recall that each structural de scription uniquely specifes a sentence, but not necessarily con versely), where both weak and strong generation a determined by the procedure f of ( 1 2iv) ¯ (1 3iv) ¯ ( 14iv). Suppose that the linguistic theory T provides the class of grammars Gl, G2,·· · , where G, weakly generates the languag L, and strongly generates the system of structural descriptions I,. Then the class {Ll' L2, · • • } constitutes the weak generative capacity of T and the class {Il, I2, · · · } constitutes the strong generative capacity of T.8G The study of strong gnerative capacity is related to the study of descriptive adequacy, in the sense defned. A grammar is de· scriptively adequate if it strongly generates the correct set of structural descriptions. A theory is descriptively adequate if its strong generative capacity includes the system of structural descriptions for each natural language; otherise, it is descrip tively inadequate. Thus inadequacy of strong generative capacity, on empirical grounds, shows that a theory of language i seriously defective. A we have observed, however, a theory of language that appears to be empirically adequate in terms of strong genera tive capacity is not necessarily of any particular theoretical in terest, since the crucial question of explanatory adequacy goes beyond any consideration of strong generative capacity. The study o f weak geneÏatiNe capacity is of rather marginal linguistic interest. It is important only in those cases where some proposed theory fails even in weak generative capacity -that is, where there is some natural language even the sentences of which cannot be enumerated by any grammar permitted by this theory. In fact, it has been shown that certain fairly elementar § 9· GENERATIVE CAPACITY AND ITS LINGUISTIC RELEVANCE 61 theories (i particular, the theory of context-free phrase-structure grammar and the even weaker theory of fnite-state grammar) do not have the weak generative capacity required for the descrip tion of natural language, and thus fail empirical tests of ade quacy in a particularly surprising way.B8 Frm this observation we must conclude that as linguistic theory progresses to a more adequate conception of grammatical structure, it will have to permit devices with a weak generative capacity that difers, in certain respcts, from that of these severely defective systems. It is important to note, however, that the fundamental defect of these systems is not their limitation in weak gnerative capacity but rather their many inadequacies in strong generative capacity. Postal's demonstration that the theory of context-free grammar (simple phrase-structure gammar) fails in weak generative capacity was preceded by over a half-dozen year of discussion of the strong generative capacity of this theory, which showed con clusively that it cannot achieve descriptive adequacy. Further more, these limitations in strong generative capacity cary over to the theory of context-sensitive phrase-structure grammar, which probably does not fail in weak gnerative capacity. Presumably, discussion of weak generative capacity marks only a very early and primitive stage of the study of generative gam mar. Questions of real linguistic interest arise only when strong gnerative capacity (descriptive adequacy) and, more important, explanator adequacy become the focus of discussion. A observed earlier, the critical factor in the development of a fully adequate theory is the limitation of the class of possible grammars. Clearly, this limitation must be such as to meet empirical conditions on strong (and, a fortiori, weak) generative capacity, and, furthermore, such as to permit the condition of explanatory adequacy to be met when an appropriate evaluation measure is developed. But beyond this, the problem is to impose sufcient structure on the schema that defnes "generative gram mar" so that relatively few hypotheses will have to be tested by the evaluation measure, given primary linguistic data. We want the hypotheses compatible with fxed data to be "scattered" in value, so that choice among tem can be made relatively easily. METHODOLOGICAL PRLIMINAES This requirement of "feasibility" is the major empirical con straint on a theory, once the conditions of descriptive and ex planatory adequacy are met. It is important to keep the require ment of explanatory adequacy and feasibility in mind when weak and strong generative capacities of theories are studied as mathematical questions. Thus one can construct hierarchies of grammatical theories in terms of weak and strong gnerative capacity, but it is important to bear in mind that these hierarchies do not necessarily correspond to what is probably the empirically most signifcant dimension of increasing power of linguistic theory. This dimension is presumably to be defned in terms of the scattering in value of gammars compatible with fxed data. Along this empirically signifcant dimension, we should like to accept the leat "powerful" theory that is empirically adequate. It might conceivably tur out that this theory is extremely powerful (perhaps even universal, that is, equivalent in genera tive capacity to the theory of Turing machines)87 along the dimension of weak gnerative capacity, and even along the dimension of strong generative capacity. It will not necessarily follow that it is very powerful (and hence to b discounted) in the dmension which is ultimately of real empirical signifcance. In brief, mathematical study of formal properties of gammars is, very likely, an area of linguistics of great potential. It has already provided some insight into questions of empirical interest and will perhaps some day provide much deeper insights. But it is important to realize that the questions presently being studied are primarily determined by feasibility of mathematical study, and it is importat not to confuse this with the question of empirical signifcance. 2 Categories and Relations in Syntactic Theor § 1. THE SCOPE OF THE BASE WE now retur to the problem of refning and elaborating the sketch (in Chapter Ï ¿ § 3) of how a generative gammar is organized. Putting of to the next chapter any question as to the adequacy of earlier accounts of grammatical transforations, we shall consider here only the formal properties of the base of the syntactic component. We are therefore concerned primarily with extremely simple sentences. The investigation of generative grammar can proftably begin with a careful analysis of the kind of information presented in traditional grammars. Adopting this as a heuristic procedure, let u Lonsider what a traditional gammar has to say about a simple English sentence such a the following: ( 1 ) sincerity may frighten the boy Concerning this sentence, a traditional grammar might provide information of the following sort: (2) (i) the string (Ï) is a Sentence (S) ; frighten the boy is a Verb Phr;se (VP) consisting of the Verb ( fighten and the Noun Phrase (NP) the boy; sincerity is also an NP; the NP the boy consists of the Determiner (Det) the, followed by a Noun (N) ; the NP sincerity consists of just an N; t he is, furthermore, an Article (Art); may i s a Verbal Auxiliary (Aux) and, furthermore, a Modal (M) . 63 64 CATEGORIES AND RELATIONS IN SYNTACTIC 1EORY (ii) the NP sincerity functions as the Subject of the sentence ( 1 ), whereas the VP frighten the boy functions as the Pred icate of this sentence; the NP the boy functions as the Object of the VP, and the V frighten as its Main Verb; the grammatical relation Subject-Verb holds of the pair (sincerity, frighten), and the grammatical relation Verb Object holds of the pair (righten, the boy) . 1 (iii) the N boy is a Count Noun (as distinct fom the Mass Noun butter and the Abstract Noun sincerity) and Û Common Noun (as distinct from the Prper Noun John and the Pronoun it) ; it is, furthermore, an Animate Noun (as distinct from book) and a Human Noun (as distinct from bee) ; frighten is a Transitive Verb (as distinct from occur), and one that does not freely permit Object dele tion (a distinct from read, eat) ; it takes Progressive Aspect freely (as distinct from know, own); it allows Abstract Subj ects (as distinct from eat, admire) and Human Ob jects (as distinct from read, wear) . It seems to me that the information presented in (2) is, with out question, substantially correct and is essential to any account of how the language is used or acquired. The main topic I should like to consider is how information of this sort can be formally presented in a structural description, and how such structural descriptions can be generated by a system of explicit rules. The next three subsections (§§ 2 . 1 , 2. 2, 2. 3) discuss these questions in connection with (2i), (2ii), and (2iii), respectively. § 2. ASPECTS OF DEEP STRUCTURE § 2. I. Categorization The remarks given in (2i) concern the subdivision of the string (1 ) into continuous substrings, each of which is assigned to a certain category. Information of this sort can be represented by a labeled bracketing of ( 1 ), or, equivalently, by a tree·diagram such a (3) . The interpretation of such a diagram is transparent, § 2. ASPECT OF DEEP STUCTUR (3) s � NP Aux VP | � N M V NP | | ` � sincerity ma y frighten Det ' N | the boy and has been discussed frequently elsewhere. If one assumes now that ( Ï) is a basic string, the structure represented as (3) can be taken as a frst approximation to its (base) Phrase-marker. A grammar that generates simple Phrase-markers such as (3) may be based on a vocabulary of symbols that includes both formatives (the, boy, etc.) and category symbol (S, NP, V, etc.). The formatives, furtherore, can be subdivided into lexical i tems (sincerity, boy) and grammatical items (Perfect, Possessive, etc. ; except possibly for the, none of these are represented in the simplifed example given). A question arises at once as to the choice of symbols in Phrase markers. That is, we must ask whether the formatives and category symbols used in Phrase-markers have some language independent characterization, or whether they are just con venient mnemonic tags, specifc to a particular grammar. In the case of the lexical formatives, the theory of phonetic distinctive features taken toether with the full set of conditions on phonological representation does, in fact, give a languag independent signifcance to the choice of symbols, though it is by no means a trivial problem to establish this fact (or to select the proper universal set of substantive phonetic features) . I shall assume, henceforth, that an appropriate phonological theory of this sort is established and that, consequently, the lexical forma tives are selected in a well-defned way from a fxed universal set. The question of substantive representation in the case of the grammatical formatives and the category symbols is, in efect, the tradi tional question of universal grammar. I shall assume that 66 CATGORIES AND RELATIONS IN SYNTACTIC TEORY these elements too are selected from a fxed, universal vocabulary, although this asumption will actually have no signifcant efect on any of the descriptive material to b presented . There is no reason to doubt the importance or rasonableness of the study of this question . It is gnerally held to involve extrasyntactic considerations of a sort that are at present only dimly perceived . This may very well be true . However, I shall later suggest several general defnitions that appear to be correct for English and for other cases with which I am acquainted.2 The natural mechanism for generating Phrase-markers such as (3) is a system of rewriting rules. A rewriting rule is a rule of the form (4) A � Z/X ¯ Y where X and Y are (possibly null) strings of symbols, A is a single category symbol, and Z is a nonnull string of symbols. This rule is interpreted as asserting that the category A is realized a the string Z when it is in the environment consisting of X to the left and Y to the right. Application of the rewriting rule (4) to a string . . . XAY . convert this to the string . . . XZY . . . . Given a grammar, we say that a sequence of strings is a W-derivation of V if W is the frst and V the last string in the sequence, and each sting of the sequence is derived from the one preceding it by application of one of the rewriting rules (with an ordering condition to be added later). Where V is a string of formatives, we say that a W-derivation of V is terminated. We call V a terminal string if there is an IS I-derivation of lVI, where S is the designated initial symbol of the grammar (representing the category "Sentence"), and I is the boundar symbol (regarded as a grammatical formative). Thus we construct a derivation of a terminal string by successively applying the rewriting rules of the grammar, beginning with the string IS/, until the fnal string of the derivation consists only of formatives and therefore no further rewriting is possible. If several other conditions are imposed on the system of rewriting rules,S it is easy to provide a simple method for assigning a unique and appropriate Phrase marker to a terinal string, given its derivation. Thus a system § 2. APECT OF DEEP STUCURE of rewriting rules, appropriately constrained, can serve as a part of a generative gammar. A un ordered set of rewriting rules, applied in the manner described loosely here (and precisely elsewhere), is called a constituent structure grammar (or phrase structure grammar). The grammar is, furthermore, called context-free (or simple) if in each rule of the form (4), X and Y are null, so that the rules apply independently of context. As noted earlier (pp. 60 £., 208) , the formal properties of constituent structure grammars have been studied fairly intensively during the past few years; and it has also been shown that almost all of the nontransforational syntactic theories that have been developed within modern lin guistics, p ure or applied, fall within this framework. In fact, such a system is apparently what is implicit in moder taxonomic ("structuralist") grammars, if these are reformulated as explicit systems for presenting grammatical information (but see note 30, Chapter 1) . The inadequacy of such systems as grammars for natural languages seems to me to have been established beyond any reasonable doubt,4 and I shall not discuss the issue here. It seems clear that certain kinds of gammatical information are presented in the most natural way by a system of rewriting rules, and we may therefore conclude that rewriting rules consti tute part of the base of the syntactic component. Furthermore, we shall assume that these rules are arrangd in a linear sequence, and shall defne a sequential derivation as a derivation formed by a series of rule applications that preserves this ordering. Thus, suppose that the grammar consists of the sequence of rules Rt, , R, and that the sequence 181 , lXI', ' ' ' , IXmf is a derivation of the terminal string Xm• For this to be a sequential derivation, it must be the case that if rule R, was used to form line IX J I from the line that precedes it, then no rule Rk (for k > Î) can have been used to form a line "X," (for I < J) from IXI_I I. We stipulate now that only sequential derivations are generated by the sequence of rules constituting this part of the base.C To provide a Phrase-marker such as (3), the base component might contain the following sequence of rewriting rules : 68 CATEGORIES AND RELATIONS IN SYNTACIC THEORY (5) (I) S- NP"Aux'" VP VP - V'"NP NP - Det'"N NP - N Det - the Aux - M (ll) M - may N - sincerity N - bo V - frighten Notice that the rules ( 5), although they do sufce to generate (3), will also generate such deviant string as boy may frighten the sincerity. This is a problem to whicl we shall turn in § 2 . 3 . There is a natural distinction in (5) between rules that in troduce lexical formatives (class (ll» and the others . In fact, we shall see in § 2. 3 that it is necessary to distinguish these sets and to assig the lexical rules to a distinct subpart of the base of the syntactic component. In the case of the inforation in ( 2i), then, we see quite clearly how it is to be formally represented, and what sorts of rules are required to generate these representations. § Ü• ×- Functional notions Turing now to (2ii), we can immediately see that the notions in question have an entirely diferent status . The notion "Sub ject," as distinct from the notion "NP, " desigates a grammatical function rather than a grammatical category . It is, in other words, an inherently relational notion. We say, in traditional terms, that in (Ï) sincerity is an NP (not that it is the NP of the sentence). and that it is (functions as) the Subject-of the sentence ( not that it i a Subject). Functional notions like "Subject," "Predicate" are to b sharply distinguished from categorial notions such as "Noun Phrae," "Verb." a distinction that is not to be obscured by the occasional use of the same term for notions of both kinds. Thus it would merely confuse the issue to attempt to represent the information presented in (2ii formally . by extending the § 2. ASPECTS OF DEEP STUCUR 69 (6) S � Subject Aux Predicate ' ' ' NP M VP ' ' � N may Main Verb Object | ' | sincerity V NP | � frighten Det N ' | the boy Phrase-marker (3) to (6), adding the necessary rewntmg rules to (51) . This approach is mistaken in two ways. For one thing, it confuses categorial and functional notions by assigning categorial status to both, and thus fails to express the relational character of the functional notions. For another, it fails to observe that both (6) and the grammar on which it is based are redundant, since the notions Subject, Predicate, Main-Verb, and Object, being relational, are already represented in the Phrase-marker (3), and no new rewriting rules are required to introduce them. It is necessary only to make explicit the relational character of these notions by defning "Subject-of," for English, as the relation holding between the NP of a sentence of the form NP'" Aux"" VP and the whole sentence,6 "Object-of" as the relation between the NP of a VP of the form V""NP and the whole VP, etc. More generally, we can regard any rewriting rule as defning a set of grammatical functions, in this way, only some of which (namely, those that involve the "higher-level, " more abstract grammatical categories) have been provided, tradi tionally, with explicit names. The fundamental error of regarding functional notions as categorial is somewhat masked in such examples as ( 6), in which there is only a single Subject, a single Object, and a single Main Verb. In thi case, the relational information can be supplied, CATEGORIES AND RELATIONS IN SYNTACTIC TEORY intuitively, by the reader. But consider such sentences as (7), in which many grammatical functions are realized, several by the same phrase: (7) ( a) John was persuaded by Bill to leave (b) John wa peruaded by Bill to be examined (c) what disturbed John was being regarded as incompetent In (7a), John is simultaneously Object.of persuade ( to leave) and Subject-of leave; in (7b), John is simultaneously Object-of per suade (to be examined) and Object-of examine; in (7c), John is simultaneously Object-of disturb, Object-of regard (a incom petent), and Subject-of the predication a incompetent . In both (7a) and (7b), Bill is the ("logical") Subject-of the Sentence, rather than John, which is the so-called "grammatical" Subject of the Sentence, that is, the Subject with respect to the surface structure (cf. note 32). In such cases as these, the impossibility of a categorial interpretation of functional notions becomes at once apparent; correspondingly, the deep structure in which the signifcant grammatical functions are represented will be very diferent from the surface structure. Examples of this sort, of course, provide the primary motivation and empirical j ustifca tion for the theory of transformational grammar. That is, each sentence of ( 7) will have a basis consisting of a sequence of base Phrase-markers, each of which represents some of the semanti cally relevant information concerning grammatical function. Returning now to the main question, let us consider the problem of presenting information about grammatical function in an explicit and adequate way, restricting ourelves now to base Phrase-markers. To develop a uniform approach to this ques tion, we may proceed as follows. Suppose that we have a sequence of rewriting rules, such as ( 5), including in particular the rule (8) A �x Associated with this rule is each grammatical function ( 9 ) [B, A] where B is a category and X ¯ YBZ, for some Y, Z (possibly null).7 Given a Phrase-marker of the terminal string W, we say § 2. ASPECTS OF DEEP STRUCTUR that the substring U of W bears the grammatical relation [B, A] to the substring V of W if V is dominated by a node labeled A which directly dominates YBZ, and U is dominated by this occurrence of B.8 Thus the Phrase-marker in question contains the subconfguration ( 1 0) . l particular, given the Phrase-marker ( 1 0) A � Y B Z / / / W' = . . . I � I v (3) generated by the rules ( 5), we should have the result that sincerity bears the relation [NP, S] to sincerity may frighten the boy, frighten the boy bears the relation [VP, S] to sincerity may frighten the boy, the boy bears the relation [NP, VP] to frighten the boy, and frighten bears the relation [, VP] to frighten the boy. SLppobe further that we propose the following general defni tions: ( l l ) (i) Subject-of: [NP, S] (ii) Predicate-of: [P, S] (iii) Direct-Object-of: [NP, VP] (iv) Main-Verb-of: [V, VP] In this case, we can now say that with respect to the Phrase marker ( 3) generated by the rules (5), sincerity is the Subject¯of the sentence sincerity may frighten the boy and frighten the boy is i ts Predicate; and the boy is the Direct-Object¯of the Verb Phrase frighten the boy and frighten is it Main-Verb. With these defnitions, the inforation presented in the redundant representation (6) is derivable directly from ( 3), that is, from the grammar (5) itelf. These defnitions must be thought of as CATGORIES AND RELTIONS IN SYNTACTIC TEORY belonging to general linguistic theory; in other words, they form part of the general procedure for assigning a full structural description to a sentence, given a grammar (the procedure f of ( 1 2iv), ( 1 3iv), ( 14iv) in § 6, Chapter 1) . In suc examples as (7), the gammatica functions will also be given directly by the system of rewriting rules that generate the base Phrase-markers that underlie these sentences, though these gramatical functions are not repreented in the confgurations of the surface structures in these cases. For example (details aside), the basis for (7a) will contain base Phrase-markers for the strings Bill pers'd ed John Sentence, John left, and these base Phrase·markers present the semantically relevant functional in formation exactly as in the case of (3). Notice that the same gammatical function may be defned by several diferent rewriting rules of the base. Thus suppose that a gramar were to contain the rewriting rules ( 1 2) (i) S � Adverbial"NP" (Naturally, John will leave) Aux"VP (ii) S � NP"Aux"VP (John will leave) (iii) VP � V"NP (examine Bill) (iv) VP � V (leave) (v) VP � V"NP"Sentence (persuade Bill that John left) (vi) VP � Copula"Predicate (be President) (vii) Predicate � N (President) Then Subject-of is defned by both (i) and (ii), so that John is Subject.of the sentences accompanying both (i) and (ii); Object of is defned by both (iii) and (v), so that Bill is the Object-of the Verb Phrases given as examples to both (iii) and (v) ; Main-Verb of is defned by (iii), (iv), and (v), so that examine, leave, per suade are the Main-Verbs of the accompanying examples. But notice that "President" is not the Object-of John is President, if the rules are as in ( 1 2). It is defnitions of this sort that were presupposed in the discussion of persuade and expect in Chap ter Ïy § 4. Notice that the general signifcance of the defnitions ( 1 1) § 2. ASPECTS OF DEEP STRUCTURE 7 3 depends on the assumption that the symbols S, NP, VP, N, and V have been caracterized as grammatical universals. We shall re turn to this question later. Quite apart from this, it is likely that these defnitions are too restricted to serve as general explications for the traditionally designated gammatical functions in tat they assume to narow a substantive specifcation of the for of grammar. They can be generalized in various ways but I do not, at the moment, se any strong empirical motivation for one or another specifc extension or refnement (but see § 2e 3 e4). In any event, these questions aside, it i s clear that information concerning gramatical functions of the sort exemplifed in (2ii) can be extracted directly from the rewiting rules of the base, without any necessity for ad hoc extensions and elabora : tions of these rules to provide specifc mention of grammatical function. Such extensions, aside from their redundancy, have the defect of failing to express properly the relational character of the functional notions and are thus useless in all but the simplest cases. However, we have not yet exhausted the information presented in (2ii). Thus it is still necessary to defne grammatical relations of the sort that hold between sincerity and frighten (Subject Verb) and between frighten and the boy (Verb-Object) in ( 1 ) . Such relations can be defned derivatively in ters of te func tional notions suggested earlier. Thus Subject-Verb can be de fned as the relation between the Subject-of a Sentence and Main-Verb-of the Predicate-of the Sentence, where Subject.of, Main-Verb-of, and Pedicate-of are the notions of (1 1 ) ; and Verb-Object can be defned as the relation between the Main Verb-of and the Direct-Object-f a VP. However, tere is still something missing in this account. Thus we have no basis, a yet, for distinguishing the legitimate and traditionally recognized gammatical relation Subject-Verb, as j ust defned, from the irrelevant pseudorelation Subject-Object, whic is defnable just as easily in the sae terms. Traditional grammar seems to defne such relations where there are selectional restrictions governing the paired categories. Thus te choice of Main-Verb is deter- 7 4 CATEGORIE AND RELTIONS IN SYNTACTIC THEORY mined by the choice of Subject and Object, tough Subject and Object are in general chosen independently of one another and, corespondingly, have no gammatical relation of the sort in question holding between them. I shall defer the discussion of selectional relations until § 4. 2, and at that point we can return to the question of grammatical relations . But in any event, it is fairly clear that nothing essentially new is involved here beyond the rules tat generate strings and Phrase-markers. In summary, then, it seems unnecessary to extend the system of rewriting rules in order to accommodate information of the sort presented in (2ii). With appropriate general defni tions of the relational notiLns involved, this inforation can be extracted directly from Phrase-markers that are generated by simple re wi ting rule such as (5) and ( 1 2) . This information is already contained, implicitly, in the system of elementary rewriting rules. Representations such as (6) and new or elaborated rewri t· ing rules to generate them are unnecessary, as well as mislead· ing and inapprpriate. Finally, I should like to call attention, once again, to the fact that various moifcations and extensions of these functional notions are pssible and that it is important to fnd empirical motivation for such improvements. For example, the char acterization might be sharpened smewhat in terms of several notions that will be useful later on. Suppse again that we have a base grammar consisting of a sequence of rewriting rules, and that (as in (5» we have distinguished lexical rules (such as (511», which introduce lexical formatives, from the others. We shall see later that this distinction is formally quite clearly marked . A categry that appears on the left in a lexical rle we shall call a lexical categor; a lexical category or a category that dominates a string . . , X . . " where X is a lexical category, we shall call a major categor. Thus in the gammar (5), the categories N, V, and M are lexical categories,9 and all categLries except Det (and possibly M and Aux ¯see note 9 ) are major categories . It would, then, be in accord with traditional usage to limit the functional notions to major categories. We shall consider a further refne ment in the fnal paragaph of § 2 . 3.4. § 2. ASPECTS OF DEEP STUCTURE 7 5 § 2 . J. Syntactic features § 2 . J. I. The problem. Information of the sort presented in (2iii) raises several difcult and rather vexing questions. Fi rst, it is not obvious to what extent this inforation should be pro vided by the syntactic component at all. Second. it is an inter est i ng question whether or to what extent semantic considera tions are relevant in deterining such sub categorizations as those involved in (2iii). These are distinct questions. though they are often confused. They are connected only in that if the basis for making the distinctions is purely syntactic. then surely the information must be presented in the syntactic component of the gmmar. We might call these the questions of presenta tion and justifcatio, respectivey. As far as the question of j ustifcation is concerned, a linguist with a serious interest in smantics will presumably attempt to deepen and extend syntactic analysis to te pint where it can provide the information concerning subcategorization, instead of relegating this to unanalyzed semantic intuition, there being. for th e moment. no other available proposal as to Û semantiL basis for making the necessary distinctions. Of course, it is an open question whether this attempt can succeed. even in part. I shall be concerned here only with th e question of presenta tion of inforation of the sort given in (2iii) . I am assuming throughout that the semantic compnent of a generative gram mar. like the phonoloical component. is purely interpretive. It follows that all information utilized in semantic interpretation must be presented in the syntactic component of the grammar (but cf. Chapter 4. § l . 2) . Some of te problems involved in presenting this information will be explored later. Although te quetion of justifcation of subcategrizations such as those of (2iii) is beyond the scope of the present discus sion. it may nevertheless be useful to touch on it briefy. What is at stake. essentially. is the status of such expressions as ( 1 3) (i) the boy may frighten sincerity (ii) sincerity may admire the boy (iii) John amazed the injustiLe of that decision CATEGORIES AND RLATIONS IN SYNTACC TEORY (iv) the boy elapsed (v) the boy was abundant (vi) the harvest was clever to agee (vii) John is owning a house (viii) the dog looks barking (ix) John solved the pipe (x) the book disprsed It is obvious to anyone who knows English that these expressions have an entirely diferent status from such sentences a ( 1 4) (i) sincerity may frighten the boy (=( 1 » (ii) the boy may admire sincerity (iii) the injustice of that decision amazed John (iv) a week elapsed (v) the harvest was abundant (vi) the boy was clever to agee (vii) John owns a house (viii) the dog looks terrifying (ix) John solved the problem (x) the boys dispersed The disti n ctio n between (1 3) and ( 1 4) is not at issue. and clearly must be accounted for somehow by an adequate theory of sen tence interpretation (a descriptively adequate grammar) . The expressions of ( 1 3) deviate in some manner (not necessarily all in the same manner) from the rules of English.1° If interpretable at all. they are surely not interpretable in the manner of the corresponding sentences of ( 1 4) . Rather. it seems that inter prtations are imposed on them by virtue of analogies that they bear to nondeviant sentences. There are fairly clear-cut cases of violation of purely syntactic rules. for example. ( 1 5) (i) sincerity frighten may boy the (ii) boy the frighten may sincerity and standard examples of purely semantic (or "pragmatic") incongruity. for example@ § 2. ASPEc OF DEEP STUCUR ( 1 6) ( i) oculists are gnerally better trained than eye-doctor (ii) both of John's parents are married to aunts of mine 7 7 (iii) I' m memorizing the score of the sonata I hope to com pose some day (iv) that ice cube that you fnally managd to melt just shattered (v) I knew you would come, but I was wrong The examples of ( 1 3), however, have a borderline character, and it is much less clear how their aberrant status is to be explained. In other words, we must face the problem of determining to what extent the results and methods of syntactic or of semantic anal y sis can be extended to account for the deviance and inter pretation of these expressions. It ges without saying that the same answer may not be appropriate in all of these cases, and that purely semantic or purely syntactic considerations may not provide the answer in some particular case. In fact, i t should not be taken for granted, necessarily, that syntactic and semantic considerations can be sharply distinguished . Several sugstions have been made a to how syntactic con siderations can provide a subclassifcation of the appropriate sort. These involve the notion of "degee of gammaticalness, " along various dimensions, and concrete proposals involve tech niques of subclassifying based on distributional similarities. Al though these notions have been advanced only very tentatively, it seems to me that they have some plausibility.ll The only sug gstion as to possible semantic grounds for these distinctions has been that they are based on language-independent semantic absolutes -that in each case, the deviance is attributable to violation of some linguistic universal that constrains the form of the semantic component of any generative grammar. It is possible that this is the right answer; furthermore, there is no reaon why some combination of these two extreme approaches should not be attempted. In any case, what is needed is a systematic account of how application of the devices and methods appropri ate to unequiv ocal cases can be extended and deepened to provide a basis for CATEGORIES AND RELATIONS IN SYNTACIC TEORY explaining the status of such expressions as those of ( 1 3), and an account of how an ideal listener might assign an interpretation to such sentences, where possible, presumably on the basis of analogy to nondeviant cases. These are real and important ques tions. A descriptively adequate gammar must account for such phenomena in terms of the structural descriptions provided by it syntactic and semantic component, and a general linguistic theory that aims for explanatory adequacy must show how such a grammar can develop on the basis of data available to the languag learer. Vage and unsupported assertions about the "semantic basis for syntax" make no contribution to the under standing of these questions. Proceeding now from the question of justifcation to the ques· tion of presentation, we must determine how a grammar can provide structural descriptions that will account for such phe nomena as those exemplifed. A prio1i there is no way to decide whether the burden of presentation should fall on the syntactic or semantic component of the generative grammar. If the forer, we must design the sytactic component so that it does not pro· vide for the sentences of ( 1 3) directly, but assigns them Phrase· marker only by virtue of their structural similarities to such perfectly well-formed sentences as those of ( 1 4), perhaps in the manner described in the references in note Ì J . Thus the syntactic component will oprate in terms of selectional restrictions in volving such categories as animateness and abstraLtness, and will characterize ( 1 3i), for example, as a string gnerated only by relaxing certain of these restrictions. Alternatively, if we conclude that the semantic component should carry the burden of accounting for these fact, we can allow the syntactic component to gnerate the sentencs of (14) as well as those of ( 1 3) , with no distinction of grammaticalness, but with lexical items specifed in such a way that rules of the semantic component will determine the incongruity of the sentences of ( 1 3) and the manner in which they can be interpreted (if at all) . Either way, we fac a well-defned problem, and it is reasonably clear how to proceed to examine it. I shall, for the present, accept the position of the references of note 1 Îy assuming that the notion "scale of gram- § 2. ASPEC OF DEEP STUCUR 7 9 maticalness" will be relevant to semantic interpretation, that a distinction should b made between ( 1 3) and ( 1 4) by rules of the syntactic component, and that the sentences of ( 1 3) are as· signed Phrase-markers only by relaation of certain syntactic conditions. Later on, I shall t to indicate the precise point at which this decision afects the form of the syntactic component, and shall discuss briefy some possible alternatives. § 2.3. 2. Some formal similarities between s ntax and phonol og. Consider now how information of the sort given in (2iii) can be presented in explicit rules. Note that this information con cerns subcategorization rather than "branching" (that is, analysis of a category into a sequence of categories, as when S is analyzed into Np Aux VP, or NP into Det N). Furthermore, it seems that the only categories involved are those containing lexical formatives as member. Hence, we are dealing with a rather restricted part of grammatical structure, and it is important to bear this in mind in exploring appropriate means for presenting these facts. The obviou suggstion is to deal with subcategorization by rewriting rules of the typ described in § 2. 2, and this was the assumption made in the frst attempts to formalize generative grammars (cf. Chomsky, 1 951 ,12 1 955, 1 957 ) . However, G. H. Matthews, in the coure of his work on a generative grammar of Grman in 1 95 7 -1 958, pointed out that this assumption was in correct and that rewriting rules are not the appropriate device to efect subcategorization of lexical categories. "" The difculty is that this subcategorization is typically not strictly hierarchic, but involves rather cross classifcation. Thus, for example, Nouns in English are either Proper (John, Egpt) or Common (bo, book) and either Human (John, boy) or non-Human (Egypt, book). Certain rules (for example, some involving Determiner) apply to the Proper/Common distinction; others (for example, rules involving choice of Relative Pronoun) to the Human/non Human distinction. But if the subcategorization is given by re writing rules, then one or the other of these distinctions will have to dominate, and the other will be unstatable in the natural 80 CATEGORIES AND RELATIONS IN SYNTACC TEORY way. Thus if we decide to take Proper/Common as the major distinction, we have such rules as ( 1 7 ) N . Proper N . Common Proper . Pr-Human Proper . Pr-nHuman Common " C -Human Common " C -nHuman where the symbols "Pr-Human," "Pr-nHuman," "C-Human, " and "C-nHuman" are entirely unrelated, as distinct from one another as the symbols "Noun, " "Verb," "Adjective," and "Modal. " In this system, al though we can easily state a rle that applies only to Proper Nouns or only to Common Nouns, a rule that applies to Human Nouns must be stated in terms of the unrelated categories Pr-Human and C-Human. This obviously indicates that a generalization is being missed, since this rule would now be no simpler or better motivated than, for example, a rule applying to the unrelated categories Pr-Human and Abstract Nouns. A the depth of the analysis increases, problems of this sort mount to the point where they indicate a serious inadequacy in a grammar that consists only of rewriting rules. Nor is this particular difculty overcome, as many other are, when we add transformational rules to te grammar. Formally, this problem is identical to one that is familiar on the level of phonology. Thus phonological units are also cross classifed, with respect to phonological rules. There are, for example, rules that apply to voiced consonants [b], [z], but not to unvoiced consonants [p], [s], and there are other rules that apply to continuants [s], [z] , but not to stops [p], [b], and so on. For this reason it is necessary to regard each phonological unit as a set of features, and to design the phonological component in such a way that each rule applies to all segments containing a certain feature or constellation of features. The same solution suggests itself in the case of the sytactic problem that we are now facing, and it is this method of dealing with the problem that I shall elaborate here. § 2. ASPECIS OF DEEP STRUCTURE Before we turn to the use of features on the syntactic level. let us review briefy the operation of the phonological com ponent (cf. Halle. 1 959a. 1 959b. 1 962a. 1 964. for discusion of this question). Each lexical formative is represented as a sequence ef segments. each segent being a set of features. In other words. each lexical formative is represented by a distinctive-feature matrix in which the columns stand for successive segments. and the rows for particular features. An entry in the ith column and it h row of such a matrix indicates how the ith segment is specifed with respect to the i t h feature. A particular entry may indicate that the segent in question is unspecifed with rspect to the feature in question. or that it is positively specifed with respect to this feature. or that it is negatively specifed with respect to this feature. We say that two segments are distinct j ust in case one is positively specifed with respect to a feature with respect to which the other is negatively specifed. and. more generally. that two matrices with the same number of columns are distinct if the ith segent of one is distinct in this sense from the it segent of the other. for some i. Suppose that ( 1 8) A � Z/X ¯ Y is a phonological rule. where A. Z. X. and Y are matrices. and A and Z are. furthermore. segments ( matrices with just a single column). This is the typical form of a phonological rule. We shall say that the rule ( 1 8) is applicable to any string WX'A'Y' V. where X'. A'. Y' are matrices wi th the same Ïumber of columns as X. A. Y. respectively. and X' A'Y' is not distinct from XAY (actually. qualifcations are necessary that do not concern us here -cf. HalIe and Chomsky. forthcoming. for discussion). The rule ( 1 8) converts the string WX'A'Y' V to the string WX'Z'Y'V. where Z' is the segment consisting of the feature specifcations of Z together with all feature specifcations of A' for features with respect to which Z is unspecifed. A an illustration of some of these notions. consider this phonological rule: 82 CATEGORIES AND RELATONS IN SYNTACIC 'EORY ( 1 9) [+continuant] � [+voiced] j -[+voiced] This will convert [sm] into [zm], [fd] into [vd], [�g] into [zg], etc., but i t will not afect est] or [Pd], for example, 1 4 These conventions (which can be simplifed and gneralized in ways that do not concer us here) alIow us to apply rules to any class of segents specifed by a given combination of features, and thus to make use of the cross classifcation of segments provided by the feature representation. These notions can be adapted without essential change to the representation of lexical categories and their members, providing a very natural solution to the cross-classifcation problem and, at the same time, contributing to the general unity of gram matical theory. Each lexical formative will have associated with it a set of syntactic features (thus boy will have the sytactic features [+Common], [+Human], etc.) . Furthermore, the symbols representing lexical categories (N, V, etc.) will b analyzed by the rules into complex symbol, each complex symbol being Û set of specifed syntactic features, just as each phonological seg ment is a set of specifed phonological features. For example, we might have the following grammatical rules : (20) (i) N � [+N ±Common] (ii) [+Common] � [±Count] (iii) [+Count] � [±Animate] (iv) [-Common] � [±Animate] (v) [+Animate] � [±Human] (vi) [-Count] � [±Abstract] We interpret rule (20i) as asserting that the symbol N in a line of a derivation is to be replaced by one of the two complex symbols [+N, +Common] or [+N, -Common]. The rules (20ii-20vi) oprate under the conventions for phonological rules. Thus rule (20ii) asserts that any complex symbol Q that is already specifed as [+Common] is to be replaced by the complex symbol con tinin g all of te features of Q along with either the feature spcifcation [+Count] or [-Count] . The same is true of the other rules that oprate on complex sybols. § 2. ASPECT OF DEEP STRUcUR The total efect of the rules (20) can be represented by te branLhing diagram (2 1 ). In tis re p resentation, each node is (2 1 ) Common Count � Animate Abtract �- + Human book virtue dirt �- boy dog Animate � Human Egpt A - John Fido labeled by a feature, and the lines are labeled + or -. Each maximal path corresponds to a category of lexical items; an element of this category has the feature [a ([ ~ + or -) if and only if one of the lines constituting this path is labeled [ and descends from a node labeled F. Typical members of the cate gories defned by (20) are given at the terminal points of (2 1 ) . A system of complex symbol rules need not be representable by a branching diagram of this sort. For example, the categries defned by te rules (20) a also defned by the rules (22), but in this case there is no representing branching diagram. (2 2) (i) N � [+N, ±Animate, ±Common] (H) [+CommonJ � [±Count] ( . . . ) [ C ] [±Abstract ] ÌÎ1 - ount � A . - mmate (iv) [+Animate] � [±Human] If we were to require representability in a branching diagram as a forml condition on these rules, then (22) would be excluded. In this case, the rules could j ust as well be presented in the form (2 1) as the form (20). In any event, with rules of this sort that introduce and elaborate complex symbols, we can develop the full set of lexical categories. CATGORIES AND RELATIONS IN SYNTACTIC TEORY § 2. J. J. GCneral structure of the base component. We now modify the description of the base subcomponent that was presented earlier, and exemplifed by (5), in the following way. In addition to rewriting rules that apply to category symbols and that generally involve branching, there are rewriting rules such as (20) that apply to symbols for lexical categories and that introduce or operate on complex symbols (sets of specifed syntaLtic features). The gammar will now contain no rules such as those of (511) that introduce the formatives belonging to lexical categories. Instead, the base of the grammar will con tain a lexicon, which is simply an unordered list of all lexical formatives. More precisely, the lexicon is a set of lexical entries, each lexical entry being a pair (D. C), where D is a phonological distinctive feature matrix "spelling" a certain lexical formative and C is a collection of specifed syntactic features (a complex sybol).11 The system of rewriting rules will now generate derivations terminating with strings that consist of grammatical formatives and complex symbols. Such a string we call a preterminal string. A terminal string is formed from a preterminal string by insertion of Û lexical formative in accordance with the follo¾ing lexical rule: If Q i s a complex symbol of a preterminal string and (D, C) is a lexical entry, where C i s not distinct from Q, then Q can be replaced by D. We now extend the fundamental notion is a that relates strings to categries (for example, the boy is an NP in (3» in the follow ing way. We say that in the terminal string formed by replacing the complex symbol Q by the formative D of the lexical entry (D, C), the forative D is an [aF (equivalently, is dominated by [aF) if [a is part of the complex symbol Q or the complex sybol C, where a is ei ther + or ¯ and F is a feature (ut cf. note 1 5). We also extend the general notion "Phrase-marker" in such a way that the Phrase-marker of a terminal string also con tains the new information. With this extension, a Phrase-maker can naturally no longr be represented by a tree-diagram, as § �. ASPEC OF DEEP STRUCURE before, since it has an addi tional "dimension" at the level of subcategorization. A a concrete example, consider again the sentence sincerity may frighten the boy (=(1». Instead of the grammar (5) we now have a grammar containing the branching rules ( 51 ), which I repeat here as (�3), along with the subcategorization rules (20), repeated as (�4)' and containing a lexicon with the entries (25). It is to be undertood, here and later on, that the i talicized i tems stand for phonological distinctive feature matrices, that is, "spelling" of fonatives. ( �3) S � NP"Aux"VP VP � V"NP NP � Det"N NP � N Det � the Aux � M (2 4 ) ( i) N � [+N, ±Common] (ii) [+Common] � [±Count] (iii) [+Count] � [±Animate] (iv) [-Common] � [±Animate] ( v) [+Animate] � [±Human] (vi) [-Count] � [±Abstract] (25) (sincerity, [+N, -Count, +Abstract]) (boy, [+N, -Count, +Common, +Animate, +Human]) (may, [+M]) We shall have more to say about these rules and lexical entries later, and tey will still undergo signifcant revision. These rules allow us to generate the preteninal string (26) [+N, -Count, +Abstract]"M"Q the"[+N, +Count, +Animate, +Human], where Q is the complex symbol into which V is analyzed by rules that we shall discuss directly . The lexical rule (which, since it is perfectly general, need not be stated in any grammar -in other words, it constitutes part of the defnition of "derivation") 86 CATEGORIES AND RELATIONS IN SYNTACTIC TEORY now allows us to insert sincerity for the frst complex symbol and boy for the last complex symbol of (26) and, a we shall see, to in sert frighten for Q (and may for M ¯cf. note 9) . Except for the case of frighten, the information about the sentence ( 1 ) that is given in (2) is now explicitly provided in full by the Phrase marker generated by the grammar consisting of the rles (23), (24), and the lexicon (25). We might represent this Phrase marker in the form shown in (27). If the lexicon includes ad- NP I N � [-Count] [ +Common] [ + A J � sincerity s T Aux VP I __ M V NP I I __ may Q Det N I I -- fTighten the [+Count] [+Common] I [+Animate] I {+Human] boy ditional specifc information about the lexical items that appear in (26), this information will also appear in the Phrase-marker, represented in terms of features that appar in the Phrase marker in a position dominated by the lexical categories N and V and dominating the formative in question. Given this Phrase-maker, we can derive all of the information (2i) and (2iii), which concerns assignment of substrings to categories, in terms of the relation is a; and the functional in formation (2ii) is derivable from the Phrase-marker in the manner described in § 2 . 2. We shall retur in Chapter 4 , § 2 to questions concering the proper formulation of lexical entries. However, we can see im mediately that separating the lexicon from the system of re witing rules has quite a number of advantages . For one ting, many of the grammatical properties of formatives can now be specifed directly in the lexicon, by association of syntactic features with lexical formatives, and thus need not be represented § 2. ASPEC OF DEEP STUCUR in the rewriting rules at all. In particular, morphological prop erties of various kinds can be treated in this way -for example, membership of lexical items in derivational classes (declensional clases, strong or weak verbs, nominalizable adjectives, etc.) . Since many such properties are entirely irrelevant to the functioning of the rules of the base and are, furthermore, highly idiosyncratic, the gamar can be signifcantly simplifed if they are excluded from the rewriting rules and listed in lexical entries, where they most naturally belong. Or, returning to (2iii), notice that it is now unnecessary to use rewriting rules to classify Transitive Verbs into those that do and those that do not normally permit Object deletion. Instead, the lexical entries for read, eat, on the one hand, and frighten, keep, on the other, will difer in specifcation for the particular syntactic feature of Object deletion, which is not mentioned in the rewriting rules at all. The transformational rule that deletes Objects will now be applicable only to those words positively specifed with respct to this feature, this information now being contained in the Phrase-marker of the string in which these words appear . Any attempt to construct a careful grammar will quickly reveal that many formatives have unique or almost unique gammatical characteristics, so that the simplifcation of the grammar that can be efected in these ways will certainly be substantial. In general, all properties of a formative that are essentially idiosyncratic will be specifed in the lexicon. IS In particular, the lexical entry must specify: (a) aspects of phonetic structure that are not predictable by general rule (for example, in the case of bee, the phonological matrix of the lexical entry will specify that the frst segent is a voiced labial stop and the second an acute vowel, but it will not specify the degree of aspiration of the stop or the fact that the vowel is voiced, tense, and unrounded);17 (b) proprties relevant to the functioning of transformational rules (as the example of the preceding paragraph, and many others) ; (L) properties of the forative that are relevant for semantic interpretation (that is, components of the dictionary defnition) ; (d) lexical features indicating the positions in which Ü lexical formative can be inserted (by the lexical rule) in a 88 CATEGORIES AND RELATIONS IN SYNTACTIC TEORY preterminal string . In short, it contains information that is re quired by the phonological and semantic components of the grammar and by the transforational part of the syntactic com ponent of the grammar, as well as inforation that determines the proper placement of lexical entries in sentences, and hence, by implication, the degree and manner of deviation of string that are not directly generated (see § 2. 3 . 1 and Chapter 4, § 1 . 1 ). Notice, incidentally, that the purely semantic lexical features constitute a well-defned set, in a given grammar. A feature belongs to this set just in case it is not referred to by any rule of the phonological or syntactic component. This may be im portant for the theory of semantic interpretation. See Katz (1 964b) . It is important to observe that the base system no longer is, strictly speaking, a phrase structure (constituent structure) gram mar. As described inforally in § 2. 3. 1 and more carefully in the references cited there, a phrase structure grammar consists of an unordered set of rewriting rules, and assigns a structural de scription that can be represented as a tree-diagam with nodes labeled by symbols of the vocabulary. This theory foralizes a conception of linguistic structure that is substantive and interest ing, and that has been quite infuential for at least half a century, namely the "taxonomic" view that syntactic structure is deter mined exclusively by operations of segmentations and classifca tion (see § 2. 3. 1 ; Postal, 1 964a; and Chomsky, 1 964). Of course, we have alrady departed from this theory by assuming that the rewriting rules apply in a prescribed sequence to generate a restricted set of (base) strings, rather than freely to generate the full set of actual sentences. This modifcation restricted the role of the phrase structure gammar. But introduction of complex symbols constitutes another radical departure from this theory, and the separate treatment of the lexicon just sugested is agin an essential revision. These modifcations afect the strong gen erative capacity of the theory. It is no longer true that a Phrase marker can be represented as a labeled tre-diagam, where each label stands for a category of strings. Furthermore, the con- § 2. ASPECTS OF DEEP STRUCTURE 8 9 ventions for the use of complex symbols in efect allow the use of quasi-transformational rules in the base component . To see why this is so, notice that a derivation involving only phrase structure rules (rewriting rules) has a strict "Markovian" character. That is, in a derivation consisting of the successive lines e1' . . . , en (el ¯ ISI; en ~ fal . . . akl, where each a, is a terinal or non terminal symbol of the vocabulary on which the grammar is based), the rules that can be applied to form the next line en +l are independent of el' . . , en-l and depend com pletely on the string en• A grammaticl transformation, on the other hand, typically applies to a string with a particular struc tural description. Thus application of such a rule to the last line of a derivation depends in part on earlier lines. A grammatical transformation is, in other words, a rule that applies to Phrase markers rather than to strings in the terinal and non terminal vocabularly of the grammar. Suppose, however, that we were to include labeled brackets in the strings that constitute a derivation and were to allow the "rewritng rules" to refer to these symbols. We should now have a kind of transformational grammar, and we should have entrely lost the intuition about language structure that motivated the development of phrase structure grammar. In fact, incorporation of brackets into strings provides the most appropriate notation for the transformational rules of the phonological component (see Halle and Chomsky, 1 960, forthcoming; Chomsky and Miller, 1 963, § 6), though not for the transformational rules of the sytactic component, which are not "local transformations" of the sort that appear, exclusively, in the transformational cycle in phonolo g . 1 8 But with the availability of complex symbols, aspects of the earlier steps of a derivation can also be carried along to later steps, j ust as in the case of the notation for transformational rLles that involves carrying along labeled brackets in lines of a derivation; and, to some extent, global operations on strings can be coded into complex categry symbols and caried along in derivations until the point of "application" of these operations. Consequently, rules applying go CATEGORIES AND RLATIONS IN SYNTACC TEORY to complex symbols are, in efect, transformational rules, and a grammar using complex sybols is a kind of transformational gammar rather than a phrase stucture grammar. Notice, incidentally, that the conventions established for the use of complex symbols do not provide systems with greater weak generative capacity than phrase structure grammar (even if apprpriate conventions are established to permit complex symbols to appear at any point in a derivation, rather than only in lexical categories -see note 4). This fact, of course, has no bearing on the observation that such a theory is no longer a verion of the theory of phrase structure grammar. § 2.3+ Context-sensitive subcategorization rules. We have not yet considered how the categry V is analyzed into a complex symbol. Thus suppose that we have the grammar (2 3 )-(25). We must still give rules to determine whether a V may or may not be transitive@ and so on, and must add to the lexicon ap propriate entries for individual verbal formatives. It would not do simply to add to the grammar the rule (28). analogus to (24) : (28) V � [+V. ±Progressive, ±Transitive, ±Abstract-Subject, ±Animate-Object] The problem is that an occurrence of the category symbol V can be rplaced by a complex sybol containing the feature [+ Transitive] just in case it is in the environment -NP. Similarly. the Verb can be positively specifed for the feature [Abstract-Subject] just in case it is the environment [+Abstract] ' " -; and it can be positively specifed for the feature [Animate Object] just in case it is in the environment - . . . [+Animate] ; and so on. in the case of all (f those lexical features that are in volved in the statement of � ontextual restrictions . Hence, the features [Transitive]. [Abstract-Subject], [Animate-Object] must be intduc by rewriting rules that are restricted with respct to context, as distinct from the context-free rules (u) that sub categorize Nouns.1D A a frst approximaton. we might consider rules of the following sort, for the analysis of V: § 2. ASPETS OF DEEP STUCUR (29) (i) V - [+V, +Transitive] / -NP (H) V - [+V, -TransitiveJ / -1 (30) (i) [+V] - [+[+Abstract]-SubjectJ/[+N, +Abstract] Aux (ii) [+V - [+[-Abstract].Subject]/[+N, -Abstract] Aux (iii) [+V - [+[+Animate] -Object]/ -Det [+N, +Animate] (iv) [+V] - [+[-Animate] -Obj ect]I -Det [+N, -Animate] We can now introduce the standard conventions for ex pressing generalizations in te case of context-sensitive rewriting rules such a (4) , (29), (30) (cf., for example, Chomsky, 1 957, Appendix; cf. § 7, Chapter Ï¿ for discussion of the role of these conventions in linguistic theory), in particular, the convention that X, -Y, is an abbreviation for the sequence of rules (32) (i) A - ZIX1 -Y 1 (n) A - Z / X, -Y, and other familiar related conventions. These allow us to restate (29) and (30) as (33) and (34), respctively. (3 3) (i) } V [ V {+ Transitive] / - (ii) - + , -TransitiveJ/ -- (34) (i) [+[+Abstract]-Subject]1 1 [+N, +Abstract] Aux - (ii) [+[-AbstractJ-Subject]1 [+N, -Abstract] Aux (iii) [+V [+[+Animate]-Object]1 -Det [+N, +Animate] (iv) [+[ -Animate]-Object]l -Det [+N, -Animate] CATEGORIES AND RLATIONS IN SYNTACTIC THEORY It is immediately apparent that the rules (33) and (34), though formally adequate, are extremely clumsy and leave important gneralizations unexpressed. This becomes still more obvious when we observe that alongide of (34) there are many other rules of the same kind; and that alongside of (33) there are rules specifying various other choices of subcategories of Verbs, for example, in such environments as : -Adjective [e.g . , grow (old), feel (sad)], - Predicate-Nominal [becoTe (president)], - like" Predicate-Nominal [look (like a nice person), act (like a foo1] , -S' [think (that he will come), believe (it to be unlikely)], where S' is a variant of a sentence, -NP'"S' [persuade (John that it is unlikely» (omitting certain refnements) . In other words, the schema for grammatical description that we have so far developed still does not permit us to state the actual processes at work in determining the for of sentences. In the present case, there is a large set of rules (of which (34) men tions j ust four) that, in efect, assign features of the Subject and Object to the Verb, somewhat in the manner of ordinary rules of agreement in many languages; and there ar also many rules (of which (33) presents just two) that impose a subclassifcation on the categr Verb in term of the set of frames in which this categry appears at the stage of a derivation where it is to be subcategorized. These generalizations are not expressible in terms of the schema for gammatical desLription so far developd, an inadequacy that reveals itself in the redundancy and clumsi ness of the systems of rules of which (33) and (34) are samples. Our present difculty can be seen clearly by comparing the rules (34) with the hypothetical set (35) : (35) (i) ¸ [+ F 1] / [+N, + Abstract] AUX - ¸ (ii) [+V [+ F 2 ] / [+ N , -Abstract] Aux - (iii) 7 [+F1]/ -Det [+N, + Animate] (iv) [-F 2 ]/ -Det [+N, -Animate] where Fl and F2 ar certain syntactic features. Rules such as (34) systematically select the Verb in terms of the choice of Sub ject and Object, whereas the rules (35) deterine the sub categorization of Verbs in some essentially haphazard way in § 2. APEC OF DEEP STUCTR 9 3 terms of choice of Subject and Object. However. the system (34) is not. in our present terms. more highly valued than (35) ; in fact. the opposite would be true in this case if the familiar notational conventions are applied to evaluate these systems. In other words, the linguistically signifcant generalization underlying (34) is not expressible within our present framework. which is therefore shown to be inadequate (in this case, at the level of explanatory adequacy) . Let u consider how a more natural and revealing expression of these processes can be developed. Observe that the feature specifcation [+Transitive] can be regarded as merely a notation indicating occurence in the environment -NP. A more ex pressive notation would be simply the symbol -NP" itself.20 Gneralizing. let us allow certain features to be designated in the form [X -Y, where X and Y are strings (perhaps null) of symbols. We shall henceforth call these contextual features. Let us regard Transitive Verbs as positively specifed for the con textual feature [ -NP]. pre-Adjectival Verbs such as grow, feel, as positively specifed for the contextual feature [ -Adjective]. and so on. We then have a general rule of subcategorization to the efect that a Verb is positively specifed with respect to the contextual feature associated with the context in which it occurs . We thus introduce the notation (3 6 ) A � X-CS-Y/Z - W a an abbreviation for the rewriting rule (37) A � X""[+A, +Z - W]""Y/Z - W, where " C S" stands for "complex symbol. " Utilizing the bracket conventions. we can now have Z" -w, a an abbreviation for the sequence of rules 94 CATEGORIES AND RELATIONS IN SYNTACIC TEORY A � X [+A. +Z" -W,] ..y/Z, -W, The notation introduced in ( 35 ) allows us to express the fact that a set of frames in which the symbol A occur imposes a cor rsponding subclassifcation on A. wit one subdiÂision cor responding to each listed context. Thus in the case of Verb subclassifcation. we shall have. instead of (33). the rule (40). as a better approximation: NP # Adjective Predicate-Nominal (40) V � CS/ - like ..Predicate-Nominal] ZI Prepositional-Phrase that S' NP (of ..Det ..N) S' etc. The lexicon might now contain the i tems (41 ) eat. [+V. + -NP] elapse. [+ v. + -#] grow . [+V. + -NP, + -#. + -Adjective] become, [+V. + -Adjective. + -Predicate-Nominal] seem, [+V, + -Adjective, + - like Pedicate-Nominal] look, [+V, + -(Prepositional-Phrase) #, + -Adjective, + - like Predicate-Nominal] believe, [+V, + -NP, + ¯that-S1 persuade, [+V, + -NP (oDet N) S1 and so on . 2 The rules (40) supplemented by the lexicon (41 ) will permit such expressions as John eats food. a week elapsed, John grew a beard, John grew. John grew sad. John became sad. John became president. John seems sad. John seems like a nice fellow. § 2. ASPECS OF DEEP STRUCTURE 9 5 John looked, John looked at Bill, John looks sad, John looks like a nice fellow, John believes me, John believes that it is un likely, John persuaded Bill that we should leave, John persuaded Bill of the necessity for M to leave. We see that with a slight extension of conventional notations the systematic use of complex symbols permit a fairly simple and informative statement of one of the basic processes of sub classifca tion. We can use the same notational device to express the kinds of selectional restriction expressed in such rules as (34 ) , which assign featurs of the Subject and Object to the Verb . Thus we can replace (34) by the rules (4 2 ) �� ¸ [+ Abstract] Aux ¸ (11) [ V ] CS! [-Abstract] Aux - (iii) + � -Det [ + Animate] (iv) -Det [-Animate] where now [[+Abstract] Aux ¯] is the feature denoted in (34) as [[ +Abstract]-Subject], etc. The notational convention (36)-(37) shows in what respect a system of rules such a (34), but not (35), expresses a linguistically signifcant generalization. The rules of (40) and (42) analyze a categor into a complex symbol in terms of the frame in which this category appear. The rules difer in that in the case of (40) the frame is stated in ter of category symbols, whereas in the case of (42) the frame is stated in terms of syntactic features . Rules such as (40), which analyze a symbol in ters of its categorial context, I shall henceforth call strict subcategorization rules. Rules such as (42), which analyze a symbol (generally, a complex symbol) in terms of syntactic features of the frames in which it appears, I shall call selectional rules . The latter express what are usually called "selectional restrictions" or "restrictions of cooccurrence. " We shall see later that there are important syntactic and semantic diferences between strict subcategorization rules and selectional rules with respect to both their for and function, and that consequently this distinction may be an important one. In the case of both the strict subcategorization rules (40) and 9 6 CATEGORES AND RELATONS IN SYNTACC THEORY the selectional rules (42), there are still deeper gneralizations that are not yet expressed. Consider frst the case of (40). This set of rules imposes a categorization on the symbol V in terms of a certain set of frames in which V occurs. It fails to express the fact that ever frame in which V appears, in the VP, is relevant to the strict subcategorization of V, and the further fact that no frame which is not part of the VP is relevant to the strict subcategrization of V. Thus the symbol VP will dominate such string as the following, in derivations generated by rewriting rules of the base: (43) (i) V (ii) V NP (iii) V NP that-S (iv) V Prep-Phrase (v) V Prep-Phrase Prep-Phrase (vi) V Adj (vii) V like Predicate-Nominal (viii) V NP PrepPhrase (ix) V NP Prp¯PhÎe Prep-Phrase (elapse) (bring the book) (ersuade John that there was no hope) (decide on a new course of action) (argue with John about the plan) (grow sad) (feel like a new man) (save the book for John) (trade the biccle to John for a tennis racket) and so on. Corresponding to each such string dominated by VP, there is a strict subcategrization of Verbs. On the other hand, Verbs are not strictly subcategrized in tens of types of Subject NPs or type of Auxiliary@ apparently." This observa tion sugests that at a certain point in the sequence of base rewriting rules, we introduce the rule that strictly subcategorizes Verbs in the following fon: (44) V � CS! ¯a, where a is a string such that Va is a VP The rule schema (44) expresses the actual gneralization that detenines strict categrization of Verbs in term of the set of syntactic frames in which V appear. § 2. ASPECT OF DEEP STRUCTURE 97 We have now discussed the problem of formulating the gen eralizations that actually underlie the strict subcategrization rules (40), and have presented informally a device that would accomplish this result. It remains to consider the selectional rules, of which (42) presents a sample. Here too it is evident that there are linguistically signifcant generalizations that are not expressed in the rules as given in this for. Thus the rules (42) do not make use of the fact that ever syntactic feature of the Subject and Object imposes a corresponding classifcation on the Verb, 2 4 not j ust certain arbitrarily chosen features. Once again, a certain extension of the notational devices for formulating rules is called for so that the evaluation measure will operate correctly. In this case, te most natural way to formulate the underlying generalization would be by such rule schemata as { OAUX -} (45) [+V � CS! D , where [ is an N, - et [ [ being a variable ranging over specifed features. We interpret these schemata as abbreviating the sequence of all rules derived from (45) by replacing [ by a symbol meeting the stated condi tion, namely dominance by N (with some ordering that is ap parently inconsequential) . The rules abbreviated by the schemata (45) assert, simply, that each feature of the preceding and follow ing Noun is assigned to the Verb and determines an appropriate selectional subclassifcation of it. Thus if the rule (45) appears in the sequence of base rules after the rules (20), then each of the lexical features that was introduced by the rules of (20) would determine a coresponding subclassifcation of the complex sym bol [+V. The rule schemata (44) and (45) deal with a situation in which an element (in this case, the Verb) is subcategorized in ters of the contexts in which this element appears, where these context all meet some syntactic condition. In all cases, an important generalization would be missed if the relevant contexts were merely listed. The theory of grammar would fail to express the fact that a grammar is obviously more highly valued if sub categorization is determined by a set of contexts that is syntacti- 98 CATEGORIES AND RELATIONS IN SYNTACC TEORY cally defnable. The appropriate sense of "syntactically defnable" is sugested by the examples just given. A precise account of "syntactically defnable" can be given quite readily within the framework of transformational grammar. At the conclusion of § 2. 3. 3 we pointed out that a system of rewriting rules that makes use of complex symbols is no longer a phrase structure grammar (though it does not difer from such a grammar in weak generative capacity), but rather is more properly regarded as a kind of transformational grammar. The rule schemata ( 44 ) and (45) take on the character of transforma tional rules even more clearly. Rules of this typ are essentially of the for (46) A � CS/X ¯Y, where XAY is analyzable as Zl, . , Z", where the expression "X is analyzable a Y _y • • • , Y means that X can b segmented into X :: Xl · · · X, in such a way that X, is dominated by Y" in the Phrase-marker of the derivation under construction. Analyzability, in this sense, is the basic predicate in terms of which the theory of transforational gam mar is developed (see Chomsky, 1 955, 1 956, and many other references). Thus, for example, we can often restate the rules in question with the use of labeled brackets (regarding these as caried along in the course of a derivation), or by allowing complex symbols to appear at arbitrary points of a derivation, with certain features being carried over to certain of the "de scendants" of a particular category symbol in the manner of Matthews's system referred to in note 1 3, or in various other similar ways .26 Along with a lexicon, then, the base component of the gram mar contains: (i) rewriting rules that typically involve branching and that utilize only catgorial (noncomplex) symbols and (ii) rule scemata that involve on 1 y lexical categories, except in the statement of context, and that utilize complex symbols. The rules (i) are ordinary phrase structure rules, but the rules (ii) are transformational rules of an elementary sort . One might, in fact, sugest that even the rules (i) must be replaced, in part, by § 2. ASPES OF DEEP STRUCTURE 9 9 rule schemata that go beyond the range of phrase structure rules in strong generative capacity (cf., for example, Chomsky and Miller, 1963, p. 298, Chomsky and Schitzenberger, 1963, p. 133, where such operations as conjunction are discussed in terms of a framework of this sort), or by local transformations (cf. note 18). In short. it has become dear that it was a mistake, in the frst place, to suppose that the base component of a transformational grammar should be strictly limited to a system of phrase struc ture rules, although such a system does play a fundamental role as a subpart of the base component. In fact, its role is that of defning the grammatical relations that are expressed in te deep structur and that therfore determine te semantic inter pretation of a sentence. The descriptive power of the base component is geatly en riched by permitting transforational rules; consequently, it is important to see what limitations can be imposed on their use -that is, to see to what extent freedom to use such devices is actually empirically motivated. From the examples just given, it seems that there are indeed heavy restrictions. Thus the strict subcategorization of V involves only frames that are dominated by the symbol VP, and there are also obvious restrictions (to which we retur in § 4.2) involved in the use of selectional rules. Putting these aside for the moment, let u continue with the investigation of strict subcategorization rules . The symbol V is introduced by rules of the form: VP � V . . . , and it is frames dominated by VP that deterine strict sub categorization of Verbs. This suggests that we impose the follow ing gneral condition on strict subcategorization rules: each such rule must be of the for (47) A � CS/a -p, where aAP is a (, where, furthermore, ( is the categor smbol that appears on the l eft in the ru l e ( � • • • A . . . that introduces A. Thus (47), reformulated within the framework of the theory of grammatical transformations, would be what we have called a "local trans formation." Cf. note 18. The italicized condition guarantees that Ï LL CATEGORES AND RELATIONS IN SYNTACIC TEORY the transformation is, furtermore, "strictly local" in the sense of note 1 8. If this condition of strict local subcategorization is adopted as a general condition on the form of grammar, then the strict subcategorization rules can simply be given in the form the rest being supplied automatically by a convention. In other words, the only characteristic of these rules that must be explicitly indicated in the grammar is their position in the sequence of rules. This position fxes the set of frames that determine sub categorization. Suppose that the rule that introduces Nouns into the gammar is, essentially, the following: (49) NP " (Det) N(S') In this case, we should expct strict subcategorization of Nouns into the categres [Det -S'], [Det -], [ -S'] , and [ -] (con tinuing with the notational conventions for features introduced earlier). The category [Det -S'] is the category of Nouns with sentential Complements (such as "the idea that he might suc ceed," "the fact that he was guilty, " "the opportunit for him to leave, " "te habit of working hard" -the latter involving a sentential Complement with an obligatorily deleted Subject). The category [Det -] is simply the category of Common Nouns. The category [ -] is the category of Proper Nouns, that is, Nouns with no Determiner (or, as in the case of "The Hague," "The Nile," with a fxed Determiner that may just as well be taken as part of the Noun itself, rather than as part of a freely and independently selected Determiner system) . 26 If this is cor rect, then the Proper/Common distinction is strict subcategrial, and does not fall togther with the other features introduced in (20). The categor [ -S'] is not realized i n so obvious a way a the others. Perhaps one should utilize this category to account for "quotes contexts" and, more importantly, for the impersonal it of such sentences as "it strikes me that he had no choice, " "it surprised me that he left," "it is obvious that the attempt must § 2. ASPECTS OF DEEP STUCTUR 1 01 fail, " which derive from underlying string with NP' s of the form: it'Sentence (the Sentence Complement either being sep arated from it by a transformation, as in the examples cited, or substituting for it by a strictly local transformation in the man ner described in note 1 8). Returning, once again, to Verb subcategorization, we note one further consequence of accepting the gneral condition sug gested in connection with (47) . It is well known that in Verb Prepositional-Phrase constructions one can distinguish various degrees of "cohesion" between the Verb and the accompanying Prepositional· Phrase. The point can be illustrated clearly by such ambiguous constructions as (50) he decided on the boat which may mean "he chose the boat" or "he made his decision while on the boat. " Both kinds of phrase appear in (5 1 ) he decided on the boat on the train that is, "he chose the boat while on the train." Clearly, the second Prepositional-Phrase in (51 ) is simply a Place Adverbial . which, like a Time AdNerbial, has no particular connection with the Verb. but in fact modifes the entire Verb Phrase or perhaps the entire sentence. It can. in fact. be optionally preposed to the sentence, although the frst Prepositional-Phrase of (5 1). which is in close construction to the Verb. cannot -that is. the sen tence "on the train. he decided" is unambiguous. There are many other examples of the same kind (for example. "he worked at the ofce" versus "he worked at the job"; "he laughed at ten o' clock" versus "he laughed at the clown"; "he ran after dinner" versus "he ran after John"). Clearly. Time and Place Adverbials can occur quite freely with various types of Verb Phrase, on the one hand, whereas may types of Prepositional.Phrase appear in much closer construction to Verbs. This observation suggests that we modify slightly the frt several rules of the base. Ïeplac ing them by 102 CTEGORES AND RELATONS IN SYNTACC THEORY (52) (i) S � NP""Predicate-Phrase (ii) Predicate-Phrase � Aux""VP (Place) (Time) be Predicate (iii) V � Adj ¸(NP) (Prep-Phrase) (Prep-Phrase) (Manner)¸ V S' (like) Predicate-Nominal Direction Duration (iv) Prep-Phrase � Place Frequenc y etc. (v) V ¯ CS The conventions govering complex symbols will interpret (v) as strictly subcategorizing Verbs with respect to all contexts intro duced in the second part of rule (iii) and in rule (iv). It will follow. then, that Verbs are subcategorized with respect to the Prepositional-Phrases introduced by (50iii) but not with respect to those introduced by (50ii) -namely. the Place and Time Adverbials that are associated with the full Predicate-Phrase. and that might. in fact. be in part more closely associated with the Auxiliary (cf. note 23) or with Sentence Adverbials which form a "pre-Sentence" unit in the underlying structure. Thus Verbs will be subcategorized with respect to Verbal Complements. but not with rspect to Verb Phrase Complements. That this is essentially the case is clear from the examples given. To illutrate. once again. in connection with the four types of Adverbials listed in (52iv). we have such phrases as (53). but not (54) : 27 (53) dash -into the room (V -Direction) last -for three hours (ViDuration) remain -in England ( -Place) win -three times a week (V -Frequency) (54) dash -in England last -three times a week remain -into the room win -for three hours § 2. ASPECS OF DEP STUCUR 103 Similarly, the i talicized phrases in "he argued with John (about politics)," "he aimed (the gn) at John," "he talked about Greece," "he ran after John, " "he decided on a new course of action, " and so on, are of types that induce a subcategorization of Verbs, whereas the italicized phrases in "John died in Eng land," "John played Othello in England," "John always runs after dinner," and so on, do not play a role in Verb sub categrization, since they are introduced by a rule (namely (5 2 ii» the left-hand symbol of which does not directly dominate V. Similarly, the other contexts introduced in (5 2 iii) will play a role in strict sub categrization of Verbs . In particular, the Manner Adverbial participates in Verb subcategorization . Thus Verbs generally take Manner Adverbials freely, but there are some that do not -for example: resemble, have, marr (in the sense of "John married Mary, " not "the preacher married John and Mary," which. does take Manner Adverbials freely); ft (in the sense of "the suit fts me, " not "the tailor ftted me," which does take Manner Adverbials freely) ; cost, weigh (in the sense of "the car weighed two tons, " not "John weighed the letter, " which does take Manner Adverbials freely) ; and so on. The Verbs that do not take Manner Adverbials freely Lees has called "middle Verbs" (Lees, 1 960a, p. 8), and he has also observed that these are, characteristically, the Verbs with following NP' s that do not undergo the passive transformation. Thus we do not have "John is resembled by Bill," "a good book is had by John, " "John was married by Mary," "I am ftted by the suit, " "ten dollars is cost by this book," "two tons is weighed by this car, " and so on (although of course "John was married by Mary" is accept able in the sense of "John was married by the preacher," and we can have "I was ftted by the tailor, " "the letter was weighed by John, " etc.). 28 These observations sugest that the Manner Adverbial should have as one of its realizations a "dummy element" signifng that the passive transforation must obligatorily apply. That is, we may have the rule (55) as a rewriting rule of the base and may formulate the passive transforation so as to apply to strings of the form (56), with an elementary transforation that sub- CATEGORIES AND RLATIONS IN SYNTACC TEORY stitutes the frst NP for the dummy element passive and places the second N P in the position of the frst NP: (55) Manner � by .. pasive (56) NP - Aux - V * . . . - NP ^ . . . - by - passive - . " " (where the leftmost . . . in (56) requires further specifcation e . g., it cannot contain an NP) . This formulation has several advantags over that presented in earlier work on transformational grammar (such as Chomsky, 1 957) . First of all, it accounts automatically for the restriction of passivization to Verbs that take Manner Adverbials freely . That is, a V erb will appear in the frame (56) and thus undergo the passive transformation only if it is positively specifed, in the lexicon, for the strict subcategorization feature [-NP-Manner] , in which case it will also take Manner Adverbials freely. Second, with this formulation it is possible to account for the derived Phrase-marker of the passive by the rules for substitution trans formations . This makes it possible to dispense entirely with an ad h oc rule of derived constituent structure that, in fact, was motivated solely by the passive construction (cf . Chomsky, 1 957, pp . 73-74) . Third, it is now possible to account for "pseudo passives," such as "the proposal was vehemently argued against," "the new course of action was ageed on," "John is looked up to by everyone, " by a slight generalization of the ordinary passive transformation . In fact, the schema (56) already permits these passives. Thus "everyone looks up to John by passive" meet the condition (56), with Jo h n as the second NP, and it will be converted into "john is looked up to by everyone" by the same elementary transforation that form "john was seen by every one" from "everyone saw jo�n. " In the earlier formulation (cf . Chomsky, 1 955, Chapter IX), it was necessary to treat pseudo passives by a new transformation. The reason was that V of (56) had to be limited to tran s itive Verbs, for the ordinary passive transformation, so as to exclude the "middle" Verbs have, resemble, etc . But if passivization is determined by a Manner Adverbial, as just sugsted, then V in (56) can be quite fre, and § 1. APEC OF DEP STUcR can be an intransitive as well a a transitive Verb. Thus "John is looked up to" and "John was seen" are formed by the same rule despite the fact that only in the latter case is John the Direct Object of the deep structure . Notice, however, that the Adverbial introduced by (52ii) is not subject to the passive transformation as defned by (56), since it will follow the Adverbial by'passive. This accounts for the fact that we can have "this j ob is being worked at quite seriously" from "Unspcifed-Subject is working at this job quite seriously," where "at this job" is a Verb-Complement introduced by (52iii), but not "the ofce is being worked at" from "Unspecifed-Subject is working at the ofce, " where the phrase "at the ofce" is a VP-Complement introduced by (52ii) and therefore follows the Manner Adverbial. Similarly, we can have "the boat was decided on" in the sense of "he chose the boat," but not in the sense of "he decided while on the boat." Thus the passive sentence corresponding to (50) is unambiguous, though (50) itelf is ambiguous. Many other fact can be explained in the same way . The fact that we are able, in this way, to account for the nonambiguity of "the boat wa decided on by John" as con trasted with the ambiguity of "John decided on the boat, " along with many similar examples, provi d es an indirect j ustifcation for the proposal (cf. p. 99) that strict subcategorization rules be limited to strictly local transforations. It is perhaps worth while to trace through the argument again to see why this is so. By the "strictly local subcategorization" principle we know that certain categories must be internal to the VP and other must be exteral to i t . One of te element that must be internal to the VP, in accordance with this principle, is the marker for " passiviza tion, since it plays a role in strict subcategorization of the Verb. Furtherore, the marker for passivization is associated with the presence of the Manner Adverbial, which is internal to the VP by the strictly local subcategorization principle. Since the passive transforation must be formulated with the structure index (56), it follows that NP' s in VP-Complements are not subject to "pseudopassivization" while NP's in V-Complements may be subject to this operation. In particular, where "on the boat" is a 1 06 CATEGORIES AND RLATIONS IN SYNTACTIC TEORY V-Complement in "John decided on the boat" (meaning "John chose the boat"), it is subject to pseudopassivization by the passive transformation; but where "on the boat" is a VP Complement in "John decided on the boat" (meaning "John decided while he was on the boat," equivalently, "on the boat, John decided"), it is not subject to pseudopassivization since it does not meet the condition (56). Therefore, observing that "the boat was decided on by John" is unambiguous and means only that John chose the boat, we conclude that the premise of this arment -namely the assumption that strict subcategoriza tion is limited to strictly local transforations -has empirical support, The reanalyis (52) requires that the defnitions of functional notions proposed in § 2. 2 (cf. ( Ï Ï » be slightly altered. Thus we might perhaps defne the notion "Predicate-of" as [Predicate Phrae, S] rather than as [P, S] . This revised forulation of the rules, incidentaIly, illustrates another proprty of te tradi tional functional notions. We observed in § 2. 2 that these notions are defned only for what we caIled "major categries. " Further more, it seems that tey ar defned only for those major cate gories A that appear in rules of the form X � . . . A . . . B· . . or X � . . · B· · · A · · · , where B is also a major category. This seems quite natural, considering the relational character of these notions. § 3. AN ILLUSTRATIVE FRAGMENT OF THE BASE COMPONENT Let us now summarize this discussion by returing to the original problem, posed in § Ïy of presenting structural informa tion of the sort illustrated in (2) of § 1 in a set of rules that are designed to express precisely the basic linguistic processes involed. We may now consider a generative gammar with a base component containing, among many other, the rules and rule schemata (57) and the lexicon (58) : (57) (i) S � NP ..Predicate-Phrase (ii) Pricate-Phrase � Aux .. VP (Place) (Time) § 3 . AN ILLUSTTIVE FRGMENT OF TE BASE COMPONENT 1 07 ¸COpula ..predicate (iii) VP � V l��P) . (PrepPhrase) (Prep-Phrase) (Manner») Predicate . {AdjeCtive } (iv) Predicate � (l ' k ) P d ' N . I , e re Icate- omma (v) PrepPhrase � Direction, Duration, P lace, FrequeÏcy, etc. (vi) V � CS (vii) NP � (Det) N (5, (viii) N � CS (ix) [+Det -] � [±Count] (x) [+Co1nt] � [±Animate] (xi) [+ N , + -] � [±Animate] (xii) [+Animate] � [±Human] (xiii) [-Count] � [± Abstract] (xiv) [+V] � CS/a .. A¼7 -(Det ..p) } , where [ is an N and (xv) Adjectiv� � CS/a . . . - p is an N (xvi) Aux � Tense (M) (Aspect) (xvii) Det � (pre-Article ..of Article (post-Article) (xviii) Article � [±Defnite] (58) (sincerity, [+N, +Det -, - Count, +Abstract, . . . ]) (boy, [+N, +Det -, +Count, +Animate, +Human, . . . ]) (frighten, [+V, + -NP, +[+Abstract] Aux -Det [ + Animate], +Object-deletion, . . . ]) (may, [+M . . . . ]) This system of rules will gnerate the Phase-marker (5 9 ). Adding the rules that realize Defnite as the and non-Defnite as null before a following non-Count Noun, we derive the sen tence "sincerity may frighten the boy" of § 1 , with the Phrase marker (59). Notice that this fragment of the base is "sequential" in the sense of § 2 . 1 . We have only sketched the procedure for constructing a Phrae-marker of te reqÎired sort fom a derivation. However . this is a relatively minor matter of appropriate formalization and involves nothing of principle . In particular, (59) represents not only all information involving the relation "is a," holding ( 5 9 ) N P � D e t N | | A r t i c l e F I [ - D e f n i t e ] F s i n c e r i t y 1 - 5 - # A u x | M I m a y P r e d i c a t e - P h r a s e V P � V N P I � G D e t N � I I C l ' C 2 ' - - - A r t i c l e H " f r i g h t e n | [ + D e f n i t e ] H � | • • • H " b o y . . Û O I P ' � � o z ' Z ' > Z � + ¯ � t o � F ~ j + ~ ] ~ C o m e n F 1 ' ¬ C e u n F " ~ j + A b s t r a c t ] G ~ j + ] ¬ T r a n s i t i v c G 1 ' ~ + + b s t r a c t u ~ c t A n i m a ] ] G 2 ' ¬ j + O b j c c t - d c l c t i e n j H ~ _ + D e t ~ ] ~ F H 1 ' ~ _ + C o u n t ] ~ ¯ F l H " ¯ ¡ + A n i m a æ ] H ' " ~ _ + H u m a n ] O � � ; i � � i W � P i . o Z P � Þ C U 110 CTGORIES AND RLATONS IN SYNTAcC TEORY between string and the categries (many of them now repre sented by features) to which they belong but also the hierarchic relation among these categories that is provided by the rules and mirrored precisely in the derivation. The Phrase-marker (59) provides directly all inforation of the sort spcifed in (2i) and (2iii); and, as we have observed, functional information of the sort specifed in (2ii) is derivable from this Phrase·marker as wen. If the analysis that we have given i correct, then it is devices of the sort just exhibited that ar implici t in the inforal statement of tradi tional grammar sumrized in (2), with one exception, to which we shall turn in the next section. Notice that neither the lexicon (58) nor the Phrase·marker (59) is fully specifed. There are clearly other syntactic features that must be indicated, and we have given no semantic features in either (58) or (59). In part, it is clear how these gaps can be flled, but it would b a seriou mistake, in this case, to suppose that this is in general merely a question of added detail. One fnal comment is necessary in connection wit te lexicon (58). Given a lexical entry (D,C), where D is a phonological feature matrix and C a complex symbol, the lexical rule (d. p. 84) permits substitution of D for any complex symbol K that is not distinct from C. Consequently, lexical entries must be specifed negatively for features corresponding to contexts in which they may not occur. Thus in (58), for example, boy must be specifed a [-V, so as to exclude it from the position of fri g hten in "sincerity may frighten the boy," and not only must fri g hten be specifed as [-N], to exclude it from the position of boy in this sentence, but it must also be specifed negatively for the feature [ -Adjective] , sL as to exclude it from the position of turn in "his hair tured gray," and so on. These negative specifcations were not actually given in (58) . We can deal with this matter by adopting several additonal conventions govering the base component. Firt of all, we may assume that a base rule that analyzes the lexical category A into a complex symbol automatically includes the feature [+A] as one of the elements of this complex sybol (see (20), § 2. 3. 2). § 4. TPES OF BAE RULS Ï Î Î Second, we may assume that each lexiLal entry automatically, by convention, contains the feature [-A] for every lexical cate gory A, unless it is explicitly provided with the feature [+A]. Thus in (58), the entry for boy contains [-V], [-Adjective] , [-M] (cf. note 9).29 Third, in the case of features introduced by strict subcategrization or selectional rules (what we have called the "contextual features"), we may adopt onC of the following conventions. (i) list in the lexicon only the features corresponding to fames in which the item in question cannot appear (rater than, a in (58), those coresponding to features in which it can appear) (ii) list only the features corsponding to fames in which the item can appear, as in (58) (in cae (i) or case (ii) we add the further convention that Ü i tem is specifed in the opposite way for every contextual feature not mentioned in its lexical entry) (iii) adopt (i) for the strict subcategrization features and (ii) for the selectional features (iv) adopt (ii) for the strict subcategorization features and (i) for the selectional features. In any case, the distinctness rquire ment of the lexical rule will now exclude items from certain contexts, and permit them in others. These conventions embody alterative empirical hypotheses concering vÜluation of gammar. Thus (i) is correct if the most highly valued gammar is that in which the distribution of item i least constrained, and (ii) is corect if the most highly valued grammar is that in which the distribution of items is most constrained (similarly, (iii) and (iv». For the time being, I have no strong Cxamples to support one or another of these assumptions, and tus prfer to leave the queston open. We shall retur briefy to the problem in Chapter 4 . § 4. TYPES OF BASE RULES § 4. I. Summar The fagment presented in § 3 illustrates the kinds of rules that apparntly are to be found in the base component. There is a fundamental distinction between the rewriting rules (57) and 1 1 2 CATEGORIES AND RLATIONS IN SYNTACTIC TIEORY the lexicon (58) . The lexical rule need not be stated in the gamar since it is universal and hence part of the theory of gram mar. It status i j ust like that of the principles that defne "derivation" in terms of a system of rewriting rules, for example . It thus has the status of a convention determining the interpreta tion of the grammar, rather than the status of a rule of the grammar. In terms of the framework of § 6, Chapter 1 , we may say that the lexical rule in fact constitutes part of the gneral, language-independent defnition of the function f of ( 1 4iv), § 6, Chapter 1 . Among the rewriting rules of the base component we can distinguish branching rules, such as (i), (ii), (iii), (iv), (v), (vii), (xvi), (xvii), from subcategorization rules, such as all others of (57). All rewriting rules are of the form (60) A � Z/X-W The branching rules are those rLles of the form (60) in which neither A nor Z involves any complex symbols . Thus a branching rule analyzes a category symbol A into a string of (one or more) symbols each of which is either a terminal symbol or a non terminal category symbol. A subcategorization rule, on the other hand, introduces syntactic features, and thus forms or extends a complex symbol. We have, so far, restricted the subcategorization rules to lexical categories. In particular, we have not permi tted rules of the form (60) in which A is a complex symbol and Z a terinal or category symbol or a string of more than one symbol. This restriction may be a bit too severe, and we must appaently weaken it slightly. See Chapter 4, § 2. Notice that these two sets of rules (branching and subcategorization) are not ordered with respect to one another, although once a subcategorization rule has been applied to a certain category symbol Ç no branching rule can be applied to any of the symbols that ae derived from (. Branching rules and subcategorization rules may be context free (such as all of the branching rules of (57) and (x), (xi), (xii), (xiii), (xviii» or context-sensitive (such a (vi), (viii), (xiv), (xv». Notice that (57) contains no context-sensitive branching rules. Moreover, the subcategorization rules that are context-sensitive § 4 . TYPES OF BASE RULS are, in efect, strictly local transformatiLnal rules (cf. p. 99). These are important facts, to which we return in Chapter 3. Among the context-sensitive subcategorization rules we have, furthermore, distinguished two important sub types, namely strict subcategorization rules (such as (57vi) and (57viii», which sub categorize a lexical category in ters of the frame of category symbols in which it appears, and selectional rules (such as (57xiv), (57xv», which subcategorize a lexical category in terms of syntactic features that appear in specifed positions in the sentence. We noted that subcategorization rules may follow branching rules in the sequence of rules constituting the base, but that once a subcategorization rule has applied to form Ü complex symbol I, no branLhing rule can later apply to I (but cf. Chapter 4, § 2). The same relation apparently holds between strict subcategorization rules and selectional rules. That is, these may be interpersed in the base, but once a selectional rule has applied to for the complex symbol I, no strict subcategoriza tion rule applies later to develop I further. So, at least, it appears from the examples that I have considered. Perhaps this should be imposed as a general, additional condition on the base. § 4.2. Selectional rules and gammatical relations We shall say that a selectional rule, suc as (57xiv), (57xv), defnes a selectional relation between two positions in a sentence -for example, in the case of (57xiv), the position of the Verb and that of the immediately preceding or immediately following Noun. Such selectional relations determine grammatical rela tions, in one of the senses of this traditional term. We observed earlier that the notion of grammatical function defned in § 2. 2 did not yet account for the assignment of the Subject-Verb rela tion to the pair sincerit, frighten and the Verb-Object relation to frighten, boy in sincerity may frighten the boy (=( 1 » . The suggsted defnition of gammatical relation would account for these assertions. given the grammar (57), (58) . The same notion of gramÏatical relation could. in fact, have been defned in terms of the heads of major categries (cf. § 2 . 2), but the defni- 1 1 4 CATEGORIES AD RLATIONS IN SYNTACIC TEORY tion in terms of selectional relations seems somewhat more natural and avoids the problem noted on pp. 7 3 -74. With this notion now defned, we have completed the analysis of the in foral grammatical statement ( 2) of § 1 . 30 Consider now the selectional rules (S7xiv), (S7xv), which con strain the choice of Verb and Adjective in terms of a free choice of certain features of the Noun (in this case, the Subject and Object). Suppose, instead, that we were to subcategorize the Verb by a context-free rule, and then to use a selectional rule to determine the subcategorization of the Subject and Object. We might have, for the Verb, such a rule a (61 ) V - [+V, + [+Abstract]-Subject, +[+Animate]-Object]S1 Thus we might in particular form the complex symbol (62) [+V, +[+Abstract]-Subject, +[+Animate]-Object] which can be replaced by a lexical item suc as frighten, lexically marked as allowing an Abstract Subject and an Animate Object. We must now give a context-sensitive selectional rule to deter mine the choice of Subject and Object, just a in (S7) we gave such a rule to determine the choice of Verb in terms of Subject and Object. Thus we would have such rules as N - CS/ ' where is a V j -Aux + l + Det - These rules would assign features of the Verb to the Subject and Object, just as (5 7 xiv) assigned features of the Subject and Ob ject to the Verb. For example, if the Verb is (62), the Subject would be specifed as having the featurs (64) [pre-+[ +Abstract]-Subject, pre-+[+Animate]-Object] Similarly, the Object would have the features (65) [post-+[+Abstract]-Subject, post-+[+Animate]-Object] But, clearly, the feature [pre+[+Animate]-Object] is irrelevant to choice of Subject Noun, and the feature [ost- +[+Abstract]- § § . TPES OF BAE RULS Subject] is irrelevant to choice of Object Noun. Much more serious than this, however, is the fact that a Noun must be marked in the lexicon for the feature [pre-X-Subject] if and only if it is marked for the feature [post.X-Object], where X is any feature. That is, the choice of elements for the posi tion "Subject of a Verb with Animate Subject" is the same as the choice of elements for the position "Object of a Verb with Animate Object." Animate Nouns appear in both posi tions. But the feature [Animate] is no longer available for Nouns, only the features [pre-+[+Animate]-Subject] and [post + [+Animate]-Object] . Consequently, a mass of perfectly ad hoc rules must b added to the grammar to assign to Nouns with the feature [pre-X-Subject] also the feature [post-X-Object], for each feature X, and conversely. Moreover, the features [pre-X-Subject], [pst-X-Object], for each X, are single symbols, and the fact that X occurs in bot of them cannot be refered to by a rule of the grammar (unless we complicate the mechanism further by allowing features to have a feature compsition them selves). In short, the decision to choose the complex symbol analysis of Verbs independently and to select Nouns by a selectional rule in ters of Verbs leads to a quite considerable complication of the grammar. The problems are magnifed when we bring into account the independent Noun·Adjective selectional rules. In much the same way we can rule out the possibility of allowing Subject to select Verb but Verb to select Object. We see, then, that within the framework so far· developed, there is no alternative to selecting Verbs in terms of Nouns (and, by a similar argument, Adjectives in terms of Nounb), rather than conversely. Furthermore, this framework seems to b optimal, in that it involves no more mechanism than is actually forced by the linguistic fact. One would imagne that a similar argument can be given for any language. If this is true, it is possible to take another signifcant step toward a gneral characterization of the categries Noun, Verb, Adjective, etc. (see §§ 2. 1 , 2. 2). In § 2. 2. I defned "lexical category" and "major category," 1 I 6 CATEGORES AND RELATONS IN SYNTACC TEORY the latter being a lexical categor or a category dominating a string containing a lexical category . Suppose that among the lexical categories, we label as Noun the one that is selectioall y dominant in the sense that its feature composition is determined by a context·free subcategrization rule, its features being car ried over by selectional rules to other lexical categories . Among the major categories introduced in the analysis of Sentence, we now desigate a NP the one that is analyzed as . . . N . . . A major category that directly dominates . . . NP . . . we can des ignate VP, and one that directly dominates VP, we can des ignate Predicate· Phrase . We can defne V in various ways for example, as the lexical category X that appears in a string · · · X· · · Np· · · or · · · Np· · · X· · · directly dominated by VP (assum ing that there can be only one such X) or as the lexical category that may obtain its features from selectional rules involving two or more N's (if transi tivity is a category that is universally realized) . One might now go on to attempt to characterize other lexical, major, and nonmajor categories in general terms . To the extent that we can do this, we shall have succeeded also in giving a substantive specifcation to the functional notions discussed in § 2 . 2. It will be obvious to the reader that this characterization is not intended as defnitive in any sense. The reason has already been indicated in note 2. There is no problem in principle of sharpening or generalizing these defnitions in one way or an other, and there are many formal features of the grammar that can be brought into consideration in doing so. The problem is merely that for the moment there is no strong empirical motiva tion for one or another suggstion that might be made in these directions. This is a consequence of the fact that there ar so few grammars that attempt to give an explicit characterization of the rang of sentences and structural descriptions (that is, so few gnerative grammars), even in a partial sketch. A explicit grammatical descriptions with this goal accumulate, it will no doubt be possible to give empirical justifcation for various re fnements and revisions of such loosely sketched proposals as these, and perhaps to give a substantive characterization to the § 4. TPES OF BASE RULS universal vocabulary from which grammatical descriptions are constructed. However, there is no reason to rule out, a priori, the traditional view that such subtantive characterizations ϼst ultimately refer to semantic concepts of one sort or another. Once again, as in §§ 2. 1-2. 2, it is clear that this attempt to characterize universal categories depends essentially on the fact that the base of the syntactic component does not, in itself, explicitly characterize the full rang of sentences, but only a highly restricted set of elementary structures from which actual sentences are constructed by transformational rules.82 The bae Phrase-marker may be regarded as the elementary content ele ment from which the semantic interpretations of actual sen tences are constructed.ss Therefore the observation that the semantically signifcant functional notions (grammatical rela tions) are directly represented in base structurs, and only in these, should come as no surprise; and it is, furthermore, quite natural to suppose that formal properties of the base will pro vide the framework for the characterization of universal cate gories . To say that formal properties of the base will provide the framework for the characterization of universal categories is to assume that much of the structure of the base is common to all languages. This is a way of stating a traditional view, whose origins can again be traced back at least to the Grammaire genera le et raisonnee (Lancelot et al., 1 660) . To the extent that relevant evidence is available today, it seems not unlikely that it i true. Insofar as aspects of the base structure ar not specifc to a particular languag, they need not be stated in the grammar of this language. Instead, they are to be stated only in general linguistic theory, as part of the defnition of the notion "human languag" itself . In traditional terms, they pertain to the for of language in general rather than to the form of particular languags, and thus presumably refect what the mind brings to the task of languag acquisition rather than what it discover (or invent) in the course of carrying out this task. Thus to some extent the account of the base rules sugsted here may not be long to the grammar of English any more than the defnition of 1 I 8 CATOIES AND RLATONS IN SYNTACIC TORY "derivation" or of "transformation" belongs to the grammar of English. Cf., §§ 6 and 8, Chapter 1 . It is commonly held that modem linguistic and anthro pological investigations have concluively refuted the doctrines of classical universal gammar, but this claim seems to me very much exagrated. Modem work has, indeed, shown a great diversity in the surface structures of languags. However, since the study of deep structure has not been it concer, it has not attempted to show a corresponding diversity of underlying structures, and, in fact, the evidence that has been accumulated in modem study of language does not appear to suggest anything of this sort. The fact that languages may difer from one an other quite sigifcantly in surface structure would hardly have come as a surprise to the scholar who developed traditional univeral grammar. Since the origins of this work in the Gram maire generale et raisonnee, it has been emphasized that the deep structurs for which universality is claimed may be quite distinct from the surface structures of sentences as they actually appear. C onsequently, there is no reason to expect unifoni ty of surface structures, and the fnding of modem linguistics are thus not inconsistent with the hypotheses of universal gammarians. Insofar as attention is restricted to surface structures, the most that can be expected is the discovery of statistical tendencies, such a those presented by Greenberg ( 1 963). In connection with the selectional rule (57xiv), we have now conclusively ruled out one possibility, namely that the Subject or Object may b selected in terms of an independent, or partially independent, choice of Verb. Not quite so simple is the question of whether this rule, which I now repeat in less abbreviated fon as (66), should be preferred to the alterative (67) . (66) ��) } [+V � CS/ { QAUX -'f} (11) a' Aux - (67) ��)} [+V � C S/ {a'AuX -} (11) ¯Det' In term of evaluation measures that have so far been propoed § 4. TYPE OF BAE RULE 1 1 9 (see, for example, Chomsky, 1 9 55, Chapter 3), there is no way of chosing between these. In accordance wi th the usual conventions for obligatory application of rewriting rules (cf. ibi d.), (66i) assigs certain features to Transitive Verbs and (66ii) to In transitive Verbs. On the other hand, (67i) assigns a feature of Subject selection to all Verbs, and (67ii) assigns a feature of Object selection to Transitive Verbs. If we choose (66), the lexical entry for frighten will be positively specifed for the feature [[+Abstract] Aux -Det [+Animate]] ; if we select (67), it will be positively specifed for the two features [[+Abstract] Aux -] and [ -Det [+Animate]]. It may appear at frst that this is little more than a terminological question, but, as i many such cases, this is not at all obvious. Thus consider the following context: (68) (i) he the platoon (ii) his decision to resig his commission -- the platoon (iii) his decision to resig his commission our respect In (68i) we can have the Verb com'nd (I neglect, for simplicity of ex p osition, questions of coic of Auxiliary). In (68iii) we can also have command, but in a diferent though not totally un related sense. In (68ii) we cannot have command, but we can have, for example, bafe, which can also appear in (68i) but not (68iii) . If we select the alterative (67), the Verb command will be positively marked for the features [[+Animate] Aux -] , [ -Det [+Animate]], [[+Abstract] Aux -], and [ -Det [+Ab stract]]. That is, it will be marked in such a way as to permit it to have either an Aimate or an Abstract Noun as Subject or Object. But this specifcation fails to indicate the dependency between Subject and Object illustrated by the deviance 'of (68ii), when command appars in this context. If we select the alter native (66), command will be positively marked for the features [[+Animate] Aux - Det [+Animate]] and [[+Abstract] Aux - Det [+Abstract]], but not [[+Abstract] Aux -Det [+Animate]] . Thus command would be excluded from the context (66ii), as required. It is for such reasons that I selected the alternative (66) in the grammatical sketch. It should be noted, however, that the grounds for this decision are very weak, since a crucial question 1 20 CATEGORIES AND RELATIONS IN SYNTACIC TEORY -namely, how to enter lexical items with a rang of distinct but related syntactic and semantic features -is far from settled. I have so far not been able to fnd stronger examples. It seems at frst as though a certain redundancy results from the decision to select (66) over (67), in the case of Verbs for which choice of Subject and Object is independent. However, the same number of features must be indicated in the lexicon, even in this cse. With the coice of (66), the features seem more "complicated," in some sense, but this is a misinterpretation of the notational system. Recall that the notation [+Animate] Aux -Det [+Abstract], for example, is a single symbol desigating a particular lexical feature, in our framework. Clearly, this comment does not exhaust the question, by any means. For some further related discussion, see Chapters 3 and 4. § 4.3. Further remarks on subcategorization rules We have distinguished, in the base, between branching rules and subcategrization rules and between context-fre and context sensitive rules. The context-sensitive subcategorization rules are further subdivided into strict subcategorization rules and selectional rules. These rules introduce contextual features, whereas the context-free subcategorization rules introduce in herent features. One might propose, alternatively, that the sub categorization rules be eliminated from the system of rewriting rules entirly and be assiged, in efect, to the lexicon. In fact, this is a prfectly feasible sugstion. Suppose, then, that the base is divided into two parts, a categorial component and a lexicon. The categorial component consists solely of branching rules, which are possibly all context free (see Chapter 3). In particular, the branching rules of (57) would constitute the categrial component of the base of this fragment of English gammar. The primary role of the categorial component is to defne implicitly the basic grammatical relations that function in the deep structures of the language. It may well b that to a large extent the form of the categorial component § 4. TYPE OF BASE RULES 1 21 is determined by the univeral conditions that defne "human language. " The subcategrization rules can be assigned to the lexical component of the base in the following way. First of all, the context·free subcategorization rules. such as {57 ix-xiii), can be regarded a syntactic redundancy rules. and hence assigned to the lexicon. Consider. ten. te rules that introduce contextual features. These rules select certain frames in which a symbol appears, and they assig coresponding contextual features. A lexical entry may be substituted in these positions if it con textual features match those of the symbol for which it is sub stituted. Obviouly. the contextual features must appear in lexical items. But the rules that introduce contexual features into complex symbols can be eliminated by an appropriate reformulation of the lexical rule. tat is. te rule that introduces lexical i tems into derivations (cf. p. 84) . Instead of formulating this a a context-free rule that operates by matching of complex symbols. we cn convert it to a context-sensitive rule by con ventions of the following sort. Suppose that we have a lexical entry (D. C) where D is a phonological feature matrix and C is a complex symbol containing the feature [+X -Y. We stipu lated previously tat the lexical rule perits D to replace the symbol Q of the preterminal string (Q1 provided that Q is not distinct fom C. Suppose that we now require. in addition, that this occurrence of Q actually appear in the frame X ¯Y. That is. we rquire that (QI equal ( 1 (2Q1112' where (2 is dominated by X and 11 by Y in the Phrase-marker of (Q1. This convention can be formulated precisely in terms of the notion "Analyz ability" on which the theory of transformations is based. We now eliminate all context-sensitive subcategorization rules from the grammar and rely on the formulation of lexical features, together with the principle j ust stated. to achieve their efect. Our earlier condi tions on subcate g orization rules (cf. § 2. t) become conditions on the kinds of contextual features that may appear in lexical entries. Thus strict subcategorization features for an item of the category A must involve frames that, togther II2 CATGORIES AND RELATONS IN SYNTACC 'EORY with A, form the single constituent B that immediately dominates A; and the selectional features must involve the lexical categories that are the heads of gammatically related phrases, in the sense outlined earlier . We now have no subcategorization rules in the categorial com ponent of the base. A preterminal string is generated by the branching rules of the categorial component . Lexical entries sub stitute for the lexical categories of a preterminal string by the principle just stated . This forulation brings out very clearly the sense in which our utilization of complex symbols was a device for introducing transformational rules into the base com ponent. In fact, suppose that (for uniformity of specifcation of transformational rules) we add the convention that in the cate gorial component, there is a rule A � ! for each lexical category A, where ! is a fxed "dummy symbol. " The rules of the cate grial component wll now generate Phra s e-marker of string consisting of various occurrences of ! (marking the positions of lexical categories) and grammatical formatives . A lexical entry is of the form (D, C), where D is a phonological matix and C a complex symbol . The complex symbol C contains inherent features and contextual features. We can restate this system of features C directly as the structure index I for a certain sub stitution transformation. This transforation substitutes (D, C) (now regarded as a complex terinal symbol -see note 1 5) for a certain occurrence of ! in the Phrase·marker K if K meets the condition I, which is a Boolean condition in terms of Ana lyzabiIity in the usual sense of transforational grammar. Where strict subcategorization is involved, the substitution transfora tion is, furthermore, strictly local in te sense of note 1 8. Thus the categorial component may very well be a context-free constituent structure grammar (simple phrase structure gram mar) with a reduced terminal vocabulary (that is, with all lexical items mapped into the single symbol !) . The lexicon con s ist s of entries associated with certain substitution transforations that introuce lexical item into strings generated by the categrial component. All contextual restrictions in the base are provided by tese trasformational rules of the lexicon . The function of § 4. TYPES OF BASE RULES the categorial component is to defne the system of gammatical relations and to determine the ordering of elements in deep structures. This way of developing the base component is not quite equivalent to that presented earlier. The earlier proposal was somewhat more restrictive in certain respcts. In both forula tions, the contextual features (structure indices of substitution transforations) that may appear in the lexicon are limited by the conditions on strict subcategorization and selectional rules previously discussed. But in the earlier formulation. wi th sub categorization rules given as rewriting rules, there is a further restriction. The ordering of the rewriting rule A � CS places an additional limitation on the class of contextual features that may be used. Similarly, the issue discussed in § 4. 2 regarding examples (66)-(68) does not arise in the new formulation. Because of the greater fexibility that it allows. certain Verbs can be restricted in terms of Subject and Object selection. some in terms of Subject selection, and some in terms of Object selection. It is an interesting question whether the geater fexibility permitted by the approach of this subsection is ever needed. If so, this must be the preferable formulation of the theory of the base. If not, then the other formulation, in terms of a lexical rule based on the distinctness condition. is to be preferred. We shall return to this question in Chapter 4. § 4+ The role of categorial rules We have defned the categorial component as the system of rewriting rules of the base -that is. the system of bae rules exclusive of the lexicon and the subcategorization rules that we, for the present, regard as belonging to the lexicon. The rules of the categorial component carry out two quite separate func tions : they defne the system of grammatical relations, and they determine the ordering of elements in deep structures. At least the frst of these functions appears to be carried out in a very general and perhaps universal way by these rules. The trans forational rules map deep s tructures into surface structures, CATEGORIES AND RELATIONS IN SYNACIC TEORY perhaps reordering elements in various ways in the course of this operation. It has been suggested several times that these two functions of the categorial component be more sharply separated, and that the second, perhaps, be eliminated completely. Such is the import of the proposals regarding the nature of sytactc structure to be found in Curry ( 1 961) and Saumjan and Soboleva ( 1 963).34 They propose, in essence, that in place of such rules as (69), the categorial component should contain the corresponding rules (70), where the element on the right is a set rather than a string: (69) S - NP""VP VP - V""NP (70) S - {NP, VP} VP - {V, NP} In (70), no order is assigned to the elements on the right-hand side of the rule; thus {NP, VP} ¯ {VP, NP}, although NP""VP o VP"NP. The rules (70) can be used to defne grammatical relations in exactly the way indicated for the rules (69). The rules (69) convey more information than the corresponding rules ( 70 ) , since they not only defne an abstract system of gram matical relations but also assign an abstract underlying order to the elements. The Phrase-marker gnerated by such rules as (69) will be representable as a tree-diagram with labeled nodes and labeled lines; the Phrase-marker generated by such rules as (70) will be reprsentable as a tree-diagram with labeled nodes and unlabeled lines. Proponents of set-systems such as (70) have argued that such systems are more "abstract" than concatenation-systems such as (69), and can lead to a study of grammatical relations that is independent of order, this being a phenomenon that belongs only to surface structure. The greater abstractness of set-systems, so far as grammatical relations are concerned, is a myth. Thus the grammatical relations defned by (70) are neither more nor less "abstract" or "order-independent" than those defned by (69) ; § 4. TPE OF BAE RULS in fact, the systems of grammatical relations defned in the two cases are identical. A priori, there is no way of determining which theory is correct; it is an entirely empirical question, and the evidence presently available i overwhelmingly in favor of con catenation-systems over set-systems, for the theory of the categorial component. In fact, no proponent of a set-system has given any in dication of how the abstract underlying unordered structures are converted into actual strings with surface structures. Hence, the problem of giving empirical support to this theory has not yet been faced. Presumably, the proposal that the categorial component should be a set-system entails that in a set of syntactically related struc tures with a single network of grammatical relations «(or ex ample, "for us to please John is difcult, " "it is difcult for us to please John," "to please John is difcult for us," or "Joh is difcult for us to please"), each member is directly related to the underlying abstract representation, and there is no internal organization -that is, no order of derivation -within the set of structures. But, in fact, whenever an attempt to account for such structures has actually been undertaken, it has invariably been found that there are strong reasons to assign an internal organization and an inherent order of derivation among the items constituting such a set. Furthermore, it has invariably been found that diferent sets in a single language lead to the same decision as to the abstract underlying order of elements. Hence, it seems that a set-system such a (70) must be supple mented by two sets of rules. The frst set will assign an intrinsic order to the elements of the underlying unordered Phrase markers (that is, it will label the lines of the tree-diagrams representing these structures). The second set of rules will be grammatical transformations applying in sequence to generate surface structures in the familiar way. The frst set of rules simply convert a set-system into a concatenation-system. It provides the bae Phrase-markers required for the application of the sequences of transformations that ultimately form surface structures. There is no evidence at all to sugest that either of these steps can be omitted in the cae of natural languages. Con- 1 26 CATEGORIES AND RELATIONS IN SYNTACIC TEORY sequently, there is no reason to consider the set-system, for the time being, as a possible theory of grammatical structure. The phenomenon of so-called "free word order" is sometimes mentioned as relevant to this issue, but, so far as I can see, it has no bearing on it at all. Suppose that for some language each permutation of the words of each sentence were to give a grammatical sentence that, in faci, is a paraphrase of the original. In this cae, the set-system would be much superior for the categorial component of the grammar of this language. No gam matical transformations would be needed, and the rule for realizing underlying abstrac t representations would be extremely simplee But there is no Jnown language that remotely resembles this description. In every known language the restrictions on order are quite severe, and therefore rules of realization of abstract structures are necessary. ' Until some account of such rules is suggested, the set-system simply cannot be considered seriously as a theory of grammar. Nevertheless, the free word order phenomenon is an interesting and important one, and much too little attention has been given to it. First of all, it should be emphasized that grammatical transforations do not seem to be an appropriate device for e7prssing the full range of possibilities for stylistic inversion. It seems, rather, that there are several underlying generalizations that determine when such reordering is permissible, and what its semantic functions are. For one thing, richly infected languages tolerate stylistic reordering much more extensively than languages that are poor in infection, for obvious reasons. Second, even richly infected langÎages do not seem to tolerate reordering when it lead to ambiguity. Thus in a German sentence such as "Die Mutter sieht die Tochter, " in which the infections do not sufce to indicate gammatical function, it seems that the interpretation will invariably be that "Die Mutter" is te Subject (unless it has contrative Stress, in which case it may be taken to be the Subject or the Object) . The same seems to be true in other languags a diverse as Russian (cf. Peshkovskii, 1 956, p. 4 2) and Mohawk. In the latter, the Verb contains afxes desigating the Subject and Object, but where the reference is § 4. TYPES OF BASE RULE ambiguous, the initial NP is taken to be the Subj ect, under normal intonation (I am indebted to Paul Postal for this in formation). If this is universal, it suggests the generalization that in any language, stylistic inversion of "major constituents" (in some sense to be defned) is tolerated up to ambiguity -that is, up to the point where a structure is produced that might have been generated independently by the grammatical rules. (As a special case of this, then, it will follow that infected languages will tolerate reordering much more freely than uninfected ones.) Something of this sort seems to be true, and it is not statable in terms of the theory of transformations. In general, the rules of stylistic reordering are very diferent from the grammatical transformations, which are much more deeply embedded in the grammatical system.35 It might, in fact, be argued that the former are not so much rules of grammar as rules of performance (cf. §§ Ï j 2, of Chapter ) ). In any event, though this is surely an interesting. phenomenon, it is one that has no apparent bearing, for the moment, on the theory of grammatical structure. 3 Deep Structures and Grammatical Transormations LET us adopt, tentatively, the theory of the base component sketched in § 4. 3 of Chapter 2, and continue to use te fragent of § 3, Chapter 2, appropriately modifed to exclude sub categrization rules from the categrial component of the base, as an illustrative example of a grammar. The base will now generate base Phrase·markers. In § 1 , Chapter 1 , we defned the basis of a sentence as the sequence of base Phrase·marker that underlies it. The basis of a sentence is mappd into the sentence by the transformational rules, which, furthermore, automatically asig to the sentence Ü derived Phrase-marker (ultimately, a surface structure) in the process. For concreteness, consider a base component which generates the Phrase-markers ( 1 )-(3). 1 The base Phrase-marker (3), with a diferent choice of Auxiliary, would be the basis for the sentence "John was examined by. a specialist. " The Phrase-marker ( I ) would be the basis for the sentence "the man was fred," were we to modify it by deleting S' from the Determiner associated with MÛÎ. (In this case, the passive transformation is followed by the deletion of unspcifed agent.) As it stands, however, to form the basis for some sentence, the base Phrase-marker ( Ï) must be supplemented by another Phrase-marker, a transform of which will fll the position of S' in ( Ï) and thus serve as a relative clause qualifying ÏM. Similarly, (2) alone cannot serve as a basis for a sentence because the S' appearing in the Verbal Complement 128 DEEP STRUCTRES AND GRMMATCAL TNSFORMATONS 129 ( 1 ) #-s-# NP Predicate-Phrase I � A Aux VP I � past V | fre #-s-# NP � Det N A I the S' man � NP � Det N Predicate-Phrase � Aux VP Manner � b passive I I I � the man pat V NP I I persuade N I J o hn � o f NP A N S' I b 1 30 DEEP STUGURES AND GRAMMATICAL TRANSFORMATIONS C) #-s-# � NP � Det N I I a specialist Predicate-Phrase � Aux VP I � nom V NP Manner I I / examine N by pasive I John must be replaced by the transform of some other Phrase-marker_ In fact. however. the sequence of base Phrase-markers ( Î )¿ (2). (3) is the basis for the well-formed sentence (4) the man who persuaded John to be examined by a specialist was fred The "transformational history" of (4) by which it is derived from its basis might be represented. informally. by the dia gram (5). ( 5 ) ( 1 ) _ DEEP STRUCTURES AND GRAMMATICAL TANSFORMATIONS 1 31 We interpret this as follows: First, apply the Passive trans formation T _ to the base Phrase-marker ( 3) ; embed the result in the base Phrase-marker (2), in place of S', by a generalized (double-base) substitution transformation T E, giving a Phrase marker for "the man persuaded John of A John nom be examined by a specialist"; to this apply frst T ¿ which deletes the repeated NP "John, " and then Tto, which replaces "of A nom" by "to," giving a Phrase-marker for "the man persuaded John to be examined by a specialist"; next embed this in the position of S' in ( I ) , by T E; to this apply the relative transformation T j which permutes the embedded sentence with the following N and replaces the repeated phrase "the man" by "who," giving a Phrase-marker for "A fred the man who persuaded John to be examined by a specialist by passive"; to this Phrase-marker apply the passive transformation and agent deletion (T )¿ giving (4)' I have left out of this description quite a few transformations that are necessary to give the correct form of (4), as well as other details, but these are, by and large, well known, and introduction of them changes nothing relevant to this discussion. The diagram ( 5 ) is an informal representation of what we may call a Transformation-marker. It represents the transforma tional structure of the utterance ( 5) very much in the way a Phrase-marker represents the phrase structure of a terminal string. In fact, a Transformation-marker may be formally rep resented as a set of strings in an alphabet consisting of base Phrase-markers and transformations as its elements, just as a Phrase-marker may be formally represented as a set of strings in an alphabet consisting of terminal symbols, category symbols, and with the developments of the preceding sections, specifed features.2 The deep structure of an utterance is given completely by its Transformation-marker, which contains its basis. The surface structure of the sentence is the derived Phrase-marker given as the output of the operations represented in the Transformation marker. The basis of the sentence is the sequence of base Phrase markers that constitute the terminal points of the tree-diagram (the left-hand nodes, in (5» . When Transformation-marker are 1 3. DEEP STRUCTURES AND GRAMMATICAL TANSFORMATONS represented as in (5), the branching points correspond to gn eralized transformations that embed a constituent sentence (the lower branch) in a designated position in a matrix sentence (the upper branch). A theoretical apparatus of this sort, in its essentials, is what underlies the work in transformational generative grammar that has appeared in the last ten years. However, in the course of this work, several important points have gradually emerged which suggst that a somewhat more restricted and conceptually simpler theory of transformations may be adequate. First, it has been shown that many of the optional singulary transformations of Chomsky (1 955. 1 957, 1 962) must be re formulated as obligatory transformations, whose applicability to a string is determined by presence or absence of a certain marker in the string. This was pointed out by Lees ( 1 96oa) for the negation transformation, and by Klima (personal communica tion) for the question transformation, at about the same time. In fact, it is also true for the passive transformation, as noted in § 2 . 3 .4 of Chapter 2. Katz and Postal ( 1 964) have extended these observations and formulated them in terms of a general principle, namely that the only contribution of transformations to semantic interpretation is that they interrelate Phrase-markers (i.e., combine semantic interpretations of already interpreted Phrase-markers in a fxed way).8 It follows, then, that trans formations cannot introduce meaning-bearing elements (nor can they delete lexical items unrecoverably, by the condition men tioned in note 1 ) . Generalizing these remarks to embedding transforations, they conclude also that a sentence transform embedded in a matrix sentence I must replace a dummy symbol of I. (In the foregoing discussion, adopting this suggstion, we have used S' as the dummy symbol -this assumption is also implicit in Fillmore, 1 963 . ) Katz and Postal point out that the principle just stated greatly simplifes the theory of the semantic component, since semantic interpretation will now be independent of all aspects of the Transformation-marker except insofar as this indicates how ba structures are interrelated. They have also succeeded in DEEP STRUCURES AND GRMMATICAL TANSFORMATONS 1 33 showing that in a larg variety of cases, where this general principle has not been met in syntactic description, the descrip tion was in fact incorrect on internal syntactic grounds. The principle, then, seems very plausible. Second, notice that the theory of Transformation-markers permits a great deal of latitude so far as ordering of transforma tions is concerned. Thus the grammar, in this view, must con tain rules gnerating the possible Transformation-markers by stating conditions that these objects must meet for well-formed ness (what Lees, 1 960a, calls "trafc rules"). 4 These rules may state the ordering of transformations relative to one another, and may designate certain transformations as obligatory, or obliga tory relative to certain contexts, by requiring that they appear in specifed positions in Transformation-markers. However, only some of the possibilities permitted by this general theory have been realized convincingly with actual linguistic material. In particular, there are no known cases of ordering among gen eralized embedding transforations although such ordering is peritted by the theory of Transformation-markers. Further more, there are no really convincing cases of singulary trans formations that must apply to a matrix sentence before a sen tence transform is embedded in it, though this too is a possibility, according to the theory. 1 On the other hand, there are many examples of ordering of singulary transformations, and many examples of singulary transformations that must apply to a constituent sentence before it is embedded or that must apply to a matrix sentence after embedding of a constituent structure in it. Thus the diagram (5) is typical of the kind of structure that has actually been discovered in Transformation-markers. In brief, presently available descriptive studies suggest the following restrictions on ordering of transformations. The sin gulary transformations are linearly ordered (perhaps only partially ordered). They may apply to a constituent structur before it is embedded, or to a matrix structure, and the constituent struc ture embedded in it, after this constituent structure is embedded. There is no reason for imposing an extrinsic order on the gneralized transformations.8 1 34 DEEP STRUCTURS AND GRAMMATCAL TANSFORMATONS These observations sugest a possible simplifcation of the theory of transformational grammar. Suppose that we eliminate the notions "generalized transformation" and "Transformation marker" altogether.7 In the rewriting rules of the base (in fact, in its Lategorial cmponent) the string #S# is intoduced in the positions where in the illustrative example we introduced the symbol S'. That is. wherever a base Phrase-marker contains a position in which a sentence transform is to be introduced, we fll this position with the string #S#, which initiates derivations. We now allow the rules of the base to apply cyclically. preserv ing their linear order. Thus, for example. after having generated (Î )y with #S# in place of S'. they reapply to the new occurrence of #S# in the terminal line of the derivation represented by ( Ï ). From this occurrence of #S# the rules of the base can generate the derivation rpresented by (2), with #S# in place of the occurrence of S' in (2). From te latter occurrence of ISH. the same base rules can reapply to form the derivation represented by (3 ) . In this way. the base rules will generate the generalized Phrae-marker fored from ( 1 ), (2), (3) by replacing S' in ( 1 ) by (2) and replacing S' in (2) by (3). We have thus revised the theory of the base by allowing #S# to appear on the right in certain branching rules. where pre viously the dummy symbol S' had appeared, and by allowing the rules to reapply (preserving their order) to these newly intro duced occurrences of ISH. A generalized Phrase-marker formed in this way contains all of the base Phrase-markers that constitute the basis of a s entenc�, but it contains more information than a basis in the old sense since it also indicates explicitly how these base Phrase-markers are embedded in one another. That is, the generalized Phrase-marker contains all of the information con tained in the basis, as well as the information provided by the generalized embedding transformations.8 In addition to the rules of the base. so modifed. the gammar contains a linear sequence of singulary transformations. These apply to generalized Phrase-markers cyclically, in the following manner. First. the sequence of transformational rules applies to the most deeply embedded base Phrase-marker. (For example. DEEP STRUCTURES AND GRMMATICAL TRANSFORMATIONS 1 35 it applies to (3), in the generalized Phrase-marker formed by embedding (3) in (2) and the result in ( Ï )¿ as described earlier.) Having applied to all such base Phrase-markers, the sequence of rules reapplies to a confguration dominated by S in which these base Phrase-markers are embedded (to (2), in the same example), and so on, until fnally the sequence of rules applies to the confguration dominated by the initial symbol S of the entire gneralized Phrase-marker (to ( I ), in our example). Notice that in the case of ( 1 )-(3), the efect of this convention is precisely what is described in the Transforation-marker (5). That is, singulary transformations are applied to constituent sentences before they are embedded, and to matrix sentences after embed ding has taken place. The embedding itself is now provided by the branching rules of the base rather than by gneralized trans formations. We have, in efect, converted the specifc properties of the Transformation-marker (5) into general properties of any possible transformational derivation. The grammar now consist of a base and a linear sequence of singulary transformations. These apply. in the manner just de scribed. The ordering possibilities that ar permitted by the theory of Transformation-marker but apparently never put to use are now excluded in principle. The notion of Transforma tion-marker disappears, as does the notion of gneralized transformation. The base rules form gneralized Phrase-markers that contain just the information contained in the basis and the gneralized transformations of the earlier version. But observe that in accordance wit the Katz-Postal principle discussed earlier (p. 1 32), it is precisely this information that should be relevant to semantic interpretation. Consequently, we may take a generalized Phrase-marker, in the sense j ust defned, to be the deep structure generated by the syntactic component. Thus the syntactic component consists of a base that gnerates deep structures and a transformational part that maps them into surface structures. The deep structure of a sentence is submitted to the semantic component for semantic interpretation, and its surface structure enters the phonological component and under goes phonetic interpretation. The fnal efect of a grammar, then, 1 36 DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATONS is to relate a semantic interpretation to a phonetic representation -that is, to state how a sentence is interpreted. This relation is mediated by the syntactic component of the grammar, which constitutes its sole "creative" part. The branching rules of the base (that is, its categorial com ponent) defne grammatical functions and grammatical relations and determine an abstract underlying order (cf. § 4. 4, Chapter 2) ; the lexicon characterizes the individual properties of particular lexical items that are inserted in specifed positions in base Phrase-markers. Thus when we defne "deep structures" as "structures generated by the base component," we are, in efect, assuming that the semantic interpretation of a sentence depends only on its lexical items and the grammatical functions and rela tions represented in the underlying structures in which they appear.9 This is the basic idea that has motivated the theory of transformational grammar since its inception (cf. note 33, Chap ter 2). Its frst relatively clear formulation is in Katz and Fodor (1 963), and an improved version is given in Katz and Postal ( 1 964), in terms of the modifcation of syntactic theory proposed there and briefy discussed earlier. The formulation just sug gested sharpens this idea still further. In fact, it permits a further simplifcation of the theory of semantic interpretation presented in Katz and Postal ( 1 964), since Transformation-markers and generalized transformations, as well as "projection rules" to deal with them, need no longer be considered at all. This formula tion seems to be a natural extension and summary of the develop ments of the past few years that have just been summarized. Notice that in this view one major function of the transforma tional rules is to convert an abstract deep structure that expresses the content of a sentence into a fairly concrete surface structure that indicates its form.lo Some possible reasons for such an organization of grammar, in tr of perceptual mechanis m s, are suggsted in Miller and Chomsky ( 1 963, § 2. 2). It is interesting to note, in this connection, that the grammars of the "artifcial languages" of logic or theory of programming are, apparently without exception, simple phrase structure grammars in mo s t signifcant respect. DEEP STRUCTURES AND GRAMMATICAL TANSFORMATONS 1 37 Looking more closely at the recursive property of the gram mar, we have now sugsted the following modifcation of trans formational theory. In the earlier version of the theory, the recursive property was assigned to the transformational com ponent, in particular, to the gneralized transformations and the rules for forming Transformation-markers. Now the recursive property is a feature of the base component, in particular, of the rules that introduce the initial symbol S in designated positions in strings of category symbols. There are, apparently, no other recursive rules in the base. ll The transformational component is solely interpretive. It is worth mentioning that with this formulation of the theory of transformational grammar, we have returned to a con ception of linguistic structure that marked the origins of moder syntactic theory, namely that prsented in the Grammaire ge nerale et raionnee. 1 One additional point must be emphasized in connection with the notion "deep structure. " When the base rules generate a Phrase-marker from an occurrence of S that is embedded in an already generated Phrase-marker, they cannot take account of the context in which this occurrence of S appears. For example, instead of the gneralized Phrase-marker M consisting of ( Ï)-(3) (with (3) embedded in (2) and the result embedded in ( I », we might just as well have constructed the gneralized Phrase-marker M' formed from (I), K, and (3), where K is a Phrase-marker difering from (2) only in that man in (2) is replaced by boy in K . But now, at the stag of derivation at which the relative clause transformation ( TB of (5» is applied to K with (3) embedded within i t, we shall have not the string (6) but rather (7) : (6) !. fred the man (I the man persuaded John t o be examined by a specialist I) by passive (7) !. fred the man (I the boy persuaded John to be examined by a specialist I) by passive The string (6) (with its Phrase-marker) is of the form that per mits the re l ative clause transformation to apply, replacing "the man" by "who," since the condition of identity of the two Nouns 1 38 DEEP STRUcURES AND GRAMMATICAL TANSFORMATIONS is met and we thus have a recoverable deletion (cf. note 1 ). But in the case of (7), the transformation will block. Thus the phrase "the boy" cannot be deleted from (7) because of the general condition that only recoverable deletions are permitted -that is, the identity condition of the transformation is not satisfed. ls This is precisely what we want, for obviously the generalized Phrase-marker formed from ( 1 ) , K, (3) does not provide the semantic interpretation of ( 4 ) , as it would if application of the relative clause transformation were permitted in this case. In fact, the generalized Phrase-marker formed from ( 1 ), K, and (3), although generated by the base rules, is not the deep structure underlying any surface structure. We can make this observation precise, in this case, by defning the relative clause transformation in such a way that it deletes the boundary symbol i when it applies. Thus if its application is blocked, this symbol will remain in the string. We can then establish the convention that a well-formed surface structure cannot contain internal occurrences of i. Such occurrences will indicate that certain transformations that should have applied were blocked. The same (or similar) formal devices can be used in a variety of other cases. Putting aside questions of foralization, we can see that not all generalized Phrase-markers generated by the base will under lie actual sentences and thus qualify as deep structures. What, then, is the test that determines whether a generalized Phrase marker is the deep structure of some sentence? The answer is ver simple. The transformational rules provide exactly such a test, and there is, in gneral. no simpler test. A generalized Phrase-marker MD is the deep structure underlying the sentence S, with the surface structure Ms, just in case the transformational rules generate Ms from MD. The surface structure Ms of S is well formed just in case S contains no symbols indicating the blocking of obligatory transformations. A deep structure is a generalized Phrase-marker underlying some well-formed surface structure. Thus the basic notion defned by a transformational grammar is: deep structure MD underlies well-formed surface structure Ms. The notion "deep structure" itelf is derivative from this. The DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATONS 1 39 transformational rules act a a "flter" that permits only certain generalized Phrase-markers to qualify as deep structures. Notice that this fltering function of the transformational component is not an entirely new feature specifc to the version of transformational grammar that we are developing now. In fact, it was also true of the earlier version, though this fact was never discussed in exposition . Thus a sequence of base Phrase markers might have been selected that could not serve as the basis of any sentence; furthermore, any system of rules for generating Transformation-markers would certainly permit cer tain structures that do not qualify as Transformation-markers because of inconsistencies and blocks arising in the course of carrying out the instructions that they represent . In the present version this fltering function is simply brought out more clearly. In § 4. 3 of Chapter 2 we sugsted: (a) that the distributional restrictions of lexical items be determined by contextual fea tures listed in lexical entries, and (b) that these contextual fea tures be regarded as defning certain substitution transforma tions. Thus strict subcategorial and selectional restrictions of lexical i tems are defned by transformational rules associated with these items. We have now observed that the transforma tional rules must also carry the burden of determining the distributional restrictions on base Phrase-markers. Thus the categorial rules that generate the infnite set of generalized Phrase-markers can apparently be context-free, with all distribu tional restrictions, whether of base Phrase-markers or lexical entries, being determined by the (singulary) transformations. Such a description of the for of the syntactic component may seem strange if one considers the generative rules as a model for the actual construction of a sentence by a speaker. Thus i t seems absurd t o suppose that the speaker frst forms a generalized Phrase-marker by base rules and then tests it for well-formedness by applying transformational rules to see if it gives, fnally, a well-formed sentence. But this absurdity is simply a corollary to the deeper absurdity of regarding the system of generative rules as a point-by-point model for te actual construction of a sentence by a speaker. Consider the simpler case of a phrase 1 40 DEEP STRUCTURES AND GRAMMATICAL TANSFORMATIONS structure grammar with no transformations (for example, the grammar of a programming language, or elementary arithmetic, or some small part of English that might be described in these terms) . It would clearly be absurd to suppose that the "speaker" of such a languag, in formulating an "utterance," frst selects the major categories, then the categories into which these are analyzed, and so forth, fnally, at the end of the process, select ing the words or symbols that he is going to use (deciding what he is going to talk about). To think of a generative grammar in these terms is to take it to b a model of prforance rather than a model of competence, thus totally misconceiving its nature. One can study models of performance that incorporate genera tive gr a mmars, and some results have been achieved in such studies. 14 But a generative grammar as it stands is no more a model of the speaker than it is a model of the hearer. Rather, as has been repeatedly emphasized, it can be regarded only as a characterization of the intrinsic tacit knowledge or competence that underlies actual performance. The base rules and the transformational rules set certain conditions that must be met for a structure to qualify a the deep structure expressing the semantic content of some well formed sentence. Given a grammar containing a base component and a transformational component, one can develop innumerable procedurs for actually constructing deep structures. These will vary in exhaustiveness and efciency, and in the extent to which they can be adapted to the problems of producing or under standing speech . One such constructive procedure is to run through the base rules (observin g order) so as to form a gen eralized Phrase-marker M, and then through the transfor tional rules (obsering order) so as to for a surface structure M' from M. If M' is well formed, then M was a deep structur; otherwise, it was not. All deep structures can be enumerated in this way, j ust as they can all be enumerated in many other ways, given the grammar . As noted earlier, the grammar defnes the relation "the deep structure M underlies the well-fored sur face structure M' of the sentence S" and, derivatively, it defnes the notions " M is a deep structure, " " M ' is a well-fored sur- DEEP STRUCURS AND GRAMMATICAL TRANSFORMATIONS 1 4 1 face structure," "s is a well-formed sentence," and many others (such as "s is structurally ambiguous," "s and S' are para phrases, " "s is a deviant sentence formed by violating rule R or condition C"). The grammar does not, in itelf, provide any sensible procedure for fnding te deep structure of a given sentence, or for producing a given sentence, just as it provides no sensible procedure for fnding a paraphrase to a given sentence. It merely defnes these tasks in a precise way. A per formance model must certainly incorporate a grammar; it is not to be confused with a grammar. Once this point is clear, the fact that transformations act as a kind of flter will occasion no sur prise or uneasiness. To summarize, we have now sugested that the form of gram mar may be as follows. A grammar contains a syntactic com ponent, a semantic component, and a phonological component. The latter two are purely interpretive; they play no part in the recursive generation of sentence structures. The syntactic com ponent consists of a base and a transformational component. The base, in turn, consists of a categorial subcomponent and a lexicon. The base generates deep structures. A deep structure enters the semantic component and receives a semantic interpre tation; it is mapped by the transformational rules into a surface structure, which is then given a phonetic interpretation by the rules of the phonological component. Thus the grammar assigs semantic interpretations to signals, this association being me diated by the recursive rules of the syntactic component. The categorial subcomponent of the base consists of a sequence of context-free rewriting rules. The function of these rules is, in essence, to defne a certain system of grammatical relations that determine semantic interpretation, and to specify an abstract underlying order of elements that makes possible the functioning of the trasformational rules. To a large extent, the rules of the base may be universal, and thus not, strictly speaking, part of particular grammars; or it may be that, although free in part, the choice of base rules is constrained by a universal condition on the gammatical functions that are defned. Similarly, the category symbols appearing in base rules are selected from a 1 42 DEEP STRUCTURES AND GRMMATICAL TANSFORMATONS fxed universal alphabet; in fact, the choice of symbol may be largly or prhaps completely determined by the formal role the symbol plays in the system of base rules. The infnite genera tive capacity of the grammar arises from a particular formal property of these categorial rules, namely that they may intro duce the initial symbol S into a line of a derivation . In this way, the rewriting rules can, in efect, insert base Phrase-markers into other base Phrase-markers, this process being iterable without limit. The lexicon consists of an unordered set of lexical entries and certain redundancy rules. Each lexical entry is a set of features (but see note 15 of Chapter 2). Some of these are phonological fea tures, drawn from a particular universal set of phonological features (the distinctive-feature system). The set of phonological features in a lexical entry can be extracted and represented as a phonological matrix that bears the relation "is a" to each of the specifed syntactic features belonging to the lexical entry. Some of the features are semantic features . These, too, are presumably drawn from a universal "alphabet," but little is known about this today, and nothing has been said about it here. We call a fea ture "semantic" if it is not mentioned in any syntactic rule, thus begging the question of whether semantics is involved in syn tax. 15 The redundancy rules of the lexicon add and specify fea tures wherever this can be predicted by gneral rule. Thus the lexical entries constitute the full set of irregularities of the languag. We may construct a derivation of a generalized Phrase-marker by applying the categorial rules in the specifed order, beginning with S, reapplying them to each new occurrence of S introduced in the course of the derivation. In this way, we derive a pre teninal string, which becomes a generalized Phrase-marker when lexical entries are inserted in accordance with the trans forational rules specifed by the contextual features that belong to these lexical entries. The base of the syntactic component thus generates an infnite set of generalized Phrase-markers. The transformational subcomponent consists of a sequence of singulary transformations. Each transformation i fully defned DEEP STUCTURES AND GRAMMATICAL TNSFORMATIONS 1 43 by a structure index, which is a Boolean condition on Analyza bility, and a sequence of elementary transformations. The notion "Analyzable" is determined in terms of the "is a" relation, which, in turn, is defned by the rewriting rules of the base and by the lexicon. Thus transformations may refer to specifed syntactic features as if they were categories. In fact, transforma tions must also be designed so that they can specify and add syntactic features, but we shall not go into this modifcation of the theory of transformational grammar here (see Chapter 4, § 2) . Given a generalized Phrase·marker, we construct a transforma tional derivation by applying the sequence of transformational rules sequentially, "from the bottom up" -that is, applying the sequence of rules to a given confguration only if we have already applied it to all base Phrase-markers embedded in this confguration. If none of the transformations blocks, we derive in this way a well· formed surface structure. In this and only this case, the generalized Phrase-marker to which the transfora tions were originally applied constitutes a deep structure, namely the deep structure of the sentence S, which is the terminal string of the derived surface structure. This deep structure expresses the semantic content of S, whereas the surface structure of S deterines it phonetic form. The interpretive components of a grammar have not been our concer her. Insofar as details of their structure have been worked out, they seem to function in parallel ways. The phono logical component consists of a sequence of rules that apply to a surface structure "from the bottom up" in the tree-diagram representing it. That is. these rules apply in a cycle. frst to the minimal element (formatives), then to the constituents of which they are parts (a constituent of a Phrase·marker being a sub string of its terinal string dominated by a single category sybol), then to the constituents of which these are parts, and so on. untl the maimal domain of phonological processes is reached. (See C homsky, Halle. and Lukof. 1 956; Halle and Chomsky. 1 960. forthcoming; C homsky, 1 962b; Chomsky and Miller. 1 963.) In this way a phonetic representation of the entire sentence is fored on the basis of the intrinsic abstract phono- 1 44 DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS logical properties of its formatives and the categories represented in the surface structure. In a somewhat similar way, the projection rules of the semantic component operate on the deep structure gnerated by the base, assigning a semantic interpretation (a "reading") to each constituent, on the basis of the readings assigned to its parts (ultimately, the intrinsic semantic properties of the foratives) and the categories and grammatical relations rep resented in the deep structure. (See Katz and Fodor, 1 963; Katz and Postal, 1 964; and other papers by Katz listed in the bibliog raphy.) To the extent that grammatical categories and rela tions can be described in language-independent terms, one may hope to fnd universal projection rules, which need not, ther fore, be stated as part of a specifc grammar. Throughout this discussion, we have simply been presuppos ing the theory of grammatical transformations as presented in the references cited, but it is perhaps worth mentioning that this theory, too, can apparently be simplifed in various ways. First, it appears that permutations can be eliminated from the set of elementary transformations in favor of substitutions, deletions, and adjunctions. That is, the derived Phrase-markers that would be provided by permutations may not be necessary in addition to those provided by the other elementary transformations. Elimination of permutations from the base set would greatly simplify the theory of derived constituent structure. 1 6 Second, it seems that the structural analyses that determine the domain of transformations can be limited to Boolean conditions on Aalyzability. That is, quantifers can be eliminated from the forulation of transformations in favor of a general convention on deletion, as mentioned in note 1 3 . If so, this places a severe additional restriction on the theory of transformations. The latter point deserves some further clarifcation. We shall discuss it briefy here and then return to the question in Chapter 4, § 2. 2. We are proposing the following convention to guarantee recoverability of deletion: a deletion operation can eliminate only a dummy element, or a forative explicitly mentioned in the structure index (for example, you in imperatives), or the DEEP STRUCTURES AND GRAMMATICAL TANSFORMATONS 14 5 designated representative of a category (for example. the wh question transforations that delete Noun Phrases are in fact limited to indefnite Pronouns -cf. Chomsky. 1 964 . § 2. 2) . or an element that is otherwise represented in the sentence in a fxed position. To clarify the latter point further. let us defne an erasure transformation as one that substitutes a term X of i t proper analysis for a term Y of its proper analysis (leaving X intact). and then deletes this new occurrence of X which re placed Y. In the example of relativization discussed earlier (pp. 1 28 f . ). if we have the string X g g . � . �������� [ö| thc m8u - [#wh. ¬ thc mau ¬ d cc tcd ct1ucd tO 1 the relative transformation can be formulated a an erasure operation that substitutes the frt term X of the proper analysis for the third term Y. erasing the latter 1 7 in the process. Avoiding details of formalization. which are straightforward within the general theory of transformations. we may say briefy that the erasure operation uses the term X to delete Y in such a case . We say. then. that an erasure operation can use te term X to delete Y just in case X and Y are identical. We shall investigate the exact nature of the required relation between X and Y some what more fully in Chapter 4, pp. I n£. A an additional illustration, consider the refexivization operation (see Lees and Klima, 1 963. for a detailed discussion) . It has frequently been observed that in a sentence such as "John hurt John" or "the boy hurt the boy." the two phonetically identical Noun Phrases are necessarily interpreted as difering in reference; sameness of reference requires refexivization of the second Noun Phrase (this is also true of pronominalization). Variou attempt have been made to build an account of this into the syntactic component, but none has been very convincing. The availability of lexical features sugsts a new approach that might be explored. Suppose that certain lexical i tems are designated as "referntial" and that by a general convention, each occurrence of a rferential item is assigned a marker, say. an integer, as a feature. 18 The refexivization rule can be formulated 1 46 DEEP STRUCURE AND GRAMMATICL TANSFORMATIONS as an erasure operation that uses one Noun Phrase to delete another. A in the case of relativization (cf. note 1 7), the erasure leaves a residue, in particular, the feature [±Human] , and it in troduces the new phonetic element self. Thus when applied to "I hurt I," the frst Noun Phrase is used to delete the second, fnally giving, "I hurt myself. " But by the recoverability condition on deletion, the refexivization rule (similarly, the pronominalization rule) will apply only when the integer assigned to the two items are the same. The semantic component will then interpret two referential items as having the same reference just in case they are strictly identical -in particular, in case they have been assiged the same integer in the deep structure. This gives the right answer in many cases, but there are interesting problems that arise when the referential items are plural, and of coure there are problems in specifying the notion "referential" properly . Notice, incidentally, that the refexivization rule dos not al ways apply (though pronominalization does) even when the two Nouns are strictly identical and hence coreferential. Thus we have "I kept it near me" alongide of "I aimed it at myself," and so on. The diference is that in the frt, but not the second, the repeated Noun is in a Sentence-Complement to the Verb . Thus "I kept it near me" has a deep structure of the form "I kept - it I S I," where S dominates "it is near me." But "I aimed it at myself" has a deep structure of the for "I ¯ aimed - it - at me" (there is no underlying sentence "it is at me") . The refexivization rule does not apply to a repated N dominated by an occurrence of S that does not dominate the "antecedent" occurrence of N. This particular remark about English is, apparently, a consequence of a more general condition on transforations, namely that no morphological material (in this case, self) can be introduced into a confguration dominated by S once the cycle of transformational rules has already com pleted its application to this confguration (though items can still be extracted from this constituent of a largr "matrix structure," in the next cycle of transformational rules) . There are a few examples that seem to confict with this analysis (such a "I DEEP STRUCTRES AND GRAMMATICAL TRANSFORMATIONS 1 4 7 pushed it away from me, " "I drew it toward me"), for reasons that I do not understand, but it LLNC1b Ü lÜ1gC number of convincing cases, and, in the distinction it makes between super fcially analogous cases that difer only in that one but not the other is based on an independently existing embedded sentence, it provides Ü interesting confrmation of the theory of trans formational grammar. Returning to the main theme, we can apparently defne a grammatical transformation in terms of a "structure index" that is a Boolean condition on Analyzability and a sequence of elementary transformations drawn from a base set including substitutions, deletions, and adj unctions. It seems also that these form larger repeated unit (for example, substitution-deletions, erasures) and that the limitations on their application can be given by gneral conventions of the sort just mentioned. If this is correct, then the foÏmal properties of the theory of transforma tions become fairly clear and reasonably simple, and it may be possible to undertake abstract study of them of a sort that has not been feasible in the past. 4 Some Residual Problems § I. THE BOUNDARIES OF SYNTAX AND SEMANTICS § I. I. Degrees of grammaticalness IT is quite apparent that current theories of syntax and semantics are highly fragmentary and tentative, and that they involve open questions of a fundamental nature. Furthermore, only very rudimentary gramatical descriptions are available, for any language, so that no satisfactory answers can be given for many factual questions. Consequently, the problem sugested by the title of this section can, for the present, be at best a source for speculation. Nevertheless, some of the topics of the preced ing chapters relate to the question of the proper balance between syntax and semantics in a way that deserves at least some further comment. The distinction between strict subcategorization features and selectional features, which is formally well defned, appears to correlate rather closely with an important distinction in language use. Each such contextual feature is associated with a certain rule that limit lexical entries containing this feature to certain context . 1 We can, in each case, construct a deviant sentence by breaking the rule. Thus in § 3 of Chapter 2, Verbs are strictly subcategorized into Intransitives, Transitives, pre-Adjectival, pre-Sentence, etc. In these cases, violation of the rules will give such strings as : ( 1 ) (i) John found sad (ii) John elapsed that Bill will come 148 § 1 . THE BOUNDARIES OF SYNTAX AND SEMANTICS l . I !I (iii) John compelled (iv) John became Bill to leave (v) John persuaded great authority to Bill On the other hand, failure to observe a selectional rule will give such typical examples as (2) (i) colorless green ideas sleep furiously (ii) glf play John (iii) the boy may frighten sincerity (iv) misery loves company (v) they perfor their leisure with dilignce (cf. § 2. 3. 1 of Chapter 2). Clearly, string such as ( 1 ) that break strict subcat g orization rules a#d string such as (2) that break selectional rules are deviant . It is necessary to impose an inter pretation on them somehow -this being a task that varies in difculty or challenge from case to case -whereas there is no question of imposing an interpretation in the case of such strictly well-formed sentences as (3) (i) revolutionary new ideas appear infrequently (ii) John plays golf (iii) sincerity may frighten the boy (iv) John loves company (v) they prfor their duty with diligence Nevertheless, the manner of deviation illustrated in (2) is rather diferent from that in ( 1 ) . Sentences that break selectional rules can often be interpreted metaphorically (particularly, as personifcation -cf. Bloomfeld, 1 963) or allusively in one way or another, if an appropriate context of greater or less com plexity is supplied. That is, these sentences are apparently interpreted by a direct analogy to well-formed sentences that observe the selectional rules in question. Clearly, one would proceed in quite a diferent way if forced t assig an inter pretation to sentences that break strict subcategorization rules, for example, the sentences of ( 1 ). These examples are, I think, typical of a fairly wide class of SOME REIDUAL PROBLEMS cases. A descriptively adequate grammar should make all of these distinctions on some formal grounds. and a gammar of the type j ust described seems to make them in some measure. at leat . It distinguishes prfectly well-fored sentences such as (3) from the sentences of ( 1 ) and (2) . which are not directly generated by the system of grammatical rules. It further separates the sentences of ( 1 ). gnerated by relaxing strict subcategorization rules, from sentences such as (2). which are generated when selectional rules are relaxed. Thus it taks several steps toward the development of a signifcant theory of "degree of gram maticalness." 2 It seems that sentences deviating from selectional rules that involve "higher-level" lexical features such as [Count] are much less acceptable and are more difcult to interpret than those tat involve such "lower-level" features as [Human] . At the same time. it is important to bear in mind that not all rules involving low-level syntactic features tolerate deviation as readily as do selectional rules involving these features.8 Thus both of the sentences (4) (i) the book who you read was a best seller (ii) who you met i John result from failure to observe rules involving the feature [Human], but are totally unacceptable -although of course an interpretation can easily. and no doubt uniformly, be imposed on them. Both in degee of acceptability and manner of interpreta tion, they difer completely from sentences that result from a failure to observe selectional rules involving the feature [Human] . Thus no matter how selectional rules are treated, there is no doubt that such features a [Human] play a role in purely syn tactic rules (since surely the examples of (4) are ruled out on purely syntactic grounds). Similarly, consider the selectional feature [[+Abstract] ' " - . . . [+Aimate]] assigned to such Verbs as frighten, amMe@ charm. . . . . This feature is involved in rules that are as inviolable as those that give the book whiLh you read was a best seller and what you found was my book, while excluding (4). Thus items § L - THE BOUNDARIES OF SYNTA AND SEMANTICS that are posi tively specifed with respect to this feature can ap pear in the posi tion of pure Adjectives, so that we have such sentences as a ver frightening (amusing, charming, . . . ) person suddenly appeared, but not, for example, (5) (i) a very walking person appeared (ii) a very hitting person appared These sentences, like those of (4), are immediately and perhaps uniquely interpretable, but are obviously much more seriously ungrammatical, in the intuitive sense that we are now attempting to explicate, than the examples of violation of selectional rules given earlier. Thus it seems that this selectionally introduced contextual feature is also involved in rules that cannot be violated without serious departure from gammaticalness .. Examples such as (4) and (5) therefore support two important observations. First, it is clear that features such as [Human] and [[+Abstract] . . . - . . . [+Animate]] play a role in the func tioning of the syntactic component, no matter how narrowly syntax is conceived, as long as it is agreed that (4) and (5) are syntactically deviant. The special character of the examples of (2) is not attributable to the fact that these sentences violate rules involving "low-level features, " but rather to the fact that the rules that they violate are selectional rules. Second, it is clear from such examples as (4) and (5) that the notion "grammatical ness" cannot be related to "interpretability" (ease, uniqueness, or uniformity of interpretation), in any simple way, at least. There are sentences such as (4) and (5 ) that are uniquely, uni formly, and immediately interpretable, no doubt, although they are paradigm examples of departure frm well-formedness. On the other hand, there are also perfectly well-formed sentences that may pose great difculties for interpretation, and may be subject to a variety of perhaps conficting interpretations . More generally, it is clear that the intuitive notion of grammatical well formedness is by no means a simple one and that an adequate explication of it will involve theoretical constructs of a highly abstract nature, just as it is clear that various diverse factors determine how and whether a sentence can be interpreted. SOME RIDUAL PROBLEMS The attempts descibed in the references of note 2 to give a prcise defnition to at least one dimension of degee of gram maticalness are much more plausible if limited to the question of deviation from selectional rules than if extended to the full range of examples of deviation from well-formedness. In fact, following this sugstion, we might conclude that the only function of the selectional rules is to impose a hierarchy of deviation from gammaticalness on a certain set of sentences, namely those sentences that can be gnerated by selectional constraints while otherwise keeping the grammar unchanged . Observe that the rules of the grammar impose a partial ordering in terms of dominance among the features that constitute a com plex symbol in a Phrase-marker. For example, referring again to the sample Phrase-marker (59) of Chapter 2 and the forative frighten, we have a complex symbol consisting of the features [+V, + - NP, +[+Abstract] . . . - . . . [+Animate]], and others, The rules of the grammar impose the dominance order [+ V]. [+ - NP] . [+[+Abstract] . . . - . . . [+Animate]], as indicated in ( 59) . In terms of this order, we can defne the degree 0/ deviation of a string that result from substituting a lexical item in the position of frighten in this Phrase-marker. The deviation is greater the higher in the dominance hierarchy is the feature corresponding to the rule that i relaxed. In the example given, then, deviance would b geatest if the item substituted for frighten is a non-Verb, less gat if it is a Verb but a non-Transi tive Verb, and still less geat if it is a Transitive Verb that dos not take an Abstract Subject. Thus we should have the following order of deviance: (6) (i) sincerity may virtue the boy (ii) sincerity may elapse the boy (iii) sincerity may admire the boy This seems to give a natural explication for at least one sense of the term "deviance." In this connection, compare the sug gestions of the references of note 2, which consider size of category within which substitution takes place in deterining § 1 . THE BOUNDARIES OF SYNTAX AND SEMANTICS 1 5 3 te degee of gammaticalness (the extent of sytactic deviance) of a string. At the end of § 4. 1 of Chapter 2, it was pointed out that features introduced by strict subcategorization rules dominate features introduced by selectional rules; and in the same section it was further noted that all lexical features are dominated by the symbols for lexical categories. Furthermore, deviation fom selectional rules involNing high-level features is apparently more serious than deviation from selectional rules involving lower level features. These various obserations combine to make the defnition of "degree of deviance" just proposed a rather natural one. If the distinction between strict subcategrization rules and selectional rules noted earlier is generally valid, we might go on to superimpose on the scale of deviance a split into perhaps three general types, namely the types that result from: (i) violation of lexical categor (such as (6i»; (ii) confict with a strict sub categorization feature (such as (6ii) and ( I »; and (iii) confict with a selectional feature (such as (6iii) and (2 » . There are, furtherore, subdivisions within at least the third type. Of course, there are also many other types (such as (4 ) , (5 » .5 This is not surprising, since there are rules of many kinds that can be Niolated. § I.2. Further remark on selectional rules Selectional rules play a rather marginal role in the gammar, although the features that they deal with may b involved in many purely sytactic processes (cf. (4 ) , (5 » . One might propose, therefore, that selectional rules be dropped from the synta and that their function be taken over by the semantic component. Such a change would do little violence to the structure of gam mar as described earlier. Of course, the features that are utilized and introduced by selectional rules would still appear in lexical entries for string. That is, boy would be specifed as [+Human] and frighten as perÏittiÏg an Abstract Subject and Animate Object, etc . , in the lexical entries for these items. Furtherore, if we continue to call a feature of the lexical entry a "syntactic 1 54 SOME RIDUAL PROBLEMS feature" when it is involved in a strictly syntactic rule, then these features of the lexical entry will be syntactic rather tan semantic features (cf. the discussion of (4), (5» . Nevertheless, in ac cordance with this proposal, the grammar will directly generate even such sentences as (2), though not, of course, ( 1 ), as syn tactically well formed. The syntactic component of the grammar would not, in other words, impose a hierarchy of degree of grammaticalness at these lower levels of deviation. This task would now have to be taken over by the semantic component. Lt us continue to suppose that the semantic component is an interpretive device based on projection rules of the type discussed earlier, following Katz, Fodor, and Postal. The projection rules must now be adapted to detect and interpret conficts in feature composition between grammatically related lexical items and, more generally, grammatically related constituents of base string. The earlier discussion of deviance, in particular the defnition of "degree of deviance," can be carried over with little change. The same is true of the comments regarding Noun Verb and Noun-Adjective selectional dominance. With slight reformulation, the same arguments will hold under this revision of the structure of grammar. In § 4. 3 of Chapter 2, we discussed two alternative proposals for dealing with contextual features. The frt was to introduce them by rewri ting rules and to have lexical items introduced into derivations by matching of nondistinct complex symbols (as in Chapter 2, § 3). The second was to regard the contextual features of the lexicon as defning certain substitution trans formations that insert lexical items. As noted there, this is not merely a notational question. We have, then, two open questions in connection with selectional rules, in particular: (i) Do they belong in the syn tactic or the semantic component? (ii) Should they be rewriting rules introducing complex symbols or substitution transforma tions? Without attempting any exhaustive investigation of these questions, I shall now mention briefy some considerations that seem relevant to them. Suppose that we were to introduce selectional features by § 1 . THE BOUNDARIES OF SYNTAX AND SEMANTICS 1 55 rew1iting rules, in accordance with § 3 of ChapteÏ 2 . Notice that the selectional rules difer from the strict subcategorization rules in that they typically involve irrelevant symbols standing between the items that they relate. The rule (57xiv) of Chapter 2 is characteristic of selectional rules in this respect, with its reference to the irrelevant items Aux and Det; it is atypical only in the simplicity of these elements. That this may be more than a purely notational matter is illustrated by (57xv) of Chapter 2, which assigns features of the Subject to a modifying Adjective of the Predicate. As these rules are formulatedg the Adjective would actually be assigned diferent features in these sentences: (7) the boy is sad (8) the boy grew sad In the case of (7), the Adjective would b assigned the feature [[+Human] Aux be -] by rule (57xv) of Chapter 2, whereas in the case of (8) it would be assiged the featuÏe [[+Human] Aux [+V] ] or something of this sort.8 These features have nothing in common, in our term, though they actually identify the same set of lexical items . This is as serious a defciency as the one noted in the case of a grammar that spcifcally distinguishes Animate Subj eLt from Animate Obj ect, etc. (see pp. 1 1 4-1 1 5). We may remedy it andg at the same time, eliminate the reference to irÏelevant intervening contexts in selectional rules by establishing the following convention for these rules . Suppose that we have the rule schema (9) A ? CS / raJ . . - . . . /] where [a] and /] are specifed features or are null (but ei ther one or the other is nonnull). 7 We take (9) to be a pp licable to any string (1 0) XWA VY where X ~ [a, . . . ], Y ¯ /, . . ] , 8 W � W1 [a, . . . ] W2 (or is null) and V � V1/, . . . ] V2 (or is null). The result of Üpplying ( 9) to ( 1 0) is the string SOME RIDUAL PROBLEMS ( Î Î ) XWBVY where B is the complex symbol containing the features of A (or [+A] , if A is a categor symbol) in addition to each contextual feature [+e -"]. where X = [e • • ] and Y ¯ [" . . . ]. (The reader will observe that except for the condition on W, V, the notion of "applicability" and the conventions for complex sym bols are as before. though stated somewhat diferently.) What this means is that the rule (9) assigns to A all contextual features [+e -I]. where [e] is a lexical feature of the nearest complex symbol containing [a] to the left of A, and [I] is a lexical feature of the nearest complex symbol containing [] to the right of A . Thus, in particular, we should now give the rules (57xiv) and (57xv) in the form ( 1 2) and (1 3), respectively: ( 1 2) [+V] � CS/[+N] . . -( . . . [+N]) ( 1 3) Adjective � CS/[+N] . . . - These rules would now have the efect of assiging to frighten the feature [+[+Abstract] -[+Animate]]. in particular, and to sad te feature [+[+Human] -] in the case of both (7) and (8) . In tis way we can avoid mention of irelevant intervening symbols in the statement of Lntext and, more importantly, can avoid the defciency of dual-feature assigment noted in the case of (7) and (8). Within the alterative famework involving substitution transformations, the analogous convention must be established. In this case, it is necessary only to state the condition on W, V of ( 1 0). This condition, however, is not statable directly in the form of a Boolean structure index for a transformation . This fact, though of no great importance, might be taken as suggesting that the system involving rewriting rules is preferable.9 More important are certain questions of interpretation that have some bearing on the form of selectional rules and their placement in the grammar.10 Consider such a typical case of violation of selectional rules as § Ï- THE BOUNDARES OF SYNTA AND SEMANTICS ( 1 4 ) John frightened sincerity This is a deviant sentence. formed by relaxing the restriction of frighten to Animate Direct-Objects. Nevertheless. there are frames in which this restriction can be violated with no consequent unnaturalness. a. for example. in ( 1 5) (i) it is nonsense to speak of (there is no such activity as) frightening sincerity (ii) sincerity is not the sort of thing that can be frightened (iii) one can (not) frighten sincerity Clearly. a descriptively adequate gammar must indicate that ( 1 4 ) is deviant (as in the case of the examples of (2» and that the examples of ( 1 5) are not. There are various ways to approach this problem. Suppose that the selectional rules are included in the syntax. Then ( 1 4) and ( 1 5) are only derivatively generated by the gam mar (in the sense of note 2) ; they are generated with Phrase markers indicating that they depart in a particular respect from grammaticalness. Since ( 1 4 ) nevertheless difers from ( 1 5) in "deviance" from the intuitive point of view. this intuitive notion does not correspond to grammaticalness. Rather. it is presumably a property determined by the joint operation of both the syntactic and the semantic components. Thus the projection rules of the semantic component and the lexical entries for such words as nonsense and speak must be designed in such a way that. although the constituent frighten sincerit of the gneralized Phrase-markers of ( 1 5i-iii) is marked as semantically incongruous. the incongruity is removed by the readings assigned to constit uents dominating it. and consequently the sentences ( 1 5) (but not ( 1 4» are fnally given a nondeviant interpretation. ll This seems to me not at all an unnatural or intolerable consequence. Surely it is not surprising to fnd that an intuitive concept such as "deviance" can be explicated only in terms of theoretical con structs of various sorts. which have in themselves no direct and uniform intuitive interpretation. In further support of this conclusion. one might cite the fact that even strict subcategoriza- SOME RIDUAL PROBLEMS tion rules can apparently be violated without leading necessarily to semantic incongruity, as, for example, in ( 1 6) (i) it is nonsense to speak of (there is no such activity as) elapsing a book (ii) elapsing a book is not an activity that can b performed (iii) one cannot elapse a book Here, too, one might plausibly maintain that base string that deviate signifcantly from grammaticalness are nevertheless con stituents of sentences that receive nondeviant interpretations, by virtue of the semantic properties of certain lexical items and certain constructions. In further support of the argument that grammaticalness cannot, in any event, coincide with the intuitive notion of "deviance," one can cite cases of perfectly grammatical strings that are incongruous on nonsyntactic grounds (cf., for example, p . 77). Thus it seems to me that examples such as ( 1 5) do not present a particularly strong argument for removing selectional rules from the syntactic component and assigning their function to the interpretive semantic rules. Nevertheless, if the latter course is taken, then ( 1 4 ) and ( 1 5) will be directly gnerated by the syn tactic rules, and at least in such cases as these the relation of grammaticalness to intuitive deviance will therefore be much closer. This might be cited as a slight consideration in favor of the decision to eliminate the selectional rules from the syntactic component, and to modify the theory of the semantic component in some way so as to allow it to accommodate these phenomena. We have been considering the possibility of assigning the function of selectional rules to the semantic component. Alter natively, one might raise the question whether the functions of the semantic component as described earlier should not b taken over, in toto, by the gnerative syntactic rules. More specifcally, we may ask whether the cycle of interpretive rules that assign reading to higher nodes (larger constituents) of the underlying generalized Phrase-marker should not be made to apply before some of the syntactic rules, so that the dLtinction between the § Ï - THE BOUNDARIES OF SYNTAX AND SEMANTICS 1 59 two components is, in efect, obliterated. This notion, which is by no means to be ruled out a priori, is explored by Bever and Rosenbaum (forthcoming) , who show that if it is adopted, the internal organization of the syntactic component must be revised in several essential ways. It is clear from this fragmentary and inconclusive discussion that the interrelation of semantic and syntactic rules is by no means a settled issue, and that there is quite a range of possi bilities that deserve serious exploration. The approach I have adopted in Chapter 2, § 3 , is a conservative compromise between the attempt to incorporate the semantic rules strictly within the syntactic component and the attempt to elaborate the semantic component so that it takes over the function of the selectional rules . Evidently, further insight into these questions will await a much more intensive study of semantic interpretive rules than it has yet been possible to undertake. The work of the last few years, I believe, has laid the groundwork for empirical investiga tion of this sort. There is a general theoretical framework parts of which have received empirical support. Within this framework it is possible to formulate certain reasonably clear questions, and it is also fairly clear what kind of empirical evidence would be relevant to deciding them. Alternative positions can be formu lated, but for the present any one that is adopted must be ex tremely tentative. In general, one should not expect to be able to delimit a large and complex domain before it has been thoroughly explored. A decision as to the boundary separating syntax and semantics (if there is one) is not a prerequisite for theoretical and descriptive study of syntactic and semantic rules. On the contrary, the prob lem of delimi tation will clearly remain open until these felds are much better understood than they are today. Exactly the same can be said about the boundary separating semantic sys tems from systems of knowledge and belief. That these seem to interpnetrate in obscure ways has long been noted. One can hardly achieve signifcant understanding of this matter in advance of a deep analysis of systems of semantic rules, on the ] 60 SOME RIDUAL PROBLEMS one hand, and systems of belief, on the other. Short of this, one can discuss only isolated examples within a theoretical vacuum. It is not surprising that nothing conclusive results from this. § I.3. Some additional problems of semantic theory One major qualifcation must be added to this discussion of the relation of syntax to semantics. I have described the semantic component as a system of rules that assig readings to con stituents of Phrase-markers -a system that has no intrinsic structure beyond this. But such a description is hardly sufcient . In particular, there is little doubt that the system of "dictionary defnitions" is not as atomistic as implied by this account . Concerning dictionary defnitions, two major problems are open to investigation. First, it is important to determine the universal, language-independent constraints on semantic features -in traditional terms the system of possible concepts. The very notion "lexical entry" presupposes some sort of fxed, universal vocabular in terms of which these objects are characterized, j ust as the notion "phonetic rpresentation" presupposes some sort of universal phonetic theory . It is surely our ignorance of the relevant psychological and physiological facts that makes possible the widely held belief that there is little or no a priori structure to the system of "attainable concepts." Furthermore, quite apart from the question of universal con straints, it seems obvious that in any given linguistic system lexical entries enter into intrinsic semantic relations of a much mor systematic sort than is sugested by what has been said so far. We might use the term "feld properties" to refer to these undoubtedly sigifcant though poorly understood aspects of a descriptive semantic theory .12 Thus, for example, consider Ad jectives that are mutually exclusive in some referential domain, for example, calor words. Such "antonymy sets" (cf . Katz, 1 964b) provide a simple example of Ü feld property that cannot be described naturally in terms of separate lexical entries, though it obviously plays a role in semantic interpretation . Or consider the "have a" relation, discussed in Bever and Rosenbaum (forth coming). We have § 1 . THE BOUNDARES OF SYNTAX AND SEMANTICS ( 1 7) (i) the man has an arm (ii) the arm has a fnger (iii) the fnger has a cut but not ( 1 8) (i) the arm has a man (ii) the fnger has an arm (iii) the cut has a fngr (except, irrelevantly to this point, a possible elliptic variants of entirely diferent constructions, as in "the fnger has an Û attached to it," "the arm ha a man on it," etc.). These examples, furthermore, illustrate relations of meaning rather than relations of fact . Thus there i no grammatical objection to "the ant ha a kidney, " where "the kidney ha an ant" is not false or impossible but senseless, with the irrelevant exception just noted. In this case, we have a hierarchy of terms with systematic relations that, once again, cannot in any natural way be described within the framework of independent lexical entries . Other systems of this sort can eaily be found, and, in fact, they suggest that part of the semantic component of a grammar must be a characterization of feld properties that is outside the lexicon. This matter is crucial but has been relatively unexplored within any general frame work, though there have been several valuable studies of certain of its aspects. (See note 1 2.) Suppose. furthermore, that an at tempt is made to relate "deviance" in the intuitive sense to "degree of grammaticalness" in the technical sense by excluding such examples as (1 8i-iii) from direct generation (cf. note 1). The consequences of such a decision are not easy to determine. Once again, we can do no more here than indicate problems and stress the fact that there are many unanswered questions of principle that might very well afect the formulation of even those parts of the theory of grammar that seem reasonably well established. Finally. it is important to be aware of the many other problems that face a theory of semantic interpretation of the kind referred to in the preceding discussion . It is clear, as Katz and Fodor have SOME RESIDUAL PROBLES emphasized, that the meaning of a sentence is based on the meaning of its elementary parts and the manner of their com bination. It is also clear that the manner of combination provided by the surface (immediate constituent) structure is in general almost totally irrelevant to semantic interpretation, whereas the grammatical relations expressed in the abstract deep structure are, in many cases, just those that determine the mean ing of the sentence . Cf. , for example, Chapter Ï ¿ § 4. and Chapter 2, § 2. 2. However, there are cases that sugest the need for an even more abstract notion of grÛatical function and gram matical relation than any that has been developed so far, in any systematic way . Consider, for example, these sentence pairs : ( 1 9) (i) John strikes me as pompous -I regard John as pompous (ii) I liked the play - the play pleased me (iii) John bought the book from Bill -Bill sold the book to John (iv) John struck Bill -Bill received a blow at the hands of John Clearly, there is a meaning relation, approaching a variety of paraphrase, in these cases. It is not expressible in transfora tional ters, as is possible, for example, in these cases: (10) (i) John is easy for us to please -it is easy for us to please John (ii) it was yesterday that he came ¯he came yesterday In the case of (20), the deep structures of the paired sentences are identical in all respects relevant to semantic interpretation of the sort we are considering here, so tht the transformatioal analysis accounts for the (cogitive) synonymy. This dos not seem to be true in the case of ( 1 9), however . For example, in the case of ( 1 9i), although the deep structures would show that "pompous" modifes "John" in both sentences of the pair, they would not express the relations of the two Nouns to the Verb that are (in some unclear sense) the semantically signifcant ones. Thus in some sense the relation of "John" to "strike" is § 1 . THE BOUNDARIES OF SYNTAX AND SEMANTICS the same as that of "John" to "regard:' and the relation of "strike" to "me" is the same as that of "regard" to "I. " We have no mechanism for expressing this fact, hence of accounting for the meaning relation, in terms of lexical features or grammatical relations of the deep structure. lS Consequently, it seems that beyond the notions of surface structure (such as "grammatical subject" ) and deep structure (such as "logical subject"), there is some still more abstract notion of "semantic function" still unexplained. Various formal devices for expressing these facts suggest themselves, but the general problem seems to me non trivial. Many related problems have been raised in the extensive dis cussion of the distinction between the "grammatical" Subject and Predicate of Ü sentence and its "logical" or "psychological" Subject and Predicate (see, for example, Paul, 1 886; Jespersen, 1 924; Wilson, 1 926). To mention j ust one, Cook Wilson main tains ( 1 926, pp. 1 1 9 f.) that "in the statement 'glass is elastic: if the matter of inquiry was elasticity and the question was what substances possessed the proprty of elasticity, glass . . . would no longer be subject, and the kind of stress which fell upon 'elastic' when glass was the subject, would now be transferred to 'glass: " Thus in the statement "glass is elastic," " 'glass: which has the stress, is the only word which refers to the supposed new fact in the nature of elasticity, that it is found in glass . . . [and therefore] . . . 'glass' would have to be the predicate. . . . Thus the same form of words should be analyzed diferently according as the words are the answer to one question or another, " and, in gen eral, "the subject and predicate are not necessarily word in the sentence, nor even something denoted by words in the sentence. " Whatever the force of such observations may be, it seems that they lie beyond the scope of any existing theory of language structure or language use. To conclude this highly inconclusive discussion, I shall simply point out that the syntactic and semantic structure of natural languages evidently ofers many mysteries, both of fact and of principle, and that any attempt to delimit the boundaries of these domains must certainly b quite tentative. SOME RIDUAL PROBLEMS § 2. THE STRUCTURE OF THE LEXICON § 2. I. Redundanc The lexicon wa described earlier simply as a set of lexical entries, each consisting of a ditinctive feature matrix D and a complex symbol C, the latter being a set of features of various sorts (syntactic and semantic features, features that specif which morphological or transformational processes apply to strings con taining the items in question, features that exempt items from certain phonological rules, and so on)_ 1 4 We have just seen that this account is oversimplifed in the case of semantic features, further structure being necessary in the lexicon to account for feld properties . Furthermor, in Chapter 2, § 3, we pointed out that variou general conventions can be given that perit signifcant simplifcation of such lexical entries. To explore the question of simplifcation of lexical entries somewhat further, let us, for concreteness, make a specifc choice at each point where, in the discussion, we listed alternative possibilities that seemed to deserve consideration. In particular, let u assume that the proper method for inserting lexical items is by a gneral rule that inserts the lexical entry (D, C) in Ü position . . . Q . . . in a Phrae-marker (Q being a complex symbol developed by rewiting rules), where C is not distinct from Q in the technical sense of feature theory. Thus we tentatively accept the method of § 3 of Chapter 2, rather than that suggsted in Chapter 2, § 4. 3. Furthermore, let us make the empirical assump tion that a grammar is more highly valued if the lexical entries contain few positively spcifed strict subcategorization features and many positively specifed selectional features. Thus we tenta tively accept alternative (iv) of p. 1 1 1 . 15 These choices do afect the following discussion, but analogous problems arise no matter which of the proposed alteratives is selected. We have, in efect, now adopted the following conventions: (2 1) (i) only positively specifed strict subcategorization features and only negatively specifed selectional features appear explicitly in lexical entries, the others being introduced by the auxiliary convention (ii) § 2. T STRUCURE OF T LICON (ii) if the lexical entry (D , C) is not explicitly provided with the feature specifcation [al - " ] for the contextual feature [l -"] (where a ¯ + in the case of a strict subcategorization feature and a ~ - in the case of a selectional feature), then assign it the specifed feature [- ° l - " ] We also pointed out (in Chapter 2, § 3) that a convention anal ogous to (lii) can be established in the case of features cor responding to lexical categories. In accordance with these conventions, we might give the lexical entry for frighten (cf. (58) of Chapter 2) simply as: (22) (frighten , [+V, + -NP, -[+N] -[-Animate], . . . ]) The conventions will introduce: the category features [-N, [-Adjective], [-M] ; the strict subcategorization features [- -], (- -Np ../ ..S ../), • • ; the selectional features [+[+N] [+ Animate]], [+[+N] -[+Human]], . . . . Thus frighten will be specifed (by (u) plus conventions) as a Verb, but not a Noun, Adjective, or Modal ; as insertable in the context sincerity -John but not sincerity Iö or sincerity -justice.11 We can proceed to develop an appropriate convention to simplify lexical representation of items with inherent features in the case where these are hierarchic rather than cross classifying . Let us say tat the sequence of specifed features ([alFl], . . , [a"F,]) (a, ¯ + or -) is a hierarchic sequence with respect to the grammar G i f [agFg] is the only specifed feature directly dominating [a'+lFHl], for each i < n, in G. Thus, for example, wi th respect to the illustrative grammar (57) of Chap ter 2 we have the hierarchic sequences (23) (i) ([+Animate] , [±Human]) (ii) ([+N], [+ Common] , [-Count], [±Abstract]) (H) ([+N], [±Common]) 18 Where such relationships obtain, we can utilize them to simplify lexical entries by the following rather natural convention: 19 (2 4 ) suppose that ([a1Fl], . . . , [a"F,]) is a maximal hierarchic 1 66 SOME RESIDUAL PROBLEMS sequence with respect to the grammar G, and that (D,C) is a lexical entry of G, where C contains [a . F @]. Then C is ex tended automatically to C' containing C along with all of the specifed features CalF,], for each i, 1 :; i < n. Using this convention, we can simplify the lexical entry in (58) of Chapter 2 for boy to the following: (25) (boy, [+Common, +Human, +Count, . . , ]) the features [+N], [+Animate] now being predictable.2 Let us say that the feature [aF is lexically determined in the grammar G if there is a hierarchic sequence ([+K] , . ¯ . , [aF]) with respect to G, where K is a lexical category (a ¯ + or ª). This is · to say that if (D, C) is a lexical entry and C contains [aF, then (D,C) is necessarily a member of the lexical category K, with respect to this entry, and it is unnecessary (by virtue of convention (24» to list [+K] in C. In the sample grammar (57), (58) of § 3, Chapter 2, each lexical item contains lexically deter mined features. Hence, it is unnecessary, in the lexicon of (58), to designate the lexical category for any item. If every lexical entry contains lexically determined features, as seems plausible, then the features [+C] and ["Cl, where C is a lexical category, need never receive explicit mention in the lexicon. We have thus far considered only universal notational con ventions underlying lexical representation. However, there are also many language-specifc redundancies. Thus, for example, every Verb in English that can occur with a Direct-Object and a following Manner Adverbial can occur as well with j ust a Direct-Object, though not conversely. 2 1 The strict subcategoriza tion rules of the grammatical sketch of § 3, Chapter 2, introduced the features [ -NP] and [ -NP Manner] for Verbs, among others. In accordance with the observation j ust made, we see that if a lexical item is specifed in the lexicon as [+ -NP ..Manner], then it must also be specifed as [+ -NP], though not necessarily conversely. For example, read will be specifed positively for both features, but resemble, cost will be specifed positively for [ -NP] and negatively for [ -NP ..Manner] , since we can have "he read the book (carefully, with geat enthusiasm)," "John § 2. THE STUcR OF THE LXICON resembled his father," but not "John resembled his father care fully (with great enthusiasm), " etc. Here again we have a redundancy in the lexicon and a signifcant generalization still not expressed in the grammar. Clearly, what is needed is the following rule: (26) [+ -NP ..Manner] � [+ -NP] to be interpreted in the following manner: if (D,C) is a lexical entry with distinctive feature matrix D and complex symbol C containing [+ -NP ..Manner], then C is replaced by C', which contains each specifed feature [aF of C, where F = [ -NP], and also the specifed feature [+ -NP] . Actually, the rule (26) can be further gneralized. It i s also true of Intransitive Verbs that if they can take a Manner Adverbial, then they can occur without one. What is needed is a convention prmitting a variable over string to appear in the rule gn eralizing (26), thus, in efect, allowing us to use part of the internal structure of the notations for lexical features. Using c as a string variable, we can give the rule in this form: This is to be interpreted as follows: frst, select any constant string as (; second, interpret the result in the manner described in connection with (26) . It might also b expedient to develop the obvious convention that allows (27) to be stated as a con text-sensitive rule, or to allow a condition on ( to be added, where this is well defned in terms of base rules. Let us suppose that the rule (27) applies before the conventions (2 1 ), (24). Then suc words as walk, hit will be entered in the lexicon in this form: (28) (i) (wal k, [+ -Manner, . , . ]) (ii) (hit, [+ -NP .. Manner, . . . ]) By the rule (27) followed by the convention (2 1 ), these will be automatically extended to (29) (i) (walk, [+ -Manner, + -g - - NP .... Manner, - -NP, ¯ ¯ ¯ ]) SOME RIDUAL PROBLEMS (ii) (hit, [+ -NP Manner. + -NP. - -Manner • . . . ]) Thus walk can appear with or without a Manner Adverbial. but with no Direct-Object. and hit can appear with or without Ü Manner Adverbial. but only with a Direct-Object. Rules such as (27). (28) are closely analogous to the phono logical rules that Halle has called "morpheme structure rles" (Halle. 1 959a. Î959b). and that I have been referring to here (following a suggestion of his) as phonological redundancy rules. These rules are designed to deal with te fact that certain phonological feature specifcations are predictable. given others. Thus in an initial sequence ICC in English. if the second C is a true consonant (that is. not Ü liquid or a glide). the frst must be [s] ; if the second consonant i a liquid. the frst must be an obstruent. etc. The phonological redundancy rules that state these facts are precisely of the form (26) and are interpreted in the same way. except that the features in question are phono logical rather than syntactic. and. consequently. the generalization to (27) has no analogue. We shall refer to the analogous syntactic rules (26). (27) a sntactic redundanc rules. The redundancy rules. both phonological and syntactic. state general properties of all lexical entries. and terefore make it unnecessary to provide feature specifcations in lexical entries where these are not idiosyncratic. Observe that a distinction must be made between the con ventions (2 1 ). (24) and the syntactic redundanc rules (26). ( 27 ) . though both play the role of eliminating redundant specifcations from the lexicon. The forer are universal. and therefore need no specifc statement in the grammar. They are part of the procedure for interpreting grammars (the function f of ( uiv) ( Ï 4iv). Chapter Î y § 6). The latter. on the other hand. are partic ular to a given language. and therefore must be given in the grammar.22 I have tried to emphasize this by calling the former "conventions." and the latter. "rules. " Given a lexical entry (D. C). the phonological redundancy rules give a fuller specifcation to D. and the syntactic redundancy § 2. TE STUCUR OF THE LEXICON 1 6 9 rules give a fuller specifcation to C . To this extent, the two systems are analogus. However, there is still an important diference between them, so far as the role that they play is con cerned. To see this, it is necessary to consider an aspect of the system of phonological redundancy rules that has not always been fully appreciated. The fact that there are rules for predicting certain phonological feature specifcations in terms of others ha long been known, and there are many descriptive studies that give charts or rules of one sort or another to specify the set of "phonologically admissible sequences, " "possible syllables§ " and so on. Halle's achievement was not merely to reiterate the fact that such constraints exist but to present Ü principled basis for selection of one set of rules rather than another to determine them. He showed that a very gneral and independently moti vated evaluation procedure for phonology (namely, minimization of feature specifcation) seems to provide such a basis. That is, application of this criterion selects a system of phonological redundancy rules that defnes the notion "phonologically ad missible" in Ü way that, in many crucial cases, conforms to the known fact. 2 3 He thus was able to propose an explanation for the facts of phonological admissibility, in place of a mere description ¯in other words, to give a gneral, langag-in dependent defnition of the notions "accidental gap" (such as, in English, /blik/) and "systematic gap" (such as, in English, /bnik/), in place of an ad hoc chart or list. The real function of the phonological redundancy rules is to determine the class of phonologically admissible (though perhaps nonoccurring) se quences in a principled way. To the extent that they succeed in doing this, they provide empirical support for the lingistic theory that contains the evaluation procedure that Halle proposes, Ü well as the system of constraints on phonological rules that this procedure presupposes. But there is no really con vincing analogue to the notion of "phonological admissibility" in the case of the syntactic redundancy rules. Consequently, it is an open question whether these have the signifcance of the phonological redundancy rules. This observation suggests that we seek an analogue to the SOME RIDUAL PROBLEMS distinction between accidental and systematic gaps, on the syn tactic level. In fact, from a purely formal point of view, the syn tactic redundancy rules do make a distinction between "possible, but nonoccurring lexical entry" and "impossible lexical entr," precisely as the phonological redundancy rules do. In both cases, the redundancy rules provide general constraints on all lexical entries, thus distinguishing possible from impossible lexical entries (possibility with respect to a particular language, that is, insofar as the redundancy rules are not universal conventions). But in general not all of the possibilities will be actually realized in the lexicon. What must be shown is that this formal tripartite distinction of occurring, possible but nonoccurring, and im possible has the signifcance in the syntactic case that it clearly does in the phonological case. Thus what must be shown is that the possible but nonoccuring lexical entries have the status of "accidental semantic gaps" in the sense that they corespond to lexical items that the language does not provide for specifcally but could in principle incorporate with no alteration of the general semantic system within which it functions. I have no very satisfying examples at present.24 The problem is reasonably clear, however, and merits investigation. The study of syntactic redundancy rules is a large topic in itself, but instead of continuing with additional examples, I should like to consider briefy some of the problems that arise in the attempt to deal with morphological processes within a frame work of the sort that has been outlined earlier. § 2.2. Infectional processes It is useful to compare two ways of dealing with questions of infectional morphology, namely the traditional method of para digms and the descriptivist method of morphemic analysis. Since English is to po r in infection to illustrate this diference, we shall tur to Grman for examples. In a traditional grammar, a particular occurrence of a Noun would be described in terms of its place in a system of paradigms defned by certain infectional categories, namely the categories of gender, number, case, and § 2. THE STUCUR OF THE LEXICON declensional typ. Each of these categories constitutes an in dependent "dimension" of the paradigm, and each word has a particular "value" along each of these indepndent dimensions. 26 Thus the word Bruder in the phrase der Bruder would be characterized as Masculine, Plural, Genitive, and belonging to Ü certain declensional clas along with Vater, Mutter, etc . In fact, we can restate the paradigmatic description directly in terms of syntactic featurs. Regarding each of the dimensions of the system of paradigms as a multivalued feature, with the specifcations being not + and - but, let us say, integers con ventionally associated with the traditional designations,28 we can represent the Phrase-marker of the sentence . . . der Brudef . . . as containing the subcfguration (30) . Thus, associated with this Article N defnite B,ider occurrence of Brudef, there will be a feature matrix indicating that this formative is assiged to the categories [1 Gender], [2 Number] , [2 Case], and [1 DC] (as well as to many others represented in (30) simply by ) e Notice that the specifed features [1 Gender] and [ 1 DC] are inherent to this formative (that is, they are part of the complex symbol C of the lexical entry (Bruder, C» and that [2 Number] and [2 Case] are in troduced by grammatical rules . Presumably, the specifed feature SOME RESIDUAL PROBLEMS [2 Number] is introduced by a context-free rule of the base apply ing to Nouns,2T and specifed feature [2 Case] is introduced by a rule that does not belong to the base subcomponent of the syntax at all but rather to it trÜnsforational part (cf . note 35, Chapter 2) . If so, then of these features only [2 Number] will be a feature of the preterinal sybol for which Bruder is substi tuted by the lexcal rule, and all but [2 Case] will appear in the terinal sting generated by the base rules. Notice, incidentally, that the specifcation [1 DC] might be introduced by a re dundancy rule that, in this case, takes into account both pho nological and other lexical features. A rule of the (interpretive) phonological component will operate on (30), giving the form Bruder . This rule will assert that a Vowel is fronted in a formative that is simultaneously of the categries [2 Number], [1 DC]. (A separate rule that is quite general would specify that /(n/ is sufxed if, furthermore, it belongs to the category [3 Case].) In short, the theory of syntactic featurs developed earlier can incorporate the traditional paradigmatic treatment directly. The system of paradigm is simply described as a system of features, one (or perhaps some hierarchic confguration) cor responding to each of the dimensions that defne the sytem of paradigms. InterpretiNe phonological rules, some quite specifc, some of considerable gnerality, then operate on the phonological matrix of the lexical entry, giving, fnally, a phonetic matrix. Where these features are not completely independent (a, for example, if declensional type depends on Gender), or where they are partially determined by other aspects of a formative, re dundancy rules of the kind discussed earlier will apply. The characteristic method of analysis of modem linguistics is rather diferent from the traditional approach that we have just restated in our ters. In plac of the traditional categories (our features), this approac would substitute morphemes. Thus Bruder in (30) would perhaps be represented in the manner of (3 1 ), in a completely consistent "item-and-arrangement" gram mar: § 2. THE STUCUR OF THE LEXICON (3 1 ) Bruder'DC{· Masculine .. Plural Genitive wher each of these elements is regarded as a single morpheme, DCI being a kind of "class marker." 28 Rules would then be given that would convert (3 1 ) into a sequence of phonemes. Respresentations such as (3 1 ) are clumsy for a grammar based on rewriting rules or transformations. There are several reasons for this. For one thing, many of these "morphemes" are not phonetically realized and must therefore be regarded, in par ticular context, as zero elements. In each such case a specifc context-sensitive rule must be given stating that the morpheme in question is phonetically null. But this extensive set of rules is entirely superfuous and can simply be omitted under the alternative paradigmatic analysis. Thus compare the rules that must be provided for the paradigmatic analysis (30) and for the morphemic analysis (3 1). In the case of (31 ) . we frst apply a rule stating that the Vowel is fronted in the context: °DCI ¯ ¯ ¨ Plural ' " g where the itm in question is a Noun. In the case of (30)' we have the corresponding rule that the Vowel is fronted when the item in question has the features [DC I] and [2 Num ber] . But in the case of the morphemic anal y sis we now have the additional rules stating that in such contexts a (3 1 ), all four infectional morphemes are phonetically null. With the feature analysis (30), we simply give no rule at all expressing te fact that certain features are phonetically unrealized, just as we give no rule expressing the fact that [+N], or, for that matter, NP, is phonetically unrealized. 2 9 More generally, the often suppletive character of infectional systems, as well as the fact that (as in the example) the efect of the infectional categories may be partially or even totally interal, causes cumbersome and inelegant formulation of rules when the representations to which they apply are in the form (3 1 ). How ever, suppletion and internal modifcation cause no special difculty at all in the paradigmatic formulation. Similarly, with morphemic representations, it is necessary to refer to irrelevant morphemes in many of the grammatical rules. For example, in the SOME RIDUAL PROBLEMS case of (31 ), the rule for fronting of the Vowel must refer to the morpheme Masculine, and this is te usual situation in the case of agreement rules . But in the paradigmatic representation, these elements, not being part of the terminal string, need not be referred to at all in the rules to which they are not relevant . Finally, notice that the order of morphemes is often quite arbitrary, whereas this arbitrariness is avoided in the para digmatic treatment, the features being unordered . I know of no compensating advantage for the modern de scriptivist reanalysis of traditional paradigatic formulations in terms of morpheme sequences. This seems, therefore, to be an ill-advised theoretical innovation . Within our framework, either paradigmatic analysis in terms of features or sequential morphemic analysis is available, which ever permits the optimal and most general statement of some aspect of the syntactic or phonological system. It seems that in in8ectional sytems, the paradigmatic analysis has many ad vantags and is to be preferred, though there may be cases where some compromise should be made.o It is difcult to say anything more defnite, since there have been so few at tempts to give precise and principled descriptions of in8ectional systems in a way that would have some bearing on the theoretical issues involved here . s l If we assume now that the paradigmatic solution is the correct one, it follows that we must allow the transformational com ponent to contain rules that alter and expand the matrix of features constituting a lexical item . For example, the feature (or features) of Case must in general be specifed by rules that apply after many transformational rules have already taken efect . (See note 35 of Chapter 2.) Similarly, rules of agreement dearly belong to the transformational component (cf. in this connection, Postal, 1 964a, pp . 43f . ), and these rules add to Phrase markers specifed features that enter into particular formatives, dominating their phonological matrices. In the case of (30), for example, the grammar must contain agreement rules that assign to the Article all of the feature specifcations for [Gnder], § 2 . THE STUCR OF THE LEXICON [ N umber] , and [Case] of the Noun it modifes. Thus we must have a rule that might be given in the form: aGender Ge d . a n er ¸ . ¸ Article � [ P Numbe r ] / . . . P N umber ' ' Case C ' ase where Article . . . N is an This rule is interpreted as asserting that in a string analyzable as (X, Article, Y, N, Z, where the second plus third plus fourth elements constitute an the second element is to be assigned to the categories [a Gender], [ Number], and [' Case] if the fourth element is of these categories, a, p, and ' being variables that range over integers . This rule thus asserts that the Article agrees with its Noun in Gender, Number, and Case . In particular, rule (32) assigs to the formative defnite32 in (30) the features [I Gender], [2 Number], [2 Case] . This formative, so categorized, would be converted to /der/ by rules of the phonology. The rule (32) is a transformational rule of the usual kind except that it introduces spcifed features instead of only non lexical formatives. Thus the features play a role which is intermediate between that of formatives and that of true cate gories with respct to the opration of transformational rules, as is quite natural. There is no particular difculty in extending the theory of transformations to allow for the formulation of rules such as (32), which provide an appropriate formalization for traditional rules of agreement . Regarding features as con sti tuent elements of foratives, these transformational rules will, in efect, rewrite terminal symbols in certain restricted ways . Formally, rules of agreement such a (32) are quite analogous to the rules of assimilation of the phonological component . For example, in English, as in many other languags, nasals are neutralized before stops, so that the words limp, lint, link, send, ring would be re p resented / liN p /, IliNt/, /liNk/, /seNd/, /riNg/ in lexical entries, where /N/ ¯ [+nasal] and the other symbols SOME RIDUAL PROBLEMS are also abbreviations for certain sets of µhonological features. The nasal assimilates to the following consonant with respect to the features of gavity and compactness, so that we have the rule (33) [+ nasal] � [ a gave ] I [ : ;;;nantal ] f compact f compact interpreted in the manner of rule (32).33 Thus (33) asserts that the featurs [a grave] and [ compact] are added to a [+ nasal] tat precedes an [a grave], [ compact] consonant, where a, f rang over {+, -} . It asserts, in other words, that the nasal is Iml before labials, 1nl before dentals, and IrJ before velars (where the voiced velar then drops in certain positions, giving Isi1 #I, etc. ¯I have not given the full statement of required context in (33 » . In the case of rule (32), the features added are, apparently, the only features associated with the non lexical item defnite (but cf. note 32)e Other agreement rules expand an already present matrix of featurs -for example, the rule asiging feÜtures of a Noun to a modifying Adjective. The latter, being Ü lexical item, will have an independent feature matrix of its own, which is expanded by the agreement rule. The Adjective, in this case, is introduced into the prenominal position by a transformational rule, and its features will include its inherent features (those given in it lexical entry) and those associated with the complex symbol that it replaces by the lexical rule. It seems, then, that the traditional approach to the de scription of infectional systems can be fonalized quite readily within te framework that we have established. Furthenore, this appear to be the most natural way to deal wi th infectional systems. Before turning to the much more perplexing problems of derivational morphology, I should like to mention a few addi tional problems that arise when infectional features are con sidered in further detail. We have been regarding a lexical item as a set of phonological, semantic, and syntactic features. When inserted into Ü Phrase-marker, a lexical item may acquire other § 2. THE STUCUR OF THE LEXICON features beyond those inherent to it in the lexicon. Thus if we adopt the method of lexical insertion described in § 3 of Chapter 2, then contextual features may be added to the lexical entry beyond those that it already contains; and, quite apart from this, such features as [a Number] are inhernt to the Phrase-marker rather than the lexical item, as we have just observed, and become part of the formative only after it is inserted into a Phrase-marker. Furtherore, the features involved in the case dimension are certainly added to a formative by rather late transformations (since case often depends on aspects of surface rather than deep structure -but see note 35, Chapter 2), and certain features that are inherent to Nouns (such a GendCr) are assigned to Verbs and Adjectives only by transformations. We have been a suming that these various operations simply extend the set of features constituting the formative. But various problems arise if we follow this assumption consistently. We have mentioned in several places (Chapter 3, notes 1 and 1 3, and pp. 1 44f.) that deletions must be recoverable, and have sug gested that this condi tion can be foralized by the following convention relating to what we called "erasure transformations": an erasure transformation can use a ter X of its proper analysis to erase a term Y of the proper analysis only if X and Y are identical. In the case of lexical items, "identity" might be taken to mean strict identity of feature composition. In some cases this decision has just the right conse q uences. Consider, for example, the case of the relativization transformation discussed previously (p. 1 45). Just Üð the generalized Phrase marker for the string "I saw the [1 the man was clever 1] boy" is not the deep structure underlying any well-formed surface struc ture and hence does not provide the semantic interpretation for any sentence (cf. pp. 1 37-1 38), so the generalized Phrase-marker for "I saw the [1 the boys were clever 1] boy" does not underlie a sentence. This is so because te element boys (containing the feature [+Plural]) is not identical with the element boy (contain ing the feature [-Plural]) j ust as the element man is not identical with boy. Hence, in neither case is relativization peritted. But matters do not always work out quite this smoothly. SOME RIDUAL PROBLEMS Consider the rules that provide for comparative constructions of various sorts, in particular, for such sentences as (34) John is more clever than Bill In this case, the sentence is formed from the underlying deep structure given as (35), following previous conventions. The fea- I-S-I . NP I N I+e John Prdicate· Phras ----- Aux VP ' . Preent Copula PredicaU . Compar � ~ NP Pricate· Phras I � N Aux VP l+ A e I Prent ---.. Copula Prdicate I Bill Adjective / ` cleer Adjective o c(ever tures constituting the lexical formatives of (35) are not given explicitly, but, rather, indicated by . . . . To derive (34) from (35) in the manner described earlier, the transformational rules frst apply to the most deeply embedded base Phrase-marker, namely that of "Bill is clever." Next, they reapply to the full confgura tion (35), which has, at this point (omitting various refnements), this terminal string: (36) John is more than [# Bill is clever #] clever The comparative transformation, which applies next, can b § 2. THE STUcRE OF THE LEXICON 1 79 formulated as an erasure operation that uses the Adjective of the matrix sentence to delete the corresponding Adjective of the embedded sentence . 34 Thus it applies to a string of this form: ( 3 7) 2 3 4 5 6 �æ æ > � = NP - is - . . . - . . . # NP is - Adjective # * Adjective (where · · · - . . . is a-a. more-than. etc. ). deleting 5 and #. Finally. it permutes 4 and 6 (technically. it places 4 to the right of 6, deleting 4) . This gives (38) John is more clever than Bill is A fnal option is to delete the repeated copula. giving ( 3 4). But recall that the deletion of the Adjective in the ffth posi tion of (37) by the comparative transforation is possible only when the two Adjectives are identical . Similarly. the deletion of the fnal copula in (38) requires identity of the two copulas. In the case of (34) . derived from (35). this causes no difculty. But consider the example (39). or the perfectly analogous French example (40) : ( 39) these men are more clever than Mary (40) ces hommes sont plus intelligents que Marie In the case of (39). deletion of the Adjective is straightforward. but our deletion conventions should prevent the deletion of the copula. since it has the feature [-Plural] in the embedded sentence and [+Plural] in the matrix sentence . Furthermore. in the case of (40) . the deletion of the Adjective of the embedded sentence should be blocked. since it difers from the Adjective of the matrix sentenc in gender and number . These observations sugst that it may not be correct to regard a formative simply as a set of features. some inherent and some added by transforation and as a consequence of insertion into a Phrase-marker . In particular. it seems from such examples as these that the features added to a formative by agreement trans formations are not part of the formative in the same sense as 1 80 SOME RIDUAL PROBLMS those which are inherent to it or a those which it assumes as it enters a Phrase-marker. Thus in the case of the relative trans formation, plurality of the Noun (which is a feature that the Noun assumes as it enters a Phrase-marker) is a feature that must be considered in determining whether it is identical to another Noun, a we have jut seen. However, in the case of Adjectives and the copula (also Verbs, which take part in similar rules) the infectional features that are added by agreement transforma tions are apparently not considered in determining whether the item in question is strictly identical wi th some other item.S Some further support for this conclusion is given by such examples as the following: (41) (i) John is a more clever man than Bill (ii) The Golden Notebook is a intricate a novel as Tristram Shandy (iii) Ïknow several more successful lawyers than Bill It is clear that the deep structures for these three sentences mut contain the base Phrase-markers underlying "Bill is a man," " Tritram Shandy is a novel," "Bill is a lawer," respectively. Thu (41 iii) implies that Bill is a lawyer; similarly, one cannot replace "Bill" by "Mary" in (4I i).s6 Sentences (4I i) and (4I ii) pose no problems. But consider (4I iii). However the transforma tional rules are actually forulated, it is clear that we are deleting "successful" and "a lawyer" as Predicates of "Bill" in the under lying structure. But the deletion of "a lawyer," in particular, is permitted only under the identity condition discussed earlier, and the string with which it is compared is not "a lawyer" but rather its pluralized for, "lawyers, "37 from the base string "I kow several [# S #] lawyers. " Here, then, is a case where plurality is not considered a distinguishing property of Nouns, for the purposes of a deletion operation, as contrasted with the case of relativization, discussed earlier, where a distinction in the feature of plurality was sufcient to block deletion. The cÏLcial diference apparently is that in this case, the Noun Phrase in question is in Predicate position and therefore receives i ts num ber not inherently (as in the example discussed in connection § 2. THE STUCTURE OF THE LEXICON 1 8 1 with relativization) but rather by an agreement transforation. Thus we cannot have "They are a lawyer," "Bill is several lawyers," etc., and such facts as these show that Prdicate Nominals must be neutral with regard to number. Hence, the confict in number between the i talicized Noun Phraes of "I know several lawyers" and "Bill is a lawyer" is on a par with the confict in number and gnder between the italicized Adjectives of "ces hommes sont intelligents" and "Marie est intelligente" (cf. (40». In both cases, the conficting features are introduced by agreement transformations. These examples sugest two conclusions. First, features intro duced by transformation into lexical formatives are not to be considered in determining when deletion is permitted; a forma tive, in other words, is to be regarded as a pair of sets of features, one member of the pair consisting of features that are inherent to the lexical entry or the position of lexical insertion, the second member of the pair consisting of features added by transformation. Only the frst set is considered in determining legitimacy of deletion in the manner previously described. Second, what is involved in deterining legitimacy of deletion is not identity but rather nondistinctness in the sense of distinc tive feature theory (cf. Chapter 2, § 2. 3. 2). Thus consider once again the case of "I know several lawyers" -"Bill is a lawyer." The Predicate-Nominal of the latter is not singular, in the base structure; rather, it is unspecifed with respect to number exactly a the nasal is unspecifed with respect to point of articulation in the lexical representations of the formatives king, fnd, lamp, etc. Hence, it is not identical with the corresponding nominal ele ment of "I know several lawyers"; it is, rather, nondistinct from it, and the example suggsts that this is sufcient to permit deletion. aa Notice that this analysis of a forative a a pair of sets of features need not actually be descibed or mentioned in any way in the rules of the grammar, since it is, apparently, determined by a general convention regarding the for of gammar. In other words, we are tentatively proposing it for consideration as a linguistic universal, admittedly, on rather slender evidence (but SOME RESIDUAL PROBLEMS see note 2 of Chapter 2) . If this proposal is a correct one, then the analysis of formatives that we have suggested is a general condition on the functioning of erasure transformations . The only apparent alternative to the proposal just advanced is a revision of the general conditions suggested earlier on the order of application of transformational rules. Whether this may be feasible, I do not know; but in any event, the proposal just dis cussed seems clearly preferable. Summarizing, we seem to be led to the conclusion that non distinctness rather than strict identity is what is involved in deletion, and that only those features of a formative that are inherent either to its lexical entry or to the position in the sentence where it is inserted are to be considered in determining nondistinctness. Formally, we can say that a formative must be regarded as a pair of set of features, one member consisting of the "inherent" features of the lexical entry or the sentence posi tion, the other member consisting of the "noninherent" features introduced by transformation. The gneral principle for erasure operations, then, is this: a term X of the proper analysis can be used to erase a term Y of the proper analysis just in case the inherent part of the formative X is not distinct from the inherent part of the formative Y. But notice that this is an entirely natural decision to reach. The original intuition motivating this condi tion was that deletions should, in some sense, be recoverable; and the noninherent features of the formative are precisely those that are determined by the context, hence that are recoverable even if deleted. Similarly, it is natural to base the opration on nondistinctness rather than identity, because the features un specifed in underlying structures (such as number, in predicate position) also make no independent contribution to sentence interpretation, being added by what are, in essence, redundancy rules, and are, in fact, simply a refection of context. Thus they are recoverable in the sense that the context that determined them is still present in the string after deletion of the item in question. Hence, the italicized condition formalizes a very reasonable sense of "recoverability of deletion. " Consider now one last set of questions relating to the compara- § 2. THE STUCUR OF THE LEXICON tive transformations. Suppose that we adopt the method of lexical insertion proposed in § 3 of Chapter 2 and based on non distinctness, rater than that proposed in § 4 . 3 of Chapter 2. In the Phrase-marker (35), then, each occurrence of the Adjective clever will have such features as [post-Animate] (that is, [+[+Animate] -]) added to it by selectional rules of the base component (in this case, (57xv) of Chapter 2, now revised as ( 1 3) of this chapter). But we clearly must allow such sentences as "John is heavier than this rock"; and in this case, heav will have the feature [post-Animate] in the matrix sentence and the fea ture [post-Inanimate] in the embedded sentence of the Phrase marker corresponding to (35) (this Phrase-marker will be iden tical with (35) except that each occurrence of clever in (35) will be replaced by heav; and Bill, with the features [+Ani mate], " ' , is replaced by the rock, with the features [-Animate], " ' , associated with rock) . Hence, the two occurrences of heavy that are compared when we attempt to apply the comparative transformation difer in feature composition, one containing the feature [post-Animate] and the other the feature [post-Inani mate] . As matters now stand, this diference of feature composi tion does not make the two items distinct from one another, in the technical sense of feature theory (that is, it is not the case that one of them is marked [+F and the other [-F, for some fea ture [F. Furthermore, it would be natural to regard these con textual features of the Adjective as noninherent, in the sense of the preceding paragraph; therefore deletion is permi tted. There is, however, one class of examples that sugsts that in certain cases Ü diference in the composition of two foratives with respect to such features a [post-Animate] should sufce to block deletion. Consider such sentences as (42 ) (i) John is as sad as the book he read yesterday (ii) he exploits his employees more than the opportunity to please (iii) is Brazil as independent as the continuum hypothesis? Clearly, these are deviant and must be marked as such in a descriptively adequate gammar. In each case. the deleted items SOME RIDUAL PROBLES difer in selectional features from the items with which they are compared. Thus sad is [ost-Animate] in the matrix sentence of (42i) and [post-Inanimate] in the embedded sentence, and pssibly this might b regarded as the factor that blocks the transformation and prevents deletion. The only alternative, in these cases, would be to assume that two homonymous lexical entries are involved, in each of the examples of (42) . 39 In intro ducing examples of this sort, however, we touch on problems of homonymity ad range of meaning that are cloaked in such obscurity, for the moment, that no conclusions at all can be drawn from them. § 2.3. Derivational processes Derivational processes create much more of a problem for any sort of generative (that is, explicit) grammar than do infec tional sytems. This results from the fact that they are typically sporadic and only quasi-productive. We shall consider several examples briefy, without, however, arriving at any very satis factory way of dealing with the problems that arise. Where derivational processes are productive, they in fact raise no serious difculties. Consider, for example, nominalization transformations of the sort that form the sentences "their de struction of the property . . . ," "their refusal to participate . . . ," etc. Clearly, the words destruction, refusal, etc., will not be en tered in the lexicon as such. Rather, destroy and refuse will be entered in the lexicon with a feature specifcation that deter mines the phonetic form they will assume (y later phono logical rules) when they appear in nominalized sentences. A nominalization transformation will apply at the appropriate stage of derivation to the generalized Phrase-marker containing the confguration "they destroy the property" dominated by S,40 forming ultimately the Phrase-marker ( 4 3 ) , where irrelevant details ae omitted,41 and where Fb , Fm, C l , " ' , Cn stand for specifed features . It is not at all clear that destruction or refusal should be regarded as Nouns in "their destruction of the prop erty . . . ," "their refusal to come . . . " (although refusal happens § 2. THE STUCURE OF THE LEXICON s NP �. Det I I+Defnite] I their N 1 nom VP � AI A \ T A destroy I+Defnite] G1 • • • G. I V the property Prcdkatc·Phrase I to be a Noun in "their refusal surprised me, " which derives in part from the string underlying "they refuse"). Alternatively, the nominalized Predicate-Phrase as a whole might be said to occupy the Noun posi tion. In any event, phonological rules will determine that nom"destro becomes destruction and that nom" refuðe becomes refusal, and so on.42 To have the proper efect, these rules must, of course, take account of inherent features associated with items in lexical entries, namely the features that deterine which form of nom these i tems take . In such cases as these, the proposed framework is quite adequate for formulating the syntactic generative rules as well as the rules of semantic and phonological interpretation. Notice, incidentally, that in the light of these remarks we must revise the description of the example (1) of Chapter 2 SOME RIDUAL PROBLMS (="sincerty may fghten John"), which served a a basis for discussion throughout that Chapter. In fact, sincerity would surely not be entered into the lexicon, though sincere would. Sincerity is formed by a transforation, and is a "defective Predicate" in j ust the same way as refusal is a defective Pred icate in "their refusal surprised me" or "the refusal surprised me." That is to say, there is a transformational rule that operates on "NP-is-Adjective" constructions such as "John is sincere (of manner)" and gives such nominalizations as "john's sincerity (of manner)," where "sincerity (of manner)," like "refusal (to come)," can be regarded as a Noun. The phrase sincerity ap pears as a full NP, in a manner which we shall not describe in detail here, when the underlying sentence "NP-is-sincere" has an Unspecifed Subject and the matrix sentence in which it is embedded has a non-Defnite Article. Details aside, it is clear that, contrary to what we assumed earlier, sinceri ty is not introduced in (1) of Chapter 2 by the lexical rule, so that actually even this ver simple sentence is the result of a transformational development fom a complex basis. But consider now the cse of quasi-productive processes, such as those that are involved in the formation of such words as horror, horrid, horif; terror, (·terrid), terrify; candor, candid, (·candif); or telegram, phonograph, gramophone, etc., or, for that matter, such words as frighten, in the example ( Ï ) of Chap ter 2). In these cases, there are no rules of any generality that produce the derived items, as there are in the case of sincerity, destruction, and so on. Hence, it seems that these items must be entered in the lexicon dirctly. This, however, is a very un fortunate conclusion, since it is clear that from the point of view of both the semantic and the phonological interpretation it is important to have internal structure represented in these words. Their meaning is clearly to some extent predictable (or at least limited) by the inherent semantic properties of the morphemes that they contain, and it is easy to show that interal structure mut b assigned to these items if the phonological rules are to apply properly in foring teir phonetic representations (cf. the discussion of the trasforational cycle for English in § �. THE STRUCUR OF THE LXICON Halle and Chomsky, 1 960; Chomsky. 1 962b; Chomsky and Miller, 1 963; and, for a detailed statement, Halle and Chomsky, forthcoming). This dilemma is typical of a wide class of examples with vary ing degrees of productivity, and it is not at all clear how it is to be resolved, or, in fact, whether there is any non-d hoc solution that can be achieved at all.4 s Perhaps one must regard the gaps as accidental, at least in some such cases, and incorporate in the grammar overly general rules that allow for nonoccuring as well as actual cases. Alteratively, it may be necessary to extend the theory of the lexicon to permit some "interal computation," in place of simple application of the gneral lexical rule in te manner already described. Thu telegraph, horri f, frighten, might be entered in the lexicon as ( 44 ) (i) (te l e '" Stem l' [ F l' . . . ]) (ii) (Stem { i f, [Gl , . " . ]) (iii) (Stems ..en, [HI> . . . ]) these items being entered into string by means of the general lexical rule. Furthermore, the lexicon would also contain the entries (45) (i) (graph, [+Steml' . . . ]) (ii) (horr, [+Stem 2 " ' ]) (iii) (fright, [+N, +Stem s • ]) these now being inserted in string formed by prior insertion in preterminal string of item selected fom (44). There may be several layers of such extension of base derivations within the lexicon, in the case of morphologically complex form. However, the rule that replaces categories such as Ste� by items of (45) must be formulated wit some care. There are contextual restrictions on these replacements that must be spcifed because these processes are only marinally productive. Thus Steml can be replaced by graph, scope, phone in the con text tele -. but not by scope or phone in the context phono -. The same is true in the other cases . More seriously, these exten sions of base derivations within the lexicon must in general de- 1 88 SOME RIDUAL PROBLEMS pend also on the feature composition of the item being analyzed. Thus Stem s can be rewtten fright in the context ¯en only when the features HI. Hz • of (44iii) indicate that it is a pure transitive. takes only Animate Object. etc. In other words. provi sion must be made for te fact that fr i ghten is not a Verb of the same type as redden or soften. and this can only be done by tak ing into account the feature composition of the only partially specifed lexical entries of (44 ) as well as the feature composition of the items of (45) that substitute for the categories appearing in the entries of (44). Precisely how such rules should be formulated is not at all clear to me. It may be possible to deterine the restrictions completely by feature specifcations in (44). (45). relying on a reapplication of the lexical rule to insert the items appropriately. Alteratively. it may be better to allow the lexi con to contain context-sensitive rewriting rules to efect these extensions of base derivations. The former alterative is of course preferable, since it does not afect the structure of the lexicon. The lexicon would, under this alternative. be simply a list of entries. and the lexical rule (now reapplicable) would be the only rule involving lexical entries. However. I do not know whether this approach will prove feasible when it is attempted in detail. In the examples just discussed, whichever method is chosen for extending base derivations, we shall have a complex symbol dominating a sequence of symbols. There is apparently no empirical motivation for allowing complex symbols to appear above the level of lexical categories. with the substantial enrich ment of linguistic theory and corresponding reduction in its interest and importance that this elaboration entails. Limitation of complex symbols to lexical categories implies that no complex symbol will dominate a branching confguration, within the categorial component. Now. however. we have some evidence that within Ü word, branching must be peritted in a confgura tion dominated by a complex symbo1.44 In the light of such examples. we may have to relax the re quirement (pp. 1 1 2-1 1 3) that branching is not tolerated within the dominance scop of a complex symbol. This restriction seem to § 2. TE STUCUR OF THE LEXICON hold only above the level of the word. Wi th this modifcation, we still retain the earlier restriction of complex symbols to lexical categories. Alternative analyses sugst themselves for several of these examples. In the case of such words as frighten, one might seek syntactic j ustifcation for a transformational analysis from an underlying causative construction so that "it frightens John" would derive from the structure underlying "it makes John afraid, " this in tur deriving from the abstract structure "it makes S" where S dominates "John is afraid. " Adjectives would then have to be divided in the lexicon into two classes depend ing on whether or not they undergo this transformation. Thus, afraid, red, soft would be in one category; whereas happy, green, tender would be in the other . Conceivably, we might go on to analyze such words as wizen, chasten as based on a similar analy sis, with the underlying Adjective designated lexically as one that must undergo this transformational process (in the case of chasten, the underlying form would have to be lexically dis tinguished from te homonymous Adjective that belong to the class of those that cannot undergo the transformational process in question) . Such an analysis could be extended to many other forms -for example, such Verbs as enrage, clarif. It might even be extended to account for such words as drop, grow, discussed in note 15 of Chapter 2, where it was observed that the Intransi tive occurrences cannot be derived from underlying Transitives. A general "causative" transformation might permit a derivation of "he dropped the ball, " "he grows corn," etc. , from an under lying structure of the form "he caused S," where S is the structure underlying "the ball drops, " "corn grows, " and so on. A number of syntactic arguments might be given in favor of a general "causative" operation to accommodate these and other cases. There is no doubt that items must be specifed lexically in terms of the operations that apply to them; this is particularly clear from a consideration of phonological rules, but is no less true of syntactic processes. Much of lexical structure is, in fact, simply a classifcation induced by the system of phonological and syntac tic rules. Postal has suggsted, furthermore, that there should be SOME REIDUAL PROBLEMS a general analysis of lexical items with respect to each rule R, into those which must, those whic may, and those which cannot b subject to R, and has investigated some of the consequences of this assumption. I mention these possibilities simply to indicate that there remain numerous relatively unexplored ways to deal with the problems that arise when the structure of a lexicon is considered seriously. Problems similar to those of derivational morphology are not lacking beyond the word level a well. Consider, for example, such phraes as "take for granted," which abound in English. From a semantic and distributional point of view, this phrase seems to b a single lexical item, and it therefore must be entered in the lexicon as such, with i ts unique set of syntactic and semantic features. On the other hand, its behavior with respect to transformations and morphological processes obviously shows that it is some sort of Verb-with-Complement construction. Once again, then, we have a lexical item with a rich interal structure. In such a phrase a "take ofense at" the problem is more acute. Again, distributional as well as semantic considera tions suggest that this is a lexical item, but certain transforma tions apply to this phrase as if "ofense" were a normal Noun Phrase (cf. "I didn' t think that any ofense would be taken at that remark"). Verb"Particle constructions also provide a variety of related problems. To some extent, the Particle is a fairly free "Adverbial" element, as in "I brought the book (in, out, up, down) . " Often, however, the Verb"Particle construction is (distribution ally as well as semantically) a unique lexical item (such as "look up," "bring of, " "look over"). In all cases, how ever, the syntactic structure is apparently the same, with respect to the possibility of applying familiar transforational rles. I see no way, for the present, to give a thoroughly satisfactory treatment of this general question.45 The Verb"Particle constructions "look up (the record)," "bring in (the book)," and so on, are of course not to be confused with the very diferent constructions discused in Chapter 2, § 2. 3+ There we noted that certain Verbs were in close construc- § l. THE STRUCURE OF THE LEXICON tion with certain Adverbials (for example, "decide on the boat, " in the sense of "choose the boat"), and that these Verb'"Ad verbial constructions are very diferent from others (such as "decide (while) on the boat") that involve a much looser associa tion of Verb and Adverbial. In these close constructions, the choice of Particle is often narrowly or even uniquely constrained by the choice of Verb (for example, "argue with X about Y") . We must therefore indicate in the lexical entry for such words as decide, argue, that they take certain particles and not others, as, in fact, is commonly done in dictionaries. This information can be presented in various ways. One possibility is to develop the Adverbial freely and to assign a contextual feature to the Verb (for example, to decide, the contextual feature [ -on'"NP], to argue, the feature [ ¯ with'"NP'" abourP]). If either of the methods of lexical insertion described in Chapter l, § 4. 3 is used, the Verbs in question will now be inserted only in the permitted positions, and the resulting Phrase-marker will now have the structure required for further rules. A second possibility is to develop the Adverbial freely but to give the lexical entry as a sequence of foratives, much as in the case of telescope, take for g ranted, and so on. Thus we would have the entries decide # on, argue(# about) ( # with), etc. Associated with these lexical entries, then, will b an erasure transformation that will use the freely gnerated Particles of the Prepositional-Phrases to delete the Particles of the lexical entries. Under this alterative, we rely on the fltering efect of transformations to guarantee correct insertion in well-fored deep structures, and, once again, we derive the correctly formed Phrase-markers when successful lexical insertion takes place. Still a third possibility would be to enter the lexical items in the manner j ust proposed, and to derive the Adverbials with a dummy element in the Preposition position, then distributing the Particles of the lexical entry by substitution transformations . Again, the same Phrase-markers result. There are still further possibilities. The same choices, incidentally, are also available in Verbrpar ticle constructions. In this case, however, the Phrase-marker that SOME REIDUAL PROBLEMS result from lexical insertion and associated operations must be diferent from that of the Verb"Adverbial constructions. since later rules apply very diferently in the two cases. I see little basis for choosing among these alteratives at the moment; until sharper criteria are discovered. these can only be regarded as essentially notational variant. Obviously. this discussion by no means exhausts the com plexity or variety of topic that. so far. resist systematic and re vealing grammatical description. It is possible that we are approaching her the fringe of marginal cases. to be expected in a sytem as complex a a natural language. where signifcant systematization is j ust not possible. Still. it is much too early to draw this conclusion with any confdence and. even if it is eventually justifed. we still must face the problem of extracting whatever subregularities exist in this domain. In any event. the questions we have touched on here have not yet been illuminated in any serious way by approaching them within the framework of any explicit grammatical theory. For the present. one can barely go beyond mere taxonomic arrangement of data. Wether these limitations are intrinsic. or whether a deeper analysis can succeed in unraveling some of these difculties. remains an open question. Notes NOTES TO CHAPTER 1 I . To accept traditional mentalism, in this way, is not to accept Bloomfeld's dichotomy of "mentlism" versus "mechanism. " Men· talistic linguistics is simply theoretical linguistics that uses per· formance as data (along wit oter data, for example, the data provided by introspecton) for te determination of competence, the latter being taken as the primar object of its investigation. The mentalist, in this taditional sense, need make no assumptions about te possible physiological basis for te mental reality that he studies. In particular, he need not deny that there is such a basis. One would guess, rather, that it is te mentalistic studies that will ultimately be of gratest value for the investigation of neurophysiological mechanisms, since they alone are concerned with determining abstractly the properties that such mechanisms must exhibit and the functions tey must perfor. In fact, te issue of mentalism versus antimentalism in linguistic apparently has to do only with goals and interests, and not with questions of trut or falsity, sense or nonsense. At least three issues are involved in this rather idle controversy: (0) dualism -are te rules that underlie performance represented in a nonmaterial medium?; (b) behaviorism -do the data of performance exhaust the domain of interest to the linguist, or is he also concerned with oter facts, in particular those pertaining to the deeper systems that underlie behavior?; (c) introspectionism -should one make use of introspective data in the attempt to ascertain the properties of these underlying systems? It is the dualistic position against which Bloomfeld irelevantly inveighed. The behaviorist position is not an arguable matter. It is simply an expression of lack of interest in theory and explanation. This is clear, for example, in TwaddeIJ 's critique (1 935) of Sapir's mentalistic phonolog, which used informant responses and comments as evidence bearing on the psychological reality of some abstract system of phonological elements. For Twaddell, the enterprise has no point because all 1 94 NOTS TO CHAPTER 1 that interest him is the behavior itself, "which is already available for the student of language, though in less concentrated form." Characteristically, this lack of interest in lingistic theory expresses itself in te proposal to limit te ter "theory" to "summary of data" (as in Twaddell's paper, or, to take a more recent example, in Dixon, 1 963, altough the discussion of "teories" in te latter is sufciently vague as to allow other interpretations of what he may have in mind). Perhaps tis loss of interest in theory, in the usual sense, was fostered by certain ideas (e.g., strict operationalism or strong verifcationism) that were considered briefy i n positivist philosphy of science, but rejected forthwith, in the early nine teen·thirtes. In any event, question (b) poses no substantive issue. Question (c) arises only if one rejects the behaviorist limita tions of (b). To maintain, on grounds of methodological purity, that introspective judgments of the inforant (often, the linguist himself) should be disregarded is, for the present, to condemn the study of language t utter sterility. It is difcult to imagine what possible reason might be given for this. We return to tis matter later. For further disssion, see Katz ( 1 964C). 2. This has been denied recently by several European linguists (e.g., Dixon, 1 963: Uhlenbeck, 1 963, 1 964). They ofer no reasons for their skepticism concerning traditional grammar, however. What ever evidence is available today seems to me to show that by and large the traditional views are basically correct, so far as they g, and that the suggested innovations are totally unjustifable. For example, consider Uhlenbeck's proposal that the constituent anal ysis of "the man saw the boy" is [the man saw] [the boy], a pro posal which presumably also implies tat in the sentences [the man put] [it into the box], [the man aimed [it at John], [the man persuaded] [Bill that it wa unlikely], etc., te consti tuents are as indicated. There are many considerations relevant to the deter mination of constituent structure (cf. note 7) : to my knowledge, they support the traditional analysis wi thout exception against this proposal, for which the only argument ofered is that it is the result of a "pure linguistic analysis." Cf. Uhlenbeck ( 1 964), and the discusion there. As t Dixon's objections to traditional gam mars, since he ofers neither any alternative nor any argument (beyond the corect but irelevant observation that they have been "long condemned by professional linguists"), there is nothing fur ther to discuss, in tis case. 3. Furthermore, it seems to me that speech perception is also best studied in this framework. See, for example, Halle and Stevens ( 1 962). NOTES TO CHR 1 1 95 4. Tests that seem to deterine a useful notion of this sort have been described in various places -for example, Miller and Isard ( 1 963). 5. These characterizations are equally vague, and the concepts in volved are equally obscure. The notion "likely to be produced" or "probable" is sometimes thought to be more "objective" and antecedently better defned than the oters, on the assumption that tere is some clear meaning to the notion "probability of a sentence" or "probability of a sentence type." Actually, the latter notions are objective and antecedently clear only if probability is based on an estimate of relative frequency and if sentence type means someting like "sequence of word or morpheme classes." (Furthermore, if the notion is to be at all signifcant, these classes must be extremely small and of mutually substitutable elements, or e lse unacceptable and ungrammatical sentences will be as "likely" and acceptable as grammatical ones.) But in this case, though "probability of a sentence (type)" is clear and well defned, it is an utterly useles notion, since almost all highly acceptable sentences (in the intuitive sense) will have probabilities empirically indistinguishable from zero and will belong to sentence types with probabilities empirically i ndistingishable from zero. Thus the acceptable or grammatical sentences (or sentence types) are no more likely, in any objective sense of this word, than the others. This remains true if we consider, not "likelihood," but "likelihood relative to d given situation," U long as "situations" are specifed in terms of observable physical properties and are not mentalistic constructs. It is noteworthy that linguists who talk of hardheaded objective study of use of sentences in real situations, when they actually come to citing examples, invariably describe the "situa tions" in completely mentalistic tC. Cf., e.g., Dixon ( 1 963, p. 1 01 ), where, in the only illustrtive example in te book, H sentence is described as gaining it meaning from the situation "British Culture." To describe British culture as "a situation" is, in the frst place, a category mistake; furthermore, to regard it as a pattern abstracted from observed behavior, and hence objec tively describable in purely physical ters, betrays a complete mis understanding of what might be expected from anthropological research. For further discussion, see Katz and Fodor ( 1 964). 6. That it may be true is sugested by several (for the moment, quite untested) observations. For example, in Chomsky and Miller ( 1 963, p. 286) the following example is cited: "anyone who feels that if so many more students whom we haven' t actually admitted are sitting in on the course than ones we have that the room had NOTES T CHAPR Ï to be changed, then probably auditor will have to be excluded, is likely to agree that the curiculum needs revision." This contains six nested dependencies (along with other dependencies that go beyond nesting) with no self-embedding. Though hardly a model of felicitous style, it seems fairly comprehensible, and not ex tremely low on the scale of acceptability. In comparison, self-em bedding of degree two or three seems to disturb acceptability much more severely. The matter is worth studying, since a positive result concerning (4ii) would, as noted, support a conclusion about organization of memory which is not entirely obvious. 7. It has sometimes been claimed tat te taditional coordinated structures are necessarily right-recursive (Yngve, 1 960) or left recursive (Harman, 1 963, p. 61 3, rule 3i). These conclusions seem to me equally unacceptable. Thus to assume (with Harman) that te phrase "a tall, young, handsome, intelligent man" has the structure [[[[tall young] handsome) intelligent) man] seems to me no more justifable than to assume that it has the structure [tall [oung [handsome [intelligent man)]]). In fact, there is no gram matical motivation for any internal structure, and, a I have just noted, the assumption that there is no structure is also supported on gounds of acceptability, with extremely weak and plausible assumptions about organization of memory. Notice that there are cases where further structure might be justifed (e.g., [intelligent [oung man]] or, perhaps [YOUNG [intelligent man]], with con trastive stress on "young"), but the issue is rather whether it is always necessary. The same is true if we consider the very diferent type of Adjec tive-Noun construction that we fnd in such phrases as "all the young, old, and middle-aged voter" (for an interesting discussion of these various kinds of modifcation relations, see Ornan, 1 964) . Here, too, neither the structure [[oung, old] and middle-aged) nor [young [old and middle-aged)) has any justifcation. Similarly, it is surely impossible to assume, with Yngve, that in te phrase "John, Mary, and their two children" the structure is [John) [[Mar) [and their two children)], M that "John" is cordi nated with "Mary and teir two children, " te latter being ana lyzed into the coordinated items "Mary" and "their two children. " This is entirely counter to the sense. Notice, again, that conjunc tion can have this structure (e.g., "John, as well as Mary and her child"), but surely it is false to claim that it must have this structure. In these cases all known syntactic. semantic. phonetic. and per ceptual considerations converge in support of the traditional view that tese constructions are typically coordinating (multiple- NOTES TO CHAPTER I 1 9 7 branching). Notice also that this is the weakest assumption. The burden of proof rests on one who claims additional structure be yond this. There are various ways of justifying assignment of constituent structure. For example. in suc a phrase as "all (none) of the blue, green, red, and (or) yellow pennants, " if one wanted to argue that "blue, green, red" is a constituent (i. e., that the strc ture is left-branching), or that "green, red. and (or) yellow" is a constituent (that the structure is right-branching), then he would have to show that these analyses are required for some grammatical rule, that the postulated intermediate phrases must receive a semantic interpretation, that they defne a phonetic contour, that there are perceptual grounds for the analysis, or something of this sort. All of these claims are patently false in this case, and the other cases mentioned here. Thus no semantic interpretation can be assigned to "old and middle-aged" in "young, old, and middle aged voters" or to "green, red, or yellow" in "none of the blue, green, red, or yellow pennants" or to "Mary and their two chil dren" in "John, Mary. and their two children"; the phonetic rules explicitly preclude such constituent analysis; there are no gam matical rules that require these analyses; there are no perceptual or other arguments to support them. It seems difcult, then, to see any grounds for objecting to the traditional analysis and insisting on additional interediate categorization, in such cases as these. 8. Yngve ( 1 960, and several other papers) has proposed a diferent theory to account for certain obserations such as tose of (4) . Beyond te obvious condition of fniteness of memory, his teory assumes also that order of generation is identical wit order of production -tat te speaker and hearer produce sentences "from top·to·bottom" (they frst decide on te major structures, then the substructures of these, etc., leaving to the very end of the process the choice of lexical items) . Under tis highly restrictive additional assumption, the optimal perceptual device mentioned earlier is no longer constructible, and left-branching and multiple-branching, as well as nesting and self·embedding, contribute to "depth" in Yngve's sense. hence to unacceptability. To support this hypothesis, it would be necessary to show (a) that it has some initial plausi bility, and (b) that left-branching and multiple-branching in fact contribute to un acceptability exactly as do nesting and self·em bedding. As to (a), I see no plausibility at all to te assumption that the speaker must uniformly select sentence type, then deter· mine subcategories, etc µ• fnally, at the last stage, deciding what he is going to talk about; or that te hearer should invariably make all higher-level decisions before doing any lower-level analysis. As NOTES TO CHAPTR Ï to (b), the hypothesis is supported by no evidence at all. The examples given by Ynge all involve nesting and self-embedding and hence are irrelevant to te hypotesis, since the unaccept ability in this case follows from te assumption of fniteness alone without the additional assumption of "top-to-bottom" production for speaker and hearer. Furtherore, the hypothesis is contadicted by the observation (4iii) that multiply coordinated structures (cf. note 7) are the most acceptable (rather than the least acceptable, as predicted) and that left-branching structures are far more acceptable than nested structures of equal "depth," in Yngve's sense. It also fails to explain why examples of type (4iv), such as (2i), though very low in "depth," are still unacceptable. However, Yngve makes one important point in these papers, namely, that some transforations can be used to decrease nesting, hence to reduce the perceptual load. This suggests an interesting argument as to why grdars should contain transformational rules. Some additional weight to this argument is given by the dis cussion of perforance models involving transforational gram mars in Miller and Chomsky ( 1 963, Part 2). 9. I t is astonishing to fnd that even this truism has recently been challenged. See Dixon (1 963). However, it seems that when Dixon denies that a languag has infnitely many sentences, he is using the term "infnite" in some special and rather obscure sense. Thus on the same page (p. 83) on which he objects to the assertion "that there are an infnite number of sentences in a language" he states that "we are clearly unable to say that there is any defnite num ber, N, such that no sentence contains more than N clauses" (that is, he states that the language is infnite). Either this is a blatant self-ontradiction, or else he has some new sense of te word "in fnite" in mind. For further discussion of his remarks in this con nection, see Chomsky (in press). l Ü- Aside from terminology, I follow here the exposition in Katz and Postal ( 1 964) . In particular, I shall asume throughout that the semantic component is essentially as they describe it and that te phonological component is essentidlly B desLribed in Chomsky, Halle, and Lukof ( 1 956); Halle ( 1 9594, 1 959b, 1 9624); Chomsky ( 1 962 b); Chomsky and Miller ( 1 963); Halle and Chomsky ( 1 960; forthcoming). 1 1 . I assume troughout that te syntactic component contains a lexicon, and that each lexical item is specifed in the lexicon in ters of its intinsic semantic features, whatever these may be. I shall return to this matter in te next chapter. 1 2. In place of the terms "deep structure" and "surface structure," one might use the coresponding Humboldtian notions "inner for" NOTS TO CHAPR Ï 1 99 of a sentence and "outer form" of a sentence. However, though i t seems to me that "deep structure" and "surface structure, " in the sense in which these ters will be used here, do correspond quite closely to Humboldtian "inner form" and "outer form, " respec tively (as used of a sentence), I have adopted the more neutral terminology to avoid te question, here, of textual interpretation. The tC "depth gammar" and "surface grammar" are famil iar in modern philosophy in something roughly like the sense here intended (cf. Wittgenstein's distinction of "Tiefengrammatik" and "Oberfachengrammatik," 1 953, p. 1 68) ; Hockett uses similar ter minology in his discussion of the inadequacy of taxonomic li nguis tics (Hockett, 1 958, Chapter 29) . Postal has used the ters "under lying structure" and "superfcial structure" (Postal, 1 964b) for the same notions. The distinction between deep and surface structure, in the sense in which tese ters are used here, is drawn quite clearly in te Port-Royal Grammar (Lancelot et al., 1 660). See Chomsky (1 964, pp. 1 5-1 6; forthcoming) for some discussion and references. In phil osophical discussion, it is often introduced in an attempt to show how certain philosophical positions arise from false grammatical analogies, the surface structure of certain expressions being mis takenly considered to be semantically interpretable by means ap propriate only to oter, superfcially similar sentences. Thus Thomas Reid (1 785) holds a common source of philosophical error to lie in the fact that in all languages, there are phrases whic have a distinct mean ing; while at the same time, there may be something in the structure of them tat disagees wit te analog of gammar or with the princi ples of philosophy . . . . Thus, we speak of feeling pain as if pain was something distinct from te feeling of it. We speak of pain coming and going, and removing from one place to another. Such phrases are meant by those who use them in a sense that is nei ther obscure nor false. But the philosopher puts them into his alembic. reduces them to their fst principles, draws out of them a sense that was never meant, and so imagines that he has discovered an error of the vulgar [pp. 1 67-1 68] . More generally, he criticizes te teor of ideas as based on a deviation from te "popular meaning, " in which "to have an idea of anything signifes nothing more than to think of it" (p. 1 05) . But philosophers take an idea to be "the object that the mind contemplates" (p. 1 05); t have an idea, then, is to possess in the mind such an image, picture, or representation as the immediate object of thought. It follows that there are two objects of thought: 200 NOTS TO CHAPTR 1 the idea, which is in the mind, and te ting represented by it. From this conclusion follow te absurdities, as Reid regards them, of the traditional theory of ideas. One of the sources of these absurdities is the failure of the philosopher to attend "to the dis tinction between the operations of the mind and the objects of these operations . . . althoug this distinction be familiar to the vulgar, and found in the structure of all languages . . . " (p. 1 1 0). Notice that these two senses of "having an idea" are distinguished by Descartes in the Preface to the Meditations ( 1 641 , p. 1 38). Reid's linguistic obseration is made considerably earlier by Du Marsais, in a work published posthumously in 1 769, in the follow· ing passage (pp. 1 79- 1 80) : Ainsi, comme nous avons dit j' ai un livre, f ai un diamant, j' ai une montre, nous disns par imitation, fai la jevre, j'ai envie, fai peur, fai un doute, fai pitie, fai une idee, etc. Mais livre, dia mant, montre sont autant de noms d'objects reels qui existent independamment de notre maniere de penser; au lieu que santt, fevre, peur, doute, envie, ne sont que des teres meta physiques qui ne designent que des manieres d'etres consideres par des points de vue particuliers de l'esprit. Dans cet exemple, jai une montre, jai est une expression qui doit etre prise dans le sens prpre: mais dans fai une idee, j'ai n'est dit que par une imitation. C'est une expression em pruntee. rai une idee, c'est-a-dire, je pense, je con�ois de telle ou telle maniere. rai envie, c'est-a·dire, je desire; fai la volontt, c'est-a·dire, je veux, etc. Ainsi, idee, concept, imagination, ne marquent point d'objets reels, et encore moins des etres sensibles que l' on puisse unir run avec l'autre. In more recent years, it has been widely held that the aims of philosophy should. in fact, be stictly limited to "te detection of the sources in linguistic idioms of recurent misconstructions and absurd theories" (Ryle, 1 931 ) . 1 3. These descriptions are not fully accurate. In fact, te sentential complement in ( 1 0) should, more properly, be regaded as em bedded in a Prepositional-Phrase (cf. Chapter 3); and, as Peter Rosenbaum has pointed out, the sentential complement of (l l ) should be regarded as embedded i n the Noun-Phrase Object of "expect." Furthermore, the treatment of the Verbal Auxiliaries in ( 1 0) and ( l l) is incorrect, and tere are other modifcations relat ing to the marking of the passive transformation, t which we shall return in the next chapter. 1 4. It seems clear that many children acquire frst or second languages NOTS TO CHAPTER 1 201 quite successfully even though no special care i s taken to teach them and no special attention is given to their progress. It also seems apparent that muc of te actual speech observed consist of fragment and deviant expressions of a variety of sorts. Thus i t seems that a child must have the ability to "invent" a generative grammar tat defnes well-formedness and assigns interpretations t sentences even though te primary linguistic data that he uses as a basis for this act of theory construction may. from the point of view of the theor he constructs. be defcient in various respects. In general. tere is an important element of truth in the traditional view that "the pains which everyone fnds in conversa tion e e • is not to comprehend what another thinketh. but to extricate his thought from the signs or words which often agree not with it" (Cordemoy, 1 667). and te problem this poses for speech perception is magnifed many times for the language learner. 1 5. For example. Russell ( 1 940. p. 33: "from B logical point of view. a proper name may be assigned to any continuous portion of space¯time"). if we interpret his notion of "logically pop name" U embodying an empirical hypothesis. Interpreted in this way. Russell is stating what is, no doubt. a psycological trut. Inter preted otherwise. he is giving an unmotivated defnition of "proper name." There is no logical necessity for names or other "object words" to meet any condition of spatiotemporal contiguity or to have other Gestalt qualities. and i t is a nontrivial fact that they apparently do. insofar as the designated objects are of te type that can actually be perceived (for example. it is not true of "United States" -similarly. it need not be true of somewhat more abstract and functionally defned notions suc U "barrier"). Thus there are no logical grounds for the apparent nonexistence in na tural languages of words such as "LIMB. " similar to "limb" except tat it designates te single object consisting of U dogâ four legs. so that "its LIMB is brown" (like "its head is brown") would mean that the object consisting of the four leg is brown. Similarly. there is no a priori reason why a natural language could not con tain d word "HERD. " like the collective "herd" except tat it denotes a si ngle scattered object wi th cows as parts. so that "a cow lost a leg" implies "the HERD lost a leg. " etc. 1 6. Thus for Aristotle (De Anima. 403b). the "essence of a house is assigned in such a formula as 'a shelter against destrction by wind, rain. and heat: " tough "the physicist would describe it as 'stones. bricks. and timbers. ' ' ' For interesting comments on such defnitions. see Foot ( 1 961 ) . Katz ( 1 964d). 17. By a "reasonable procedure" I mean one that does not involve 202 NOTES TO CHAPTR 1 extralinguistic inforation -that is. one tat does not incorporate an "encyclopedia." Se Bar-Hillel ( 1 960) for discussion. The possi bility of a reasonable procedure for translation between arbitrary languages depends on the sufciency of substantive universals. In fact. although there i s much reason to believe that languages are to a signifcant extent cast in the same mold. there is little reason to suppose that reasonable procedures of translation are in general possible. 1 8. Actually. a set of structural descriptions should be assigned by f to each s, (and each structural description must be assigned to exactly one sJ. given Gj• one for each way of interpreting the sen tence ð_ with respect to GJ• Thus an unambiguous sentence should receive one structural description. a doubly ambiguous sentence two structural descriptions. etc. We assume tat mappings are efctive -that there is an algoritm for enumerating sentences. structural descriptions. and grammars and (throughout tis is less obvious) for deterining the values of f and m in all cases. 19. Obviously. to construct an actual theory of language learning. i t would be necessary to face several other very serious questions in volving. for example. the gradual development of an appropriate hypothesis. simplifcation of the technique for fnding a compatible hypothesis. and te continual accretion of linguistic skill and knowledge and te deepening of the analysis of language structure that may continue long after the basic form of the language has been mastered. Wat I d describing is an idealization in which only the moment of acquisition of the correct grammar is con sidered. Introduction of these additional considerations might afect the general discussion in many ways. For example. in some limited but nevertheless real way. the preconditions (i)-(v) them selves might possibly be developed on the basis of deeper in nate structure. in ways that depend in part on primary linguistic data and the order and manner in which they are presented. Fur thermore. it might very well be true tat a series of succesively more detailed and highly structured schemata (corresponding to maturational stages. but perhaps in part themselves deterined in form by earlier steps of language acquisition) are applied to the data at successive stages of language acquisition. There H¿ a priori, many possibilities that can be considered here. 20. It is instructive to see how modern structural linguistics has at tempted to meet tese conditions. It assumes that the tecnique for discovering the correct hypothesis (gammar) must be based on procedures of succesive segmentation and classifcation of the items in the corpus (which constitutes the primary linguistic data, when supplemented, perhaps. by certain knds of semantic in- NOTES TO CHAPTR Ï foration the exact relevance of which to the problem at hand has never been clarifed). To compensate for this extremely strong demand on the procedure of grammar discovery, it was necessary to sacrifce descriptive adequacy, over a wide rang of cases. In fact, the methodological discussions of modern linguistics pay very little attention to considerations (ii)-(iv) (though they do imply certain conclusions about them) and concentrate almost solely on development of constructive, step-by·step procedures of classifca tion and segmentation. For discussion, see Lees ( 1 957), Chomsky (1 964). 21 . This point has some historical interest. In fact, as has generally been noted by commentators, Locke's attempt to refute the doc trine of innate ideas is largely vitiated by his failure to obsere the distinction we have just been discussing, although tis was clear to Descartes (and was later re-emphasized by Leibniz, in his critique of Locke's Essay). Cf. § 8. 22. See note 19. An actual acquisition model must have a strategy for fnding hypotheses. Suppose, for example, that the strategy is to consider only grammars tat have better than a certain value (in terms of the evaluation measure (v», at each stage in the process of language learning. What is required of a signifcant linguistic theory, ten, is tat given primary linguistic data D, the class of grammars compatible with D be sufciently scattered, in terms of value, so that the intersection of te class of gammars compatible with D and the clas of gramars which are highly valued be reasonably small. Only then can language learning actually take place. 23. See references of note 1 0. 24. Failure of attempts to justify an explanatory theory may be in terpreted i n various ways, of coure. It may indicate that the theory is wong, or that its consequences were incorectly de termined -in particular, that the gammar tested for descriptive adequacy was not te most highly valued one. Since any reason able evaluation measure must be a systematic measure, and since language is a tightly interconnected system, the latter possibility is not to be discounted. In short, justifcation of linguistic teory does not avoid the problems faced by justifcation of any sub stantive and non trivial empirical hypothesis. 25. Actually, it is not clear that Quine's position should be taken as in any real sense an empiricist one. Thus he goes on to propose that in te innate quality space a red ball might be less distant from a green ball than from a red kerchief, so that we have not just a pre-xperiential characterization of distance but also an innate analysis of this into distance in various respects. On the 204 NOTES TO CHAPTR Ï basis of tese few comment, one might interpret him as propos ing tat suc concepts as "ball" are innate ideas, hence as adopting an extreme for of nativism; at least, it is difcult to see wherein the cited proposal difers from this. In further support of such an anti empiricist interpretation, one may point to Quine's virtual renunciation of reinforcement theory (cf. my note 26). Unfortunately, what are intended U empiricist views have gen erally been formulated in such an indefnite way that it is next to impossible to interpret tem with any certainty, or to analyze or evaluate them. An extreme example, perhaps, is Skinner's ac count of how language is learned and used (Skinner, 1 957). There seem to be only two coherent interpretations that one can give to this account. If we interpret the terms "stimulus," "reinforce ment," "condi tioning," etc., which appear in it, as having te meanings given to them in experimental psycholog, then tis account is so grossly and obviously counter to fact that discussion is quite beside te point. Alternatively, we may interpret these terms U metaphoric extensions of the (essentially homonymous) terms used in experimental psycholog, in which case what is proposed is a mentalist account difering from traditional ones only in that many distinctions are necessarily obscured because of the poverty of te terinologicl appartus available for para phrase of the traditional mentalistic notions. What is particularly puzzling, then, is te insistent claim that tis paraphrase is some how "scientifc" in a way in which traditional mentalism is not. 26. This applicaton is perhaps mediated by "reinforcement, " though many contemporary behaviorists use this term in such a loose way that reference to reinforcement adds nothing to the account of acquisition of knowledge that they propose. For example, Quine suggst ( 1 960, pp. 82-88) tat "some basic predilection for con formity" may take the place of "ulterior values," and that society's reinforcement of the response may consist "in no more than cor roborative usage, whose resemblance to te child's efort is the sole reward." As Quine correctly notes, "this again is congenial enough t Skinner's seme, for he does not enumerate the re wards" (this being one of the contributory factors to the near vacuity of Skinner's scheme) . What this proposal comes to is that te only function of "reinforcement" may be to provide the child with information about corect usage; tus the empirical claim of "reinforcement theory" will be that learning of language can not proceed in the absence of data. Actually, Skinner's concept of "reinforcement" is apparently still weaker than this, for he does not even require that the "reinforcing stimulus" impinge on te responding organism; it is sufcient that it be hoped for NOTES TO CHAPTR 1 or imagined (for a collection of examples bearing on this matter, see Chomsky, 1 959b). 2 7. These mechanisms, as is now known, need not be at all elementary. Cf., for example, Lettvin et al. ( 1 959), Hubel and Wiesel ( 1 962), Frishkopf and Goldstein ( 1 968). This work has demonstrated that peripheral processing in the receptor system or iq lower cortical centers may provide a complex analysis of stimuli that, furter more, seems to be rather specifc to the animal's life-space and well correlated with behavior patterns. Thus it seems tat not even peripheral processing can be described within the unstruc tured and atomistic framework that has been presupposed in empiricist thinking. 28. I depart here from the Langley translation, which renders this passage inaccurately. The French original is U follows: e e • je demeure d'accord que nous apprenons les idees et les veri tees innees, soit en prenant garde a leur source, soi t en les verifant par l'experience. Ainsi je ne saurois admettre cette proposition, tout ce qu'on apprend n'est pas inne. Les verites des nombres sont en nous, et on ne laisse pas de les apprendre, sit en les tirant de leur source lorsqu'on les apprend par raison demonstrative (ce qui fait voir qu'eIles sont innees) sit en les eprouvant dans les exemples comme font les arithmeticiens vulgaires . . . . 29. Cf. Chomsky ( 1 964) for additional discussion and quotations il lustrating Humboldt's views on these questions. 80. That this is a fair interpretation of taxonomic linguistics is not at all clear. For one thing, structural linguistics has rarely been concerned with the "creative" aspect of language use, which was a dominant teme in rationalistic linguistic theory. It has, in other words, given little attention to the production and interpretation of new, previously uneard sentences -tat is, to the normal use of language. Thus te suggestion tat te various teories of immediate constituent analysis might be interpreted ð gen erative, phrase structure grammars (as in Chomsky, 1 956, 1 962a, or Postal, 1 964a) certainly goes beyond what is explicitly stated by linguist who have developed these theories, and very likely beyond their intentions as well. Hence, te central problem of descriptive adequacy is not really raised within structural lin guistics. Secondly, many "neo-Bloomfeldian" linguists, accepting Bloomfeld's behaviorism under interpretation (b) of note J (as well as Firtians and "neo-Firthians" and many others), have tereby explicitly rejected any concern for descriptive adequacy, limiting the task of grammatical description, at least in theory, to organization of the primary linguistic data. Others have held that a grammar should at least describe te "habits" or "disposi- 206 NOTES T CHAPTR 1 tions" of the speaker, though the sense in which language use might be regarded as a matter of habit or disposi tion has never been satisfactorily clarifed. To be more precise, there is no clear sense of the term "habit" or "disposition" in accordance with which it would be correct to desibe languag as a "habit structure" or a "system of dispositions." In general, it is not clear tat most behaviorist tendencies should be regarded as varieties of empiricism at all, since. as distinct fom classical empiricism, they renounce any interest in mental proceses or faculties (that is. in te problems of descrip tive or explanatory adequacy). 1 1 . This is the only respect i n which a comparison of such alternatives is relevant, apart from their relative success in accounting for the given fact of language acquisition. But this consideration ap parently ofers no information tat has any bearing on te choice among alternative theories. In general. it is important to bear in mind that an extremely specialized input-output relation does not necessarily presuppose a complex and highly structured device. Wether our assumption about the mind is that it contains te schema for transforational gammar or that it contains mechanisms for making arbitrary associations or for carying out certain kinds of inductive or taxonomic operations, tere is apparently little knowledg about the brain and little engineering insight into plausible physical systems that can be used to support these hypotheses. Similarly. there is no justifcation for the common assumption that there is an asymmetry between rationalist and empiricist views in tat the forer somehow beg the question. not showing how the postulated internal structure arises. Empiricist views leave open precisely the same question. For te moment, there is no better account of how te empiricist data-processing operations might have been developed. as innate stucture, in a species. than there is of how the rationalist schema may arise through evolutionary processes or other deternants of te structure of organisms. Nor does comparison with species other than man help the empiricist argument. On the contrary. every known species has highly spe cialized cognitive capacities. It is important to observe that com parative psychology has not caractersticlly proceeded on empiricist asumptions about knowledge and behavior. and lends no support to tese assumptions. 32. There is reason to believe tat the language-acquisition system may be fully functional only during a "critical period" of mental development or, more specifcally. that its various maturational stages (see note 1 9) have critical periods. See Lenneberg (forth- NOS T CHAPTR 1 coming) for an important and informative review of data bearing on tis question. Many oter aspects of the problem of biologically given constraints on the nature of human language are discussed here and in Lenneberg ( 1 960). Notice tat we do not, of course, imply that the functions of language acquisition are carried out by entirely separate com ponents of te abstract mind or the physical brain, just as when one studies analyzing mechanisms in perception (cf. Sutherland, 1959, 1 964), it is not implied that these are distinct and separate components of the full perceptual system. In fact, it is an impor tant problem for psychology to deterine t what extent other aspects of cognition share properties of language acquisition and language use, and to attempt, in tis way, t develop a richer and more comprehensive theory of mind. 33 . It is a curious fact tat Cmpiricism is comonly regarded as some how a "scientifc" philosophy. Actually, te empiricist approach to acquisition of knowledge has a certain dogmatic and aprioristic character tat is largely lacking in its rationalist counterpart. In the particular case of language acquisition, te empiricist approach begins it investigation with the stipulation that certain arbi trarily selected data-processing mechanisms (e.g., principles of association, taxonomic procedures) are the only ones available to te language-acquisition device. It then investigates the applica tion of tese procedures to data, without, however, attempting to show that the result of tis application coresponds to grammars that can be shown, independently, to be descriptively adequate. A non dogmatic alternative to empiricism would begin by observ ing that in studying language acquisition, what we are given is certain information about the primary data that are presented and the grammar that is the resulting product, and the problem we face is that of determining the structure of the device that mediates this input-output relation (the same is true of the more general problem of which language acquisition is a special case). There are no grounds for any specifc assumptions, empiricist or oterwise, about te internal structure of this device. Continu ing with no preconceptions, we would naturally turn to te study of uniformities in te output (foral and substantive universals), which we then must attribute to the structure of the device (or, if this can be shown, to uniformities in the input, tis alternative rarely being a serious one in te cases that are of interest) . This, in efect, has been the rationalist approach, and it is difcult to see what alternative tere can be to it if dogmatic presuppositions U to the nature of mental processes are eliminated. M. That is, a teory that meets conditions (i)-(iv) of p. 3 1 . I shall lw8 NOTS T CHAPR 1 hencefort assume, with no further comment, that any linguistic theory under discussion attempts to meet at least these conditions. 8 5. Tere has, during the past few years, been a fair amount of in vestigation into the formal properties of very simple theories of grammar. For the most part, it has been restricted to weak gen erative capacity, though tere are a few results involving stong generative capacity as well (in particular, tose refered to in § 2). The latter is, obviously, by far the more interesting notion, but i t is much more difcult to study. For surveys of tis work, see Chomsky ( 1 963), Chomsky and Schitzenberger ( 1963)ø 36. See Postal ( 1 962b, 1 964a, 1 964C) . Neither te teory of context free grammar nor the theory of fnite-state grammar is an artifact invented for matematical investigation. Each is well motivated formally and has independent interest, apart from linguistics, and each has in fact been proposed by linguists as a comprehensive theory of language. In fact, as Postal shows ( 1 964a), almost every linguistic theory that has received any substantial attention in recent years, insofar as it is clear, falls witin te framework of context-free grammar. As we shall see later, a special form of te theory of context-free grammar apparently plays a crucial role within the general theory of transformational grammar. 37. This possibility cannot be ruled out a priori, but, in fact, it seems defnitely not to be the case. In particular, it seems that, when the theory of transformational grammar is properly formulated, any such grammar must meet formal conditions tat restrict it to the enumeration of recursive sets. Cf. the conditions on base rules; also note I¿ Chapter 3, and furter discussion in Chapter 3 and Chapter 4, § 2 .2, of conditions on deletion transformations. NOTES TO CHAPTER 2 I . In detail, there is some room for discussion about both termi nology and substance throughout (2), and, particularly in the case of (2ii), alternative conventions and decisions have been applied. However, I tink tat the central facts are clear enough, and there has, in fact, been overhelming accord about most of them. For present purposes, I shall raise no furter question (except of detail) about the adequacy of tese observations, taking them simply as facts to be accounted for by a grammatical teory. 2. A teory of language must state the principles interrelating its theoretical terms (e.g., "phoneme," "morpheme," "transformation," "Noun Phrase," "Subject") and ultimately must relate this ss tem of concept to potential empirical phenomena (to primary linguistic data) . For reasons discussed in Chomsky ( 1 957) and NOTS TO CHAPTER 2 elsewhere, it seems to me that all signifcant structural notions will have to be characterized in terms of the previously defned notion "generative gammar" (whereas structural linguistics has assumed, in general, that the notion "grammar" must be developed and explained in ters of previously defned notions such as "phoneme," and "morpheme"). That is, I am assuming that the basic notion to be defned is "G is a most highly valued grammar of the language of which primary linguistic data D constitutes a sample," where D is represented in terms of primitive notions of the theory; the phonemes, morphemes, transformations, etc., of the language are, then, the elements tat play a specifed role in the derivations and representations determined by G. If so, partial gnerative grammars will provide the only empirical data critical for evaluating a theory of the for of language. For the present, then, such evidence must be drawn from grammatical descriptions of relatively few languages. This is not particularly disturbing. Wat is important is that such assumptions be supported by available evidence and formulated with enough clarity so that new or improved generative grammars will have bearing on teir correctness, as the depth and range of linguistic study increases. We must, in short, accept Humboldt's conclusion, expressed in a letter of 1 822 to Schlegel (Leitzmann, 1 908, p. 84) : "dass jede grammatische Discussion nur dann wahrhaften wissenschaftlichen Gewinn bringt, wenn sie so durchgefihrt wird, als llge in ihr allein der ganze Zweck, und wenn man jede, noch M rohe Sprache selbst, gerade mit derselben Sorgfalt behandelt als Griechisch und Lateinisch. " Study of a wide range of languages is only one of the ways to evaluate the hypothesis that some formal condition is a linguistic universal. Paradoxical as this may seem at frst glance, considera tions internal to a single language may provide signifcant sup port for the conclusion that some formal property should be attributed not to the teory of the particular language in ques tion (its grammar) but rather to the general linguistic theory on which the particular grammar is based. Study of descriptive or explanatory adequacy· may lead to such a conclusion; further more, the difcult or impossibility of formulating certain condi tions within the framework of an otherwise well-supported theory of grammar provides some evidence that these are, in reality, general conditions on the applicability of grammatical rules rather than apects of the particular language, to be expressed within the system of grammatical rules itself. Several cases of this sort will be mentioned later. . In general, it should be expected that only descriptions con- 2 10 NOTS T CHAPR 2 cered with deep structure will have serious import for proposals concerning linguistic universals. Since descriptions of this sort are few, any such proposals are hazardous, but are clearly no less in teresting or importnt for being hazardous. 3 . A weak though sufcient condition is given in Chomsky ( 1 955, Chapter 6). A stronger but rather well-motivated condition is pro· posed by Postal ( 1 964a) . Some aspects of this question are dis cussed in Chomsky and Miller ( 1 963, § 4); Chomsy ( 1 963, § 3)· 4 . For some discussion, see the references cited on p. 1 6, and many others. These demonstations of the inadequacies of phrase struc ture gramma have not been challenged, although some confu sions have been introduced by terminological equivocations. The most extreme example of this can be found in Haran ( 1 963), where many of the standard arguments against phrase stucture gammar are repeated, with approval, in an article wit the subtitle "a defense of phrase structure." This curious situation result simply from the author's redefnition of te term "phrase structure" to refer to a system far richer than that t whic the ter "phrase stucture grammar" has been universally applied in the rather ample literature on this subject (in particular, to a system in which in place of category symbols, in te sense of phrase structure grammar, we have pairs ( a, q , where @ is a category sybol and q is a set of indices used to code transfor· mations, contextual restrictions, etc.). That is, Haran in efect re states the arguments against phrase structure grammar as argu ments against limiting the ter "phrase structure grammar" to the particular systems tat have previously been defned as "phrase stucture grHar." This terinological proposal does not touch on the substantive isue as to the adequacy of the taxonomic theory of grammar for which phrase structure grammar (in the usual sense) is a model. The essental adequacy of phrase structure gammar U a model for taxonomic grammatical theor (with the posible but irelevant exception of problems involving discon tinuous constituents -see Chomsy, 1 957, Postal, 1 964a) is demon strated quite convincingly by Postal, and is not challenged by HaÏ, or anyone else, to my knowledge. The only issue that Haran raises, in tis connection, is whether the term "phrase strcture grHar" should be restricted to taxonomic models or whether the ter should be used in sme far richer sense as well, and this terinological question is of no conceivable importance. The terminological equivocation has only the efect of suggesting t the casual reader, quite eroneously, that there is some issue about the linguistic adequacy of the theory of phrase structure grammar (in the usual sense). NOTES TO CHAPTER 2 2 1 1 A further source of possible confusion. in connection with this paper. is that there is a way of interpreting the grammar pre sented tere as a phrase structure grammar. namely by regarding each complex element ( a. ') as a single. unanalyzable category symbol. Under this interpretation. what we have here is a new proposal as to the proper evaluation procedure for a phrase structure gammar. a proposal which is immediately refuted by the fact that under this interpretation. te structural description provided by the Phrase-marker of the now highest-valued grammar is invariably incorect. For example. in John saw Bill, did Tom see you', the three elements John, Bill, Tom would belong to three distinct and entirely unrelated categories. and would have no categrial assignment in common. Thus we have the following alternatives: we may interpret the paper as proposing a new evaluation measure for phrase structure grammars. in which case it is immediately refuted on grounds of descriptive inadequacy; or we may interpret it as proposing that the term "phrase struc ture grammar" be used in some entirely new sense. in whic case it has no bearing on the issue of the adequacy of phrase structure gd ar. For some furter discussion see Chomsky (in press). where this and other criticisms of transforational grammar. some real. some only apparent. are taken up. 5. This assumption is made explicitly in Chomsky ( 1 955). in the discussion of te base of a transformational grammar (Chapter 7). and. to my knowledge, in all subsequent empirical studies of transformational grðar. An analogous assumption with respect to transformational rules is made in Matthews (1 964. Appendix A, § 2). Formal properties of se q uential grammars have been studied by Ginsburg and Rice (1962) and Samir ( 1 961 ), these being context-free grammars where te sequential property is, furtherore, intinsic (in the sense of note 6, Chapter 3). rather than extrinsic, as presupposed here (for the context-snsitive case, at least). 6. As noted earlier. there are rather diferent conventions, and some substantive disagreements about the usage of these terms. Thus if we were to change the rules of (5), and, corespondingly. the Phrase-marker (3). to provide a binary analysis of the major category S into sincerity (NP) and may frighten the boy (P). ten the latter would be the Predicate-of the sentence in the sense defned in (1 1). See the fnal paragraph of § 2. 3.4 for an emendation of these suggested defnitions of functional notions. 7. Let us assume. furtherore, that Y, Z are unique. in tis case in other words, that there is only one occurrence of B in X. The defnition can be generalized to accommodate the case where this 2 1 2 NOTS TO CHAPR 2 condition is violated, but it seems to me reasonable to impÜe this condition of uniqueness on the system of base rules. 8. Notice that accurate defnitions require a precise specifcation of the notions "occurrence," "dominate," etc. This raises no difculty of principle, and throughout the informal discussion here I shall simply avoid these questions. Precise defnitions for most of te notions tat will be used here, taking occurrences into account, may be found in Chomsky ( 1 955). 9. One might question whether M should be regarded as H lexical category, or whether, alternatively, the rules M - may, can, • • • should not be included in the set (51). The signifcance of this distinction will be discussed later. This is by no means merely a terminological issue. Thus, for example, we might hope to estab· lish general conventions involving the distinction between lexical and Monle7ical categories. To illustrate the range of possibi1ities that may be relevant, I mention just two considerations. The gen eral rule for conjunction seems to be roughly this: if XZY and XZ'Y are two strings such that for some category A, Z is an A and Z' is an A, then we may form the string X'"Z'"and'Z''Y, where Z"and"Z' is an A (see Chomsky 1 957, § 5. 2, and for a much more far-reaching study, Gleitman, 1 961 ). But, clearly, A must be H category of a special type; in fact, we come much closer t charac terizing the actual range of possibilities if we limit A to major categories. By tis criterion, M should be a lexical category. Second, consider the phonological rules that assign stress in English by a transformational cycle (see Chomsky, Halle, and Lukof, 1 956; Halle and Chomsky, 1 960, forthcoming; Chomsky and Miller, 1 963). These rules assign stress in a fxed way in strings belonging to certain categories. By and large, the categories in question seem to be te major categories, in the sense just de scribed. In particular, elements of nonlexical formative categories (e.g., Articles) are unstressed. By this criterion, one might want M to be a nonlexical category, though even here the situation is unclea; cf. the well-known contrast of may-may, as in John may tr (it is permitted) and John may tr (it is possible). 1 0. Some have argued tat the distinction in question has nothing to do wi th rules of English, but only wi t statistics of usage. Wat seem to be insuperable difculties for any such analysis have been raised and frequently reiterated, and I see no point in considering this possibility any further as long as proponents of this most implausible view make no attempt to deal with these objections. Cf. Chapter 1, § X- 1 1 . For some discussion of a possible syntactic basis for such sub categrization, wit a small amount of supporting evidence, see NOT TO CHAPTER � Chomsky ( 1 955, Chapter 4 ) , summarized in part in Chomsky ( 1 961 ) and Miller and Chomsky ( 1 963). A cri tique of these and other disussions is given in Katz ( 1 964a). I thi nk that Katz's major criticisms are correct, but that tey can perhaps be met by narowing te scope of te proposals to just what is being discussed here, namely the question of subcategorization of lexical categories witin the framework of an independently justifed generative gramar. Ì- In the syntactic component of this (pretransformational) grammar, indices on category symbols were used to express agreement (and, in general, what Harris, 1 951 , calls long components) but not subcategorization and selectional restrictions. These devices be come unnecessary once grammatical transformations are intro duced. See, in tis connection, the discussion in Postal ( 1 964a). 1 3. Matthew qevised H tecni q ue of indexin g category symbols to meet the difculties that he found, and he later incorporated this technique as one of te main devices of te COMIT programming system that he developed with the collaboration of V. Yngve. Similar difculties were noted independently by R. Stockwell, T. Anderson, and P. Schachter, and tey have suggested a somewhat diferent way of handling them (see Stockwell and Scachter, 1 962; Schachter, 1 962). E. Bach has also dealt with tis question, in a somewhat diferent way (Bdch, 1 964). The method that I shall elaborate later incorporates various features of these pro posals, but difers from tem in certain respects. The problem of remedying tis defect in phrase structure grammÜr is clearly very much open, and deserves muc further study. Although this defect was pointed out quite early, there was no attempt to deal with it in most of the published work of the last several years. 14. Thus [s] is an abbreviation for thC set of features [+ consonantal, - vocalic, - voiced, + continuant, + strident, - grave] and [m] for the set of features [+ consonantal, - vocalic, + nasal, + voiced, + grave]. Rule ( 1 8) applies to any segment specifed as [+ continu ant] (hence to [s]) in a context which is specifed as - [+ voiced] (hence to the context [ -m]), converting the segment to which it applies to a voiced segent with, otherwise, the same features as before (hence converting Cs] to [z] ~ [+ consnantal, - vocalic, + voiced, + continuant, + strident, - grave] ) . I shall henceforth use te convention, customary on the phono logical level, of enclosing sets of features by square brackets. 1 5. But notice tat a phonological matrix can be regarded simply as a set of specifed phonological features, if we index each specifed feature by an iMteger indicating the column it occupies in the matrix. Thus the two·column matrix representing te formative 2 14 NOT TO CHAR 2 bee can be regarded as consisting of the features [+ consonantall• - vocalicl' - continuantl• . • . • ¯ consonantal2' + vocali�. - gave2• . . . ] . A lexical entry can now be regarded simply as a set of features. some phonological. some syntactic. Of course. a lexical entry must also contain a defnition. in a complete ga mar. and it can be plausibly argued (see Katz and Fodor. 1 9 68) tat this too consist simply of a set of features . (Actually te Katz-Fodor defnitions are not simply sets. but it does not seem that the furter structure they impose plays any role in their theory.) We might. then. take a lexical entry to be simply a set of features. some syntactic. some phonological, some semantic . However. largely for ease of exposition. we shall not follow this course but shall. rather. regard a lexical entry as a matrix complex symbol pair, as in the text. If we regard a lexical entry as a set of features. then items that are similar in sound, meaning. or syntactic function will not be related to one another in te lexicon. For example, the In transitive "grow" of "the boy grew" or "corn grows," and the Transitive "gow" of "he gows corn" would have to constitute two separate lexical entries, despite the meaning relation that holds between them. since tere is apparently no way to derive the Intransitive stuctures from the Transitive ones, as can be done in the case of "the window broke." "someone broke the window." Cf. p. 1 89. The same would be true of "drop" in "te price dropped," "he dropped te ball, " "he dropped that silly pre tense"; or of "command" in the example discussed on p. 1 1 9, and in innumerable other cases of many diferent kinds. Alternatively. such relationships can be expressed by taking a lexical entry to be a Boolean function of features. Although it is likely tat such a modifcation of the theory of lexical structure is necessary. it raises many problems of fact and principle to which I have no answer, and I therefore continue the exposition without develop ing it. 1 6. Recall Bloomfeld's characterization of a lexicon as te list of basic irregularities of a language ( 1 9 83. p. 274). The same point is made by Sweet ( 1 91 3, p. 3 1 ), who hold that "gramar deals with the general fact of language. lexicology with the special facts." 17. More generally. te phonological redundanc rules. which de termine such features as voicing of vowels or unrounding of high front vowels in English. can be supplemented by analogous syn tactic and semantic redundancy rules. Furthermore. redundancy rules may relate features of these various types. For example. if the taditional view tat syntactic categorization is in pan de NOTS TO LÎPÝÂÎÝ Z terined semantically can be substantiated in any serious way, it can be expressed by a redundancy rule determining syntatic features in terms of semantic ones. We shall return to the question of redundancy rules in § 6. Notice, incidentally, that the rules (lW) (and, in fact, all rules that establish a partial hierarchy among syntactic features) might be regarded as redundancy rules rather tan as rules of te base. Such a decision would have various consequences, to which we shall return in § 4. 3. 18. By a local transformation (with respect to A) I mean one that afects only a substring dominated by te single category symbol A. Thus all rules of te transformational cycle in p honolog are local, in this sense . There is some reason to suspect tat it might be appropriate to intersperse certain local transformations among te rewiting rules of the base. Thus Adverbial Phrases consisting of Preposition"Deteriner"Noun are in general restricted as to the choice of these elements, and tese restrictions could be stated by local transformations to the efect that Preposi tion and Noun can be rewi tten in certai n restricted ways when dominated by such category symbols as Place Adverbial and Time Adverbial. In fact, one might consider a new extension of the teory of context-free gammar, permitting rules that restrict rewriting by local transforations (i.e., in ters of the dominating categor symbol), alongside of te fairly widely studied extension of con text-free gammar to context-sensitive grammars that permit rules that restrict rewiting in ters of contiguous smbols. The example of the preceding paragraph involves a transfor mation that is local with respect to a categry A (A, in tis case, being sme type of Adverbial), and, furtherore, that introduces a string into a position dominated by te lexical categry B which is immediately dominated by A. Let us call such a transformation strictly local. The only motivation for this highly special defnition is that many of te examples of local transformations that come to mind meet this restrictive condition as well (for example, quite gnerally, nominalization transformations that give such forms as "I prsuaded John of my seriousness" from an underlying form "I persuaded John of N S," where S dominates the string under lying "I am serious" and the transformation substitutes a trans f orm o f this string for te dummy symbol occupying te pition of the lexical category N, which is immediately dominated by the category symbol NP wi th respect to which te transformation i local) . J 9 . Notice that an importart question is begged when we assume tat Noun subcategorization is independent of context and that the 216 NOTS TO CHAER 1 selectional restrictIons on Subject-Verb-Object are given com pletely by rules determining te subcategorization of Verbs in terms of previously chosen Noun subcategories. We shall return to this matter in § 4. 2. 20. This decision, as several of the oters, will be slightly modifed later in the text. u. The status of the symbol S' in this rule is unexplained at the present state of the exposition. It will indicate the position of H transfor of a sentence, as the theory of the syntactic component is extended later on. 8- Observe that in (36) such an expression as " -like"Predicate Nominal" is a single smbol, standing for a particular sntactic feature. The careful reader will notice that as tese rules are forulated, lexical items can be inserted in the wrong position by te lexical rle. We shall return to this question in § 3 , avoiding it now only so as not to overburden te exposition. Actually, a more careful analysis would revise (40) and (41 ) in detail. 13. An apparent exception to the last remark is the subcategorization of Verbs in terms of choice of the Progressive form be + Ing. To maintain the suggested generalization concerning strict sub categorization, we should have to claim that such Verbs as own, understand, and know occur freely with or without Progressive (along with all oter Verbs), but tat te Progressive for is deleted by an obligatory transformation when it precedes tese Verbs (this peculiarity would be marked by a feature that consti tutes part of te lexical entries for tese forms). But, in fact, there is good reason to assume tis, U has been pointed out to me by Barbara Hall. Tus each element of te Auxiliary has assciated with it certain characteristic Adverbials tat may (or, in the case of Present tense, must) cooccur wit this Auxiliary element, and the characteristic Adverbials of Progressive do occur with te Verbs own, understand, knoU, etc. (cf. "I know the answer right now," alongside of "I know the answer"), although such fors as I eat the apple right now," "I eat the apple," are ruled out (except, in the latter case, as "generic," which can, in fact, be treated as involving deletion of a "dummy" Adverbial). 14. Strictly speaking, this is not the case, as we have defned "syntactic feature" (cf. pp. 821.). Actually, it is only the features involved in the set of rules of which (2o)-(n) constitute a sample that de terine selectional classifcation. Idiosyncratic syntactic features of particular lexical items, not introduced by such general rules as (20)-(U) but simply listed in the lexical enties, play no rle in Verb subclassifcation. NO TO CHPTR 2 15. Notice tat tese alteratives are not strictly equivalent. Thus. for example. of the three mentioned only the one we are using perits als the free use of variables. as in the case of schema (44) . On the other hand. the use of labeled brackets is appropriate for the formulation of the transformational rules of te phono· logical component. Use of complex symbols at arbi trary nodes (as in Haran. 1 963 -cf. note 4) gives a form of transformational grammar tat is richer in some respects and poorer in oters than the formulation in terms of Boolean condi tions on Analyz· ability. as in most current work on generative grammar. Cf. Cbom sky (in press) for some discussion. 16. Proper Nouns of course can have nonrestrictive relatives (and. marginally. Adjective modifers derived from nonrestrictive rela tives -e.g., "clever Hans" or "old Tom"). But although restrictive relatives belong to the Determiner system, there are several rea sns for supposing that nonrestrictive relatives are, rather, Cm plement of the full NP (and in some cases, of a full sentence e.g., "I found John likable, which surprised me very much"). Notice that Adjective modifers can derive from eiter restrictive or nonrestrictive relatives (consider, for example, the ambiguity of the sentence "the industrious Chinese dominate the economy of Southeast Asia") . This matter is discussed in the Port-Royal Logic (Arnauld et al., 1 662). and. in more recent times, by Jesper sen ( 1 924. Chapter 8). Notice also that Proper Nouns can also be used as Common Nouns, in restricted ways (e.g., "tis cannot be the England that 1 know and love." "I once read a novel by a diferent John Smith"). Some such expressions may be derived from Proper Nouns wi t nonrestrictive relatives by transforation; oters sug gest that a redundancy rule may be needed, in te lexicon, ð signing certain of the features of Common Nouns to Proper Nouns. 1 7 . Once again, this is not to deny that an interpretation can some times be imposed on such phrases as tose of (54) . See the dis cussion of te problem of justifcation at te outset of § 2. 3. 1 , and the references o footnote 1 1 . Notice. in particular. that the relation of the Verb to the Place Adverbial in "John died in England" (¯ "in England. John died") is very diferent from tat in "John stayed in England" ("John lived in England" is. in fact, an ambiguous representative of bot constructions, being i nterpretable as either "John re sided in England, " analogous structurally to "John stayed in England" wit H Verbal Complement introduced by rule (52iii). or roughly as "in England, John really lived" or "in England. 2 1 8 NOTES TO CHAPR 2 John remained alive," with a Place Adverbial that is a Verb Phrase Complement introduced by (5lii) -cf. "John will surely die on te Continent, but he may live in England"). This difer ence of structure between "live in England" and "die in England" account for te fact (noted by Ralph Long) that "England is lived in by many people" is much more natural tan "England is died in by many people" -in fact, tis remark is true only when "live in" has the sense of "reside in" or "inhabit." Cf. p. 1 04 for furter disLussion of such "pseudopassives. " 28. There are well-known marginal exceptions to this remark (e.g., "a good time was had by all" or "recourse was had to a new plan"), and i t is also clear that the locution "take Manner Ad verbials freely" requires considerable furter analysis and clarif cation (see Lees, 1 960a, p. 26), as does the distinction between Adverbials that qualify the Verb and those which might more properly be said to qualify the Subject. (As an example of the latter, consider the Adverbial of "John married Mary with no geat enthusiasm, " which means, roughly, "John was not very enthusiastic about marrying Mary, " and therfore seems to play a role more like that of te Adverbial modifer of the Subject in "John, cleverly, stayed away yesterday" tan like that of the Ad verbial modifer of te Verb in "John laid his plans cleverly." See Austin ( 1 956) for sme discussion of such cases.) Nevertheless, te essential correctness of te comments in the text does not seem to me in doubt. It must be borne in mind that the general rules of a grammar are not invalidated by the existence of exceptions. Thus one does not eliminate the rule for forming dle past tense of Verbs from the gammar on the grounds mat many Verbs are irregular; nor is me generalization that relates Manner Adverbials to passivization invalidated by the fact that certain i tems must be listed, in the lexicon, as conficting with this generalization, if this turs out to be te case. In either the case of past tense or tat of passivi ztion, the generalization is invalidated (in the sense of "internal justifcation" -cf. Chapter I¿ § 4) only jf H more highly valued gammar can be constructed tat does not contain it. It is for tis reason that the discovery of peculiarities and exceptions (which are rarely lacking, in a system of the complexity of a natural langage) is generally so unrewarding and, in itself, has so little importance for te study of the grammatical structure of te langage in question, unless, of course, it lead to te discovery of deeper generalizations. It is also worth noting tat many of te Manner Adverbials, like many oter Adverbials, are Sentence transforms with deleted NOTES TO CHAPTER 2 2 1 9 Subjects. Thus underlying the sentence "John gave the lecture with great enthusiasm, " wit te Adverbial "wi th great enthusi asm, " is the base string "John has geat enthusiam" (note that "with" is quite generally a transform of "have"), wit the re peated NP "John" deleted, U is usual (cf. Chapter 1 and Chapter 4, § X- X)- Similarly, Place Adverbials (at least those which are VP complements) must smetimes, or perhaps always, be regarded as Sentence transforms (so that, for example, "I read te bok in England" derives from an underlying structure very much like the one tat underlies "I read the bok while (I was) in England"). Adverbials are a ric and as yet relatively unexplored system, and therefore anythiIg we say about tem must be regaÏded as quite tentative. 19. Alternatively, we may drop this condition and extend te frst convention so tat the complex symbol introduced in the analysis of H lexical category A contains not only te feature [+ A], but als te feature [- B] for any lexical category B other than A. This convention entails that a word specifed as belonging to two lexical categories must have two separate lexical entries, and it raises unanswered questions about te strctur of the lexicon. It would have the advantage of overcoming a defect in our notation for features introduced by context-sensitive subcategorization rles. Thus, in the grammar (55), the feature [-] designates both Proper Nouns and Intransitive Verbs. (This is why the featur [+ N] had to be mentioned in rul e (57iv) .) This might lead to difculty if a certain lexical i tem were bot a Noun and a Verb, since i t might be non-Proper as a Noun but Transitive as ð Verb, or Transitive as a Verb and Proper as a Noun. If te propsal of tis notC is adopted, the problem cannot aÏise. Al teInatively, it will be necessar to designate such features by a more complex notation i ndicating not only te frame in question but also the symbol tat dominates it. There may be sme point to allowing a lexical item to appear in sveral categorial positions (either by specifying it positively wit respect to several lexical categorieô, or by leaving it totally un specifed with respect . to tese categories) -for example, in the case of such words as "proof," "desire," "belief." Suppose that these are specifed as taking Sentential Complements of various forms, but are permitted to enter either the Noun or Verb position. Then the lexical insertion rle will place them in either the frame " . . . N tat S . . . " or te frame " . . . V that S . . . ," in the poåi tions of the Noun and Verb, res p ectively. Hence it will not be necessar to derive the foer by transforation from te latter, as is necessary, for example, in the case of " . . . proving that S . . . ". 220 NOT TO CHAPR 2 Under such an analysis, "John's proof that S" would derive from the structure underlying "John has a proof that S" by the sequence of transformations that derives "john's book" from the structure underlying "John has a book." One might go on t relate "John has a proof that S" to "John proves that S" (perhaps, ultimately, as "John takes a walk" is related to "John walks"), but this is another matter. In connection with this discussion, it is also necessary to estab· lish a general distinctness condition regarding te idiosyncratic, purely lexical features (e.g., the feature [Object-eletion] in (58), (59 » . For discusion of tis question, which becomes critical in case these features relate to te phonological component, see Halle and Chomsky (forthcoming). 80. It has been maintained that tese relations can be defned in terms of some notion of cooccurence, but tis seems to me duo bious, for reasons presented in various places (e.g., in Bar·Hillel, 1 954; and Chomsky, 1 964). Observe that the defnitions of gram matical relation or grammatical function tat have been suggsted here refer only to the base of the syntax and not to surface structures of actual sentences in other tan the simplest cases. The signifcant grammatical relations of an actual sentence (e.g. (7), p. 70), are those which are defned in te basis (deep structure) of this sentence. 81 . I give these informally, instead of using the notation developed earlier. to simplify the reading. There is nothing essential in· volved in this change of notation. 81. For example. if we were to adapt the defnitons of universal categories and functions so tat they apply to such sentences U "in England is where I met him, " which are often cited to show that phrases other than NP's can occur as Subjects, these proposals would fail completely. This sentence, however, is obviously trans formationally derived. It would be perfectly corect to say that "in England" is the Subject of "in England is where I met him," extending the gammatical relation Subject,of. that is, [NP, S], t te derived Phrase·marker (the surface strcture). In the basis, however. "in England" is an Adverbial of Place, associated with the VP meet him in te Predicate·Phrase "met him in England," and the sentence is interpreted in accordance wit te gram· matical relations defned in this underlying deep structure. This extension to surface structures of such functional notion as Subject-f is not an entirely straightforward matter. Thus in base structures, tere is apparently never more than H single ocurence of a category such as NP in any structure immediately dominated by a single category (cf. note 7). and our defnition NOTS TO CHAPT 2 2 2 1 of these notions relied on this fact. But this is not true of surface structures. In the sentence "tis book I really enjoyed," bot "tis book" and "I" are NP's immediately dominated by S. Ap parently, then, order is sigifcant in determining te grammatical relations defned by surface structures (not surprisingly), though it seems to play no role in the determination of gammatical rela tions in deep structures. Consequently, somewhat diferent def nitions are needed for the surface notions. It might be sugested that Topic-Coment is te basic gam matical relation of surface structure corresponding (roughly) to the fundamental Subject-Predicate relation of deep structure. Thus we might defne the Topic-f the Sentence U te leftmost NP immediately dominated by S in the surface structure, and the Comment-of te Sentence as the rest of the string. Often, of course, Topic and Subject will coincide, but not in the examples discussed. This proposal, which seems plausible, was sugested to me by Paul Kiparsky. One might refne it in various ways, for example, by defning the Topic-f the Sentence as the leftmost NP that is immediately dominated by S in te surface structure and that is, furthermore, a major category (cf. p. 74 ¯this will make John the Topic in te cleft sentence "it was John who I saw"). Other elaborations also come to mind, but I shall not go into the question any more fully here. 83. This very fruitful and important insight is as old U syntactic teory itself; it is developed quite clearly in te Grammaire generale et raisonnee of Port-Royal (cf. Chomsky, 1 964, § 1 .0; fort coming, for discussion). Wat is, in essence, the same idea was rein troduced into modern linguistics by Harris, though he has not dis cussed it in quite these terms (cf. Haris, 1 952, 1 95 4, 1 957). For further discussion of this notion, witin te framework of trans formational gnerative grammar, see Chomsky ( 1 957), and for steps toward a substantive theory of semantic interpretation based on tis asumption, see Katz and Fodor (1 963) and Katz and Postal ( 1 964). 34. Cur's prposals are s setchy that it is impossible to extract fm tem mre than a generl point of view. The position of Saumjan and Sobleva is much more explicitly worked out, but it is defective in crcial respects. Cf. Hall ( 1 965), for an analysis of this approach. It is possible that "stratifcational grammar" als adopts a similar position, but the published references t this theory (e.g., Gleasn, 1964) are much too vague for any conclusion to be drawn. 35. Notice, for example, that Cas is usually determined by the posi tion of the Noun in surface structure rater than in deep struc- NOTE TO CHAPR 2 ture, although the surface structures given by stylistic inversion do not af ect Case. Even in English, poor as it is in in8ecton, tis can be observede For example, te Pronoun in the sentences "he was struc by a bullet," "he is eay to please," he frightens easily" is, in each case, the "logical Object, " that is, the Direct-Object o f Verbs strike, pleae, frighten, respectively, in te underlying deep structures. Neverteless, the form is he rather than him. But stylistic inversion of te type we have just been discusing gives such forms as "him I really like," "him I would defnitely t not to antagonize." Were in8ections are richer, this phenomenon, whic illustrates the peripheral character of tese processes o f inversion, is much more apparent. The relation between infection, ambiguityg and word order was discussed at some length in traditional linguistic theory. See Chomsky, forthcoming, for some references. NOTES TO CHAPTER J I . Some details irrelevant to te problem under discusion are omit ted in these examples. We here regard each lexical i tem as stand ing for a complex of features, namely those that constitute it lexical entry in addition to tose entered by redundancy rules. The use of te dummy symbol a has been extended here to the case of various unspecifed elements that will be deleted by oblig ator transforations. There is, in fact, good reasn to require tat only "recoverable deletions" be permitted in the gammar. For discussion of this very important question, see Chomsky, 1 9 64, § 2. 2. We shall return to it at te end of this chapter and in Chapter 4, § 2. 2. The formative nom in (!) is one of several that might be a signed to the Tense ..Modal position of the Auxiliary, and tat determine the form of the Nominalization (for-to, possessive-;ng, etc.). X- The detils of this, bot for Transformation-markers and Phrase markers, are worked out in Chomsky ( 1 955), witin the following general framework. Linguistic teory provides a (universal) system of levels of representation. Each level L is a system based on a st of primes (minimal element -i. e. , an alphabet); the operation of concatenation, which form strings of primes of arbitrary fnite length (te terms and notions all being borowed from the theory of concatenation algebras -cf. e_g., Rosenbloom, 1 950); various Ï^ lations; a designated class of strings (or sets of strings) of primes called L-markers; a mapping of L-markers onto V-markers, where L' is the next "lower" level (thus levels are aranged in a hier arcy). In particular, on te level P o f phrase stucture and te NOT TO CHAPTER 3 223 level T of transformations we have P-markers and T-markers in te ôM just descibed informally. A hierarchy of linguistic levels (phonetic, phonological, word, morphological, phrase structure, transforational structure) can be developed within a uniform framework in this way. For details, see Chomsky ( 1 955) . For a discussion of T-markers, see Katz and Postal ( 1 964). 3. For discussion of negation, see Klima ( 1 964), Katz ( 1 964b). The formation of questions and imperatives and the semantic inter pretation of the question and imperative markers are discussed in Katz and Postal ( 1 964). In Rockett ( 1 96 1 ) the proposal is made that the passive transformation be condi tional on a marker in the underlying form, but no supporting argument is given for what, in the context of that paper, is no mor than a notational innovation. Notice that the reformulation of the passive transformation as obligatory, relative to choice of an optional marker in the under lying string, is independent of te principle that we have just cited, since the passive marker, as distinct from the question, nega tion, and imperative markers, has no independent semantic inter pretation. Furtherore, we have noted in § 4.4 of Chapter X that there are good reasns to distinguish such transformations as pas sive from purely stylistic i nversion operations. These obsrations suggest that we attempt to forulate a more general condition of whic the principle just cited i s i tself a consequence, namely that "nonstylistic transformations" are all signaled by optional markers drawn from a fxed, universal, language-independent set. This attempt presupposes a deeper analysis of the notion "nonstylistic tansforation" than we have been able to provide here, however. 4. For illuminating discussion of this question, and several others tat we are considering here, see Fillmore (1 963) and Fraser ( 1 963). 5. Both of these observations are due to Fillmore ( 1 963). 6. In connection with ordering of rules, it is necessary to distinguis extrinsic order, imposed by te explicit ordering of rules, from intrinsic order, which is simply a consequence of how rules are forulated. Thus if te rule R I introduces the symbol A and R2 analyzes A, tere is an " intrinsic order relating R I and R2, but not necessarily any extrinsic order. Similarly, if a certain transforma tion T I applies to a certain structure tat is formed only by application of T2, there is an intrinsic order T1, T2• Taxonomic linguistic disallows extrinsic ordering, but has not been clear about te status of intrinsic ordering. Generative grammars have ordinarily required bot. For some discussion of this matter, see Chomsky ( 1 964). 2 24 NOTE TO CHAPTER 3 7. We are discussing only embedding transformations here, but should extend the discussion to various generalized transformations that form coordinate constructions (e.g., conjunction). There are certain problems concerning these, but I believe tat tey can be incorporated quite readily in the present scheme by permitting rule schemata (in the sense of Chomsky and Miller, 1 963, p. 298; Chomsky and Schitzenberger, 1 963, p. 1 33) introducing coordi· nated element tat are then modifed, rearanged, and appropri. ately interelated by singulary transformations. If the suggestion of note 9, Chapter X¡ is workable, ten such rule scemata need not be stated in the gammar at all. Rather, by a general con· vention we can asociate suc a schema with each major categor. This approach to coordination relies heavily on the fltering efect of transforations, discussed later. Thus wherever we have co· ordination, some category is coordinated M times in the matrix sentence, and n occurences of matched sentences are independ. ently generated by the base rules. 8. Notice, incidentally, that we can now eliminate Complement from te set of category symbols. We could g on, at this point, to defne "Complement" as a functional notion (to be more precise, as a cover ter for several functional notions), in the manner of pp. 707 1 • 9. As it stands, this claim seems to me somewhat too strong, though it is true in one important sense of semantic interpretation. For example, it seems clear that the order of "quantifers" in surface structures sometimes plays a role in semantic interpretation. Thus for many speakers -in particular, for me -te sentences "ever. one in the room knows at least two languages" and "at least two languages are known by everyone in the room" are not synony mous. Still, we might maintain that in such examples bot in terpretations are latent (as would be indicated by the identity of te deep structures of the two sentences in all respects relevant to semantic interpretation), and that te reason for the opposing interpretations is an extraneous factor -an overiding considera tion involving order of quantifers in surace structures -that flter out certain latent interpretations provided by the deep structures. In support of this view, it may be pointed out that other sentences that derive from these (e.g., "there are two lan guages that everyone in the room knows") may switch interpre tations, indicating that these interpretations must have been latent all along. Tere are other examples that sugst something similar. For example, Grice has sugsted that the temporal order im· plied in conjunction may be regarded as a feature of discoure rather than U part of the meaning of "and," and Jakobson has NOTE TO CHAPTER 3 225 also discussed "iconic" features of discourse involving relations between temporal order in surface structure and order of im portance, etc. Also relevant in this connection is the notion of Topic-Comment mentioned in note 32, Chapter 2. For some references to remarks in the Port-Royal Logic on the efect of gammatical transformations on meaning, see Chomsky (forth coming). 10. The other function of the transformational component is to express restrictions on distribution for lexical i tems and for sen tence s tructures. 1 1 . Formally speaking, what we are suggesting is this. Suppose tat the symbol A immediately dominates XBY (where B is a symbol) in the Phrase-marker K; that is, A * XBY was one of te cate gorial rules used in generating this Phrase-marker. Then (A,B) constitutes a branch of K. Furthermore, if this occurence of B immediately dominates ZCW (where C is a symbol), so tat (Bg C) is a branch, then (A,B, C) is H branch, etc. Suppose now that (A I ' ¯ ¯ ¯ , A . ) is a branch of the generalized Phrase-marker K formed by base rules, and that Al ¯ A". Then it must be that for some i, l :;i :;M¿ At ¯ S. In other words, the only way to for new deep structures is to insert elementary "propositions" technically, base Phrase-markers -in oter Phrase-markers. This is by no means a logically necessary feature of phrase structure gammars. Notice tat the schemata that underlie coordination (cf. note 7) also provide infnite generative capacity, but here too te true recursive property can apparently be limited to the schema S - S#S# • • • IS, hence t rles introducing "propositions." This forulation leaves unexplained some rather margnal phenomena (e.g., the source of such expressions as "very, very, . . . , very Adjective" and some more signifcant ones (e.g., the possibility of iterating Adverbials and various kinds of parenthetic elements, the status of whic in general is unclear). For some discussion of Adverbial sequences, see Matthews ( 1 961 ) . u. Cf. pp. 1 1 7-1 1 8. For some discussion, see Chomsky (1 964, § 1 .0, and forthcoming). . 1 3 . Notice, i ncidentally, tat tis identity condition need never be stated in the grammar, since it is a general condition on the func tioning of gammars. This is important, since (as was pointed out by Lees, 1 960a), the condition is not really identity of strings but rather total identity of structurs, in all cases in which identity conditions appear in transformations. But to defne identity of strctures in terms of Analyzability it is necessar to use quanti fers; in fact, tis may be te only case in which quantifers must NOTES TO CHAPTER 3 appear in te structural analyses tat defne transforations. Ex tracting the identity condition from grammars, we are terefore able to formulate the structural analyses that defne transfora tions strictly as Boolean conditions on Analyzability, thus greatly restricting the power of te teory of transforational gram mar. 14. For discussion see Miller and Chomsky ( 1 963) ; Schlesinger ( 1 964): Miller dnd Isard ( 1 964): and te rbume in Chapter 1 , § 2. 1 5. See § 2. 3. 1 of Chapter 2, and § I of Chapter 4. A serious discusion of this question, as well B the question of dependency of syntax on semantics, awaits ð development of the teory of universal semantics, that is, an account of the nature of semantic representa tion. Although various posi tions about these questions have been stated wit great confdence and authority, the only serious work tat I know of on the relation of tese domains is tat of Katz, Fodor, and Postal (see bibliography; for discussion of other claims that have been made, see Chomsky, 1 957, and many oter publica tions). For the moment, I see no reason to modify the view, ex pressed in Chomsky (1 957) and elswhere, tat although, obviouåly, semantic considerations aIe relevant to the coMstruction of gneral linguistic theory (that is, obviously the theory of syntax should be designed so tat te syntactic structures exhibited for particular languages will support semantic interpretation), tere is, at pres ent, no way to show that semantic considerations play a rle in the coice of te syntactic or phonological component of a gam mar or that smantic features (in any signifcant sense of this ter) play a role in the functioning of te åyntactiC or phonological rules. Thus no serious proposal has been advanced to show how semantic considerations can contribute to an evaluation procedure for such systems or provide some of the primary linguistic data on the basis of whic they are selected. See Chdpter 1 , § 6, and Chap ter 4, § 1 , for some additional related discussion. 1 6. Some of the details of this modifcation are worked out in Fraser (fortcoming). Te extent to which te complexity of the theor of derived constituent structure depends on te presence of per mutations is quite clear, for example, from the analysis of thes notions in Chomsky (1 955, Chapter 8). 1 7. Notice tat in this case the tird term of the proper analysis is not strictly deleted. Rater, this ter is deleted except for te feature [± Human] , which then assumes its phonological shap ( giving who, which, or that) by later rules. This is often tue of what we are here calling erasure operations. 1 8. A natural notational decision would be to restict te integers one and two to frst and second person, respectively. NOT TO CHAP 4 227 NOTES TO CHAPTER 4 l . Wether the rule is a rewriting rule or a substitution transforma tion -cf. Chapter 2, § 4. 3 -does not concern us here; for con venience of exposition, we shall assume the latter. 2. To avoid what has been a persistent misunderstandi ng. it must be emphasized again tat "grammaticalness" is being used here as a tecnical ter, with no implication that deviant sentences are being "legislated against" as "wi thout a function" or "illegi timate. " Quite the contrary is true, as has repeatedly been stressed and illustrated, in discussions of generative grammar. For discussion, see Chomsky ( 196 1 ) and many other references. The question as to whether the grammar should generate deviant sentences is purely terminological. having to do with noting more than the technical sense of "generate. " A descriptively adequate grammar mUSt aååigM to eac string H structural descri ption that indicates te manner of its deviation from strict well-formedness (if any) . A natural terinological decision would be to say that te grammar diTectly geneTates the language consisting of just te sentences that do not deviate at all (such as (3», with their structural descrip tions. The grammar deTivatively geneTates all other strings (suc as ( 1 ) and (2». wit teir structural descriptions. These structural desriptions will indicate te manner and degree of deviance of the derivatively generated sentences. The principles that deter mine how interpretations can be imposed on deviant sentences may be universal (as sugested in Chomsky, 1 955. 1 961 ; Miller and Chomsky, 1 963; and again here) or specifc to a given language (as suggested in Katz. 1 964a). This is a substantive issue, but many of the other questions that have been debated concerning these notions seem to me quite empty, having to do only wit termino logical decisions. 3 . Recall that selectional rules, as illustrated earlier, are rules that insert Verbs and Adjectives into generalized Phrase-markers on the basis of the intrinsic syntactic features of the Nouns tat appear in various posi tions. But not all of te rules referring to intrinsic syntactic features of Nouns are selectional rules; in particular, the rules violated in the formation of (4) involve such features but are not selectional rles. 4. Many of the Verbs of thC category [+[ + Abstract] . . . - . . . [+ Animate]] do not have Adjectival forms with ing. but these seem invariably to have other afxes as variants of ing (botheTsome for botheTing, scar for scU7ing, impressive for impressing. etc. ). 5. These examples do not begin to exhaust the range of possibilities that must be considered in a full study of interpretation of deviant sentences. For one ting, tey do not illustrate the use of order- 228 NOTES TO CHAPR 4 inversion a a stylistic device (d. Chapter 2, § 4.4, for some discus sion). The discussion of deviation from grammaticalness that ha been carried on here ofers no insight into this phenomenon. For example, consider te following line: "Me up at does/out of te Hoor/quietly Stare/a poisoned mouse/still who alive/is asking What/have i done that/You wouldn' t have" (E. E. Cummings ). This poses not te slightest difculty or ambiguity of interpreta tion, and it would surely be quite beside the point to try to assig it a degee of deviation in terms of te number or kind of rule of te grammar tat are violated in geneating it. 6. Notice that the formulation given previously left an ambiguity in the latter case, whic is resolved only by the convention that we now state. 7 . We are, in efect, assuming the convention e ¯ [e, . ], where C is te null element. Notice that features are unordered in a com plex symbol. As elsewhere in tis discussion, I make no attempt here to present an absolutely precise account or to give thes defnitions in their simplest and most general fors. S . Thus X is null if [a] is null; Y is null if [�] is null. 9 . This difculty would, in fact, not arise if we were to give a some what diferent analysis of post-Yerbal Adjectives in Englishg deriv¯ ing them from underlying strings with Sentence-Complements to the Yerbs. In some cass, this is surely correct (e.g., "John seems sad" from an underlying structure containing the base string "John is sad, " which becomes John seems to be sad," and then John seems sad" by further transformations -similarly, i n the case of "become" tis analysis is well motivated, in particular, because it can provide a basis for excluding "become" from passivization), and i t may be correct to extend it to many or all such cases. For sme other proposals for derivation of certain of these fors, se Zierer ( 1 964). It is worth noting that d condition like that imposed on W and V in the discussion of the schema (9) is probably necessary in te theory of transformations, although this problem has never been disussed explicitly. 1 0. I am indebted to Thomas Bever and Peter Rosenbaum for many interesting and sugestive comment relating to this question. 1 1 . In many or all such cases, some notion of "generic" seems to be involved critically (I owe this observation to Barbara Hall). One might therefore try to show that part of the semantic efect of generic is to cancel semantic conficts of certain sorts. Notice, incidentally, tat the deep structure of each of te sentences of ( 1 5) will contain a string wit sincerity as the Direct-Object of te Main Yerb frighten (and wit an unspecifed Subject). NOS TO CHAPTR 4 229 Ì- Interest i n these questions can be traced to Humboldt ( 1 836) ; for representative statements of his, see Chomsky ( 1 964) . See Ullmann ( 1 959) for discussion of much related descriptive work. Also relevant are some psychological studies that have attempted to place a lingistic item in Ü context of somehow related items, suc as Luria and Vinogradova ( 1 959), and much current work in "componential analysis. " I l. Although te sentences of ( 1 9i) are near-paraphrases, still it is by no means true that a "cooccurrence relation" of the sort that has been discussed by Harris (1 957), Hif ( 1 961 ), and others holds be tween them. Thus pompous can be replaced quite naturally by Ü friend in "I regard John as ¯ , " but hardly in "John strikes me as ¯" (I owe this observation to J. Katz) . It is dear, then, that te dose meaning relation between regard and strike (involv ing_ in particular, inversion of the Subject-Verb-Object relations) does not deterine a corresponding similari ty of distributional restrictions. The rules involving contextual features, in other words, may be partially independent of semantic properties. Such examples must be borne in mind if any attempt is made to give sme substance to the widely voiced (but, for the moment, totally empty) claim that semantic considerations smehow determine syntactic structure or distributional properties. I have been assuming, in discussing ( 1 9i) that the Subject-of strikes in te deep structure is John, but it should be noted that tis is not at all obvious. One alternative would be to take the underlying structure to be it"S -strikes me, where it"S is an NP and S dominates te structure underlying "John is pompous." An obligatory transformation would giVe the structure underlying it strikes me that John is pompous," and a furter optional trans formation would give "John strikes me as pompous." The lexical item strike of ( 1 9i) would then have very diferent strict subcate gorization features from te phonetically identical item of "it struck me blind, " while both would difer in strict subcategoriza tion from strike in "he struck me," "he struck an outlandish pose," etc. (cf. note 1 5, Chapter 2). If tis analysis can be justifed on syntactic grounds, then the deep structures will be somewhat more appropriate for the semantic i nterpretation tan assumed in the text. As several people have observed, tere are other relevant syntactic diferences between te paired examples of ( I gi). For example, such sentences U "John strikes me as pompous, " "his remarks impress me U unintelligible" do not passivize, although te sentences "I regard John as pompous," "it struck me blind," and so on, are freely subject to passivization. In connection with ( 1 9iii), Haris has sugested ( 1 952, pp. 24-25) NOTES T CHAPR 4 that it may be possible to express the meaning relation on dis tributional gounds, but his suggestions as to how tis might be possible have not yet been developed to the point where teir merits can be evaluated. Notice . that the problems mentioned here admit of no merely terminological solution. Thus we could perfectly well state te facts relating to ( 1 9 ) in terms of such new notions as "semantic subject, " "semantic object," various kinds of "sememes," etc., but such proliferation of terminology contributes noting toward clarifying the serious issues raised by such examples. 14. A pointed out in note 1 5, Chapter X¡ a distinctive-feature matrix is simply a way of representing a set of abstract phonological features, so tat a lexical entry (a formative) may be regarded simply as a set of features, wit further structure defned on tem in the manner suggested informally in this discusion. 15. With respect to se1ectional features, alternative (iv) is well moti vated. See note 10. To say tat a feature is positively (negatively) specifed is to say that it is marked + (respectively, ¯) . Notice that these or any anal ogous conventions make a distinction amounting to the marked! unmarked distinction that has often been discussed, though quite inconclusively, in connection with features and categories. 1 6. Such examples ð "sincerity frightens" can be found, of course, but only as (rater mannered) transforms of "sincerity frightens Un specifed-Object," and so on. The possibilities for tis are, in fact, quite limi ted -for example, no one would interpret "his sincerity was frightening" as ambiguous. Notice that words of the categor of "frighten" do appear quite naturally as Intransi tives in surface structures, as in "John frightens easily" (this in fact is much more general -cf. "the book reads easily," etc.). But this is irelevant here. In such a case, the "grammatical Subject" is the "logical Object" -tat is, the Direct Object of the deep structure "Un specifed-Subject frightens John easily." The often obligatory Adverbial of Manner, in these cases, suggests tat one might seek a generaliztion involving also te passive transformation. 1 7. The latter would be interpretable only as a deviant sentence. 1 8. One might question the factual correctness of tis, particularly in the case of {[- Count] , [± Abstract] }. I have been assuming tat the features {[- Count]. [+ Abstract]} characterize the true Ab stract Nouns such as virtue, jutice, while the features {[- Count], [- Abstract] } characterize the Mass Nouns such as water, dirt. But there is a subdivision of Inanimate Count Nouns tat seems to corespond to this, namely the distinction into [+ Concrete], suc NOTS TO CHAPTR 4 as table, mountain, and [- Concete], suc as problem, efort. I f it tuns out that the features [± Concrete] and [: Abstract] (as subfeatures of [- Animate] and [- Count], respectively) should b identifed, then te feature [Abstract] would be cros-lassifying rater tan hierHcic wit respect to [+ Count]. This queston is not easy to resolve without much more empirical study, hoever. 19. The desirability of suc a conventon wa pointed out by Paul Postal. 20. Notice tat if we were explicitly to list positively specifed rater than negatively specifed selectional features in the lexicon, then this convention would have to be extended to selectional features as well. Thus we should not want to have t list both te features coresponding to "takes Human Subject" and "takes Animate Subject" for "run," for .example. Suc a convention would, in efct, treat a selectional feature a itelf being a kind of complex symbol. 2 1 . A always, there are a few exceptions tat require separate state ment. Recall that we have presented some reasons for regarding te phrae by"" passive (where pasive is a dummy terinal symbol, replaceable, in fact, by te universal dummy symbol A) a a Man ner Adverbial. A Verb that can appear only in the passive would therefore be an exception to tis rule (e.g . , "he is said t b a rater decent fellow," or, perhaps, such for as "he was shorn of all dignity"). XX- The phonological redundancy rules are alM subject t certain uni versal constraints, and there is no doubt that, for all features, tese constraints g well beyond what ha been illustrated here. A these are forulated, they will also play te role of general con ventions (i.e., apect of the general defnition of "human lan guag") that can be rlied on to reduce te specifcity of paticular gd ars. 2 1 . See Halle ( 1 959a, 1 959b), 1 961 , 1962a, 1 964. Cf. also the discussion of evauation proedures and explanator adequacy in Chapter 1y §§ 6, 7, and in the references given tere. Notice that Halle's defnition of the notion "phonologcally admisible" (i.e-, "acci dental" versus "systematic gap") sugests what in Chapter I wa called a "foral" rater than a "substantive" linguistic univeral, though tere are, no doubt, also substantive constaints to be dis covered here. 24. As pssible examples of "accidental gaps" we might pint to te nonexistence of a Verb X taking as Direct-Object exprssions designating ' animals and having otherwis the same meaning as te trnsitive "gw," M that "he X's dogs" is parallel in meaning 232 NOTS TO CHAPTER 4 t "he grows cor" ("raise" appears to cover both senses); or the absence of a word that bears to plants the relation tat "corpse" bears to animals (tis example was sugested by T. G. Bever). 15. Thus we can regard te category of case in German as a four valued, gender as a three-valued, and number as a two-valued dimension, and ¼C can consider all Nouns as being arrayed in H single multivalued dimension of dedensional classes. Presumably, this is not the optimal analysis, and further structure must be imposed along these "dimensions. " It is also possible to try to give a language-independent characterization of these categories. These are important matters and have been the subject of much study tat, however, ges well beyond the scope of this discussion. I shall therefore consider only an unstructured description in these illustrative examples. 16. Simply for expsitory purposes, let us take the integers in the order of conventional presentations, so that [1 Gender] is Mas culine, [1 Number] is Plural, [1 Case] is Genitive, and Bruder is assigned to Class 1 along the "dimension" of declensional class. Notice that we have assumed all along that features are "binary" -that they simply partition their domain of applicability into two disjoint classes. There was no logical necesity for this. In phonology, it seems clear that the distinctive features ae, in fact, best regarded as binary in their phonological function (d., e.g., Halle, 1 957), though obviously not always in their phonetic func tion. Thus in the case of the feature Stress, we can easily fnd fve or more degrees that must be marked in English, and other phonetic features would also have to be regarded as multivalued in a detailed grammar. It has been maintained (cf. Jakobson, 1 936) that such "dimensions" as Case should also be analyzed into a hierarchy of binary features (like phonological distinctive fea tures), but we shall not consider this question here. 27. That is, the categorial rule that develops Nouns will not be N � I (d. p. U2), but rather N � [I, a Number] (a ¯ + or - for English or German, tough it may have more values or a diferent oraniza tion of values ¯d. note 15 -for other systems). 18. Actually, in descriptivist grammars of the item-and-arrangement type the latter migt be omitted, since its only function is to per mit some generality to be introduced into the "morphophonemic" rules and sinc tese gammar are, in fact, designed in such H way as to exclude the possibility of all but the most elementary gneral rules. See Chomsky ( 1 964, pp. 3 d.) for discussion. 29. This defect of morphemic analysis of infectional systems, which is quite serious, in practice, was pointed out to me by Moris Halle. NOTE TO CHAPTER 4 30. Thus an alternative to the analysis presented in (30) would be to regard a lexical item such as Bruder a consisting of a Stem fol lowed by an Ending, and to regard the Ending as belonging to the paradigmatic categories. 3 1 . In the last few years, there has been ver intensive and fruitful study of the transforational cycle of Russian and Latvian phonol ogy (for references, see Chomsky, 1 964, note 6, p. 1 4). The rules that constitute this system apply to Phrase-markers, and conse quently their formulation depends very heavily on answer to the questions being considered here. There has, so far, been no serious investigation of how a transformational cycle applies to a feature system and to Phrase-markers such a (30). When tis is claifed, it will be possible to bring phonological evidence to bear on the question of morphemic versus paradigmatic representation of i nHectional systems. For the moment, the empirical evidence sug gests that te ordering of the transformational cycle in phonology is deterined completely by categories, not features (though of course certain rules may be restricted in application in ters of syntactic features). This is, furthermore, te most natural assump tion, if we regard the features as actually constituting the terminal symbol (the formative) . 32. This forative might, in fact, be regarded as consisting of te feature [+ Defnite], hence as a degenerate complex symbol tat is expanded by the rule into the full complex symbol [+ Defnite, @ Gender, � Number, y Case] . See note 38 for sme support for this assumption. 33. Variables over feature specifcations were used in Chomsky, Halle, and Lukof ( 1 956) and Halle and Chomsky ( 1 960), in developing the transformational stress cycle. The idea of using them to deal with assimilation is due to Halle ( 1 962 b). T. G. Bever has pointed out that the same device can be applied to a description of various kinds of alternations that involve feature shift (e.g., Ablaut) . Cf. Bever ( 1 963), Bever and Langendoen ( 1 963). 34. See Lees ( 1 961 ) and Smit ( 1 961 ) . When the two Adjectves are paired in a rather special way that is for the present poorly under stood, the transformation is not blocked even when they are dis tinct. Thus we have such forms as "this is taller than that is wide." Cr. Haris ( 1 957), p. lP 4- 35. Notice tat te distinction tat is emergng in this discussion is not coincident with that suggested in note 30. It is interesting to note that the correctness of such examples as (40) has been questioned. In one of the earliest descriptive studies of French, Vaugelas ( 1 647, pp. 461 -462) maintains that such a fafon de parler cannot be considered either "abslument 2 3 4 NOT TO CHAPR 4 mauvais" or "fort bonne," and sugests that it be avoided when masuline and feminine fors of the Adjective difer. Thus; a man speaking to a woman should not sy je suis plu beau que vou, but suld rater ("pour parler reguli�rement") reört to te parphrase je suis plus beau que vous n'etes belle, although it would be perfectly all right for him to say je suis plus Tiche que vou. 16. Tis fact, pointed out to me by Brandon Qualls, raises various dfculties for te analysis of comparatives. In partiLlar, if such sentences as (41 iii) are regarded as derived from "I know several lawyer (who are) more successful tan Bill" by Noun-Adjective inverion following deletion of "who are," as seem quite plausible, we must somehow account for such facts as the following: te impossibility of "I know a more clever man than Mary" or "I have never seen a heavier book tan tis rock," altough the pre sumed sources of tese (namely, "I know a man (who is) more clever tan Mary" and "I have never seen a book (which is) heavier tan tis rok") are perfectly all right; te fact that the sntence "I have never read a more intricate pom tan TTistTam Shandy" implies that the latter is a poem, whereas te sentence "I have never read a poem (which is) more intricate tan TTistTam Shandy," which, in this view, is taken to be its source, dos not imply tat TTitTam Shandy is a pom; etc. Again, a troughout this discussion, I should like to emphasize that there is no paticular difculty in foÏulating an ad hoC sys tem of tanforational rules that will have the desired properties. The problem, rther, is to provide some explanation for suc phenomena a those of te preceding paragraph. 1 7. Te deletion of the pluralized non-Defnite Article is automatic, in tis psition. 1 8. Similar considerations may account for another apparent violation of te genera condition on recoverability of deletions. As has fre quently been obsered, the identity condition for relativization involves only te Noun, and not the Determiner of the deleted Noun Phrd. Thus frm "I have a [# the friend is from England #] fend" we cn for, by relativization, "I have a friend (who is) from England" in the usual way. The deleted Noun Phrae is "te friend," and te problem is the deletion of te Article, which difr frm te Article tat is used to erase it by te relative trans foration. The embedded sentence could not be "a friend is from England," in which case the problem would not arise, since def initenes of the Article is automatic in this position. But the fact tat defniteness is Çbligatory suggests that in the underlying Phrase- NOTES TO CHAPTER 4 marker the Article be left unspecifed for defnitenes, tis being added by a "redundancy rule" (in this cae. an obligator tans formation). If this is the corect analysis, then by te prinCiple just estabJished. deletion of te Aticle will be perissible, sinC in its underlying for it is nondistinct from te Article of the Noun Phrae of te matrix sentnce. Note that this decision requires a feature analysis for Articles. with [± Defnite] taken as a syntactic featue. S 9 . Notice tat altough sad, for example, need not b marked in te lexicon for pst-Animatenes (if we decide tat what is involved here i not Ü matter of homonymity), it may very well be asigned contextual features corresponding to vaious subfeatures of [- Ani mate]. so as to characterize as deviant suc sentences as "te pencil is sad," which cannot receive an interpretation analogous to that of "te bok was sad." This matter ha no relevance to te pint at issue, tough it rdises non tivial problems of a diferent srt. 40. We oversimplify somewhat. Thus the constituent base Phrase marker, in this case, might contain a certain nominalization mor pheme in place of te pre-Aspect pat of the Auxiliary. 41 . These constructions Ü interesting in many respect. See Les (1 960a. pp. 64f.), Chomky ( 1 964, pp. 47f.), and Katz and Postal (1 964, pp. 1 2of.) for discussion. 41. Here, too. we might raise the question wheter the nominalization element should be represented as a morpheme MUH or as one of the features F _¿ • • ¯ , Fm - in tis case, a feature added by te transforation. 4 S ' A detailed study of one system of essentially this sort, namely foration of cmpound nouns, is presnte in Lees ( 1 960, Chap ter 4. and appendices). See now als Zimmer (1 964). 44. Cf. also note 30. Perhaps it will be possible t rephras this con vention as part of a general dfnition of the notion "word. " That is. one might tr to state a general rule determining placement of word boundaies in tC of lexical categories and branching wi tin te spe of complex sybols. Tis psibili ty wa sug gsted by some obserations of Paul Postal's. and should be furter explored. 45. A related clÜ of problems is examined brie8y by Haris ( 19 57. § 4. 5 ). in his discussion of "quasi-transformations." Bolingr, in various articles ( e.g., Bolinger. 1 961 ). has listed many examples of porly understood quai-productive processes. Suc list simply indicate areas where all presently known theories of language have failed t provide any substantial insight, and they can be ex tnded in many ways. wit little difcult. Bolinger sugst that NOTES TO CHAPTR 4 his examples support an alternative theory of gammar, but this seems to me an entirely unwarranted conclusion, for reaSons dis cussed elsewhere (in particular, Chomsky, 1 964, p. 54). Bibliography Aristotle. De Anima. Translated by J. A. Smith. In R. McKeon (ed.), The Basic Works of Aristotle. New York: Random House, 1 941 . Arnauld, A. , and P. Nicole ( 1 662). La Logique, ou l'art de pener. Austin, J. L. ( 1 956). "A plea for excuses. " Proceedings of the Aristote lian Society. Reprinted in J. O. Urmson and G. J. Warnock (eds.), Philosophical Papers of J. L. Autin. London: Oxford University Press, 1 96 1 . Bach, E. ( 1 964) . "Subcategories i n transformational grammars." I n H. Lunt (ed.), Proceedings of the Ninth International Congress of Linguists. The Hague: Mouton S Co. Bar-Hillel. Y. ( L 954) . "Logical syntax and semantics. " Language, 3 0, pp. 230237. ¬ ( 1 960). "The present status of automatic translation of lan guages." In F. L. Alt (ed.), Advances in Computers, Vol. I, pp. 91 - 1 63. New York: Academic Pres. ¬ , A. Kasher, and E. Shamir ( 1 963) . Meaures of Syntactic Com plexity. Report for U.S. Ofce of Naval Research, Inforation Sys tems Branch. Jerusalem. Beattie, J. ( 1 788). Theor of Language. London. Bever, T. G. ( J 963). "The e-o Ablaut in Old English. " Quarterly Progress Report, No. 69, Research Laboratory of Electronic, M.LT., pp. 203-207. ¬ , and T. Langendoen ( 1 963). "The reciprocating cycle of te IndO-European e-o Ablaut. " Quarterly Progress Report, No. 69, Re search Laboratory of Electronics, M.LT., pp. 202-203. -- , and P. Rosenbaum (forthcoming). Two Studies on Syntax and Semantics. Bedford, Mass. : Mitre ' Corporation Technical Re ports. Bloch, B. (1 950) . "Studies in colloquial Japanese IV: Phonemics." Lan guage, 26, pp. 86-1 l5. Reprinted in M. Joos (ed.), Readings in Linguistics. Washington, 1 957. Bloomfeld, L. (1933). Language. New York: Holt. 23 7 BIBLIORHY Blomfeld, M. ( 1 963). "A gamatical approach to personifcation allegor." Modem Philolog, 60, pp. 161 -1 7 1 . Bolingr, D. L. (1 961). "Syntactic blends and oter matters." Language, 3 7. pp. 366-381 • Breland, K., and M. Breland ( 1 961 ). "The misbehavior of organism." American Psychologist, I6, pp. 681-84. Chomsky, N. ( 1 951 ) . Morphophonemics of Modem Hebrew. Unpub· lished Master's tesis, Univerity of Pennsylvania. -( 1 955). The Logical Structure of Linguistic Theor. Mimeo· gaphed, M.!.T. Library. Cambridge, Mass. -(1956). "Three models for the description of language. " I.R.E. Transactions on Information Theor, Vot. IT'2, pp. 1 1 3-1 24. Re printed, wt corrections, in R. D. Luce, R. Bus, and E. Galanter (eds.). Readings in Mathematical Pscholog, Vot. 11. New York: Wiley, 1 965. -(1 957). Syntactic Structures. The Hague: Mouton 8 Co. -( 1959a). "On certan foral properties of grammars:' Informa- tion and Control, 2¿ pp. 1 37-1 67. Reprinted in R. D. Luce, R. Bush. and E. Galanter (eds.), Readings in Mathematical Pscholog, Vol . 11. New York: Wiley, 1 965. -(1959b). Review of Skinner ( 1 957). Language, 3 5, pp. 26-58. Re printed in Fodor and Kau (1 964). -(1961). "Some metodological remaks on generative grammar:' Word, I7, pp. 2 1 g239. Reprinted in part in Fodor and Katz ( 1 964). -(1962a). "A tansforational approach to syntax:' In A. A. Hill (ed.), Pmceedings of the I958 Conference on Problems of Linguistic Analysi in English, pp. 124-1 48. Austin. Texas. Reprinted in Fodor and Kau (1964). -( 1 96ab). "Explanatory models in linguistics:' In E. Nagel, P. Suppes, and A. Tarki, Logic, Methodolog and Philosophy of Science. Stanford, California: Stanford University Press. -(1 963). "Formal properties of grammars:' In R. D. Luce, R. Bush, and E. Galanter (ed.). Handbook of Mathematical Psychol og, Vol. 11. pp. 323-41 8. New York: Wiley. -(1 g64). Current Issues in Linguistic Theor. The Hague: Mouton 8 Co. A slightly earlier version appear in Fodor and Katz ( 1 964). Tis is H revised and expanded version of a paper presented to te session "The logical basis of linguistic teory." at the Nint Inter national Cngress of Linguist, Cambridg, Mass., 1 962. It appear under te title of the session in H. Lunt (ed.), Proceedings of te Congress. The Hague: Mouton 8 Co., 1 964. -(in press). "Topics in the theory of generative grammar." In T. A. Sebeok (ed.). Current Trends in Linguistics. Vot. Ill. Lin guistic Theor. The Hague: Mouton 8 C. BIBLIORHY 1 �9 ¬(forthcoming). "Cartesian Linguistic." ¬ M. Halle, and F. Lukof (1956). "On accent and junctur in English." In M. Halle, H. Lunt, and H. MacLean (eds.), For Roman ]akobson, pp. 65-80. The Hague: Mouton & Co. ¬, and G. A. Miller (1 963). "Introuction to the foral analysis of natua languag." In R. D. Luce. R. Bus, and E. Galanter (e.), Handbook of Mathematical Pscholog. Vol. II, pp. 269-322. New York: Wiley. ¬, and M. P. Schitzenbergr (1963). "The algebraic teor of con text-free langags." In P. Brafort and D. Hirsberg (e.), Com puter Programming and Formal System, pp. 1 191 61 , Studies in Logic Series. Aster: North-Holland. Cordemy, G. de ( 1 66 7). A Philosophicall Dicourse Concerning Speech. Te Englis translation is dated 1 668. Cudworth, R. ( 1 731 ). A Treatise Concerning Eternal and Immutable Moralit. Edited by E. Chandler. Curr, H. B. (1961 ). "Some logical aspects of gmaticl stcture." In R. Jakobson (ed.), Structure of Language and Its Mathematical Aspects, Proceedings of the Twelfth Symposium in Applied Mathe matics, pp. 56-68. Providence, R. I.: American Mathematical So ciet. Desrtes, R. ( 1 641 ). Meditations. ¬(1647). "Notes directed against a certain programme. " Both works by Descartes translated by E. S. Haldane and G. T. Ros in The Philosophical Works of Descartes, Vol. I. New York: Dover, 1 955- Didert, D. ( 1 751). Lettre sur les Sourd et Muets. Pag references are to J. Asswt (ed.), Oeuvres Completes de Diderot, Vol. I (1 875). Paris: Garnier Freres. Dixon, R_ W. (1 963). Linguistic Science and Logic. The Hague: Mouton & Co. Du Marais, C. C. (1 729). Les vhitables principes d la grammaire. On the dating of this manusipt, see Sahlin (1 928), p. i. ¬( 1 76g). Logique et principes de grammaire. Fillmore, C. J. (1 963). "Te psition of embedding transforations in a grmmar." Word, I9, pp. 208-231 . Fodor, J. A., and J. J. Katz (eds.) (1 964). The Structure of Language: Readings in the Philosophy of Language. Englewood Clif, N. J. : Prentice-Hall. Fot P. (1961). "Godness and choice. " Proceedings of the Aristotelian Society, Supplementar Volume 3 5, pp. 45-80. Frasr, B. (1 963). "The position of conjoining transforations in a grammar." Mimeographed. Bedford, Mass. : Mitr Corpration. -(forthcoming) "On te notion 'derived constituent structure.' " BIBLIORPHY Proceedings of the 1964 Magdeburg Symposium, Zeichen und Sys tem der Sprache. Frishkopf, L. S., and M. H. Goldstein ( 1 963). "Responses to acoustic stimuli from single units in the eight nere of the bullfrog." Journal of the Acoutical Society of America, J5, pp. l ugu28. Ginsburg, S., and H. G. Rice ( 1 962). "Two families of languages related to ALGOL. Journal of the Association for Computing Machiner, 1U¿ pp. 350371 . Gleason, H. A. ( 1 961 ) . Introduction to Descriptive Linguitics, second edition. New York: Holt, Rinehart S Winston. ¯ (1 964). "The organiztion of language: a stratifcational view." In C. I. J. M. Stuart (ed.), Report of the Fifteenth Annual Round Table Meeting on Linguitics and Language Studies, pp. 75-5. Washington, D. C.: Georgetown Univerity Pres. Greenberg, J. H. ( 1 963). "Some universals of gammar wit particula reference to te order of meaningful element." In J. H. Greenberg (ed.), Universal of Language, pp. 580. Cambridge: M.I.T. Press. Gleitman, L. (1 96 1). "Conjunction with and:' Transformations and Dis course Analysis Projects, No. 40, mimeographed. Philadelphia: Uni verity of Pennsylvania. Gross, M. ( 1 964). "On te equivalence of models of language used in the felds of mecanical translation and inforation retrieval." Informa tion Storage and Retrieval, 2, pp. 43-57. Hall, B. (1964). Review of Saumjan and Soboleva (1 963). Language 40, pp. 397-4 10. Halle, M. (1957). "In defense of the number two." In E. Pulgram (ed.), Studies Pesented to Joshua Whatmough. The Hague: Mouton S Co. ¯ ( 1959a). "Questions of linguistics. " Nuovo Cimento, IJ, pp. 494- 5 1 7. ¯( 1 959b). The Sound Pattern of Rusian. The Hague: Mouton S Co. -- ( 1 961 ). "On the role of the simplicity in linguistic description." In R. Jakobson (ed.), Structure of Language and Its Mathematical Aspects, Proceedings of the Twelfth Symposium in Applied Mathe matics, pp. 894. Providence, R. I. : American Mathematical Societ. ¯ (1g62a). "Phonology in generative grammar." Word, 18, pp. 54- 72. Reprinted in Fodor and Katz ( 1 g64). -- (1 962b) . "A descriptive convention (or treating assimilation and dissimilation." Quarterly Progress Report, No. 66, Researc Labora tory of Electronics, M.LT., pp. 295-296. -( 1 964) . "On the bases o( phonolog." In Fodor and Katz ( 1 964). BIBLIORPHY ¯, and N. Chomsky (1 960). "The morphophonemics of English." Quarterly Progress Report, No. 58, Research Laboratory of Elec tronics, M.I.T., pp. 275-281 . -(in preparation). The Sound Pattern 0/ English. New York: Harper S Row. ¯ , and K. Stevens (1962). "Speech recognition: a model and a pro gam for research." I.R.E. Transactions in Information Theor, VoI. IT¯8, pp. 1 55 - 1 59. Reprinted in Fodor and Katz (1964). Harman, G. H. (1 968). "Generative grammars without transformational rules: a de£ense of phrase strcture." Language, 3 9, pp. 5 97- 1 6. Haris, Z. S. (1951). Method in Structural Linguistics. Chicago: Univer- sity of Chicago Press. ¯ ( 1 952). "Discourse analysis." Language, 28, pp. 1 828. ¯ (1 954). "Distributional structure." Word, IO, pp. 1 46-162. -- ( 1957). "Coccurrence and transformation i n linguistic struc- ture. " Language, 33, pp. 29 8 -840. Held, R., and S. J. Freedman ( 1 968). "Plasticity in human sensorimotor contrl. " Science, I12, pp. 455-462. -- , and A. Hein (1 968), "Movement-produced stimulation in the development of visually gided behavior." Journal of Comparative and Physiological Psycholor, 56, pp. 872-76. Herbert of Cherbury (1 624). De Veritate. Translated by M. H. Carre ( 1 987). University of Bristol Studies, No. 6. Hiz, H. (1961). "Congammaticality, batteries of transformations and gammatical categries." In R. Jakobson (ed.), Structure of Lan guage and Its Mathematical Aspects, Proceedings of the Twelfth Symposium in Applied Mathematics, pp. 48-50. Providence, R.I. : American Mathematical Society. Hockett, C. F. (1 958). A Course in Modern Linguistics. New York: Mac millan. -- ( 1 961 ). "Linguistic elements and their relations. " Language, 3 7, pp. 2 9 - 5 8 . Hubel, D. H., and T. N. Wiesel ( 1 962). "Receptive felds, binocular in teraction and functional architecture in the eat's visual cortex." Journal of Physiolog, I6U, pp. 1 06-1 54. Hull, C. L. (1 948). Principles of Behavior. New York: Appleton-entury Crofts. Humboldt, W. von. (1 886)ø Ober die Verschiedenheit des Mensch lichen Sprachbaues. Berlin. Hume, D. ( 1 748). An Enquir Concerning Human Understanding. Jakobson, R. (1 986) . "Beitrag zur allgemeinen Kasuslehre." Travaux du Cercle Linguitique de Prague, 6, pp. 240288. Jesperson, O. (1924). Philosophy of Grammar. London: AlIen S Unwin. BIBUORPHY Katz, J. J. ( 1964a). Semimentence- In Foor and Katz (1964). ¬( 1 964b). "Analyticity and contradicton in natural languag." In Fodor and Kat (1 964). ¬( 1964C). "Mentalism in linguistc." Language, 40, pp. 1 24-1 37. "(1 964d. "Smantc teor and te meaning of 'good. ' ' ' Journal of Philosophy. ¬(fortcomng). "Innate ideas. *, and J. A. Fodor. "The structure of a semantic theory. " Lan· guage, J9, pp. 1 702 10. Reprinted in Fodor B Katz (1964). *, and J. A. Fodor (1 964). A reply to Dixon's 'A trend in seman tics.' " Linguistics, 3 , pp. 19-29. -, and P. Postal (1964). An Integrated Theor of Linguistic Descrip- tions. CambIidge, Mass.: M. I.T. Press. Kima, E. S. ( 1 964). "Negation in Englis." In Fodor and Katz (1 964). Lncelot, C., A. Arnauld, et al. ( 1 660). Grammaire generale et raionnee. Lees, R. B. ( 1 957). Review of Chomsky ( 1 957 ). Language, 33 , p p . 3 75 - 4 7 · -(1 960a). The Grammar of English Nominaliztions. The Hague: Mouton B Co . -(1960b). "A multiply ambiguous adjectival construction in Eng· lish." Language, 3 6, pp. 207-221 . -(1961 ). "Grammatical analysis of te English comparative con struction." Word, 17, pp. 1 7 1 -1 85. -, and E. S. Klima (1 963). "Rules for English pronominalization. " Language, 3 9, pp. 1 7¬X8. Leibniz, G. W. New Esays Concerning Human Understanding. Trans lated by A. G. Langley. LaSalle, Ill.: Open Court, 1 949. Leitzmann, A. ( 1 908). Briefwechsel %wischen W. von Humboldt und A. W. Schlegel. Halle: Niemeyer. Lemmon, W. B., and G. H. Patteron ( 1 964). "Depth perception in sheep. Science, 145, p. 835. Lenneber, E. (1 960). "Languag, evolution, and purposive behavior." In S. Diamond (ed.), Culture in Histor: Esays in Honor of Paul Radin. New York: Clumbia UniveIsity Press. Reprinted in a re vised and extended version under te title "The capacity for lan guage acquisition" in Fodor and Katz (1 964). ¬(in pICparaÎion). The Biological Baes of Language. Lettin, J. Y., H. R. Maturana, W. S. McCulloh¿ and W. H. Pius (1 959). ''Wat te fIogs eye tells the fro's brain." Proceeding of the I.R.E., 47, pp. 1 9401 951 . Lura, A. R., Hnd O. S. Vinogadova ( 1959). "An objective investigation of the dyndic of seantic systems- Britih Journal of Pscholog, 5 0, p p . 89105. Mattews, G. H. (1964). Hidatsa Syntax. Te HagueÎ Mouton B Co. BIBLIORPHY Matthews, P. H. ( 1 961 ). "Tranformational gHa." Archivum Lin guiticum, IJ, pp. 1 9 6- 209. Miller, G. A. , and N. Chomky ( 1 963). Finitar models of langag users." In R. D. Luce, R. Bu, and E. Galanter (ed.), Handbook of Mathematical Psycholog, Vol. 11, C. 1 3, pp. 41 9492 . New York: WHey. ¯ , E. Galanter, and K. H. Pribram (1960). Plan and the Structure of Behavior. New York: Henry Halt. ¯ , and S. Isard ( 1963). "Some perceptual consequences of linguistic rules." Journal of Verbal Learning and Verbal Behavior, 2, No. 3, pp. 21 7-228• ¯, and S. Isard ( 1 964) . "Free recall of self-mbedded English sen tences." Information and Control, 7, pp. 29t-303' -, and D. A. Noran ( 1 964). Research on the Use of Formal Lan guages in the Behavioral Sciences. Smi-annual Technical Reprt Department of Defense, Advanced Research Project Agency, Janu ary-June 1 964, pp. 101 1 . Cambridg: Harvard Univerity, Center for Cogni tive Studies. ¯ , and M. Stein (1 963). Grammarama. Scientifc Report No. CS-t, December. Cambridge: Harvard Univerity, Center for Cogitive Studies. Oman, U. ( 1 964). Nominal Compound in Modem Literar Hebrew. UnpOblished doctoral dissertation, Jerualem, Hebrew Univerity. Paul, H. ( 1 886). Prinzipien der Sprachgeschichte, second edition. Trans lated into English by H. A. Strong. London: Lngans, Green B Co._ 1 891 . Peshkovskii, A. M. ( 1 956). Russkii Sintaksi v Nauchnom Osveshchenii. Moscow. Postal, P. M. (1 96ta). Some Syntactic Rules in Mohawk. Unpublished doctoral dissertation, New Haven, Yale University. ¯ ( 1 962b). "On te limitations of context-free phrase-structure description." Quarterly Progress Report No. 64, Researc Labora tor of Electronics, M. I.T., pp. t3 1-238. ¯ ( 1 964a). Constituent Structure: A Study of Contemporar Models of Syntactic Description. The Hague: Mouton B Co. -- ( 1 964b). "Underlying and superfcial linguistic structure." Har vard Educational Review, J4, pp. t46-266. -- ( 1 964C). "Limitations of phrase stucture gd ars." In Fodor and Katz ( 1 964). Quine. W. V. ( 1 960). Word and Object. Cambridg. Mas. : M. I.T. Pres, and New York: Wiley. Reichling. A. ( 1 961 ). "Principles and methods of syntax: cryptanalyti cal foralism." Lingua, 1O¿ pp. 1-1 7. Reid, T. (1 785). Esays on the Intellectual Powers of Man. Page refer- 244 BIBUORPHY ences are to the abridged edition by A. D. Woozley, 1 941 . London: Macmillan and Co. Rosenblom, P. ( 1 95O). The Elements of Mathematical Logic. New York: Dover. RusselI, B. (1 940). An Inquir into Meaning and Truth. London: AlIen S Unwin. Ryle, G. (193 I ). "Systematically misleading expressions. " Proceedings of the Aritotelian Societ. Reprinted in A. G. N. Flew (ed.), Logic and Language, frst series. Oxford: Blackwell, 1 951 . ¯(1 953). "Ordinary language." Philosophical Review, 62 , pp. 167- 1 86. Sahlin, G. ( 1 918). Cesar Chesneau du Marsais et son role dans /evolu tion de la grammaire generale. Paris: Presses Universitaires. Saumjan, S. K., and P. A. Soboleva (1 9 63)ø Applikativnaja porozda juJcaja model' i icilenie trsformacij v Tskom jazke. Moscow: Izdatel'stvo Akademii Nauk SSSR. Schachter, P. ( I 96X). Review: R. B. Lees, "Grammar of English nominal izations. " International Journal of American Linguistics, 28, pp. 1 34 - 1 45. Schlesinger, I. (1 964). The Infuence of Sentence Structure on the Reading Process. Unpublished doctoral dissertation, Jerusalem, Hebrew University. Shamir, E. ( 1 961 ). "On sequential grammars. " Technical Report No. 7, O.N.R. Information Systems Branch, November 1 961 e To appear in Zeitschrift fur Phonetik, Sprachwissenschaft and Kommunikations forschung. Skinner, B. F. (I957)e Verbal Behavior. New York: Appleton-Century Crofts. Smith, C. S. ( 196I ) . "A class of complex modifers in English." Language, 3 7, pp. 341 - 865. Stockwell, R., and P. Schacter (1 961). "Rules for a segment of English syntax." Mimeographed, Los Angeles, University of California. Sutherland, N. S. ( 1 959). "Stimulus analyzing mechanisms." Mechaniza tion of Thought Processes, Vol. Il, National Physical Laborator Symposium No. 10, London. ¯ (1 964). "Visual discrimination in animals." British Medical Bulletin, ×Ü) pp . 54 - 59. Sweet, H. (191 8)e Collected Papers, arranged by H. C. Wyld. Oxford: Clarendon Pres. TwaddelI, W. F. (1 985) . On Defning the Phoneme. Language Mono graph No. I6. Reprinted in part in M. Joos (ed.), Reading in Linguistics. Washington: 1957. Uhlenbeck, E. M. (1 963). "An appraisal of transforation theor." Lingua, I2, pp. 1-1 8. BIBLIORPHY 245 ¯ ( 1 964) . Discussion in te session "Logical basis of linguistic theory." In H. Lunt (ed. ), Proceedings of the Ninth Congress of Linguists, pp. 981-983. The Hague: Mouton B Co. Ullmann, S. ( 1959). The Principles of Semantics. Second edition. Glas gow: J ackson, Son B Co. Vauglas, C. F. de ( 1 647). Remarques sur la langue Franfaise. Facsimile edition, Paris: Librairie E. Droz, 1 934. Wilsn, J. C. (1 926). Statement and Inference, Vol. I. Oxford: Clar endon Press. Wittgnstein, L. ( 1 953). Philosophical Investigations. Oxford: Black well' s. Ynvge, V. ( 1 960) . "A model and a hypothesis for language structure." Proceedings of the American Philosophical Society, IO. pp. 444-466. Zierer, E. ( 1 964). "Linking verbs and non-linking verbs. " Languaje " Ciencias, 12y pp. 1 3-20. Zimmer, K. E. ( 1 964). Afxal Negation in English and Other Languages. Monograph No. 5, Supplement to Word, ×Ü- Index Acquisition moel, see Langage learing, acuisition moel Adequacy, empirical descptive. 27. 3038 • 4, 41 • 5 9 62, 76. 78. 115. 150. 157. 203. 205.20,20.211.227 explanatory. 25-27. 3(8, 40• 41• 4, 46• 5g2. 78• 93. 97. 203. 206, 231. 234 Adverbials, 101-105. 191. 215-219 Analyzability. 56. 98, 121, 122. 143, 144, 147. 217,225, 226 Anderson. T., In3 Assimilation, rule of. 175,133 Bac. E .• 113 Base component general structure. 84-, 98, 9. 120. 12 3 . 128. 136. 1 3 7. 140. 141 ilustrated. 10111 transformational rles in, 122 Base rules. types of. 111-127 Basis of a sntence, 17. 128. 130 Beattie, James. 5 Behavioris, 193. 194.104.20 Bever. T. G., 16. 228. 131. 133 Bloomfeld, L .• 105, 114 Bolinger. D. L., 235 Boundary symbol, 6 Cas. III Caterie grammaticl. 64-69. 86. 14 lexical. 74. 82. 84. 115. 116, 16, 212. 113, 119 subcategoriztion of, 79 major. 74. 1. 115. 116.212 Competence. 3. 4. 6. 8-10. 15. 16. 18. 19. 11. 24. 27, 32. 34. 140• 193 and perorance. 10. 25. 139. 197 Complex symbols. 81-, 95. 98. lOl. 1 10. Ill, 121. 122. 188. 189, 119. 233 Compnent bas, see Bas component cateoria. 120. 122-126. 128. 134. 136. 141 transforational. 140. 142 see alo Grammar. generative Concateation·systems. 114. 115 Constructions cusatve. 189 comparative. 178-180. 182. 183,233. 234 coordinated. 196, 198, 212. 224. 225 left-branching. 12-14. 196198 mUltple-branching. 12-14, 19. 197 nested. 12-14. 198 right· branching. 12-14. 19. 197 self·embedded, 12-14. 196198 Cooccurence, 95. 220, 229 Cudworth, R .• 49 Curr. H. B., 221 Deep strcture. 1618. 23. 24. 29. 70. 99. 118, 120. 123. 162. 163. 178. 199, lO. 220221. 224. 119 aspect of, 64-106 248 Deep structure (continued defned, 136, 198, 199 and grammatical transformations, 128-147 Derivation, 66, 85, 88, 98, 142 sequential, 6 transformational, 135, 143 Descartes, R., 48, 20, 203 Dictionary defnitions, 16 Diderot, D., 7 Direct-object, logical, 23 Distinctive-feature tery, 28, 55, 65, 80,81,87 Distinctness of feature matrices, 81, 84,85, 110, 123, 181, 182,220 Dixon, R. W., 194 Du Marsais, C. Ch., 5, 200 Dummy symbol, 122, 132, 144, 222 Empiricis m, 47, 51 - 54, 58, 59, 203, 206,207 Etholog, 57, 206 Evaluation measures, 32, 34-47, 61, 97, 111, 164, 16, 203,211, 226, 231 and notational conventions, 42-46 Exceptions, 192, 218, 231 Feature contextual, 93, l l l, 121, 123, 119, 148, 151, 154, 156, 165, 229, 235 distinctive or phonological, 142, 213, 214, 230, 232 selectional, 122, 148, 164, 165,230 smantic, 88, 1I0, 120, 142, 154, 164, 19, 198, 214 strict subcategorization, 148, 164, 165 syntactic, 75- 9, 82 - 87, 95, 97, ll O, 112, 113, 120, 142, 153, 154, 164, 171, 172, 175, 190 , 214- 216,227, 233, 235 Field properties, 160, 161, 229 Firthian linguistics, 205 Fodor, J. A., 154, 161,214 Foratives, 3, 14, 16, 65, 85, 143, 144, 181, 230, 233 grammatical, 65, 122 INDEX lexical, 65, 6, 74, 79, 81, 82, 84, 86, 87 see also Lexical entries Functional notions, 68-4 Functions, gammaticl, 23, 68-4, 86, 113, 126, 136, 141, 162, 220, 224 Generative capacity strong, 39, 62, 8, 99, 208 weak, 62,9°,98, 20 Grammaire genera le et raisonnee, by Lncelot et al., 6, 117, 118, 137, 199, 221 Grammar, generative bas cmponent in, 105- 111 defned,4,8,9,61,2D delimitation of, 35 frst attempts at, 79 organization of, 15-18, 63 in performance models, 10, 15 phonological component of, 16, 28, 29,35,40,45,75,80,81,88,89, 135, 141, 143, 175, 198 smantic cmponet of, 16, 28, 75, 77, 78, 88, 132 , 135, 141 , 146, 153, 154, 157, 158, 159, 16 164, 198 projection rles of, 144 syntactic component of, 3, 16, 17, 28, 78, 79, 88, 89, 117, 135, 136, 139, 141, 145, 151, 154, 157- 159, 198 ba of, 17, 63 defned, 141 transforational component of, 17, 132 Grammar, particular, 5- opposed to uiversal gammar, 6 Grammar, philophical, 5 Grammar and philosphicl error, 199, 20 Grammar, types of constituent structure, 67, 122 context-free, 61, 6, 139,208,215 context-sensitive, 61, 215 selectional rules, 95"7 subcategoriztion rules, g106, ll3 fnite-state, 208 INDE Grammar, types of (continued) phrase structure, 61, & 88-o, 98, 99, 122, 116, 119, 140, 205, 110, Ill, 211 squential, 211 stratifctional, 221 structuralist, 5, 6 traditional, 5, 6, 8, 61, 64, 71, 110, 172, 194, 221 transforational, 16, 54, 59, 70, 89, g, 98, 122, 198, 211, 217 theory of, 55, 134, 116, 117, 141, 208 Grammar, universal, 5-7, 28, 65, 115- 118, 141, 142, 144, 211 Grammars justifcation of, 18-27, 11, 40, 41 recursive property of, 116, 117, 142, 225 simplicity of, 17, 40 Grammaticalness, 1, 11, 19, 75-79, 88, 195, 212, 211, 227,228 dee of, 148-151 Gric, A. P., 224 Hall, Barbara, 216, 228 Halle, M., 45, 212 Harman, G. H., 210 Held, Ricard, 11 Herbert of Cherbury, 49 Humboldt, W. von, 4, 8, 9, 51, 198, 199, 205, 20 Hu· me, David, 51 Iconic eleents of discourse, 11, 225 Ideas, theory of, 199, 20, 201 Immediate cnstituent analysis, 17, 205 Infection processs, 170184 systems, 171, 174, 176,212 Inner form, 198 Inversion, stylistic, 222, 221, 228 Jakobson, R., 28, 55, 224 Katz, J. J., 112, 115, 154, 161,211, 11 4 Kerel sentences, 17, 18 Lancelot, C., 6 2 4 9 Language learing, 25, 27, 28, 16• 17, 41, 4 5 -47, 51, 5 4 , 57. 58, 20, 201, 201, 206, 207 acquisition model, 1011. 15, 18, 47, 51-58, 117, 202, 201, 207 and linguistic theory, 47-59 Language use, creative aspect of. 6, 8, 57, 58, 116, 205 Languages, artifcial, 116, 140 Lngue·parole, 4 Learing theory, 204 Leibniz, G., 4952, 201 Lexical entries, 87, 122. 19. 214 see also Formatives Lexical ru l e, 84, 85, 87, lIO, 112, 121, 121, 186, 188 Lexicon, 84-88, 9, g, 107, lIO, 112, 120, 122, 121, 116, 141, 142. 154, 19 B , 214 structure of, 164-192, 219, 222 Linguistic data, primary. 25, 1 0 15. 17, 18, 46, 47, 201, 201. 205, 207, 2oS, 226 Linguistic thery, 1-, 9. 201 Linguistics mathematical, 62, 20S structural, 16, 47, 51, 52, 54, 57, 88, 172-174, 202, 205, 208-210, 221 Locke. John, 49. 201 Long components, 211 Main·verb, 71, 72 Matthews, G. H., 79, lUl Memory fnitenes of, 14, 197 organization of. 14, 19 Mentalism, 4. 193, 204, 20 Models acuisition, 25 performance, 9, 15, 19 see alo Perceptual model Morphological properties. 87 Natural order, 6, 7 Nesting, 197, 198 Notational conventions, 42-45 Nouns, proper, 217 Object words, 201 25° Opertional proedure. teted. 1!21 Operationalism, 194 Ordering of rule. 39. 40• 6. 107. 123. 133. Ill. 223 Outer or exteral for, 199 Perceptual model. 9. 13-15. 51. 136• 1 9 4. 201, 207 Perforance. 3. 4. 9, 18. 127 theory of. 10- 15 see also Models. perorance Phones. 16 Phonetic iterpretation, 135, 136, 141• 197 Phonetic. universal. 28 Phonologicl admisibility, 45. 169, 231 Phonological interpretation, 186 Phonology. 793 Phrase-marker. 65-7, 84. 85, 88, 1l0, 117. 124. 131. 222 base, 17. 18. 65. 70. 72, 125, 128, 130, 131• 134-136 • 139, 142, 143. 235 derive. 104, 128, 131. 144 generalized, 13 4 . 135 , 137-140, 142, 143 , 177, 227 Plato. 24, 51 Post, E .• 9 Postal, P. M., 127. 132, 135, 154, 210 Predicate, 71, 1, 211, 11 Procsss derivational, 184-192 q uasi-productive, 186. 235 Proper name, 201 Psycholog. physioloicl. 205 Quantifers. 224 Quine, W_ V., 51. 203, 1 Rationalism. 48- 5 4. 205-20 7 Redundac, 164-170 Refexiviztion, 145. 146 Reid. Thomas, 199. 20 Relations, grammatical. 78, 74. 99. 113-120. 123-125, 136, 141, 144. 162. 163, 220. 221 slection aI, 74. 113 Relative clause. 137, 138. 145. 177, 180 retrictive, 107 INDE Represntation. levels of. 212 Restriction. slectional, 95. 139. 216 strict subcategorial. 139 Rosenbaum, P .• 16. 20, 228 Rule shemata. 9. 224. 225 Rules branching, 112. 120, 122, 134. 135, 136. 188 cteorial, 113-127. 142 context-fee, 112, 114. 120. 121 context-sensitive. 112, 120, 121 projection. 136, 144. 154. 157 redundanc. 121. 142. 182. 215. 222 phoologcl. 16170. 214, 231 syntactic. 1 68-170 rewriting. 66, 6, 70 , 72, 74, 7 9 , 8 4, 86, 98 , 111, 112, 119. 123. 134. 142, 154. 155 context-free. gf. 120. 121. 141 context-sensitive. 91, 188 selection al. 95"97. 99. Ill. 113- 120, 123. 14916, 227 context-sensitive, 114 subcategorization. g- l06. 113. 120-123 context-free. 121 context-snsitive, 219 ecluded from ctegorial com ponent, 128 strict, 95-10. 10 3 . 105. lo, Ill. 112. 120. 123. 149. 150• 15 l . 1 55, 157, 158, 106 , 229 trasforational, 87. 89. g, g. 117, 1l2. 1l8. 141. 217 Rues of agreement. 174. 175. 179. 180 Ryle. G_, 8 Sahlin, G .• 5 Sapir. E., 191 Saumjan, S. K .• 221 Sausure, F. de .• 4, 8, 4 7 Schacter. P., 211 SchlegI. A. W_, 20 Smantic iterretation, 70. 71, 75. 7 9, 87, 88 . 99. 1 17, 132 • I l5. 136, 138, 140• 1 41, 1 4 4. 151, 159, 161. 177. 186. 197. 220, ll. 22l. 224. U9 INDE Sentences acceptability of, 10, 11, 195, 19 and perforance, 11-14 cnstituent, 182, 188, 185 deviant, 10, 77, 141, 148, 157, 217, 230, 135 matrix, 132, 13!, 185 Seuenc, hierarcic, 165, 166 Set-systes, 124-126 Simplicity measure, see Evaluation measures; Grammars, sim plicity of Sinner, B. F., 51, 204 Soboleva. P. A., 221 Stocwell, R., 218 Strings basic. 1 7. 18 preterminal, 84, 85, 88, 122, 141 terial, 66, 84, 14! Strcture index, 121, 143, 144, 147, 156 Strcture. inate, 25-27, !a. !2-84. 36, 37. 48-4, 58, 59, 117. 16, 101, 104, 20. 20 7 Subject gaaticl. 23. 70. 163 loic. 13. 70. 1 6 3. III Surfac structure. 1&18. 13. 14, 29, 70, 11 8. 123. 11 5. 118. 1!1. 135 . 138• 140. 141• 162. 16!. 199. 111.114 defned. 198. 19. 220. 121 Synta. semantic basis fr. 32. !!. 7 5 . 78. 203, 114. 220, 226 Systems of belief, 16 Topicmment. Ill. 225 Trsformation-mare. 1!1. 1 8 2. 1 ! !. 134. 135. 18 6 • 1 87. 139. III Transforational cycle. 29. !5. 89. 14!. 186. 212. 115. 2!8 Transforational histor. 130 Transforations elementar. 14!. 14. 147 gammaticl. 55. 6!. 89. 125. 127- 147. 175 deletion. 138. 144. 177. 179181. 208. 222. 2!4 erasure. 145. 146. 177. 179. 182. 191. u6 flterig function of, 189. 141. 191. 224 generalized, 132-137. 224 local, a. 99. 215 pasive. 108-106. 223 relative clause. 234 s g la r . 13! - 1!5. 13 9 , 142 strictly local. 10. 101. 105. 10. 118. 122. 215 and meaning. 132. 225 Translation. 80 Turing machines. 62 Twaddell. W. F .• 19!. 194 Uhlenbec. E. M .• 194 Universals. lingistic, !5. 36• 41• 55. 73. 118. 181, log. 210 fral. 17-!o. !6. 46 • 53. 55. 20 7 , 2!1 phonetic, 16 semantic. 28. 77. 160.201 substantive. 27-80. 4 6 • 5!. 55. 65. 73. 117. 207. 231 Usage, statistics of. 195, 211 Verfcationism. 194 Wilson. J. C .• 1 63 Wittgenstein. L .• 51 Word.I!5 Word order. 126 Yngve. V .• 19'. 198.113 ASPECTS OF THE THEORY OF SYNTAX ASPECTS OF THE THEORY OF SYNTAX Noam Chomsky THE M.LT. PRESS Massachusetts Institute of Technology Cambridge, Massachusetts 1 1 111 ACKNOWLEDGMENT This is Special Technical Report Number II of the Research Labora tory of Electronics of the Massachusetts Institute of Technology. The Research Laboratory of Electronics is an interdepartmental laboratory in which faculty members and graduate students from numer ous academic departments conduct research. The research reported in this document was made possible in part by support extended the Massachusetts Institute of Technology. Research Laboratory of Electronics. by the JOINT SERVICES ELECfRONICS PROGRAMS (U.S. Army. U.S. Navy. and U.S. Air Force) under Contract No. DAS6o39-AMC-03llOo(E); additional support was received from the U.S. Air Force (Electronic Systems Division under Contract AFI9(628)-2487). the National Science Foundation (Grant GP-2495). the National Insti tutes of Health (Grant MH-D4737-D4). and The National Aeronautics and Space Administration (Grant NsG-496). Reproduction in whole or in part is permitted for any purpose of the United States Government. Copy right @) I965 by The Massachwetts Instit ute Of Technology All Rights Reserved Library of Congress Catalog Card Number: 65-I9080 Printed in the United States of America Preface The idea that a language is based on a system of rules deter mining the interpretation of its infinitely many sentences is by no means novel. Well over a century ago. it was expressed with reasonable clarity by Wilhelm von Humbold� in his famous but rarely studied introduction to general linguistics (Humboldt. 183 6). His view that a language "makes infinite use of finite means" and that its grammar must describe the processes that make this possible is. furthermore. an outgrowth of a persistent concern. within rationalistic philosophy of language and mind, with this "creative" aspect of language use (for discussion, see Chomsky. 1964. forthcoming). What is more, it seems that even Panini's grammar can be interpreted as a fragment of such a "generative grammar," in essentially the contemporary sense of this term. Nevertheless, within modem linguistics. it is chiefly within the last few years that fairly substantial attempts have been made to construct explicit generative grammars for particular lan guages and to explore their consequences. No great surprise should be occasioned by the extensive discussion and debate concerning the proper formulation of the theory of generative grammar and the correct description of the languages that have been most intensively studied. The tentative charac�er of any conclusions that can now be advanced concerning linguistic theory, or, for that matter, English grammar, should certainly be obvious to anyone working in this area. (I� is sufficient to v with ques tions that are at the border of research in transformational gram mar. then. Chapter 11 is concerned with the base of the syntactic component. Lees ( 1 9600). What is at issue here is precisely how this theory should be formulated. This monograph is an exploratory study of various problems that have arisen in the course of work on transformational gram mar. The position discussed is that of Chomsky ( 1 957). the central role of grammatical transformations in any empirically adequate gen erative grammar seems to me to be established quite firmly. though there remain many questions as to the proper form of the theory of transformational grammar. For some. and with difficulties that arise from the assumption that it is. I shall sketch briefly what seems to me. strictly speaking.vi PREFACE consider the vast range of linguistic phenomena that have re sisted insightful formulation in any terms. Chapter 3 suggests a revi sion of the transformational component and its relation to base . but more often the discussion will merely raise issues and consider pos sible approaches to them without reaching any definite conclu sion. but aims only to summarize and to clarify certain points that are essential and that in some instances have been repeatedly misunderstood. and a system of transformations that map structures generated by the base into actual sentences . which is presupposed throughout as a general framework for the discussion. and many others. In particular.) Still. But I should like to reiter ate that this can be only a highly tentative proposal . The monograph is organized in the following way . a phrase structure grammar. Chapter 1 sketches background assumptions. the most promising direction for the theory of generative grammar to take. Chapters 11 and 3 deal with a variety of defects in earlier versions of the theory of transformational grammar. This study deals. These writers take the syntactic component of a transformational grammar to consist of a phrase structure grammar as its base. definite answers will be proposed. This posi tion is restated briefly at the beginning of Chapter 3. it seems that certain fairly substantial conclusions are emerging and receiving continually increased support. in the light of this discussion. In Chapter 3. It contains little that is new. who have sug gested many valuable improvements. I should like to acknowledge wit. and discussed briefly and quite inconclusively. and to many students whose reactions and ideas when this material has been presented have led to quite substantial modifications. H. MH 05120-04 and -05 from the National Institutes of Health to Harvard University. lames McC awley. The writing of this book was completed while I was at Harvard University. George Miller. and in part by a fellowship of the American Council of Learned Societies. supported in part by Grant No. various residual problems are raised. Matthews. In particular. I am indebted to Morris Halle and Paul Postal. The notion of "grammatical transformation" itself is taken over without change (though with some simplifications). In Chapter 4.PREFACE vii structures. as well as to lerrold Katz. and G. Massachusetts October Ig64 . NOAM CHOMSKY Cambridge.h gratitude the very helpful comments of many friends and colleagues who have taken the trouble to read earlier versions of this manuscript. Center for Cognitive Studies. . § 9. THE SCOPE OF THE BASE ASPEcrS OF DEEP STRUcrURE §��. 5.�� § 2. PLANATORY THEORIES ON EVALUATION PROCEDURES LINGUISTIC THEORY AND LANGUAGE LEARNING GENERATIVE CAPACITY AND ITS LINGUISTIC RELEVANCE 30 37 47 60 2 Categories and Relations in Syntactic Theory 63 § 1.2. 2.Contents Preface 1 Methodological Preliminaries § § § § § § 1. 4. § 8. § 2. Functional notions 63 64 68 � ix . GENERATIVE GRAMMARS AS THEORIES OF LINGUISTIC COMPETENCE TOWARD A THEORY OF PERFORMANCE THE ORGANIZATION OF A GENERATIVE GRAMMAR JUSTIFICATION OF GRAMMARS FORMAL AND SUBSTANTIVE UNIVERSALS FURTHER REMARKS v 3 3 3. ON 15 18 27 10 DESCRIPTIVE AND EX- § 7. 6. I.2. Selectional rules and grammatical relations § 4.J.2. 2.J. AND SEMANTICS Degrees of grammaticalness Further remarks on selectional rules Some additional problems of semantic M� 148 I5J 148 § 2. THE BOUNDARIES OF SYNTAX § § § 1. Further remarks on subcategorization rules § 4+ The role of categorial rules 120 I2J 3 Deep Structures and Grammatical Transformations I28 I48 4 Som e Residual Problems § l. 2.1.1.J.J+ Context-sensitive subcategori%tJtion rules 84 90 75 § 3. AN ILLUSTRATIVE FRAGMENT OF THE BASE PONENT TYPES OF BASE RULES COM- III 106 HI HJ § 4.J. 1.I. § § § THE STRUCTURE OF THE LEXICON 2. Some formal similarities between syntax and phonology 79 § 2.2.2 Chapter 3 Chapter 4 � 193 208 222 227 Bibliography 237 247 Index .x CONTENTS § 2.2. Redundancy Inflectional processes Derivational processes 164 164 170 184 � &� Notes Notes Notes Notes to to to to Chapter 1 Chapter.J. Summary § 4. The problem 75 § 2.J.J.1. General structure of the base component § 2.J. Syntactic features § 2. § 4. THEORY OF SYNTAX ASPECTS OF THE . . in a completely homogeneous speech-community. and none exhaustively. This seems to me to have been the position of the founders of modem general linguistics. It will be concerned with the syntactic component of a generative grammar. that is. The general framework within which this investigation will proceed has been presented in many places. and no cogent reason for . making no serious attempt here to justify them but only to sketch them clearly. distractions. GENERA TIVE GRAMMARS AS THEORIES OF LINGUISTIC COMPETENCE THIS study will touch on a variety of topics in syntactic theory and English syntax. and some familiarity with the theoretical and descriptive studies listed in the bibliog raphy is presupposed. I shall survey briefly some of the main background assumptions. In this chapter. Linguistic theory is concerned primarily with an ideal speaker listener. several quite superficially. and errors (random or character istic) in applying his knowledge of the language in actual per formance.I Methodological Preliminaries § I. who knows its language perfectly and is unaffected by such grammati cally irrelevant conditions as memory limitations. with the rules that specify the well-formed strings of minimal syn tactically functioning units (Jormatives) and assign structural information of various kinds both to these strings and to strings that deviate from well-formedness in certain respects. shifts of attention and interest. a few in some detail. it obviously could not directly reflect competence. if it does not rely on the intelligence of the understanding reader but rather provides an explicit analysis of his contribution . A grammar of a language purports to be a description of the ideal speaker-hearer's intrinsic competence. is to determine from the data of performance the underlying system of rules that has been mastered by the speaker-hearer and that he puts to use in actual performance. if this is to be a serious discipline. We thus make a fundamental distinction between competence (the speaker-hearer's knowledge of his language) and performance (the actual use of language in concrete situations). but it is necessary to reject his concept of langue as merely a systematic inventory of items and to return rather to the Humboldtian conception of underlying competence as a system of generative processes. furthermore. If the grammar is. The problem for the linguist. In this respect. The distinction I am noting here is related to the langue-parole distinction of Saussure. study of language is no different from empirical investigation of other complex phenomena. as well as for the child learning the language. since it is concerned with discovering a mental reality underlying actual behavior. To study actual linguistic per formance. and so on. For discussion. In actual fact. changes of plan in mid·course. and so on. see Chomsky ( 1 964). we must consider the interaction of a variety of factors. Hence. of which the underlying competence of the speaker-hearer is only one. Only under the idealization set forth in the preceding paragraph is per formance a direct reflection of competence. perfectly explicit .l Ob served use of language or hypothesized dispositions to respond.we may (somewhat redundantly) call it a generative grammar. linguistic theory is mentalistic. may provide evidence as to the nature of this mental reality. A fully adequate grammar must assign to each of an infinite range of sentences a structural description indicating how this . but surely cannot constitute the actual subject matter of linguistics. habits.4 METHODOLOGICAL PRELIMINARIES modifying it has been offered. deviations from rules.in other words. in the technical sense. A record of natural speech will show numerous false starts. The peculiarities of individual tongues are explained in their respective grammars and dictionaries. yet all have certain qualities in common. are treated of in a science. however informally. resemble men in this respect. Those things. 1 928. However. or that are necessary to every language. Traditional linguistic theory was not unaware of this fact.§ I. James Beattie (1 788) remarks that Languages. ces observations forment ce qU'on appelle la grammaire . and traditional grammars give a wealth of information concerning structural descriptions of sentences. that. An analysis of the best existing grammars will quickly reveal that this is a defect of principle. valuable as they obviously are. it seems obvious that the attempt to explore this largely uncharted territory can most profitably begin with a study of the kind of structural information presented by traditional grammars and the kind of linguistic processes that have been exhibited. Although such grammars may contain full and explicit lists of exceptions and irregularities. they provide only examples and hints concerning the regular and productive syntactic processes. where no traditional or structuralist grammar goes beyond classification of particular examples to the stage of formulation of generative rules on any significant scale. For example. pp. therefore.2 The limitations of traditional and structuralist grammars should be clearly appreciated. Du Marsais defines universal and particular grammar in the following way ( 1 729. though each has peculiarities. whereby it is distinguished from every other. Nevertheless. in these grammars. quoted in Sahlin. GENERATIVE GRAMMARS AND LINGUISTIC COMPETENCE 5 sentence is understood by the ideal speaker-hearer. that all languages have in common. which some have called Universal or Philosophical grammar. Somewhat earlier. traditional grammars are deficient in that they leave un expressed many of the basic regularities of the language with which they are concerned. not just a matter of empirical detail or logical preciseness. This is the traditional problem of descriptive linguistics. This fact is particularly clear on the level of syntax. 29-30) : 11 y a dans la grammaire des observations qui conviennent a toutes les langues. Therefore it is quite proper for a grammar to discuss only exceptions and irregularities in any detail. Modern linguistics. cf. and. characteristically rejected the study of universal grammar as misguided. it was clearly understood that one of the qualities that all languages have in common is their "creative" aspect. Within traditional linguistic theory.6 METHODOLOGICAL PREUMINARlES generale: telles sont les remarques que ron a faites sur les sons articules. being universal. et sur les differentes manieres dont ils doivent �tre ou arranges ou termines pour faire un sens. forth coming). as noted before. has not explicitly recognized the necessity for supplementing a "particular grammar" of a lan guage by a universal grammar if it is to achieve descriptive adequacy. Thus an essential property of language is that it provides the means for expressing indefinitely many thoughts and for reacting appropriately in an indefinite range of new situations (for references. in fact. 1 964. is to be supplemented by a universal grammar that accommodates the creative aspect of language use and expresses the deep-seated regularities which. It thus suggests no way to overcome the fundamental descriptive inadequacy of structuralist grammars. furthermore. it has not attempted to deal with the creative aspect of language use. to attempt a precise statement of regular processes of sentence formation and sentence interpretation lay in the widely held belief that there is a "natural order of thoughts" that is mirrored by the order of words. sur les lettres qui sont les signes de ces sons. It has. particular or universal. are omitted from the grammar itself. It is only when supplemented by a universal grammar that the grammar of a language provides a full account of the speaker-hearer's competence. Chomsky. Thus in the Grammaire genera le et raisonnee (Lancelot . et c'est ce qui forme les gram maires particulieres de chaque langue. il y en a qui ne sont propres qu'a une langue particuliere. however. sur la nature des mots. Outre ces observations generales. The grammar of a particular language. Another reason for the failure of traditional grammars. Hence. the rules of sentence formation do not really belong to grammar but to some other subject in which the "order of thoughts" is studied. then. la latine et les autres seront les langues de la fable et du mensonge. Thus "quel que soit l'ordre des termes dans une langue ancienne ou moderne. l'imagination et les passions donneront la preference aux langues anciennes et a celIes de nos voisins . . it is not necessary to describe it in a particular or universal grammar. whereas Greek. emouvoir et tromper: parlez grec. dans les chaires et sur les theA tres.§ 1. . . and so on. et grec. latin. le grec. la grecque. si jamais elle revient sur la terre. italien au peuple. in an interesting essay devoted largely to the question of how the simultaneous and sequential array of ideas is reflected in the order of words. few gram matical rules need be formulated beyond the rules of ellipsis. 37 1). Diderot concludes that French is unique among languages in the degree to which the order of words corresponds to the natural order of thoughts and ideas (Diderot. 372). l' esprit de l'ecrivain a suivi 1'0rdre didactique de la syntaxe fran�aise" (p. Italian. which determine the figurative use of lan guage. It is worth noting that this naive view of language structure persists to modem times in . . il faut parler fran�ais dans la societe et dans les «oles de philosophie. 37 1 -372) In any event. mais . and we therefore have principled grounds for excluding an explicit formulation of syntactic processes from grammar. the sequence of words follows an "ordre naturel. . . anglais. 1 75 1 ). 1 660) it is asserted that." Moreover. and English "sont plus avanta geuses pour les lettres." which conforms "a l'expression naturelle de nos pensees. latin. With admirable consistency he goes on to conclude that "notre langue pedestre a sur les au tres l'avantage de l'utile sur l'agreable" (p. l'italien. GENERATIVE GRAMMARS AND LINGUISTIC COMPETENCE 7 et al . . le bons sens choisirait la langue fran�aise. pour persuader. l'anglais. Latin. et . . 390) . comme l'esprit est force de les considerer en quelque langue qu'on ecrive" (p. . mais parlez fran�is au sage. thus French is appropriate for the sciences. . eclairer et con vaincre. To mention j ust one additional example. (pp. insofar as the order of words is determined by factors independent of language. notre langue sera ceIle de la verite. le latin. inversion. aside from figurative speech. The same view appears in many forms and variants." Consequently. "Nous disons les choses en fran�ais. Le fran�ais est fait pour instruire. Now that these insights are readily available it is possible to return to the problems that were raised. for example. in short. furthermore. though these state- . a theory of visual perception would attempt to account for what a person actually sees and the mechanisms that determine this rather than his statements about what he sees and why. But the fundamental reason for this inadequacy of traditional grammars is a more technical one. Returning to the main theme. Any interesting generative grammar will be dealing. it is quite apparent that a speaker's reports and viewpoints about his behavior and his competence may be in error. but not solved. by a generative grammar I mean simply a system of rules that in some explicit and well defined way assigns structural descriptions to sentences. every speaker of a language has mastered and internalized a gen erative grammar that expresses his knowledge of his language. not what he may report about his knowledge. 1953 ). no longer a technical barrier to the full-scale study of generative grammars. Obviously. or that his statements about his intuitive knowledge of the language are necessarily accurate. Although it was well under stood that linguistic processes are in some sense "creative.8 METHODOLOGICAL PRELIMINARIES various forms. for the most part. a real understanding of how a language can (in Humboldt's words) "make infinite use of finite means" has developed only within the last thirty years. Thus a generative grammar attempts to specify what the speaker actually knows. There is." the technical devices for expressing a system of recursive processes were simply not available until much more recently. This is not to say that he is aware of the rules of the grammar or even that he can become aware of them. in the course of studies in the foundations of mathematics. Similarly. In fact. in Saussure's image of a sequence of expressions corresponding to an amorphous sequence of concepts or in the common characterization of language use as merely a matter of use of words and phrases (for example. with mental processes that are far beyond the level of actual or even potential consciousness. in traditional linguistic theory. and to attempt an explicit formulation of the "creative" processes of language. Ryle. the theory of per formance. GENERATIVE GRAMMARS AND LINGUISTIC COMPETENCE 9 ments may provide useful. as a basic component. No doubt. The term "generate" is familiar in the sense intended here in logic. For various attempts to clarify this point. Since this use of the term "generate" is well established both in logic and in the tradition of linguistic theory. Never theless. I think that the term "generative grammar" is completely appropriate. the generative grammar that expresses the speaker-hearer's knowledge of the language. it seems. and many other publications. Confusion over this matter has been sufficiently persistent to suggest that a terminological change might be in order. in itself. in some practical or efficient way. particularly in Post's theory of combinatorial systems. prescribe the char acter or functioning of a perceptual model or a model of speech production. When we say that a sentence has a certain derivation with respect to a particular generative grammar.§ 1. see Chomsky ( 1957). and have therefore continued to use it. in essentially the sense here intended. Miller and Chomsky ( 1963 ) . we mean simply that the grammar assigns this structural description to the sentence. compelling evidence for such a theory. I can see no reason for a revision of terminology. "generate" seems to be the most appropriate translation for Humboldt's term erzeugen. we say nothing about how the speaker or hearer might proceed. but this generative grammar does not. a reasonable model of language use will incorporate. It attempts to characterize in the most neutral possible terms the knowledge of the language that provides the basis for actual use of language by a speaker hearer. it is perhaps worth while to reiterate that a generative grammar is not a model for a speaker or a hearer. Further more. Gleason ( 196 1). which he frequently uses. to construct such a derivation. in fact. To avoid what has been a continuing misunderstanding. . When we speak of a grammar as generating a sentence with a certain structural description. These questions belong to the theory of language use . correctness.that is. acceptability will be a matter of degree. and access. let us use the term "ac ceptable" to refer to utterances that are perfectly natural and immediately comprehensible without paper-and-pencil analysis.3 In particular. To my knowledge. have come from studies of performance models that incorporate generative grammars of specific kinds . To illus trate. the only concrete results that have been achieved and the only clear suggestions that have been put forth concerning the theory of performance. normalcy of intonation).10 METHODOLOGICAL PREUMINARIES § 2. The latter question is one to which we shall return in Chapters 2 and 4. the sentences of ( I ) are somewhat more acceptable. along various dimensions. To clarify further the distinction between competence and performance. in the intended sense. and in no way bizarre or outlandish. there are some suggestive observations concerning limitations on performance imposed by organization of memory and bounds on memory. it is unnecessary to delimit it more carefully. recent work on performance seems to give new support to this assumption. and uniformity of recall and recognition. rapidity. One could go on to propose various operational tests to specify the notion more precisely (for example. outside of phonetics. from studies that have been based on assumptions about underlying competence. Obviously. Further more. and concerning the ex ploitation of grammatical devices to form deviant sentences of various types. time. than those of (2) : ( I ) (i) I called up the man who wrote the book that you told me about (ii) quite a few of the students who you met who come from New York are friends of mine ." For present purposes. TO WARD A THEOR Y OF PERFORMANCE There seems to be little reason to question the traditional view that investigation of performance will proceed only so far as understanding of underlying competence permits. For the purposes of this discussion. it may be useful to summarize briefly some of the suggestions and results that have appeared in the last few years in the study of performance models with limita tions of memory. more easily understood. the generative rules of the language assign an interpretation to them in exactly the way in which they assign an interpretation to the somewhat more ac ceptable sentences of ( 1 ) . but the scales of grammaticalness and acceptability do not coincide. Chapter 3 ) . The notion "acceptable" is not to be confused with "gram matical. and several of their friends visited last night (ll) (i) I called the man who wrote the book that you told me about up (ii) the man who the boy who the students recognized pointed out is a friend of mine The more acceptable sentences are those that are more likely to be produced. Bill. for reasons having to do. 1 96 1). Grammaticalness is only one of many factors that interact to determine acceptability. and so on. in the technical sense of this term. it is unlikely that a necessary and sufficient operational criterion might be invented for the much more abstract and· far more important notion of grammaticalness. 1 957. in actual discourse. Note that it would be quite impossible to characterize the un acceptable sentences in grammatical terms. a tendency to place logical subject and object early rather than late. a matter of degree (cf.1i The unacceptable sentences one would tend to avoid and replace by more acceptable variants. whereas grammaticalness belongs to the study of competence. wherever possible. we can not formulate particular rules of the grammar in such a way as . Chapter ll.§ ll. That is. less clumsy. "iconic" elements of discourse (for example. TOWARD A THEORY OF PERFORMANCE 11 us (iii) John. not with grammar. For example." Acceptability is a concept that belongs to the study of · performance. grammaticalness is. but rather with memory limitations. note 311. no doubt. intonational and stylistic factors. although one might propose various operational tests for acceptability. Correspondingly. 1 955. The sentences of (ll) are low on the scale of ac ceptability but high on the scale of grammaticalness. Chomsky. and note 9. Tom. cf. The unacceptable grammatical sen tences often cannot be used. Like acceptability. and in some sense more natural. not to a particular rule. Thus the phrase "the man who wrote the book that you told me about" is nested in the phrase "called the man who wrote the book that you told me about up. Nor." in (2ii). those to which we give the follow ing conventional technical names. can we exclude them by limiting the number of reapplications of grammatical rules in the gen eration of a sentence. for the purposes of this discussion: (3) (i) nested constructions (ii) (iii) (iv) (v) self-embedded constructions multiple-branching constructions left-branching constructions right-branching constructions The phrases A and B form a nested construction if A falls totally within B. it is clear that we can characterize unacceptable sentences only in terms of some "global" property of derivations and the structures they define . The phrase A is self-embedded in B if A is nested in B and.for example. In (liii). the Subject Noun Phrase is multiple-branch- . A is a phrase of the same type as B . This observation suggests that the study of performance could profitably begin with an investigation of the acceptability of the simplest formal structures in grammatical sentences. but rather to the way in which the rules inter relate in a derivation. Thus nesting has to do with bracketing. Thus "who the students recognized" is self-embedded in "who the boy who the students recognized pointed out. The most obvious formal property of utterances is their bracketing into constituents of various types. the "tree structure" as sociated with them. with some non null element to its left within B and some nonnull element to its right within B . and self-embedding with labeling of brackets as well. since unacceptability can just as well arise from application of distinct rules.a property that is attributable. A multiple-branching construction is one with no internal structure. obviously.METHODOLOGICAL PRELIMINARIES to exclude them. Among such structures we can distinguish various kinds . that is. each being applied only once." in (2i). since both are relative clauses. furthermore. In fact. in reading the right-branching construction "this is the cat that caught the rat that stole the cheese.thus. after "cat" and "rat. TOWARD A THEORY OF PERFORMANCE ing. such indefinitely iterable structures as [[[[1ohn]'s brother]. although these constructions are unnatural in other ways . Chomsky. and Miller and Chomsky. the Direct-Object of (Ii) or [this is [the cat that caught [the rat that stole the cheese]]] ." "Bill." instead of where the main brackets appear) In some measure. . Right-branching structures are those with the opposite prop erty . .s father]. or (Iii). A left-branching struc ture is of the form [[[ . 1 959a. and for discussion. since "John. 1 96 1 . in English. . Summarizing this work briefly. 1 96 3 ) that an optimal perceptual . . . for example. In terms of bracketing. ] ." and "several of their friends" are its immediate constituents. and have no further association among themselves. ] . which combines several kinds of left-branching. a multiple-branching construction has the form [[A][B]· · . their role in limiting performance). and there are some suggestive observations concerning their role in determin ing acceptability (that is." the intonation breaks are ordinarily inserted in the wrong places (that is. the following observations seem plausible: (4) (i) repeated nesting contributes to unacceptability (ii) self-embedding contributes still more radically to unac ceptability (iii) multiple-branching constructions are optimal in accepta bility (iv) nesting of a long and complex element reduces accepta bility (v) there are no clear examples of unacceptability involving only left-branching or only right-branching. . Chomsky. The effect of these superficial aspects of sentence structure on performance has been a topic of study since almost the very inception of recent work on generative grammar.§ 2.for example.s uncle] or [[[the man who you met] from Boston] who was on the train]." "Tom. Thus it is known (cf.for example. [M]] . ] . these phenomena are easily explained . it would impose the least strain on memory. To account for the greater unacceptability of self-embedding (assuming this to be a fact). Thus multiple coordination would be the simplest kind of construction for an analytic device . can accept unbounded left-branching and right-branching structures. 1961. The high acceptability of multiple-branching. for example.and right branching structures should become unnatural after a certain point. even with a bounded memory. But it is unclear why left. See. for a discussion of this point). If (4ii) is correct. though nested (hence ultimately self-embedded) structures go beyond its memory capacity. and the unacceptability of such examples as (aii) raises no problem. if they actually do. in this connection. Case (4iv) suggests decay of memory. and Shamir. one corresponding to each kind of phrase. but it is a rather plausible one. We might aSsume. that the perceptual device has a stock of analytic procedures available to it.METHODOLOGICAL PRELIMINARIES device. An optimal finite perceptual device of the type discussed in Chomsky (19594) need have no more difficulty with self-embedding than with other kinds of nesting (see Bar HilIel. Kasher. as in case (4iii). This is not a necessary feature of a perceptual model. in a tree-diagram of a sentence) is a rough measure of the amount of computation that has to be performed in analysis.7 For discussion. and that it is organized in such a way that it is unable (or finds it difficult) to utilize a procedure rp while it is in the course of executing rp. but raises un solved problems (see ChoIDSky. note 19). Thus case (4i) is simply a consequence of finiteness of memory. see Miller and Chomsky (1963). Case (4v) follows from the result about optimal perceptual models mentioned earlier. Miller and Isard (1964).a then we have evidence for a conclusion about organization of memory that goes beyond the triviality that it must be finite in size. 1963. and it would account for (4ii). -perhaps. is easily explained on the rather plausible assumption that the ratio of number of phrases to number of formatives (the node-to terminal node ratio.8 . we must add other conditions on the perceptual device beyond mere limitation of memory. as a s tylis tic device.One might ask whether attention to less superficial aspects of gramma tical structure than those of (3) could lead to somewhat deeper conclusions about performance models. furthermore. a generative grammar must be a system of rules tha t can iterate to generate an inerative . The facts. seem to be tha t the only studies of performance. it seems that the study of performance models incorpora ting generative grammars may be a fruitful study. Furthermore. the study of mem ory limitations j ust summarized and the study of deviation from rules. and. to describing "speech habits" or "habit structures. to which we return in Chapters 2 and 4. furthermore. we stress again that knowledge of a language involves the implicit ability to under stand indefinitely many sentences. it seems that these. THE ORGANIZATION OF A GENERATIVE GRAMMAR § 3. For example. etc. that precludes the development of a theory of actual performance. in Miller and Chomsky (1963) some syntactic and percep tual considerations are adduced in support of a suggestion (which is. It is the descriptivist limitation-in-principle to classification and organization of data. THE ORGANIZATION OF A GENERATIVE GRAMMAR Returni ng now to the question of competence and the gen grammars that purport to describe it. to "extracting p atterns" from a corpus of observed speech. There has been a fair amount of cri ticism of work in generative grammar on the grounds that it slights study of per formance in favor of study of underlying competence. it is difficult to im agine any other basis on which a theory of performan ce might develop." insofar as these may exist. In general. this criticism is unwarran ted . to be sure. highly speculative) as to the somewhat more detailed organization of a perceptual device. This seems entirely possible. outside of phonetics (but see note 3).8 Hence. lines of investigation can provide some insight into per formance.. are those carried out as a by-product of work in generative grammar. completely misdirected. In particular. have developed in this way. Conse quently. however. § 3. In fact.METHODOLOGICAL PRELIMINARIES definitely large number of structures. and semantic com ponents. their inherent properties. distinct and that the surface structure is deter mined by repeated application of certain formal operations called "grammatical transformations" to objects of a more . Ig64a. The phonological component of a grammar determines the phonetic form of a sentence generated by the syntactic rules. That is. I shall use the term "sentence" to refer to strings of formatives rather than to strings of phones. it relates a structure generated by the syntactic com ponent to a phonetically represented signal. in general. This system of rules can be analyzed into the three major components of a generative grammar: the syntactic. Both the phono logical and semantic components are therefore purely inter pretive. Chomsky. That is. and their interrelations in a given sentence. Postal. by the phonological component.10 The syntactic component specifies an infinite set of abstract formal objects. The first of these is interpreted by the semantic component. Each utilizes information provided by the syntactic component concerning formatives.12 It might be supposed that surface structure and deep structure will always be identical. Consequently. but not conversely. the second. Ig64). one might briefly characterize the syntactic theories that have arisen in modern structural (taxonomic) linguistics as based on the assumption that deep and surface structures are actually the same (cf.ll Since I shall be concerned here only with the syntactic com ponent. phonological. The semantic com ponent determines the semantic interpretation of a sentence. it relates a structure generated by the syntactic com ponent to a certain semantic representation. It will be recalled that a string of formatives specifies a string of phones uniquely (up to free variation). a deep structure that determines its semantic interpretation and a surface structure that determines its phonetic interpretation. the syn tactic component of a grammar must specify. for each sentence. The central idea of transformational grammar is that they are. each of which incorporates all information relevant to a single interpretation of a particular sentence. To clarify exposition. it is sufficient to observe that although the Immediate Constituent analysis (labeled bracketing) of an actual string of formatives may be adequate as an account of surface structure. each with an associated structural description called a base Phrase-marker. new proposal as to precisely how it should be formulated. in part. This assumption seems to me correct. The base of the syntactic component is a system of rules that generate a highly restricted (perhaps finite) set of basic strings. most sentences will have a sequence of such objects as an underlying basis. the syntactic component of a generative grammar contains a transformational sub component. For the moment. I shall use the following terminology. and must interrelate them. I shall present a specific and. with the elementary objects of which deep structure is constituted. we can delimit a proper subset called "kernel sentences. then the syntactic component must generate deep and surface structures. In Chapter 3. in particular. I shall refer to this sequence as the basis of the sentence that it underlies. This idea has been clarified substantially in recent work. henceforth). in ways that will be described later. but has no important conse quences for what follows here. from its basis. with occasional revisions as the discussion proceeds. In addition to its base. If this is true (as I assume.§ 3. Among the sentences with a single base Phrase-marker as basis. though it simplifies exposition. Underlying each sentence of the language there is a sequence of base Phrase-markers. THE ORGANIZATION OF A GENERATIVE GRAMMAR elementary sort. with its surface structure." These are sentences of a particularly simple . it is certainly not adequate as an account of deep structure. each generated by the base of the syntactic component. This is concerned with generating a sentence. Some familiarity with the operation and effects of transformational rules is henceforth presupposed. These base Phrase-markers are the elementary units of which deep structures are constituted. My concern in this book is primarily with deep struc ture and. for each sentence. Since the base generates only a restricted set of base Phrase markers. I shall assume that no ambiguity is introduced by rules of the base. METHODOLOGICAL PREUMINAlUFJI sort that involve a minimum of transformational apparatus in their generation. Since transformations will not be considered here in detail. this is neither presented for direct observation nor extractable from data by inductive procedures of any known sort. although there are methodological discus- . no careful distinction will be made. first of all. the actual data of linguistic performance will provide much evidence for determining the correctness of hypotheses about underlying linguistic structure. The notion "kernel sentence" has. I shall say nothing more about them here. an important intuitive significance. between the basic string underlying this sentence and the sentence itself. JUS TIFICA TION OF GRAMMARS Before entering directly into an investigation of the syntactic component of a generative grammar. I think. in the case of a sentence with a single element in its basis. and that the base Phrase-marker is the surface structure as well as the deep structure. In other words. about his knowledge of the language. the question of how one is to obtain information about the speaker-hearer's competence. Clearly. it seems. § 4. but the simplifying assumption should be borne in mind through out. it is important to give some thought to several methodological questions of justification and adequacy. or the linguist who has learned the language) . The basic strings and base Phrase-markers do. in this case. at many points in the exposition I shall make the tacit simplifying (and contrary-to-fact) assumption that the underlying basic string is the sentence. One must be careful not to confuse kernel sentences with the basic strings that under lie them. but since kernel sentences play no distinctive role in generation or interpretation of sentences. along with introspective reports (by the native speaker. Like most facts of interest and importance. There is. This is the position that is universally adopted in practice. play a distinctive and crucial role in language use. I shall try to select examples in such a way as to minimize possible confusion. however. Even though few reliable operational procedures have been developed. in other words. JUSTIFICATION OF GRAMMARS sions that seem to imply a reluctance to use observed perform ance or introspective reports as evidence for some underlying reality. it is unfortunately the case that no adequate for malizable techniques are known for obtaining reliable informa tion concerning the facts of linguistic structure (nor is this particularly surprising). proceed. This is the objective situa tion of present-day linguistic work.a grammar . we might be justified in relying on their results in unclear and difficult cases. There are.must be tested for adequacy) by measuring it against the standard provided by the tacit knowledge that it attempts to specify and describe_ Thus a proposed operational test for. must meet the empirical condition of conform ing. it is without value. in a mass of crucial and clear cases. to the linguistic intuition of the native speaker concerning such elements. It is important to bear in mind that when an operational procedure is pl'Oposed. If operational procedures were available that met this test. very few reliable experimental or data-processing procedures for obtaining significant information concerning the linguistic intuition of the native speaker. Furthermore. the theoretical (that is. is true in the case of any proposed operational procedure or any proposed grammatical description. This remains a hope for the future rather than a present reality. for the present. there is no reason to expect that reliable operational criteria for the deeper and more important theoretical notions of linguistics (such as "gram maticalness" and "paraphrase") will ever be forthcoming. grammatical) investigation of the knowledge of the native speaker can proceed perfectly well. The same. allusions to presumably well known "procedures of elicitation" or "objective methods" simply obscure the actual situation in which linguistic work must. The critical problem for grammatical theory today is not a paucity of evidence but rather the inadequacy of present theories of language to account for masses of evidence that are hardly . say. In brief. obviously.§ 4. segmenta tion into words. it must be tested for adequacy (exactly as a theory of linguistic intuition . Otherwise. but they must obviously converge on the tacit knowledge of the native speaker if they are to be of any significance. One who disagrees with this estimate of the present situation in linguistics can . it determines how we shall denote the kind of research that can be effectively carried out in the present state of our technique and understanding. the problem for one concerned with operational procedures is to develop tests that give the correct results and make relevant distinctions . One may ask whether the necessity for present·day linguistics to give such priority to introspective evidence and to the linguistic intuition of the native speaker excludes it from the domain of science. On the other hand. by and large. himself). However. a good case can be made for the view that the natural sciences have. The answer to this essentially terminological question seems to have no bearing at all on any serious issue. this terminological question actually does relate to a different issue of some interest. At most. an explanation for the enormous mass of unquestionable data concerning the linguistic intuition of the native speaker (often. The problem for the grammarian is to construct a description and. where possible. We may h ope that these efforts will converge. The social and behavioral sciences provide ample evidence that objectivity can be pursued with little consequent gain in insight and understanding. sought objectivity primarily insofar as it is a tool for gaining insight (for providing phenomena that can suggest or test deeper explanatory hypotheses).20 METHODOLOGICAL PRELIMINARIES open to serious question. In any event. one whose con cern is for insight and understanding (rather than for objectivity as a goal in itself) must ask whether or to what extent a wider range and more exact description of phenomena is relevant to solving the problems that he faces. for the present . at a given stage of investigation. it seems to me that sharpening of the data by more objective tests is a matter of small importance for the" problems at hand. In linguistics. Neither"the study of grammar nor the attempt to develop useful tests is hampered by lack of evidence with which to check results. namely the question whether the important feature of the success ful sciences has been their search for insight or their concern for objectivity. Perhaps the day will come when the kinds of data that we now can obtain in abundance will be insufficient to resolve deeper questions con cerning the structure of language. it must be emphasized. To eliminate what has seemed to some an air of paradox in this remark. when this is pointed out to him. this fact can be brought to consciousness by consideration of slight elaborations of sentence . and will fail to detect the ambiguity. many questions that can realistically and significantly be formulated today do not demand evidence of a kind that is unavailable or unattainable without significant improvements in objectivity of experimental technique. once again. or operational test. his intuitive knowledge of the language is clearly such that both of the inter pretations (corresponding to "Hying planes are dangerous" and "Hying planes is dangerous") are assigned to the sentence by the grammar he has internalized in some form . In fact. the ambiguity may be fairly trans parent. Although there is no way to avoid the traditional assumption that the speaker-hearer's linguistic intuition is the ultimate standard that determines the accuracy of any proposed grammar. that this tacit knowledge may very well not be immediately available to the user of the language . let me illustrate with a few examples. as forced or un natural (independently of which interpretation he originally selected under contextual pressure). But consider such a sentence as (5) I had a book stolen Few hearers may be aware of the fact that their internalized grammar in fact provides at least three structural descriptions for this sentence. In the case just mentioned.§ 4. he may reject the second inter pretation. If a sentence such as "Hying planes can be dangerous" is presented in an appropriately constructed context. Nevertheless. However. the listener will interpret it immediately in a unique way. Nevertheless. JUSTIFICATION OF GRAMMARS u justify his belief in the current importance of more objective operational tests by showing how they can lead to new and deeper understanding of linguistic structure. linguistic theory. METHODOLOGICAL PREUMINARIES (5), for example: (i) "I had a book stolen from my car when I stupidly left the window open," that is, "someone stole a book from my car"; (ii) "I had a book stolen from his library by a professional thief who I hired to do the job," that is, "I had some one steal a book"; (iii) "I almost had a book stolen, but they caught me leaving the library with it," that is, "I had almost succeeded in stealing a book." In bringing to consciousness the triple ambiguity of (5) in this way, we present no new informa tion to the hearer and teach him nothing new about his language but simply arrange matters in such a way that his linguistic intuition, previously obscured, becomes evident to him. As a final ill us tra tion, consider the sentences (6) I persuaded John to leave The first impression of the hearer may be that these sentences receive the same structural analysis. Even fairly careful thought may fail to show him that his internalized grammar assigns very different syntactic descriptions to these sentences. In fact, so far as I have been able to discover, no English grammar has pointed out the fundamental distinction between these two constructions (in particular, my own sketches of English grammar in Chomsky, Ig55 , 1 962a, failed to note this). However, it is clear that the sentences (6) and (7) are not parallel in structure. The difference can be brought out by consideration of the sentences (7) I expected John to leave (S) (i) I persuaded a specialist to examine John (ii) I persuaded John to be examined by a specialist (9) (i) I expected a specialist to examine John (ii) I expected John to be examined by a specialist The sentences (gi) and (gii) are "cognitively synonymous" : one is true if and only if the other is true. But no variety of even weak paraphrase holds between (Si) and (8ii). Thus (Si) can be true or false quite independendy of the truth or falsity of (Sii). What ever difference of connotation or "topic" or emphasis one may find between (gi) and (gii) is just the difference that exists be- § 4 . JUSTIFICATION OF GRAMMARS tween the active sentence "a specialist will examine John " and its passive counterpart "John will be examined by a specialist." This is not at all the case with respect to (8), however. In fact, the underlying deep structure for (6) a nd (8ii) must show that "John" is the Direct-Object of the Verb Phrase as well as the grammatical Subject of the embedded sentence. Furthermore, in (8ii) "John" is the logical Direct-Object of the embedded sentence, whereas in (8i) the phrase "a specialist" is the Direct-Object of the Verb Phrase and the logical Subject of the embedded sentence. In (7 ) , (gi), and (gii), however, the Noun Phrases "John ," "a specialist," and "John," respectively, have no grammatical functions other than those that are internal to the embedded sentence; in par ticular, "John" is the logical Direct-Object and "a specialist" the logical Subject in the embedded sentences of (g). Thus the under lying deep structures for (8i), (8ii), (gi), and (gii) are, respectively, the following: 1 8 (1 0) (i) Noun Phrase - Verb - Noun Phrase - Sentence (I - persuaded - a specialist - a specialist will examine John) (ii) Noun Phrase - Verb - Noun Phrase - Sentence (I - persuaded - John - a specialist will examine John) ( 1 1) (i) Noun Phrase - Verb - Sentence (I - expected - a specialist w ill examine John) (ii) Noun Phrase - Verb - Sentence (I - expected - a specialist w ill examine John) In the case of ( I oii) and ( I l ii), the passive transformation will apply to the embedded sentence, and in all four cases other ope rations will gi ve the final surface forms of (8) and (g) . The important point in the present connection is that (8i) differs from (8ii) in underlying structure, although (9i) and (gii) are essentially the same in underlying structure. This accounts for the difference in meaning. Notice, in support of this difference in analysis, that we can have "I persuaded John that (of the fact that) Sentence," but not "I expected John that (of the fact that) Sentence." METHODOLOGICAL PRELIMINARIES The example (6)-(7) serves to illustrate two important points. First, it shows how unrevealing surface structure may be as to underlying deep structure. Thus (6) and (7) are the same in surface structure, but very different in the deep structure that underlies them and determines their semantic interpretations. Second, it illustrates the elusiveness of the speaker's tacit knowl edge. Until such examples as (8) and (9) are adduced, it may not be in the least clear to a speaker of English that the grammar that he has internalized in fact assigns very different syntactic analyses to the superficially analogous sentences (6) and (7). In short, we must be careful not to overlook the fact that surface similarities may hide underlying distinctions of a funda mental nature, and that it may be necessary to guide and draw out the speaker's intuition in perhaps fairly subtle ways before we can determine what is the actual character of his knowledge of his language or of anything else. Neither point is new (the former is a commonplace of traditional linguistic theory and analytic philosophy; the latter is as old as Plato's Meno) ; both are too often overlooked. A grammar can be regarded as a theory of a language; it is descriptively adequate to the extent that it correctly describes the intrinsic competence of the idealized native speaker. The struc tural descriptions assigned to sentences by the grammar, the distinctions that it makes between well-formed and deviant, and so on, must, for descriptive adequacy, correspond to the linguistic intuition of the native speaker (whether or not he may be immediately aware of this) in a substantial and significant class of crucial cases. A linguistic theory must contain a definition of "grammar," that is, a specification of the class of potential grammars. We may, correspondingly, say that a linguistic theory is descriptively adequate if it makes a descriptively adequate grammar available for each natural language. Although even descriptive adequacy on a large scale is by no means easy to approach, it is crucial for the productive develop ment of linguistic theory that much higher goals than this be pursued. To facilitate the clear formulation of deeper questions, § 4. JUSTIFICATION OF GRAMMARS 25 it is useful to consider the abstract problem of constructing an "acquisition model" for language, that is, a theory of language learning or grammar construction. Clearly, a child who has learned a language has developed an internal representation of a system of rules that determine how sentences are to be formed, used, and understood. Using the term "grammar" with a sys tematic ambiguity (to refer, first, to the native speaker's internally represented "theory of his language" and, second, to the linguist's account of this), we can say that the child has developed and internally represented a generative grammar, in the sense de scribed. He has done this on the basis of observation of what we may call primary linguistic data . This must include examples of linguistic performance that are taken to be well-formed sen tences, and may include also examples designated as non sentences, and no doubt much other information of the sort that is required for language learning, whatever this may be (see pp. 3 1-3 2). On the basis of such data, the child constructs a grammar - that is, a theory of the language of which the well-formed sentences of the primary linguistic data constitute a small sample.14 To learn a language, then, the child must have a method for devising an appropriate grammar, given primary linguistic data. As a precondition for language learning, he must possess, first, a linguistic theory that specifies the form of the grammar of a possible human language, and, second, a strategy for selecting a grammar of the appropriate form that is com patible with the primary linguistic data. As a long-range task for general linguistics, we might set the problem of developing an account of this innate linguistic theory that provides the basis for language learning. (Note that we are again using the term "theory" - in this case "theory of language" rather than "theory of a particular language" - with a systematic ambiguity, to refer both to the child's innate predisposition to learn a language of a certain type and to the linguist'S account of this.) To the extent that a linguistic theory succeeds in selecting a descriptively adequate grammar on the basis of primary linguistic data, we can say that it meets the condition of explanatory ade quacy. That is, to this extent, it offers an explanation for the METHODOLOGICAL PREUMINARIES intuition of the native speaker on the basis of an empirical hypothesis concerning the innate predisposition of the child to develop a certain kind of theory to deal with the evidence presented to him. Any such hypothesis can be falsified (all too easily, in actual fact) by showing that it fails to provide a descriptively adequate grammar for primary linguistic data from SOme other language - evidently the child .is not pre disposed to learn one language rather than another. It is sup ported when it does provide an adequate explanation for some aspect of linguistic structure, an account of the way in which such knowledge might have been obtained. Clearly, it would be utopian to expect to achieve explanatory adequacy on a large scale in the present state of linguistics. Nevertheless, considerations of explanatory adequacy are often critical for advancing linguistic theory. Gross coverage of a large mass of data can often be attained by conflicting theories; for precisely this reason it is not, in itself, an achievement of any particular theoretical interest or importance. As in any other field, the important problem in linguistics is to discover a complex of data that differentiates between conflicting concep tions of linguistic structure i n that one of these conflicting theories can describe these data only by ad hoc means whereas the other can explain it on the basis of some empirical assump tion about the form of language. Such small-scale studies of explanatory adequacy have, in fact, provided most of the evi dence that has any serious bearing on the nature of linguistic structure. Thus whether we are comparing radically different theories of grammar or trying to determine the correctness of some particular aspect of one such theory, it is questions of explanatory adequacy that must, quite often, bear the burden of justification. This remark is in no way inconsistent with the fact that explanatory adequacy on a large scale is out of reach, for the present. It simply brings out the highly tentative character of any attempt to justify an empirical claim about linguistic structure. To summarize briefly, there are two respects in which one can speak of "justifying a generative grammar." On one level (that In this sense. and specific enough to account for the fact of language acquisition. It proposes. namely the linguistic intui tion . his problem being to determine which of the (humanly) possible languages is that of the com munity in which he is placed. FORMAL AND SUBSTANTIVE UNIVERSALS of descriptive adequacy). a grammar is justified to the extent that it is a principled descrip tively adequate system. On a much deeper and hence much more rarely attainable level (that of explanatory adequacy). detailed. in that the linguistic theory with which it is associated selects this grammar over others. an account of the specific innate abilities that make this achieve ment possible. the main . FORMAL AND S UBSTA N TI VE UNIVERSALS A theory of linguistic structure tha t aims for explanatory adequacy incorporates an account of linguistic universals. the grammar is justified to the extent that it correctly describes its object. then. The important question is: What are the initial assumptions concerning the nature of language that the child brings to language learning. Consequently.is essentially the problem of constructing a theory of language acquisition. and it attributes tacit knowledge of these universals to the child. on grounds of correspondence to linguistic fact. § 5. the grammar is justified on ex ternal grounds.the tacit compe tence . the grammar is justified on internal grounds.of the native speaker. given primary linguistic data with which all are compatible. The problem of internal justification . Language learning would be impossible unless this were the case. on grounds of its rela ti on to a linguis tic theory that constitutes an explanatory hypo th esis about the form of language as such. In this sense.§ 5.of explanatory adequacy . that the child approaches the data with the presumption that they are drawn from a language of a certain antecedently well-defined type. and how detailed and specific is the innate schema (the general definition of "grammar") that gradually becomes more explicit and differ entiated as the child learns the language? For the present we cannot come at all dose to making a hypothesis about innate schemata that is rich. feelings. It not only put forth interesting views as to the nature of universal phonetics. Verb. but also advanced the position that certain fixed syntactic categories (Noun. however. The study of linguistic universals is the study of the prop erties of any generative grammar for a natural language. A theory of substantive semantic universals might hold for ex ample. and so on. and the remarkable com plexity and range of the generative grammars that are the product of language learning.METHODOLOGICAL PRELIMINARIES task of linguistic theory must be to develop an account of linguistic universals that. on the one hand. on the other. It is also possible. It is useful to classify linguistic universals as formal or sub stan tive . universal. etc.) can be found in the syntactic representations of the sentences of any language. or to interrelations among the three components. will be sufficiently rich and explicit to account for the rapidity and uniformity of language learning. Traditional universal grammar was also a theory of substantive universals. phonetic features (perhaps on the order of fifteen or twenty). in this sense. Jakobson's theory of distinctive features can be interpreted as making an assertion about substantive universals with respect to the phonological component of a generative grammar. A theory of substantive universals claims that items of a particular kind in any language must be drawn from a fixed class of items. Partic ular assumptions about linguistic universals may pertain to either the syntactic. For example. It asserts that each output of this component consists of elements that are characterized in terms of some small number of fixed. each of which has a substantive acoustic-articulatory characterization independent of any partic ular language. semantic. or phonological component. and that these provide the general underlying syntactic structure of each language. Thus it might assert that each language will contain terms that designate persons or lexical items referring to certain specific kinds of objects. to search for universal properties . will not be falsified by the actual diversity of languages and. be havior. that certain designative functions must be carried out in a specified way in each language. for example. Substantive universals such as these concern the vocabulary for the description of language.15 and that the same is true of other terms designating objects. too. Consider a claim that the grammar of every language meets certain specified formal conditions. formal universals involve rather the character of the rules that appear in grammars and the ways in which they can be interconnected. consider the proposal that the syntactic component of a grammar must contain transformational rules (these being operations of a highly special kind) mapping semantically interpreted deep structures into phonetically inter preted surface structures. Recent attempts to specify the abstract conditions that a generative grammar must meet have produced a variety of proposals concerning formal uni versals. a subset of which may apply cyclically to successively more dominant con stituents of the surface structure (a transformational cycle. On the semantic level. or that certain semantic features or categories provide a universal framework for semantic description. The truth of this hypothesis would not in itself imply that any particular rule must appear in all or even in any two grammars. FORMAL AND SUBSTANTIVE UNIVERSALS 29 of a more abstract sort. or the proposal that the phonological component of a grammar consists of a sequence of rules. The property of having a grammar meeting a certain abstract condition might be called a formal linguistic universal. in this sense. and functions instead of solely in terms of physical qualities. or the condition that artifacts are defined in terms of certain human goals.§ 5. must designate objects meeting a condi tion of spatiotemporal contiguity. Consider. in any language. Such proposals make claims of a quite different sort from the claim that certain sub stantive phonetic elements are available for phonetic representa tion in all languages. or that certain specific categories must be central to the syntax of all languages. it is possible to search for what might be called formal universals. if shown to be a general property of natural languages. For example. needs. or the condition that the color words of any language must subdivide the col or spectrum into continuous segments. in the sense of much recent work on phonology).lo Formal con- . in essentially the sense j ust described. the assumption that proper names. FUR THER REMARKS ON DESCRIPTI VE AND EXPLANA TOR Y THEORIES Let us consider with somewhat greater care just what is involved in the construction of an "acquisition model" for language. They seem to offer extremely rich and varied possibilities for study in all aspects of grammar. A child who is capable of language learning must have ( 1 2 ) (i) a technique for representing input signals (ii) a way of representing structural information about these signals (iii) some initial delimitation of a class of possible hypotheses about language structure (iv) a method for determining what each such hypothesis im plies with respect to each sentence (v). in the sense suggested by such examples as these. there is no doubt that a theory of language. should concern itself with both substantive and formal universals. It does not. re garded as a hypothesis about the innate "language-forming capacity" of humans. The existence of deep-seated formal universals. but does not imply that there is any point by point correspondence between particular languages. imply that there must be some reason able procedure for translating between languages. infinitely many) hypotheses that are allowed by (iii) and are com patible with the given primary linguistic data . 1 7 In general. § 6. implies that all languages are cut to the same pattern. for example. But whereas substantive universals have been the traditional concern of general linguistic theory. given primary linguistic data. investi gations of the abstract conditions that must be satisfied by any generative grammar have been undertaken only quite recently. a method for selecting one of the (presumably. or the linguist) of a descriptive grammar.METHODOLOGICAL PRELIMINARIES straints of this sort on a system of concepts may severely limit the choice (by the child. § 6. DESCRIPTIVE AND EXPLANATORY THEORIES Correspondingly, a theory of linguistic structure that aims for explanatory adequacy must contain ( 1 3) (i) a universal phonetic theory that defines the notion "possi ble sentence" (ii) a definition of "structural description" (iii) a definition of "generative grammar" (iv) a method for determining the structural description of a sentence, given a grammar (v) a way of evaluating alternative proposed grammars Putting the same requirements in somewhat different terms, we must require of such a linguistic theory that it provide for of possible sentences ( 1 4) (i) an enumeration of the class S1 ' S2' of possible (ii) an enumeration of the class SD 1 , SD2, structural descriptions of possible genera (iii) an enumeration of the class G1 , G2, tive grammars (iv) specification of a function f such that SD,«(,J ) is the struc tural description assigned to sentence s, by grammar Gi, for arbitrary i,j 18 (v) specification of a function m such that m ( z) is an integer associated with the grammar G, as its value (with, let us say, lower value indicated by higher number) • • • • • • • • • Conditions of at least this strength are entailed by the decision to aim for explanatory adequacy. A theory meeting these conditions would attempt to account for language learning in the following way. Consider first the nature of primary linguistic data. This consists of a finite amount of information about sentences, which, furthermore, must be rather restricted in scope, considering the time limitations that are in effect, and fairly degenerate in quality (cf. note 1 4). For example, certain signals might be accepted as properly formed sentences, while others are classed as nonsentences, as a result of correction of the learner's attempts on the part of the linguistic community. Furthermore, the conditions of use might be such METHODOLOGICAL PRELIMINARIES as to require that structural descriptions be assigned to these objects in certain ways. That the latter is a prerequisite for language acquisition seems to follow from the widely accepted (but. for the moment. quite unsupported) view that there must be a partially semantic basis for the acquisition of syntax or for the justification of hypotheses about the syntactic component of a grammar. Incidentally. it is often not realized how strong a claim this is about the innate concept-forming abilities of the child and the system of linguistic universals that these abilities imply. Thus what is maintained. presumably. is that the child has an innate theory of potential structural descriptions that is sufficiently rich and fully developed so that he is able to deter mine. from a real situation in which a signal occurs, which struc tural descriptions may be appropriate to this signal, and also that he is able to do this in part in advance of any assumption as to the linguistic structure of this signal. To say that the assumption about innate capacity is extremely strong is, of course. not to say that it is incorrect. Let us, in any event, assume tentatively that the primary linguistic data consist of signals classified as sentences and nonsentences. and a partial and tenta tive pairing of signals with structural descriptions. A language-acquisition device that meets conditions (i)-(iv) is capable of utilizing such primary linguistic data as the empirical basis for language learning. This device must search through the set of possible hypotheses GI• G2• • • • which are available to it by virtue of condition (Hi), and must select grammars that are compatible with the primary linguistic data, represented in terms of (i) and (ii). It is possible to test compatibility by virtue of the fact that the device meets condition (iv). The device would then select one of these potential grammars by the evaluation measure guaranteed by (V).19 The selected grammar now pro vides the device with a method for interpreting an arbitrary sentence, by virtue of (H) and (iv) . That is to say, the device has now constructed a theory of the language of which the primary linguistic data are a sample. The theory that the device has now selected and internally represented specifies its tacit competence, its knowledge of the language. The child who acquires a language • § 6. DESCRIPTIVE AND EXPLANATORY THEORIES 33 in this way of course knows a great deal more than he has "learned." His knowledge of the language, as this is determined by his internalized grammar, goes far beyond the presented primary linguistic data and is in no sense an " inductive gen eralization" from these data. This account of language learning can, obviously, be para phrased directly as a description of how the linguist whose work is guided by a linguistic theory meeting conditions (i)-(v) would justify a grammar that he constructs for a language on the basis of given primary linguistic data.20 Notice, incidentally, that care must be taken to distinguish several different ways in which primary linguistic data may be necessary for language learning. In part, such data determine to which of the possible languages (that is, the languages pro vided with grammars in accordance with the a priori constraint (iii» the language learner is being exposed, and it is this function of the primary linguistic data that we are considering here. But such data may play an entirely different role as well; namely, certain kinds of data and experience may be required in order to set the language-acquisition device into operation, although they may not affect the manner of its functioning in the least. Thus it has been found that semantic reference may greatly facilitate performance in a syntax-learning experimen t, even though it does not, apparently, affect the manner in which acquisition of syntax proceeds; that is, it plays no role in deter mining which hypotheses are selected by the learner (Miller and Norman, 1 964). Similarly, it would not be at all surprising to find that normal language learning requires use of language in real-life situations, in some way. But this, if true, would not be sufficient to show that information regarding situational context (i n particular, a pairing of signals wi th structural descrip tions that is at least in part prior to assumptions about syntactic structure) plays any role in determining how language is acquired, once the mechanism is put to work and the task of language learning is undertaken by the child. This distinction is quite familiar outside of the domain of language acquisition. For example, Richard Held has shown in numerous experiments 34 METHODOLOGICAL PRELIMINARIES that under certain circumstances reafferent stimulation (that is, stimulation resulting from voluntary activity) is a prerequisite to the development of a concept of visual space, although it may not detennine the character of this concept (cf. Held and Hein, 1 963 ; Held and Freedman, 1 963, and references cited there). Or, to take one of innumerable examples from studies of animal learning, it has been observed (Lemmon and Patterson, 1 964) that depth perception in lambs is considerably facilitated by mother-neonate contact, although again there is no reason to suppose that the nature of the lamb's "theory of visual space" depends on this contact. In studying the actual character of learning, linguistic or otherwise, it is of course necessary to distinguish carefully be tween these two functions of external data - the function of initiating or facilitating the operation of innate mechanisms and the function of detennining in part the direction that learning will take.21 Returning now to the main theme, we shall call a theory of linguistic structure that meets conditions (i)-(v) an explanatory theory, and a theory that meets conditions (i)-(iv) a descriptive theory. In fact, a linguistic theory that is concerned only with descriptive adequacy will limit its attentlon to topics (i)-(iv). Such a theory must, in other words, make available a class of generative grammars containing. for each language. a descrip tively adequate grammar of this language - a grammar that (by means of (iv» assigns structural descriptions to sentences in accordance with the linguistic competence of the native speaker. A theory of language is empirically significant only to the extent that it meets conditions (i)-(iv). The further question of explana tory adequacy arises only in connection with a theory that also meets condition (v) (but see p. 36). In other words, it ari se s only to the extent that the theory provides a principled basis for selecting a descriptively adequate grammar on the basis of primary linguistic data by the use of a well·defined evaluation measure. This account is misleading in one important respect. It sug gests that to raise a descriptively adequate theory to the level § 6. DESCRIPTIVE AND EXPLANATORY THEORIES 35 of explanatory adequacy one needs only to define an appropriate evaluation measure. This is incorrect, however. A theory may be descriptively adequate, in the sense just defined, and yet provide such a wide range of potential grammars that there is no possi bility of discovering a formal property distinguishing the de scriptively adequate grammars, in general. from among the mass of grammars compatible with whatever data are available. In fact, the real problem is almost always to restrict the range of possible hypotheses by adding additional structure to the notion "generative grammar." For the construction of a reasonable acquisition model, it is necessary to reduce the class of attain able22 grammars compatible with given primary linguistic data to the point where selection among them can be made by a formal evaluation measure. This requires a precise and narrow delimitation of the notion "generative grammar" a restrictive and rich hypothesis concerning the universal properties that determine the form of language, in the traditional sense of this term. The same point can be put in a somewhat different way. Given a variety of descriptively adequate grammars for natural languages, we are interested in determining to what extent they are unique and to what extent there are deep underlying similari ties among them that are attributable to the form of language as such. Real progress in linguistics consists in the discovery that certain features of given languages can be reduced to universal properties of language, and explained in terms of these deeper aspects of linguistic form. Thus the major endeavor of the linguist must be to enrich the theory of linguistic form by for mulating more specific constraints and conditions on the notion "generative grammar." Where this can be done. particular gram mars can be simplified by eliminating from them descriptive statements that are attributable to the general theory of grammar (cf. § 5). For example, if we conclude that the transformational cycle23 is a universal feature of the phonological component, it is unnecessary, in the grammar of English, to describe the man ner of functioning of those phonological rules that involve syntactic structure. This description will now have been ab- are almost always those involving explanatory adequacy with respect to particular aspects of language structure. notes 1 9 and 2 2). if false. To acquire language. In this case.that is. an assumption that can be refuted. since it would then have been shown that what appears to be a peculiarity of English is actually explicable in terms of a general and deep empirical assumption about the nature of language. the most serious problem that arises in the attempt to achieve explanatory adequacy is that of characterizing the notion "generative grammar" in a sufficiently rich. In short. the crucial questions. It is not necessary to achieve descriptive ade quacy before raising questions of explanatory adequacy. no evaluation procedure will be necessary as a part of linguistic theory . if justified.is so crucial a t every stage o f understanding o f linguistic structure. the questions that have the greatest bearing on our concept of language and on descriptive practice as well. a child must devise a hypothesis compa tible with presented data . It is logically possible that the data might be sufficiently rich and the class of potential grammars sufficiently limited so that no more than a single permitted grammar will be compatible with the available data at the moment of success ful language acquisition.METHODOLOGICAL PRELIMINARIES stracted from the grammar of English and stated as a formal linguistic universal.he must select from the store of potential grammars a specific one that is appropriate to the data available to him. by study of descriptively adequate grammars of other languages. this conclusion. Obviously. would represent an im portant advance in the theory of language.the attempt to discover linguistic universals . It is for just this reason that the attempt to achieve explanatory adequacy . as part of the theory of generative grammar. as an innate property of an organism or a device capable of language acquisi- . detailed. despite the fact that even descriptive adequacy on a broad scale may be an unrealized goal. and highly structured way. in our idealized "instantaneous" model (cf. A theory of grammar may be descriptively adequate and yet leave unexpressed major features that are defining properties of natural language and that distinguish natural languages from arbitrary symbolic systems. On the contrary. " etc. and all concrete attempts to formulate an empirically adequate linguistic theory certainly leave ample room for mutually inconsistent grammars. that this is an empirical fact about the innate human facu lte de langage and consequently about general linguistic theory as well. all com patible with primary data of any conceivable sort. It is rather difficult to imagine how in detail this logical possibility might be realized. ON EVALUATION PROCEDURES 37 tion. in some manner. This is a misconception. Given primarily linguistic data D. and will therefore lead to entirely differen t predictions as to how a person who learns a language on the basis of D will interpret new sentences not in D . In the context of this discussion. and particular proposals are correct or incorrect. All such theories therefore require supplementation by an evaluation measure if language acquisition is to be accounted for and selec tion of specific grammars is to be justified. and I shall continue to assume tentatively. any proposal concerning such a me asu re is an empi rical hypothesis about the nature of language. § 7. Perhaps confusion about this matter can be traced to the use of the term "simplicity measure" for particular proposed evalua tion measures. It must first of all be kept clearly in mind that such a measure is not given a priori. ON EVA L UA TION PR O CEDURES The status of an evaluation procedure for grammars (see condi tion (v) of ( 1 2)-(1 4» has often been misconstrued." "phoneme. This is evident from the preceding discussion. Rather.§ 7. Con sequently. it being assumed that "simplicity" is a general notion somehow understood in advance outside of linguistic theory. as heretofore. however. different choices of an evaluation measure will assign quite different ranks to alternative hypotheses (alternative grammars) as to the lan guage of which D is a sample. meeting conditions (i)-(iv) of ( 1 2)-(1 4) in some fixed way. choice of an evaluation measure is an empirical matter. Suppose that we have a descriptive theory. the evaluation measure m of (v» is a notion to be defined within linguistic theory along with "grammar. "simplicity" (that is. Choice of a simplici ty measure is . We are given.. What we have been discussing here. It is also apparent that evaluation measures of the kinds that have been discussed in the literature on generative grammar cannot be used to compare different theories of grammar. by such a measure. comparison of a grammar from one class of proposed grammars with a grammar from another class. This is. of primary linguistic data and descriptively adequate grammars are given. and if pairs (D1. All of this has been said before. in part. a matter of formulating an explanatory theory of language. I repeat it at such length be cause it has been so grossly misunderstood.two grammars of this language . A proposed simplicity measure. for each i.METHODOLOGICAL PREUMINARIES rather like determination of the value of a physical constant. an evaluation measure of this kind is an essential part of a particular theory of grammar that aims at explanatory adequacy. the problem of defining "simplicity" is just the problem of discovering how G. that we regard an acquisition model for language as an input-output device that determines a particular generative grammar as "output. is utterly without sense. Choice of a simplicity measure is there fore an empirical matter with empirical consequences. Rather. G2). A pro posed simplicity measure constitutes part of the attempt to deter mine precisely the nature of this association. If a particular formulation of (i)-(iv) is assumed. Suppose. is not this general ques tion but rather the problem of comparing two theories of a language . (D2 . in other words. It is true that there is a sense in which alternative theories of language (or alternative theories in other domains) can be compared as to simplicity and elegance. taken together with a specification of (i)-(iv). it is not to be confused with the problem of choosing among competing theories of language. is determined by D. constitutes a hypothesis concerning the nature of such a device.G 1). then. Choice among competing theories of language is of course a fundamental question and should also be • • • ." given certain primary linguistic data as input.in terms of a particular general linguistic theory. however. an empirical pairing of certain kinds of primary linguistic data with certain grammars that are in fact constructed by people presented with such data. Consequently. consider the question of whether the rules of a grammar should be unordered (let us call this the linguistic theory Tu) or ordered in some specific way (the theory To) . then it can be shown that TuB and T OB are incomparable in descriptive power (in "strong generative capacity" . or leads to explanatory adequacy. These measures do not apply. in neither case Will this concept have any known justification. on empirical grounds of descriptive and explanatory adequacy. if TuB is the familiar theory of phrase struc ture grammar and TOB is the same theory. 1 96 1 . insofar as possible. we might ask whether natural languages in fact fall under T Jl or TOB. Chapters 6 and 7. in accordance with which Tu and To can be compared. A priori. minimization of feature specification (as discussed in Halle. there is no way to decide which of the two is correct. for some discus sion of such systems). One can easily invent a general concept of "simplicity" that will prefer Tu to To. We must ask whether Tu or To provides descrip tively adequate grammars for natural languages." developed within linguistic theory or general epis temology. or To to Tu. It is quite meaningless. therefore. 1 955. For example. 1 962a.see § 9 . Certain measures of evaluation have been proposed and in part empirically justified within linguistics . and Chomsky. There is no known absolute sense of "simplicity" or "elegance. 1 956. because they are internal to a specific linguistic theory and their empirical justification relies essentially on this fact. with the further condi tion that the rules are linearly ordered and apply cyclically.) . these being non- .§ 7. This is a perfectly meaningful empirical question if the theories in question are stated with sufficient care. ON EVALUATION PROCEDURES 39 settled. see Chomsky. to maintain that in some absolute sense Tu is "simpler" than To or conversely. 42f. so as to guarantee that each cycle is nonvacuous.for example. As a concrete illustration. To choose between Tu and To. we must proceed in an entirely different way. with at least one rule A � X being obligatory for each category A. But it is not the question involved in the use of an evaluation measure in the attempt to achieve explanatory adequacy. 1 964) or the measure based on abbreviatory notations (discussed on pp. 1 95 9a. Chomsky. not on any presumed absolute sense of "simplicity" that might rank Tu and To relative to one another. the grammar can be justified on external grounds of descriptive adequacy . HaIIe. 1 964).we may ask whether it states the facts about the language cor rectly. and it is therefore justified on much deeper empirical grounds. In both cases. 1 959b. To repeat the major point: on the one hand. 26-27. it can be shown that this grammar is the highest-valued grammar permitted by the theory and com� patible with given primary linguistic data. or conversely. having to do with the properties of natural languages. In the latter case. whether it predicts correctly how the idealized native speaker would understand arbitrary sentences and gives a conect account of the basis for this achievement. We can therefore try to determine whether there are significant gen eralizations that are expressible in terms of one but not the other theory in the case of empirically given languages. on the other hand. as a theory of phonological processes (cf. Confusion about these questions may also have been engen dered by the fact that there are several different senses in which one can talk of "justifying" a grammar. as noted on pp. a grammar can be justified on internal grounds if. 1 964. and strong arguments have been offered to show that ToP is conect and T � is wrong. given an ex planatory linguistic theory. either result is possible. Or. sup posing TuP and ToP to be theories of the phonological component (where TuP holds phonological rules to be unordered and ToP holds them to be partially ordered). it is easy to invent hypo thetical "languages" for which significant generalizations are expressible in terms of ToP but not TuP. it is an entirely factual question. 1 962a. Both kinds of justification are of course necessary. a principled basis is presented for the construction of this grammar. 1 959a. it is im- .METHODOLOGICAL PRELIMINARIES equivalent and empirically distinguishable theories. the argument turns on the factual question of expressibiIity of linguistically significant generalizations in terms of one or the other theory. We shall see later that T08 is rather well motivated as a theory of the base. Failure to appreciate this fact has led to a great deal of vacuous and pointless discussion. In principle. 1 95 1 . If the linguist is content to formulate descriptions one way or another with little concern for justification.namely. In the case of a linguistic theory that is merely descriptive. since interest in j ustification has been abandoned. It is also apparent that the discussion as to whether an evaluation measure is a "necessary" part of linguistic theory is quite without substance (see. 24 It is only when all of the conditions (i)-(v) of (1 2)-( 1 4) are met that the deeper question of internal justification can be raised. he must concern himself with the problem of develop ing an explanatory theory of the form of grammar. choice of a grammar for a particular language L will always be much underdetermined by the data drawn from L alone. not to confuse them. On the other hand. however. Hence. Such a theory limits the choice of grammar by the dual method of imposing formal conditions on grammar and providing an evaluation procedure to be applied for the language L with which he is now concerned. since this provides one of the main tools for arriving at a descriptively adequate grammar in any particular case. if he wishes to achieve descriptive adequacy in his account of language structure. and if he does not intend to proceed from the study of facts about particular languages to an investigation of the characteristic properties of natural language as such. neither evidence nor argumen t (beyond minimal requirements of consistency) has any bearing on what the linguist presents as a linguistic description. In this case. only one kind of justification can be given . Moreover. any far-reaching concern for descriptive adequacy must lead to an attempt to develop an explanatory theory that fulfills these dual functions. and concern with ex- . successful grammars for other languages or successful fragments for other subparts of L) will be available to the linguist only if he possesses an explanatory theory. Both the formal conditions and the evaluation procedure can be empirically justified by their success in other cases. In other words. pp. then construction of an evaluation procedure and the associated concerns that relate to explanatory adequacy need not concern him.§ 7. we can show that it permits grammars that meet the external condition of descriptive adequacy. 36-37). ON EVALUATION PROCEDURES portant. other relevant data (namely. however. where the elements must appear in the order just given. Thus it is the notational conventions used in presenting a grammar that define "significant generalization. consider the analysis of the English Verbal Auxili ary. This is. we can state this rule in the following form: . The problem is to devise a procedure that will assign a numerical measure of valuation to a grammar in terms of the degree of linguistically significant generalization that this grammar achieves. choice of an evaluation measure constitutes a decision as to what are "similar processes" and "natural classes" . The major problem in constructing an evaluation measure for grammars is that of determining which generalizations about a language are significant ones. the rationale behind the conventions for use of parentheses. partially identical rules) about the whole set. For a detailed discussion of these. in fact. But if this is to be a meaningful measure. generative) grammars. Postal ( 1 962a). or when it can be shown that a "natural class" of items undergoes a certain process or set of similar processes.in short. Using familiar notational conventions. etc that have been adopted in explicit (that is. Past or Presen t)." if the evalua tion measure is taken as length. 1 955). brackets.• . The facts are that such a phrase must contain Tense (which is. so that real generalizations shorten the grammar and spurious ones do not. in terms of number of symbols.METHODOLOGICAL PRELIMINARIES planatory adequacy surely requires an investigation of evaluation procedures . and Matthews ( 1 964). We have a generalization when a set of rules about distinct items can be replaced by a single rule (or. more generally. it is necessary to devise notations and to restrict the form of rules in such a way that significant considerations of complexity and gen erality are converted into considerations of length. what are significant generalizations. an evaluation measure must be selected in such a way as to favor these. The obvious numerical measure to be applied to a grammar is length. Thus. and then may or may not contain a Modal and either the Perfect or Progressive Aspect (or both). To take just one example. see Chomsky ( 1 95 1 . furthermore. whereas cyclic regularities o f the type exemplified i n ( 1 7). are not of the type for which children will intuitively . has the following meaning. adoption of the familiar notational conventions involving the use of parentheses amounts to a claim forms in list that there is a the set of linguistically significant generalization underlying to the empirical hypothesis that regularities of the type ex emplified in ( 1 6) are those found in natural languages. ON EVALUATION PROCEDURES 43 ( 1 5) Aux � Tense (Modal) (Perfect) (Progressive) (omitting details that are not relevant here) . The parenthesis notation. these eight rules would involve twenty symbols. Tense---Per/ect. in both cases). Progressive In the case involved.Modal---Per/ect"""Progressive. Tense ""'Perfect---Progressive.ms. though perfectly genuine. Tense""'Modal""'Per/ect. It amounts of the type that children learning a language will expect . in this case. for example. the forms given in list ( 1 7) as the representatives of this category: ( 1 6) Tense. as compared with a language that has. list ( 1 7) cannot be abbreviated by this convention. Tense""'Modal""'Progressive. Tense""'Progressive. abstractly. twenty symbols are rule (1 5) by the notational convention. whereas rule ( 1 5) involves four (not counting Aux. It asserts that the difference between four and twenty symbols is a measure of the degree of linguistically significant generalization achieved in a language that has the forms given in list (16). Tense""'Per/ect. Tense--. and are (16) but not the set of forms in lis t ( 1 7) .Modal---Perfect. Per/ect""'Progressive'·"'Tense--.Modal---Perfect---Progressive ( 1 7) Tense--.§ 7.Modal. List of both list ( 1 6) and (16) abbreviates to list ( 1 7). Modal---Per/ect""'Pro Pro Modal-- gressive""'Tense. Hence. Tense""'Modal. and are much more difficul t for the language-learner to construct on the basis of scattered data language. for the Auxiliary Phrase. are not characteristic of natural search in language materials. Rule ( 1 5) is an abbreviation for eight rules that analyze the element Aux into its eight possible for. Stated in full. gressive---Tense--. One might easily propose a different notational convention that would abbreviate list ( 1 7) to a shorter rule than list ( 1 6). The extent to which particular parts of the grammar can be selected independently of others is an empirical matter about which very little is known. conclude from the existence of "yesterday John arrived" and "John arrived yesterday" that there is a third form "arrived yesterday John.44 METHODOLOGICAL PRELIMINARIES or to use. The effect of modifications on a larger . descriptions of particular subsystems of the grammar must be evaluated in terms of their effect on the entire system of rules." or from the existence of "is John here" and "here is John" that there is a third form "John here is. To my knowledge. at present. it s i mpl y embodies a factual claim about the structure of natural language and the predisposition of the child to search for certain types of regularity in natural language. Although alternatives can be clearly formulated. in the manner outlined earlier. the only attempt to evaluate a fairly full and complex subsystem of a grammar is in Chomsky ( 1 95 1)." etc. the language learner will construct the rule ( 1 5) generating the full set with their semantic interpretations.he will not. then. There is no a priori reason for preferring the usual convention. What is claimed. It is the full set of notational con ventions that constitute an evaluation procedure. The illustrative examples of the preceding paragraph must be regarded with some caution. Hence. he will not incorporate this "generalization" in his grammar . deeper studies of particular languages than are presently available are needed to settle the questions that immediately arise when these extremely important issues are raised. but even here all that is shown is that the value of the system is a "local maximum" in the sense that interchange of adjacent rules decreases value. thus making a different empirical assumption about what constitutes a linguistically significant generalization. for example. on the basis of given data. The factual con ten t of an explanatory theory lies in its claim that the most highly valued grammar of the permitted form will be selected. whereas when given scattered examples that could be subsumed under a cyclic rule. is that when given scattered examples from ( 1 6). § 7. a certain empirical claim is made. concerning natural language . 1 96 1 . as Halle shows. It is implied that a person learning a language will attempt to formulate generalizations that can easily be expressed (that is. It is clear. . 1 964). It is important to observe that the effectiveness of this parti cular evaluation measure is completely dependent on a strong assumption about the form of grammar. rather. These may be very strong claims. then. implicitly. and that he will select grammars containing these generalizations over other grammars that are also compatible with the given data but that contain different sorts of generaliza tion." and so on. namely. relating to lexical and phonological structure. 1 962a. If phonemic notation is allowed in addition to feature notation. and that determine the intuitively given distinction between "phonologically pos sible" and "phonologically impossible" nonsense forms. Certain aspects of the general question. and need by no means be true on any a priori grounds. if length is to be taken as the measure of valuation for a grammar. that choice of notations and other conventions is not an arbi tra ry or "merely technical" matter. the measure gives absurd consequences. To avoid any possible lingering confusion on this matter. A very plausible argument can be given to the effect that this convention defines the notions of "natural class" and "significant generalization" that have been relied on implicitly in descriptive and com parative-historical phonological investigations. the assumption that only feature notation is per mitted. see Halle ( 1 959a. different concepts of "natural class. One special case of this general approach to evaluation that has been worked out in a particularly convincing way is the condition of minimization of distinctive feature specifications in the phonological component of the grammar . When particular notational devices are in corporated into a linguistic theory of the sort we are discussing. 1 959b. For discussion. Halle and Chomsky (forthcoming). a matter that has immediate and perhaps quite drastic empirical consequences. are discussed in Halle and Chomsky (forthcoming). ON EVALUATION PROCEDURES 45 scale is not investigated. It is. with few symbols) in terms of the notations available in this theory. In brief. does not refer to conscious formulation and expression of these but rather to the process of arriving at an internal representation of a gen erative system. an assertion about a particular language is replaced by a corresponding assertion. about language in general.. In other words. linguistic theory may move toward explanatory adequacy and contribute to the study of human mental processes and intellectual capacity more specifically. In short. we are very far from being able to present a system of formal and substantive linguistic universals that will be sufficiently rich and detailed to account for the facts of language learning. I am making the obvious comment that. it is clear that no present-day theory of language can hope to attain explanatory adequacy beyond very restricted domains. this fact should become evident when its effect on the d es crip tion of other aspects of the language or the description of other languages is ascertained . There can be no doubt that present theories of grammar require modification in both of these ways. in general. thus enriching this theory and imposing more structure on the schema for grammatical description. from which the first follows. In this way. to attribute them to the general theory of linguistic structure. Whenever this is done. - . To advance linguistic theory in the direction of explanatory adequacy. to the determination of the abilities that make . we can attempt to refine the evaluation measure for grammars or to tighten the formal con straints on grammars so th a t it becomes more difficult to find a highly valued hypothesis compatible with primary linguistic data. wherever possible. which can be appropriately described in these terms. where ver possible. hypotheses. etc. being the more promising · Thus the most crucial problem for linguistic theory seems to be to abstract statements and gen eralizations from particular descriptively adequate grammars and. general assumptions about the nature of language shoul d be formulated from which particular features of the grammars of individual languages can be deduced. If this formulation of a deeper hypothesis is incorrect.METHODOLOGICAL PRELIMINARIES let me repeat once more that this discussion of language learning in terms of formulation of rules. the latter. and the theorist has the problem of determining the intrinsic properties of a device capable of mediating this input-output relation. of which the problem of acquisition of language is a special and particularly informative case. LING UIS TIC THEOR Y AND LANG UA GE LEARNING In the preceding discussion. certain principles of association. in our case. it assumes that the device has certain analytical data-processing mechanisms or inductive principles of a very elementary sort. UNGUISTIC THEORY AND LANGUAGE LEARNING 47 language learning possible under the empirically given limita tions of time and data. Beyond this. in recent versions. pp. certain problems of linguistic theory have been formulated as questions about the construction of a hypothetical language-acquisition device. 1 950). We may think of the theorist as given an empirical pairing of collections of primary linguistic data as sociated with grammars that are constructed by the device on the basis of such data.25 a set of primitive unconditioned reflexes (Hull. 1 960. the set of all "aurally distin guishable components" of the full "auditory impression" (Bloch.for example. The empiricist approach has assumed that the structure of the acquisi tion device is limited to certain elementary "peripheral proc essing mechanisms" .). in accordance with the Saussurian emphasis . an innate "quality space" with an innate "distance" defined on it (Quine. 83f. for example. 1 943).§ 8. § 8. Much information can be obtained about both the primary data that constitute the input and the grammar that is the "output" of such a device. It may be of some interest to set this discussion in a somewhat more general and traditional framework. taxonomic principles of segmentation and classification such as those that have been developed with some care in modem linguistics. weak principles of "generalization" involving gradi ents along the dimensions of the given quality space. we can distinguish two general lines of approach to the problem of acquisition of knowledge. Historically. in the case of language. or. This seems a use ful and suggestive framework within which to pose and consider these problems. or. A condition for innate mechanisms to become activated is that apprQpriate stimulation be presented. and the figures which arise from them. nothing reaches our mind from external objects through the organs of sense beyond certain corporeal movements . but even these move ments. for existence in any faculty is not actual but merely potential existence. . . on occasion of certain corporeal move ments. the innate ideas are those arising from the faculty of thinking rather than from external objects : . . . beyond this. beyond these voices or pictures. envisage these ideas. . sound and the like be innate. since the very word "fac ulty" designates nothing more or less than a potentiality. Hence it follows that the ideas of the movements and figures are themselves innate in us. that our mind may. . It is then assumed that a preliminary analysis of experience is provided by the peripheral processing mechanisms. [po 443]. . . Thus for Descartes ( 1 647 ). are presented to us by means of ideas which come from no other source than our faculty of thinking. and are accordingly together with that faculty innate in us. for they have no likeness to the corporeal movements .27 there are innate ideas and principles of various kinds that determine the form of the acquired knowledge in what may be a rather restricted and highly organized way. such notions as that things equal to the same thing are equal to each other are innate. So much the more must the ideas of pain. so that all these things that we think of. always existing in us potentially. . since they cannot arise as necessary principles from "particular movements: ' In general. sight . . and hearing nothing be yond voices or sounds. A rather different approach to the problem of acquisition of knowledge has been characteristic of rationalist speculation about mental processes. [Thus .26 Such views can be formulated dearly in one way or another as empirical hypotheses about the nature of mind. presents nothing beyond pictures.METHODOLOGICAL PRELIMINARIES on the fundamental character of such principles. colour. as being symbolized by them. Similarly. . that is. . are acquired by application of the available inductive principles to this initially analyzed ex perience . are not conceived by us in the shape they assume in the organs of sense . . The rationalist approach holds that beyond the peripheral processing mechanisms. and that one's concepts and knowledge. . [Book IV]. therefore. . 1 662). in others certain diseases like gout or gravel. LINGUISTIC THEORY AND LANGUAGE LEARNING 49 ideas are innate in the sense that] in some families generosity is innate. the same point of view is expressed in the following way: It is false." These notions are extensively developed throughout seventeenth-century rationalist philosophy . though these ideas have very rarely any resemblance to what takes place in the sense and in the brain. having no connection with any . In the Port-Royal Logic (Arnauld. To mention j ust one example. which though the cogitations of them be often oc casionally invited from the motion or appulse of sensible objects without made upon our bodies. . "we could have no experience at all nor be capable of observations". yet notwithstanding the ideas themselves could not possibly be stamped or impressed upon the soul from them. or to grasp any general nature . except on occasion of those movements which are made in the brain through sense. and even disappear and give no sign of their existence". [po 442]. Lord Herbert ( 1 624) maintains that innate ideas and principles "remain latent when their corresponding objects are not present. because sense takes no cognizance at all of any such things in those corporeal objects. Cudworth ( 1 7 3 1 ) gives an extensive argument in support of his view that "there are many ideas of the mind. . but because they are born with a certain disposition or propen sity for contracting them . Still earlier. they "must be deemed not so much the outcome of experience as principles without which we should have no ex perience at all . . as was pointed out by Leibniz and many com mentators since. the impulse from sense giving oc casion to the mind to form differen t ideas which it would not have formed without it. . . that all our ideas come through sense. On the con trary. . and there are at least a very great number of ideas which. it may be affirmed that no idea which we have in our minds has taken its rise from sense." Even in Locke one finds essentially the same conception. . [po 1 05]. "we should never come to distinguish between things. not that on this account the babes of these families suffer from these diseases in their mother's womb. [po 1 32].§ 8." Without these principles. and therefore they must needs arise from the innate vigour and activity of the mind itself . not even when we need it [po 77] . [p o 74]. . . [pp. . for which the senses furnish occasions. . i. In the same vein. particular or individual truths. (as. • • . [New Essays. without manifest absurdity. . p. so that we can find them there if we consider attentively and set in order what we already have in the mind [po 7 8]. although they have not invented this). . Thus it is that one possesses many things without knowing it . . [There are innate general princi· pIes that] enter into our thoughts. and therefore the basis for alphabetic writing. for example. whatever their number.e. do not suffice to establish the univenal necessity of that same truth . must have principles whose proof does not depend on examples. [Thus] all arithmetic and all geometry are in us virtually. the Chinese possess articulate sounds. Leibniz refuses to accept a sharp distinction between innate and learned: I agree that we learn ideas and innate truths either in considering their source or in verifying them through experience. . Now all the examples which confirm a general truth. and successful experience serves to confirm reason [p o 44]. • . yet nonetheless one learns them. be referred to sense . cannot.METHODOLOGICAL PRELIMINARIES • bodily image.28 either by drawing them from their source when we learn them through demonstrative proof (which shows that they are innate). nor consequently upon the testimony of the senses. since the senses never give us anything but examples. 75 ] . It is true that we must not imagine that these eternal laws of the reason can be read in the soul as in an open book but it is sufficient that they can be discovered in us by dint of at tention. • • • . [Chapter 1]. . . . . The mind leans upon these principles every moment. The truths of numb en are in us. . of which they form the soul and the connection. although we do not at all think of them. although without the senses it would never have occurred to us to think of them. . Necessary truths . The senses.] we have an infinite amount of knowledge of which we are not always conscious. 4 2-4S] . are not sufficient to give it all to us. . as do ordinary arithmeticians . because that demands great attention to its acts.. . . And I cannot admit this proposition: all that one learns is not innate. They are as necessary thereto as the muscles and sinews are for walking. [In general. but it does not come so easily to distinguish them and to represent them distinctly and separately. although necessary for all our actual knowl edge. . . or by testing them in exam ples. of drawing out what is innate in the mind. for the individual. the method of "experimental reasoning" is a basic instinct in animals and humans.29 This view contrasts sharply with the empiricist notion (the prevailing modern view) that language is essentially an ad ventitious construct. Like Leibniz. relatively independent in its structure of any innate mental faculties. for example. the art of incubation. in any event. 1 748. once "activated. and the function of experience is to cause this general . or built up by elementary "data-processing" procedures (as modern linguistics typically maintains). is to a large ex tent given. empiricist speculation has characteristically assumed that only the procedures and mechanisms for the acquisition of knowledge constitute an innate property of the mind. On the other hand. Humboldt ( 1 836) concludes that one cannot really teach language but can only present the conditions under which it will develop spontaneously in the mind in its own way. by Skinner or Quine) or by drill and explicit explanation (as was claimed by Wittgenstein). on a par with the instinct "which teaches a bird. and the whole economy and order of its nursery" . Thus the form of a language. that is. however. though it will not be available for use without ap propriate experience to set the language-forming processes into operation." are then available for interpretation of the data of sense in a way in which they were not previously. learning is largely a matter of Wiederer zeugung. no sharp distinction is made between perception and acquisition. but. rationalist speculation has assumed that the general form of a system of knowledge is fixed in advance as a disposition of the mind.§ 8. is otherwise quite free. incidentally. he reiterates the Platonistic view that. with such exactness. the schema for its grammar. § IX). although there would be no inconsistency in the assumption that latent innate mental structures. In short. The form of knowledge. that throughout these classical discussions of the interplay between sense and mind in the formation of ideas. Thus for Hume. LINGUISTIC THEORY AND LANGUAGE LEARNING Notice. taught by "conditioning" (as would be maintained.it is derived "from the original hand of nature" (Hume. Applying this rationalist view to the special case of language learning. New Es says. truths would be in us as the figure of Hercules is in the marble. rather than a block of marble wholly even.. habits. For if the soul resembled these blank tablets. data processing approach of modem linguistics as an empiricist view that contrasts with the essentially rationalist alternative proposed in recent theories of transformational grammar. often insensible. and not as actions. Taxonomic linguistics is empiricist in its assumption that general linguistic theory consists only of a body of procedures for determining the grammar of a language from a corpus of data. the form of language being unspecified except insofar as restrictions on possible grammars are determined by this set of procedures. the comparison of a block of marble which has veins. or of blank tablets. although these potentialities are always accompanied by some actions. Particular empiricist and rationalist views can be made qui te precise and can then be presented as explicit hypotheses about acquisition of knowledge. pp. to clear them by polishing. in particular. dispositions. To follow Leibniz's enlightening analogy. this block would be more determined thereto. whi ch correspond to them [Leibniz. Bo . as inclinations. it is historically ac curate as well as heuristically valuable to distinguish these two very different approaches to the problem of acquisition of knowledge. about the innate structure of a language-acquisition device. If we interpret taxonomic linguistics as making an empirical claim. we may make . when the marble is wholly indifferent to the reception of this figure or some other. and Hercules would be in it as in some sense innate. necessary to assume that empiricist and rationalist views can always be sharply distinguished and that these currents cannot cross. although it would be needful to labor to discover these veins. In fact. it would not be inaccurate to describe the taxonomic. i. . But if there were veins in the block which should i ndicate the figure of Hercules rather than other figures. 45-46J.METHODOLOGICAL PRELIMINARIES schematic structure to be realized and more fully differentiated. or natural potentialities. . of course.e. of what is called among philosophers a tabula rasa. Thus it is that ideas and truths are for us innate. and by cutting away what prevents them from appearing. It is not. Nevertheless. for example. In contrast. we must bear in mind that any concrete empi ri cis t proposal does impose certain conditions on the form of the grammars that can result from application of its inductive principles to primary data. must therefore be regarded as a specific hypothesis. it cannot.in other words. it would have no bearing on what is completely a factual issue. as an empirical question. in part. Where empiricist and rationalist views have been presented with sufficient care so that the question of correctness can be seriously raised. we may ask. these provi ding a schema that is applied to data and that determines in a highly restricted way the general form and. 1 964) and Katz (forthcoming) for some further discussion of this point. In particular. in principle.§ 8. We may therefore ask whether the grammars that these principles can provide. are at all close to those which we in fact discover when we investigate . There is no a priori way to settle this issue. as to the nature of mental structures and processes. the discussion of language acquisition in preceding sections was rationalistic in its assumption that various formal and sub stantive universals are intrinsic properties of the language-acqui sition system. and elaborated in more detail in the following chapters and in other studies of transformational grammar. that the set of procedures can be regarded as constituting a hypothesis about the innate language-acquisition system. even the substantive features of the grammar that may emerge upon presentation of appropriate data. UNGUISTIC THEORY AND LANGUAGE LEARNING 53 this claim must be that the grammars that result from application of the postulated procedures to a sufficiently rich selection of data will be descriptively adequate . When such constrasting views are clearly formulated. one way or the other. This factual question can be approached in several ways. which (if either) is correct. be maintained that in any clear sense one is "simpler" than the other in terms of its potential physical realization. restricting ourselves now to the question of language acquisition. A general linguistic theory of the sort roughly described earlier. See Chomsky ( 1 959b. of an essentially rationalist cast.3 1 and even if this could be shown. 1 962 b. and many other publications for discussion of these questions that seems un answerable and is.54 METIlODOLOGICAL PRELIMINARIES real languages. Katz and Postal. 1 960. for the moment. quite apart from any question of feasibility. we can then turn to the question of feasibility: can the inductive procedures (in the empiricist case) or the mechanisms of elaboration and realization of innate schemata (in the rationalist case) succeed in producing grammars within the given constraints of time and access. and to offer at least some hope of providing a hypothesis about the intrinsic structure of a language-acquisition system that will meet the condition of adequacy-in-principle and do so in a sufficiently . It seems to have been demonstrated beyond any reasonable doubt that. since study of the first question has been sufficient to rule out whatever explicit proposals of an essentially empiricist character have emerged in modem discussions of language acquisition. 145-1 48. 1 957. and within the range of observed uniformity of output? In fact. proved to be a useful way to subject such hypotheses to one sort of empirical test. the rationalist approach ex emplified by recent work in the theory of transformational grammar seems to have proved fairly productive. I g64a. then. in either case. p. and that further empiricist speculations have been quite empty and uninformative. The only proposals that are explicit enough to support serious study are those that have been developed within taxonomic linguistics. 1 964. On the other hand. Miller. 1 963. and Pribram. This has. I g64c. 1 956. Galanter. 1£ the answer to this question of adequacy-in-principle is positive. 430. not challenged). to be fully in accord with what is known about language. I g62 b. methods of the sort that have been studied in taxonomic linguistics are intrinsically incapable of yielding the systems of grammatical knowledge that must be attributed to the speaker of a language (cf. pp. In general.5. I g64' § 5. in the past. the second question has rarely been raised in any serious way in connection with empiricist views (but cf. Chomsky. for some comments). The same question can be asked about a concrete rationalist proposal. Postal. and Miller and Chomsky. it seems to me correct to say that empiricist theories about language acquisition are refutable wherever they are clear. It is impossible.§ 8. to formulate as a transformation such a simple operation as reflection of an arbitrary string (that is. implies that only certain kinds of symbolic systems can be acquired and used as languages by this device. where each a. according to the theory of transformational grammar. Systems can certainly be invented that fail the conditions. Specifically.l)th word with the 2 n th word throughout a string of . or interchange of the ( 2n . As a concrete example. is a single symbol. however. have no a priori justification. while many of the operations that do qualify are far from elementary. LINGUISTIC THEORY AND LANGUAGE LEARNING 55 narrow and interesting way so that the question of feasibility can. For example. In principle.aft. the permitted operations cannot be shown in any sense to be the most "simple" or "elementary" ones that might be invented. One might seek other ways of testing particular hypotheses about a language-acquisition device. that have been proposed as tentative linguistic uni versals in. furthermore. by aft " ' al) . only certain kinds of formal operations on strings can appear in grammars . for the first time. A theory that attributes possession of certain linguistic universals to a language-acquisition system. and do fall beyond the domain for which the language acquisition system is designed.operations that. as a property to be realized under appropriate external conditions. in any general sense. independently of what the length or internal complexity of the strings belonging to these categories may be. grammatical transformations are necessarily "struc ture-dependent" in that they manipulate substrings only in terms of their assignment to categories. be seriously raised. replacement of any string a l " . consider the fact that. Jakobsonian distinctive-feature theory or the theory of transformational grammar. In fact. Others should be beyond its language-acquisition capacity. what might in general be considered "elementary operations" on strings do not qualify as grammatical transformations at all. Thus it is possible to formulate a transformation that can insert all or part of the Auxiliary Verb to the left of a Noun Phrase that precedes it. formal and sub stantive. for example. one might try to determine whether invented systems that fail these condi tions do pose inordinately difficult problems for language learn ing. for example. the faculte de langage is only one of the faculties of the mind. short of listing all categories of the grammar in the structural analysis). see Miller and Stein ( 1 963). What one would expect. we do not imply that this system cannot be mastered by a human in some other way. can be deduced from any sufficiently explicit theory of linguistic universals that is attributed to a language-acquisition device as an intrinsic property. or interchange of odd and even words. For some initial approaches to the very difficult but tantalizing problem of investigating ques tions of this sort. or an operation that will iterate a symbol that belongs to as many rightmost as leftmost categories). The language-acquisition device is only one component of the total system of intellectual structures that can be applied to problem solving and concept formation. however. is that there should be a qualitative difference in the way in which an organism with a functional language-acquisition system82 will approach and deal with systems that are languagelike and others that are not. Many other such predictions. in other words. Miller and Norman ( 1 964). or insertion of a symbol in the middle of a string of even length. as suggested later. it will be impossible to formulate many "structure-dependent" operations as transformations . an operation that will iterate a symbol that is the left mOst member of a category (impossible. by interchanging the order of certain categories (as in English). or insertion of a marker in the middle of the sentence. if treated as a puzzle or intellectual exercise of some sort. Hence. for example. it could not form inter rogatives by reflection. one who proposes this theory would have to predict that although a language might form interrogatives. none of them at all obvious in any a priori sense. Similarly.METHODOLOGICAL PREUMINARIES arbitrary length. Notice that when we maintain that a system is not learnable by a language-acquisition device that mirrors human capacities. The problem of mapping the intrinsic cognitive capacities of . if the structural analyses that define transformations are restricted to Boolean conditions on A na lyzability. such speculations have not provided any way to account for or even to express the fundamental fact about the normal use of language. some incidental information has been obtained about the effect of intrinsic cognitive structure and intrinsic organization of behavior on what is learned. The general question and its many ramifications. Learning theory has. indirect. The sporadic exceptions to this observation (see.§ 8. psy chology. Consequently. the discussion of "instinctual drift" in Breland and Breland. as are many ethological studies of lower organisms.tasks that must be approached in a devious. remain in a primitive state.) of any sort that have yet been developed within linguistics. namely the question of species-independent regularities in acquisition of items of a "behavioral repertoire" under experimentally manipulable conditions. This evidence shows clearly that taxonomic views of linguistic structure are inadequate and that knowledge of grammatical structure cannot arise by application of step-by step inductive operations (segmentation. LINGUISTIC THEORY AND LANGUAGE LEARNING 57 an organism and identifying the systems of belief and the organization of behavior that it can readily attain should be central to experimental psychology. association. the field has not developed in this way. Further empiricist speculations contribute nothing that even faintly suggests a way of overcoming the intrinsic limitations of the methods that have so far been proposed and elaborated. However. namely the speaker's ability to produce and understand instantly new . etc. In brief. for the most part. and piecemeal fashion. filling of slots in frames. it seems clear that the present situation with regard to the study of language learning is essentially as follows . concentrated on what seems a much more marginal topic. We have a certain amount of evidence about the character of the generative grammars that must be the "output" of an acquisition model for language. for example. substitu tion procedures. but this has rarely been the focus of serious attention (outside of ethology). classification. however. 1 96 1 ) are quite suggestive. it has necessarily directed its attention to tasks that are extrinsic to an organism's cognitive capacities . or philosophy. In the course of this work. In particular. the striking uniformity of the resulting grammars. innate structure rich enough to account for the fact that grammatical knowledge is attained on the basis of the evidence available to the learner. Consequently.many of the concepts and principles of which are only remotely related to experience by long and intricate chains of unconscious quasi-inferential steps. The real problem is that of developing a hypothesis about initial structure that is sufficiently rich to account for acquisition of language. over wide ranges of variation. the empiricist effort to show how the assumptions. about a language·acquisition device can be reduced to a conceptual minimum38 is quite mis placed. the degenerate quality and narrowly limited extent of the available data. and emotional state. It seems plain that language acquisition is based on the child's discovery of what from a formal point of view is a deep and abstract theory a generative grammar of his language . for the present. nor obtainable from them by any sort of "gen eralization" known to psychology or philosophy. These preconceptions are not only quite implausible. It is. impossible to formulate an assumption about initial. but are without factual support and are hardly consistent with what little is known about how animals or humans construct a "theory of the external world." It is clear why the view that all knowledge derives solely from the senses by elementary operations of association and "gen- . and their independence of intelligence. nor associated with those previously heard by conditioning. motivation. It is a matter of no concern and of only historical interest that such a hypothesis will evidently not satisfy the preconceptions about learning that derive from centuries of empiricist doctrine. yet not so rich as to be inconsistent with the known diversity of language. to begin with. leave little hope that much of the structure of the language can be learned by an organism initially uninformed as to its general character.METHODOLOGICAL PRELIMINARIES sentences that are not similar to those previously heard in any physically defined sense or in terms of any notion of frames or classes of elements. A consideration of the character of the grammar that is acquired. § 8. LINGUISTIC THEORY AND LANGUAGE LEARNING 59 eralization" should have had much appeal in the context of eighteenth-century struggles for scientific naturalism. However, there is surely no reason today for taking seriously a position that attributes a complex human achievement entirely to months (or at most years) of experience, rather than to millions of years of evolution or to principles of neural organization that may be even more deeply grounded in physical law - a position that would, furthermore, yield the conclusion that man is, apparently, unique among animals in the way in which he acquires knowl edge. Such a position is particularly implausible with regard to language, an aspect of the child's world that is a human creation and would naturally be expected to reflect intrinsic human capacity in its internal organization. In short, the structure of particular languages may very well be largely determined by factors over which the individual has no conscious control and concerning which society may have little choice or freedom. On the basis of the best information now available, it seems reasonable to suppose that a child cannot help constructing a particular sort of transformational grammar to account for the data presented to him, any more than he can control his perception of solid objects or his attention to line and angle. Thus it may well be that the general features of language structure reflect, not so much the course of one's experience, but rather the general character of one's capacity to acquire knowl edge-in the tradit.ional sense, one's innate ideas and innate principles. It seems to me that the problem of clarifying this issue and sharpening our understanding of its many facets provides the most interesting and important reason for the study of descriptively adequate grammars and, beyond this, the formulation and justification of a general linguistic theory that meets the condition of explanatory adequacy. By pursuing this investigation, one may hope to give some real substance to the traditional belief that "the principles of grammar form an im portant, and very curious, part of the philosophy of the human mind" (Beattie, 1 788). 60 METHODOLOGICAL PRELIMINARIES § 9. GENERA TIVE CAPA CITY AND ITS LING UISTIC RELE VANCE It may be useful to make one additional methodological obser vation in connection with the topi cs discussed in the last few sec tions. Given a descriptive theory of language structure,84 we can distinguish its weak generative capacity from its strong genera tive capacity in the following way. Let us say that a grammar weakly generates a set of sentences and that it strongly generates a set of structural descriptions (recall that each structural de scription uniquely specifies a sentence, but not necessarily con versely), where both weak and strong generation are determined by the procedure f of ( 1 2iv) = (1 3iv) = ( 14iv). Suppose that the linguistic theory T provides the class of grammars Gl, G2,· · · , where G, weakly generates the language L, and strongly generates the system of structural descriptions I,. Then the class { Ll' L2, · } constitutes the weak generative capacity of T and the class {Il, I2, · · · } constitutes the strong generative capacity of T.8G The study of strong generative capacity is related to the study of descriptive adequacy, in the sense defined. A grammar is de· scriptively adequate if it strongly generates the correct set of structural descriptions. A theory is descriptively adequate if its strong generative capacity includes the system of structural descriptions for each natural language; otherwise, it is descrip tively inadequate. Thus inadequacy of strong generative capacity, on empirical grounds, shows that a theory of language is seriously defective. As we have observed, however, a theory of language that appears to be empirically adequate in terms of strong genera tive capacity is not necessarily of any particular theoretical in terest, since the crucial question of explanatory adequacy goes beyond any consideration of strong generative capacity. The study of weak genera tive capacity is of rather marginal linguistic interest. It is important only in those cases where some proposed theory fails even in weak generative capacity - that is, where there is some natural language even the sentences of which cannot be enumerated by any grammar permitted by this theory. In fact, it has been shown that certain fairly elementary • • § 9· GENERATIVE CAPACITY AND ITS LINGUISTIC RELEVANCE 61 theories (in particular, the theory of context-free phrase-structure grammar and the even weaker theory of finite-state grammar) do not have the weak generative capacity required for the descrip tion of natural language, and thus fail empirical tests of ade quacy in a particularly surprising way.B8 From this observation we must conclude that as linguistic theory progresses to a more adequate conception of grammatical structure, it will have to permit devices with a weak generative capacity that differs, in certain respects, from that of these severely defective systems. It is important to note, however, that the fundamental defect of these systems is not their limitation in weak generative capacity but rather their many inadequacies in strong generative capacity. Postal's demonstration that the theory of context-free grammar (simple phrase-structure grammar) fails in weak generative capacity was preceded by over a half-dozen years of discussion of the strong generative capacity of this theory, which showed con clusively that it cannot achieve descriptive adequacy. Further more, these limitations in strong generative capacity carry over to the theory of context-sensitive phrase-structure grammar, which probably does not fail in weak generative capacity. Presumably, discussion of weak generative capacity marks only a very early and primitive stage of the study of generative gram mar. Questions of real linguistic interest arise only when strong generative capacity (descriptive adequacy) and, more important, explanatory adequacy become the focus of discussion. As observed earlier, the critical factor in the development of a fully adequate theory is the limitation of the class of possible grammars. Clearly, this limitation must be such as to meet empirical conditions on strong (and, a fortiori, weak) generative capacity, and, furthermore, such as to permit the condition of explanatory adequacy to be met when an appropriate evaluation measure is developed. But beyond this, the problem is to impose sufficient structure on the schema that defines "generative gram mar" so that relatively few hypotheses will have to be tested by the evaluation measure, given primary linguistic data. We want the hypotheses compatible with fixed data to be "scattered" in value, so that choice among them can be made relatively easily. METHODOLOGICAL PRELIMINARIES This requirement of "feasibility" is the major empirical con straint on a theory, once the conditions of descriptive and ex planatory adequacy are met. It is important to keep the require ments of explanatory adequacy and feasibility in mind when weak and strong generative capacities of theories are studied as mathematical questions. Thus one can construct hierarchies of grammatical theories in terms of weak and strong generative capacity, but it is important to bear in mind that these hierarchies do not necessarily correspond to what is probably the empirically most significant dimension of increasing power of linguistic theory. This dimension is presumably to be defined in terms of the scattering in value of grammars compatible with fixed data. Along this empirically significant dimension, we should like to accept the least "powerful" theory that is empirically adequate. It might conceivably turn out that this theory is extremely powerful (perhaps even universal, that is, equivalent in genera tive capacity to the theory of Turing machines)87 along the dimension of weak generative capacity, and even along the dimension of strong generative capacity. It will not necessarily follow that it is very powerful (and hence to be discounted) in the dimension which is ultimately of real empirical significance. In brief, mathematical study of formal properties of grammars is, very likely, an area of linguistics of great potential. It has already provided some insight into questions of empirical interest and will perhaps some day provide much deeper insights. But it is important to realize that the questions presently being studied are primarily determined by feasibility of mathematical study, and it is important not to confuse this with the question of empirical significance. 2 Categories and R e latio ns in S yntactic Theory WE now return to the problem of refining and elaborating the sketch (in Chapter 1 , § 3) of how a generative grammar is organized. Putting off to the next chapter any question as to the adequacy of earlier accounts of grammatical transformations, we shall consider here only the formal properties of the base of the syntactic component. We are therefore concerned primarily with extremely simple sentences. The investigation of generative grammar can profitably begin with a careful analysis of the kind of information presented in traditional grammars. Adopting this as a heuristic procedure, let us consider what a traditional grammar has to say about a simple English sentence such as the following: § THE SCOPE OF THE BA SE I. ( 1 ) sincerity may frighten the boy Concerning this sentence, a traditional grammar might provide information of the following sort: (2) (i) the string ( 1 ) is a Sentence (S) ; frighten the boy is a Verb Phr;tse (VP) consisting of the Verb (V) frigh ten and the Noun Phrase (NP) the boy; sincerity is also an NP; the NP the boy consists of the Determiner (Det) the, followed by a Noun (N) ; the NP sincerity consists of just an N; the is, furthermore, an Article (Art); may is a Verbal Auxiliary (Aux) and, furthermore, a Modal (M). 63 and the V frighten as its Main Verb. . and one that does not freely permit Object dele tion (as distinct from read. equivalently. 1 .I. the grammatical relation Subject-Verb holds of the pair (sincerity. frighten is a Transitive Verb (as distinct from occur). substantially correct and is essential to any account of how the language is used or acquired. 2. by a tree·diagram such as (3) . Categorization The remarks given in (2i) concern the subdivision of the string (1) into continuous substrings. each of which is assigned to a certain category. The interpretation of such a diagram is transparent. The main topic I should like to consider is how information of this sort can be formally presented in a structural description. 1 (iii) the N boy is a Count Noun (as distinct from the Mass Noun butter and the Abstract Noun sincerity) and a Common Noun (as distinct from the Proper Noun John and the Pronoun it) . or. it allows Abstract Subjects (as distinct from eat. and (2iii).2. eat) . whereas the VP frighten the boy functions as the Pred icate of this sentence. frighten). wear) . § 2. and the grammatical relation Verb Object holds of the pair (frighten. It seems to me that the information presented in (2) is.64 CATEGORIES AND RELATIONS IN SYNTACTIC 1HEORY (ii) the NP sincerity functions as the Subject of the sentence ( 1). it is. the NP the boy functions as the Object of the VP. 2 . ASPECTS OF DEEP STR UC T URE § 2. the boy) . admire) and Human Ob jects (as distinct from read. and how such structural descriptions can be generated by a system of explicit rules. The next three subsections (§§ 2 . an Animate Noun (as distinct from book) and a Human Noun (as distinct from bee). respectively. it takes Progressive Aspect freely (as distinct from know. own). Information of this sort can be represented by a labeled bracketing of ( 1 ). furthermore. (2ii).3) discuss these questions in connection with (2i). with out question. we must ask whether the formatives and category symbols used in Phrase-markers have some language independent characterization. consequently.§ 2. The formatives. except possibly for the. furthermore. That is. or whether they are just con venient mnemonic tags. I shall assume that . boy) and grammatical items (Perfect. none of these are represented in the simplified example given). If one assumes now that ( I ) is a basic string. the lexical forma tives are selected in a well-defined way from a fixed universal set. etc. that an appropriate phonological theory of this sort is established and that. can be subdivided into lexical items (sincerity. henceforth. in effect. A question arises at once as to the choice of symbols in Phrase markers.) and category symbols (S. The question of substantive representation in the case of the grammatical formatives and the category symbols is. A grammar that generates simple Phrase-markers such as (3) may be based on a vocabulary of symbols that includes both form atives (the. the traditional question of universal grammar.). specific to a particular grammar. V. in fact. etc. boy. ASPECTS OF DEEP STRUCTURE (3) s NP � I Aux M V VP N � De t NP sincerity I may I frighten I � the I N boy I and has been discussed frequently elsewhere. Possessive.. etc. the theory of phonetic distinctive features taken together with the full set of conditions on phonological representation does. In the case of the lexical formatives. give a language independent significance to the choice of symbols. though it is by no means a trivial problem to establish this fact (or to select the proper universal set of substantive phonetic features). the structure represented as (3) can be taken as a first approximation to its (base) Phrase-marker. NP. I shall assume. Given a grammar. A rewriting rule is a rule of the form (4) A � Z/X - Y where X and Y are (possibly null) strings of symbols. and Z is a nonnull string of symbols. This rule is interpreted as asserting that the category A is realized as the string Z when it is in the environment consisting of X to the left and Y to the right. I shall later suggest several general definitions that appear to be correct for English and for other cases with which I am acquainted. . we say that a sequence of strings is a W-derivation of V if W is the first and V the last string in the sequence. although this assumption will actually have no significant effect on any of the descriptive material to be presented . until the final string of the derivation consists only of formatives and therefore no further rewriting is possible. XZY . we say that a W-derivation of V is terminated. . Where V is a string of formatives. . and I is the boundary symbol (regarded as a grammatical formative). However. universal vocabulary. XAY . Application of the rewriting rule (4) to a string . . where S is the designated initial symbol of the grammar (representing the category "Sentence"). . .2 The natural mechanism for generating Phrase-markers such as (3) is a system of rewriting rules. Thus we construct a derivation of a terminal string by successively applying the rewriting rules of the grammar.66 CATEGORIES AND RELATIONS IN SYNTACTIC THEORY these elements too are selected from a fixed. converts this to the string . given its derivation. beginning with the string IS/. If several other conditions are imposed on the system of rewriting rules. . . . There is no reason to doubt the importance or reasonableness of the study of this question .S it is easy to provide a simple method for assigning a unique and appropriate Phrase marker to a terminal string. and each stting of the sequence is derived from the one preceding it by application of one of the rewriting rules (with an ordering condition to be added later). Thus a system . It is generally held to involve extrasyntactic considerations of a sort that are at present only dimly perceived . This may very well be true . We call V a terminal string if there is an IS I-derivation of lVI. A is a single category symbol. if these are reformulated as explicit systems for presenting grammatical information (but see note 30. It seems clear that certain kinds of grammatical information are presented in the most natural way by a system of rewriting rules. and it has also been shown that almost all of the nontransformational syntactic theories that have been developed within modern lin guistics. then no rule Rk (for k > . 60 £. X and Y are null. it must be the case that if rule R. Furthermore. and that the sequence 118'11. Thus.§ 2... IIXmfl is a derivation of the terminal string Xm• For this to be a sequential derivation. can serve as a part of a generative grammar. is called a constituen t structure grammar (or phrase structure grammar). lXI 'I. and we may therefore conclude that rewriting rules consti tute part of the base of the syntactic component. 208). furthermore.4 and I shall not discuss the issue here. applied in the manner described loosely here (and precisely elsewhere). . we shall assume that these rules are arranged in a linear sequence. We stipulate now that only sequential derivations are generated by the sequence of rules constituting this part of the base. fall within this framework. Chapter 1 ) . ASPECTS OF DEEP STRUCTURE of rewriting rules. The inadequacy of such systems as grammars for natural languages seems to me to have been established beyond any reasonable doubt. called con text-free (or simple) if in each rule of the form (4)." (for I < J) from IXI_I I. so that the rules apply independently of context. R. ' ' ' . such a system is apparently what is implicit in modern taxonomic ("structuralist") grammars. p ure or applied. was used to form line IXJ I from the line that precedes it. and shall define a sequen tial derivation as a derivation formed by a series of rule applications that preserves this ordering. suppose that the grammar consists of the sequence of rules Rt.CI To provide a Phrase-marker such as (3). the base component might contain the following sequence of rewriting rules : " ' . appropriately constrained. The grammar is. the formal properties of constituent structure grammars have been studied fairly intensively during the past few years. An un ordered set of rewriting rules. In fact.) can have been used to form a line "X. As noted earlier (pp. we shall see in § 2 . by extending the ." "Verb.2. This is a problem to whicli we shall turn in § 2 ." as distinct from the notion "NP. we see quite clearly how it is to be formally represented. Functional notions Turning now to (2ii). then." a distinction that is not to be obscured by the occasional use of the same term for notions of both kinds. Functional notions like "Subject. will also generate such deviant strings as boy may frigh ten the sincerity. we can immediately see that the notions in question have an entirely different status . Thus it would merely confuse the issue to attempt to represent the information presented in (2ii) formally . an inherently relational notion. It is. • § . In fact. and that it is (functions as) the Su bject-of the sentence (not that it is a Subject)." designates a grammatical function rather than a grammatical category . 3 . We say.2 . The notion "Sub ject. although they do suffice to generate (3)." "Predicate" are to be sharply distinguished from categorial notions such as "Noun Phrase.68 CATEGORIES A ND RELATIONS I N SYNTACTIC THEORY (5) (I) S-') NP""'Aux'""'VP VP -') V'""'N P N P -') Det'""' N NP -') N Det -') the Aux -') M (ll) M -') may N -') sincerity N -') boy V -') frigh ten Notice that the rules ( 5). that in ( I ) sincerity is an NP (not that it is the NP of the sentence). In the case of the information in ( 2 i).3 that it is necessary to distinguish these sets and to assign the lexical rules to a distinct subpart of the base of the syntactic component. There is a natural distinction in (5) between rules that in troduce lexical formatives (class (ll» and the others . in other words. in traditional terms. and what sorts of rules are required to generate these representations. a single Object. in this way. in which there is only a single Subject. being relational. we can regard any rewriting rule as defining a set of grammatical functions." for English. For one thing. and thus fails to express the relational character of the functional notions. the relational information can be supplied. only some of which (namely. Main-Verb. as the relation holding between the NP of a sentence of the form NP'""'Aux"""VP and the whole sentence. with explicit names. More generally. and no new rewriting rules are required to introduce them. adding the necessary rewntmg rules to (51) . It is necessary only to make explicit the relational character of these notions by defining "Subject-of. it confuses categorial and functional notions by assigning categorial status to both. those that involve the "higher-level.§ 2. This approach is mistaken in two ways.6 "Object-of" as the relation between the NP of a VP of the form V"""N P and the whole VP. since the notions Subject. For another. tradi tionally. . ASPECTS OF DEEP STRUCTURE 69 (6) Subject NP � Aux M S I I Predicate VP I N I sincerity I may I Main Verb V � I I Object NP I frigh ten Det the � I N boy I Phrase-marker (3) to (6)." more abstract grammatical categories) have been provided. are already represented in the Phrase-marker (3). it fails to observe that both (6) and the grammar on which it is based are redundant. In this case. etc. Predicate. and a single Main Verb. and Object. The fundamental error of regarding functional notions as categorial is somewhat masked in such examples as ( 6). correspondingly. each sentence of ( 7) will have a basis consisting of a sequence of base Phrase-markers. John is simultaneously Object. in (7b). John is simultaneously Object-of disturb. To develop a uniform approach to this ques tion. That is.7 Given a Phrase-marker of the terminal string W. for some Y. note 32). let us consider the problem of presenting information about grammatical function in an explicit and adequate way. Examples of this sort. Returning now to the main question. A] where B is a category and X = YBZ. including in particular the rule (8) A �x Associated with this rule is each grammatical function (9) [B. several by the same phrase: (7) (a) John was persuaded by Bill to leave (b) John was persuaded by Bill to be examined (c) what disturbed John was being regarded as incompetent In (7a). Z (possibly null). Suppose that we have a sequence of rewriting rules. the Subject with respect to the surface structure (cf. But consider such sentences as (7). that is.CATEGORIES AND RELATIONS IN SYNTACTIC THEORY intuitively. by the reader. which is the so-called "grammatical" Subject of the Sentence. of course. restricting ourselves now to base Phrase-markers. the impossibility of a categorial interpretation of functional notions becomes at once apparent. we may proceed as follows. In both (7a) and (7 b).of persuade ( to leave) and Subject-of leave. rather than John. in (7c). Object-of regard (as incom petent). each of which represents some of the semanti cally relevant information concerning grammatical function. in which many grammatical functions are realized. we say . provide the primary motivation and empirical j ustifica tion for the theory of transformational grammar. such as (5). and Subject-of the predication as incompetent . John is simultaneously Object-of per suade (to be examined) and Object-of examine. the deep structure in which the significant grammatical functions are represented will be very different from the surface structure. In such cases as these. Bill is the ("logical") Subject-of the Sentence. S] to sincerity may frigh ten the boy. we can now say that with respect to the Phrase marker (3 ) generated by the rules (5). With these definitions. S] Direct-Object-of: [NP. S] to sincerity may frighten the boy. These definitions must be thought of as - .§ 2. we should have the result that sincerity bears the relation [NP. A] to the substring V of W if V is dominated by a node labeled A which directly dominates YBZ.. S] Predicate-of: [VP. from the grammar (5) itself. frigh ten the boy bears the relation [VP. VP] Main-Verb-of: [V. that is. VP] to frighten the boy. VP] - I � I v In this case. S uppos e further that we propose the following general defini tions: ( l l ) (i) (ii) (iii) (iv) Subject-of: [NP. given the Phrase-marker ( 1 0) A W '= . and U is dominated by this occurrence of B . ASPECTS OF DEEP STRUCTURE that the substring U of W bears the grammatical relation [B.8 Thus the Phrase-marker in question contains the subconfiguration ( 1 0) . VP] to frighten the boy. l1l particular. and the boy is the D irect-Objec t of the Verb Phrase frigh ten the boy and frighten is its Main-Verb. sincerity is the S u bject of the sentence sincerity may frigh ten the boy and frighten the boy is its Predicate. the boy bears the relation [NP. and frighten bears the relation [V. the information presented in the redundant representation (6) is derivable directly from ( 3).. /\ /\ /\ Y � B Z (3) generated by the rules (5). so that examine. Object of is defined by both (iii) and (v). the basis for (7a) will contain base Phrase-markers for the strings Bill pers'/UJ. Chapter 1 ) .CATEGORIES AND RELATIONS IN SYNTACTIC THEORY belonging to general linguistic theory.ded John Sentence. But notice that "President" is not the Object-of John is President. John will leave) (John will leave) (examine Bill) (persuade Bill that John left) ( be President) (President) (leave) Then Subject-of is defined by both (i) and (ii). Notice that the same grammatical function may be defined by several different rewriting rules of the base. leave. Thus suppose that a grammar were to contain the rewriting rules ( 1 2) (i) S (ii) (iii) (iv) (v) (vi) (vii) Adverbial""'NP""' Aux""'VP S � NP""'Aux""'VP VP � V""'NP VP � V VP � V""'NP""'Sentence VP � Copula""'Predicate Predicate � N � (Naturally. (iv).of the sentences accompanying both (i) and (ii). John left. and these base Phrase·markers present the semantically relevant functional in formation exactly as in the case of (3). ( 14iv) in § 6. In such examples as (7). ( 1 3iv). For example (details aside). given a grammar (the procedure f of ( 1 2iv). and (v). i f the rules are as in ( 1 2). It is definitions of this sort that were presupposed in the discussion of persuade and expect in Chap ter I. though these grammatical functions are not represented in the configurations of the surface structures in these cases. the grammatical functions will also be given directly by the system of rewriting rules that generate the base Phrase-markers that underlie these sentences. § 4. so that John is Subject. Main-Verb of is defined by (iii). in other words. per suade are the Main-Verbs of the accompanying examples. they form part of the general procedure for assigning a full structural description to a sentence. Notice that the general significance of the definitions ( 1 1) . so that Bill is the Object-of the Verb Phrases given as examples to both (iii) and (v) . of. ASPECTS OF DEEP STRUCTURE 73 depends on the assumption that the symbols S. Such relations can be defined derivatively in terms of the func tional notions suggested earlier. Quite apart from this. and Predicate-of are the notions of (1 1) . They can be generalized in various ways.§ 2. where Subject. and V have been characterized as grammatical universals. these questions aside. which is definable just as easily in the same terms. as j ust defined. and Verb-Object can be defined as the relation between the Main Verb-of and the Direct-Object-of a VP. at the moment. it is likely that these definitions are too restricted to serve as general explications for the traditionally designated grammatical functions in that they assume too narrow a substantive specification of the form of grammar. . but I do not. VP. Main-Verb-of. NP. it is clear that information concerning grammatical functions of the sort exemplified in (2ii) can be extracted directly from the rewriting rules of the base. without any necessity for ad hoc extensions and elabora: tions of these rules to provide specific mention of grammatical function. Thus Subject-Verb can be de fined as the relation between the Subject-of a Sentence and Main-Verb-of the Predicate-of the Sentence. We shall re turn to this question later. Thus the choice of Main-Verb is deter. N. Thus it is still necessary to define grammatical relations of the sort that hold between sincerity and frighten (Subject Verb) and between frigh ten and the boy (Verb-Object) in (1). aside from their redundancy. However. we have not yet exhausted the information presented in (2ii). have the defect of failing to express properly the relational character of the functional notions and are thus useless in all but the simplest cases. . as yet. Thus we have no basis. there is still something missing in this account. from the irrelevant pseudorelation Subject-Object. Such extensions. for distinguishing the legitimate and traditionally recognized grammatical relation Subject-Verb. However. Traditional grammar seems to define such relations where there are selectional restrictions governing the paired categories. In any event. see any strong empirical motivation for one or another specific extension or refinement (but see § 2 3 4). then. It would. this information can be extracted directly from Phrase-markers that are generated by simple re wri ting rules such as (5) and ( 1 2). once again.9 and all ca t egori es except Det (and poss ibly M and Aux see note 9) are major categories . the char acterization might be sharpened somewhat in terms of several notions that will be useful later on. Representations such as (6) and new or elaborated rewrit· ing rules to generate them are unnecessary. and at that point we can return to the question of grammatical relations. the categories N. to the fact that various modifications and extensions of these functional notions are poss ible and that it is important to find empirical motivation for such improvements. For example. V. it is fairly clear that n o thing essentially new is involved here beyond the rules that generate stri ngs and Phrase-markers. Thus in the grammar (5). A ca t egory that appears on the left in a lexical rule we shall call a lex ical category. it seems unnecessary to extend the system of rewriting rules in order to accommodate information of the sort presented in (2ii). as well as mislead· ing and inappropriate.4. This information is already contained.3. I shall defer the discussion of selectional relations until § 4. - . and that (as in (5» we have distinguished lexical rules (such as (511». from the others.74 CATEGORIES AND RELATIONS IN SYNTACTIC THEORY mined by the choice of Subject and Object. a lexical category or a category that dominates a s tring . have no grammatical relation of the sort in question holding between them. . implicitly. though Subject and Object are in general chosen independently of one another and. which introduce lexical formatives. X . But in any event. Finally. . We shall consider a further refine ment in the final paragraph of § 2 . In summary. . we shall call a major category. then. and M are lexical categories. I should like to call attention. be in accord with traditional usage to limit the functional n otions to major categories. . Suppose again that we have a base grammar consisting of a sequence of rewriting rules. With appropriate general definitions of the relational noti ons involved. We shall see later that this distinction is formally quite clearly marked . " where X is a lexical category. correspondingly. in the system of elementary rewriting rules .2. The problem. even in part. What is at stake. Of course . there being. F i rst. though they are often confused. it is an open q u e s tio n whether this attempt can succeed. it may nevertheless be useful to touch on it briefly .§ 2. Chapter 4. is purely interpret ive . Information of the sort presented in ( 2i ii) raises sever al difficul t and rather vex i ng q uest ion s . is the status of such expressions as ( 1 3) (i) the boy may frighten sincerity (i i) sinceri ty may admire the boy (iii) John amazed the inj us tice of that decision . essentially. then surely the information must be presented in the syntactic component of the grammar. These are distinct questions. We migh t call these the questions of presenta tion and justification. I shall be concerned here only with the question of presenta tion of information of the sort given in (2iii) . no other available proposal as to a semanti c basis for making the necessary distinctions. a linguist with a serious interest in semantics will presumably attempt to deepen and extend syntactic analysis to the poin t where it can provide the information con cerning subcategorization.2). for the moment. respectively. it is an inter est i ng question whether or to wha t extent semantic considera tions are relevant in determining such sub categorizations as those involved in (2iii).I. Some of the problems involved in prese n ting this information will be explored later. ASPECTS OF DEEP STRUCTURE 75 § 2 . I am as suming throughout that the semantic component of a generative gram mar.J. As far as the qu estion of j ustification is concerned. S econd. it is not obvious to wha t extent this information should be pro vided by the syntactic component at all. like the phonological component. Although the qu es ti on of justification of s u bca tegori zat i on s such as those of (2iii) is beyond the scope of the present discus sion. They are connected only in that if the basis for mak ing the distinctions is purely syntactic.J. It follows that all information utilized in semantic i nterpretat ion must be presented in the syntactic component of the grammar (but cf. § l . instead of relegatin g this to unan alyzed semantic intuition. Syn ta ctic fea t ures § 2 . for example. The expressions of ( 1 3) deviate in some manner (not necessarily all in the same manner) from the rules of English.1° If interpretable at all. There are fairly clear-cut cases of violation of purely syntactic rules. ( 1 5) (i) sincerity frighten may boy the (ii) boy the frighten may sincerity and standard examples of purely seman tic (or "pragmatic") incongruity. for e xample . .CATEGORIES AND RELATIONS IN SYNTACTIC THEORY (iv) (v) (vi) (vii) (viii) (ix) (x) the boy elapsed the boy was abundant the harvest was clever to agree John is owning a house the dog looks barking John solved the pipe the book dispersed It is obvious to anyone who knows English that these expressions have an entirely different status from such sentences as (14) (i) (ii) (iii) (iv) (v) (vi) (vii) (viii) (ix) (x) sincerity may frighten the boy (=(1» the boy may admire sincerity the injustice of that decision amazed John a week elapsed the harvest was abundant the boy was clever to agree John owns a house the dog looks terrifying John solved the problem the boys dispersed The disti nction between (1 3) and ( 1 4) is not at issue. Rather. they are surely not interpretable in the manner of the corresponding sentences of (14). it seems that inter pre ta t ions are imposed on them by virtue of analogies that they bear to nondeviant sentences. and clearly must be accounted for somehow by an adequate theory of sen tence interpretation (a descriptively adequate grammar). and that purely semantic or purely syntactic considerations may not provide the answer in some particular case. In fact. it seems to me that they have some plausibility. furthermore. it should not be taken for granted.ll The only sug gestion as to possible semantic grounds for these distinctions has been that they are based on language-independent semantic absolutes . Several suggestions have been made as to how syntactic con siderations can provide a subclassification of the appropriate sort. what is needed is a systematic account of how application of the devices and methods appropriate to unequiv ocal cases can be extended and deepened to provide a basis for . the deviance is attributable to violation of some linguistic universal that constrains the form of the semantic component of any generative grammar. there is no reason why some combination of these two extreme approaches should not be attempted. It goes without saying that the same answer may not be appropriate in all of these cases. ASPEcr5 OF DEEP STRUCTURE 77 ( 1 6) ( i) oculists are generally better trained than eye-doctors (ii) both of John's parents are married to aunts of mine (iii) I'm memorizing the score of the sonata I hope to com pose some day (iv) that ice cube that you finally managed to melt just shattered (v) I knew you would come. and it is much less clear how their aberrant status is to be explained. It is possible that this is the right answer. In other words.that in each case.§ 2. have a borderline character. we must face the problem of determining to what extent the results and methods of syntactic or of semantic anal ysis can be extended to account for the deviance and inter pretation of these expressions. that syntactic and semantic considerations can be sharply distinguished. necessarily. Al though these notions have been advanced only very tentatively. however. These involve the notion of "degree of grammaticalness. and concrete proposals involve tech niques of subclassifying based on distributional similarities." along various dimensions. but I was wrong The examples of ( 1 3). In any case. but assigns them Phrase· markers only by virtue of their structural similarities to such perfectly well-formed sentences as those of ( 1 4). if we conclude that the semantic component should carry the burden of accounting for these facts. Vague and unsupported assertions about the "semantic basis for syntax" make no contribution to the under standing of these questions. as a string generated only by relaxing certain of these restrictions. assuming that the notion "scale of gram- . we must determine how a grammar can provide structural descriptions that will account for such phe nomena as those exemplified. we face a well-defined problem. perhaps in the manner described in the references in note 1 1 .CATEGORIES AND RELATIONS IN SYNTACTIC TIlEORY explaining the status of such expressions as those of ( 1 3). we must design the syntactic component so that it does not pro· vide for the sentences of ( 1 3) directly. I shall. presumably on the basis of analogy to nondeviant cases. with no distinction of grammaticalness. Alternatively. Thus the syntactic component will operate in terms of selectional restrictions in volving such categories as animateness and abs trac tn ess . we can allow the syntactic component to generate the sentences of (14) as well as those of ( 1 3) . accept the position of the references of note 1 1 . Either way. and it is reasonably clear how to proceed to examine it. A descriptively adequate grammar must account for such phenomena in terms of the structural descriptions provided by its syntactic and semantic components. for the present. and wi l l characterize ( 1 3i). where possible. and an account of how an ideal listener might assign an interpretation to such sentences. These are real and important ques tions. A priori there is no way to decide whether the burden of presentation should fall on the syntactic or semantic component of the generative grammar. but with lexical items specified in such a way that rules of the semantic component will determine the incongruity of the sentences of ( 1 3) and the manner in which they can be interpreted (if at all) . Proceeding now from the question of justification to the ques· tion of presentation. and a general linguistic theory that aims for explanatory adequacy must show how such a grammar can develop on the basis of data available to the language learner. If the former. for example. for example. Some formal similarities between sy n tax and phonol ogy.... others (for example. boo k) and either Human (John. Consider now how information of the sort given in (2iii) can be presented in explicit rules. G.. but involves rather cross classification. Certain rules (for example. and shall discuss briefly some possible alternatives. H.. or NP into Det N). But if the subcategorization is given by re writing rules. in the course of his work on a generative grammar of German in 1 95 7-1 958. and the other will be unstatable in the natural .12 1 955.. § 2. Later on. 2. .. then one or the other of these distinctions will have to dominate. Hence. The obvious suggestion is to deal with subcategorization by rewriting rules of the type described in § 2. it seems that the only categories involved are those containing lexical formatives as members.3. ASPECTS OF DEEP STRUCTURE 79 maticalness" will be relevant to semantic interpretation.. pointed out that this assumption was in correct and that rewriting rules are not the appropriate device to effect subcategorization of lexical categories. some involving Determiners) apply to the Proper/Common distinction. Chomsky. and that the sentences of ( 1 3) are as· signed Phrase-markers only by relaxation of certain syntactic conditions.. 1 957) . Thus.. Nouns in English are either Proper (John... book).. Egypt) or Common (boy.. analysis of a category into a sequence of categories. and it is important to bear this in mind in exploring appropriate means for presenting these facts. Furthermore. rules involving choice of Relative Pronoun) to the Human/non Human distinction.. Note that this information con cerns subcategorization rather than "branching" (that is.§ 2. Matthews... . and this was the assumption made in the first attempts to formalize generative grammars (cf. we are dealing with a rather restricted part of grammatical structure..2. as when S is analyzed into Np Aux VP. However. that a distinction should be made between ( 1 3) and ( 1 4) by rules of the syntactic component. 1 95 1 . . I shall try to indicate the precise point at which this decision affects the form of the syntactic component. boy) or non-Human (Egypt. IS The difficulty is that this subcategorization is typically not strictly hierarchic. . and there are other rules that apply to continuants [s]. when we add transformational rules to the grammar. [s]." "Adjective. The same solution suggests itself in the case of the syntactic problem that we are now facing. rules that apply to voiced consonants [b]. Thus if we decide to take Proper/Common as the major distinction... but not to unvoiced consonants [p]. although we can easily state a rule that applies only to Proper Nouns or only to Common Nouns. a rule that applies to Human Nouns must be stated in terms of the unrelated categories Pr-Human and C-Human." and "Modal..80 CATEGORIES AND RELATIONS IN SYNTACTIC THEORY way. Common Proper . and so on. Proper N . since this rule would now be no simpler or better motivated than.." "C-Human. but not to stops [p].. problems of this sort mount to the point where they indicate a serious inadequacy in a grammar that consists only of rewriting rules. Pr-nHuman Common "" C-Human Common "" C-nHuman where the symbols "Pr-Human. As the depth of the analysis increases. and it is this method of dealing with the problem that I shall elaborate here. Pr-Human Proper ... as many others are." In this system. with respect to phonological rules. as distinct from one another as the symbols "Noun. we have such rules as (1 7 ) N .. Nor is this particular difficulty overcome." and "C-nHuman" are entirely unrelated. This obviously indicates that a generalization is being missed. Formally. [z]. . for example.. a rule applying to the unrelated categories Pr-Human and Abstract Nouns.. [b]. for example. There are." "Verb. For this reason it is necessary to regard each phonological unit as a set of features. [z].. Thus phonological units are also cross classified." "Pr-nHuman. and to design the phonological component in such a way that each rule applies to all segments containing a certain feature or constellation of features. this problem is identical to one that is familiar on the level of phonology. The rule ( 1 8) converts the string WX'A'Y' V to the string WX'Z'Y'V. 1 964. each lexical formative is represented by a distinctive-feature matrix in which the columns stand for successive segments. In other words. This is the typical form of a phonological rule. and Y are matrices. for discussion of this question). Y' are matrices wi th the same n um ber of columns as X. each segment being a set of features. Suppose that ( 1 8) A � Z/X - Y is a phonological rule. for discussion). 1 959a.§ 2. furthermore. Halle. respectively. An entry in the ith column and ith row of such a matrix indicates how the ith segment is specified with respect to the ith feature. where A. forthcoming. or that it is negatively specified with respect to this feature. 1 959b. ASPECI'S OF DEEP STRUCTURE Before we turn to the use of features on the syntactic level. that two matrices with the same number of columns are distinct if the ith segment of one is distinct in this sense from the ith segment of the other. HalIe and Chomsky. A '. A. for some i. consider this phonological rule : . more generally. or that it is positively specified with respect to this feature. X. qualifications are necessary that do not concern us here . and. 1 962a. Y. As an illustration of some of these notions. and X'A'Y' is not distinct from XA Y (actually. We shall say that the rule ( 1 8) is applicable to any string WX'A 'Y' V. where X'. Each lexical formative is represented as a sequence ef segments. Z. We say that two segments are distinct just in case one is positively specified with respect to a feature with respect to which the other is negatively specified. and A and Z are. where Z' is the segment consisting of the feature specifications of Z together with all feature specifications of A' for features with respect to which Z is unspecified. segments (matrices with just a single column). let us review briefly the operation of the phonological com ponent (cf. A particular entry may indicate that the segment in question is unspecified with respect to the feature in question.cf. and the rows for particular features. For example.) will be analyzed by the rules into complex symbols. the symbols representing lexical categories (N. .) . at the same time. Furthermore. contributing to the general unity of gram matical theory. we might have the following grammatical rules : (20) (i) (ii) (iii) (iv) (v) (vi) N � [+ N ±Common] [+Common] � [±Count] [+Count] � [±Animate] [-Common] � [±Animate] [+Animate] � [±Human] [-Count] � [±Abstract] . [+Human]. each complex symbol being a set of specified syntactic features. just as each phonological seg ment is a set of specified phonological features. providing a very natural solution to the cross-classification problem and. [fd] into [vd]. Each lexical formative will have associated with it a set of syn tactic features (thus boy will have the syntactic features [+Common]. We interpret rule (20i) as asserting that the symbol N in a line of a derivation is to be replaced by one of the two complex symbols [+N. -Common].. etc. Thus rule (20ii) asserts that any complex symbol Q that is already specified as [+Common] is to be replaced by the complex symbol con taining all of the features of Q along with either the feature specification [+Count] or [-Count] . These notions can be adapted without essential change to the representation of lexical categories and their members.[+voiced] This will convert [sm] into [zm].82 ( 1 9) CATEGORIES AND RELATIONS IN SYNTACTIC 'IHEORY [+continuant] � [+voiced]j . The same is true of the other rules that operate on complex symbols. [�g] into [zg]. +Common] or [+N. 1 4 These conventions (which can be simplified and generalized in ways that do not concern us here) alIow us to apply rules to any class of segments specified by a given combination of features. for example. V. but it will not affect est] or [Pd]. etc. and thus to make use of the cross classification of segments provided by the feature representation. The rules (20ii-20vi) operate under the conventions for phonological rules. etc. ±Abstract 1ll) [. In this represen tation. . with rules of this sort that introduce and elaborate complex symbols. -Ammate (iv) [+Animate] � [ ±Human] [ ] If we were to require representability in a branching diagram as a formal condition on these rules. For example. In any event. ( 2 2 ) (i) N � [+N. ±Animate. the rules could just as well be presented in the form (2 1) as the form (20). each nod e is (2 1) Common � An imate Abstract Human Coun t �-ok b o virtue +Adirt � Egypt AHuman Animate John Fido boy �dog labeled by a feature. an element of this category has the feature [aF] (a = + or -) if and only if one of the lines constituting this path is labeled a and descends from a node labeled F.C ount] � . ASPECTS OF DEEP STRUcrURE The total effect of the rules (20) can be represented by the bran chi ng diagram (2 1 ). Each maximal path corresponds to a category of lexical items. A system of complex symbol rules need not be representable by a branching diagram of this sort. In this case. . Typical members of the cate gories defined by (20) are given at the terminal points of (2 1). . the categories defined by the rules (20) are also defined by the rules (22). then (22) would be excluded. ±Common] (H) [+CommonJ � [±Count] (. we can develop the full set of lexical categories. but in this case there is no representing branching diagram. and the lines are labeled + or -.§ 2. J. is dominated by [aF]) if [aF] is part of the complex symbol Q or the complex symbol C. Genera l structure of the base componen t. The grammar will now contain no rules such as those of (511) that introduce the formatives belonging to lexical categories. We now modify the description of the base subcomponent that was presented earlier. With this extension. We also extend the general notion "Phrase-marker" in such a way that the Phrase-marker of a terminal string also con tains the new information. We say that in the terminal string formed by replacing the complex symbol Q by the formative D of the lexical entry (D. the base of the grammar will con tain a lexicon. where C is not distinct from Q. the boy is an NP in (3» in the follow We now extend the fundamental notion is a that relates strings ing way. where D is a phonological distinctive feature matrix "spelling" a certain lexical formative and C is a collection of specified syntactic features (a complex symbol). the formative D is an [aF] (equivalently. In addition to rewriting rules that apply to category symbols and that generally involve branching. Instead. C). there are rewriting rules such as (20) that apply to symbols for lexical categories and that introduce or operate on complex symbols (sets of specified s yn tac tic features). where a is either + or and F is a feature (but cf. each lexical entry being a pair (D. the lexicon is a set of lexical entries. which is simply an unordered list of all lexical formatives. then Q can If Q is a complex symbol of a preterminal string and (D. Such a string we call a preterminal string.J. a Phrase-maker can naturally no longer be represented by a tree-diagram. More precisely. note 1 5). and exemplified by (5). C). C) be replaced by D. A terminal string is formed from a preterminal string by insertion of a lexical formative in accordance with the fol lowing lexical rule: is a lexical entry. in the following way. to categories (for example.CATEGORIES AND RELATIONS IN SYNTACTIC TIlEORY § 2. as - .111 The system of rewriting rules will now generate derivations terminating with strings that consist of grammatical formatives and complex symbols. since it is perfectly general. "spellings" of fonnatives. which I repeat here as (�3). -Count.. -Count. where Q is the complex symbol into which V is analyzed by rules that we shall discuss directly .in other words. Instead of the grammar (5) we now have a grammar containing the branching rules ( 51) . along with the subcategorization rules (20). As a concrete example. +Abstract]"'M"'Q the"'[+N. +Animate. The lexical rule (which. [+N. +Count. ASPECTS OF DEEP STRUCTURE before. +Abstract]) (boy. since it has an addi tional "dimension" at the level of subcategorization.. +Human]) (may. + Common.. -Count. +Animate.. [+M]) We shall have more to say about these rules and lexical entries later. it constitutes part of the definition of "derivation") . [+N. that is. It is to be understood. +Human]. . here and later on.. that the i talicized i tems stand for phonological distinctive feature matrices. consider again the sentence sincerity may frigh ten the boy (= (1».. ±Common] [+Common] � [±Count] [+Count] � [±Animate] [-Common] � [±Animate] [+Animate] � [±Human] [-Count] � [±Abstract] (25) (sincerity. need not be stated in any grammar . and they will still undergo significant revision. These rules allow us to generate the pretenninal string (26) [+N. ( � 3) S � NP"'Aux"'VP VP � V"'NP NP � Det"'N NP � N Det � the Aux � M (2 4) ( i) (ii) (iii) (iv) ( v) (vi) N � [+N.§ �. repeated as (�4)' and containing a lexicon with the entries (25). represented in terms of features that appear in the Phrase marker in a position dominated by the lexical categories N and V and dominating the formative in question. many of the grammatical properties of formatives can now be specified directly in the lexicon. note 9). Except for the case of frighten. and the functional in formation (2ii) is derivable from the Phrase-marker in the manner described in § 2 . We shall return in Chapter 4. I N sin cerity boy ditional specific information about the lexical items that appear in (26). and thus need not be represented . to in sert frighten for Q (and may for M cf. If the lexicon includes ad- s NP [-Count][+Common] [+AJ� � N I Aux M may T I I I Q fTigh ten I V ____ Det I VP _____ I NP the [+Count] [+Common] [+Animate] {+Human] ---. However. For one thing. we can see im mediately that separating the lexicon from the system of re writing rules has quite a number of advantages . and the lexicon (25). (24). Given this Phrase-maker. the information about the sentence ( 1 ) that is given in (2) is now explicitly provided in full by the Phrase marker generated by the grammar consis ting of the rules (23). we can derive all of the information (2i) and (2iii). this information will also appear in the Phrase-marker... by association of syntactic features with lexical formatives.86 CATEGORIES AND RELATIONS IN SYNTACTIC THEORY now allows us to insert sincerity for the first complex symbol and boy for the last complex symbol of (26) and. 2. § 2 to questions concerning the proper formulation of lexical entries. as we shall see. in terms of the relation is a. which concerns assignment of substrings to categories. We might represent this Phrase marker in the form shown in (27). notice that it is now unnecessary to use rewriting rules to classify Transitive Verbs into those that do and those that do not normally permit Object deletion. IS In particular. Instead. on the other.§ 2. The transformational rule that deletes Objects will now be applicable only to those words positively specified with respect to this feature. components of the dictionary definition) . ASPECTS OF DEEP STRUCTURE in the rewriting rules at all. morphological prop erties of various kinds can be treated in this way . the lexical entries for read. In general. furthermore. Or. where they most naturally belong. the grammar can be significantly simplified if they are excluded from the rewriting rules and listed in lexical entries. this information now being contained in the Phrase-marker of the strings in which these words appear. eat. in the case of bee. returning to (2iii). In particular. (d) lexical features indicating the positions in which a lexical formative can be inserted (by the lexical rule) in a . strong or weak verbs. so that the simplification of the grammar that can be effected in these ways will certainly be substantial. etc. nominalizable adjectives.for example. and unrounded). which is not mentioned in the rewriting rules at all.17 (b) properties relevant to the functioning of transformational rules (as the example of the preceding paragraph. all properties of a formative that are essentially idiosyncratic will be specified in the lexicon. keep. (c) properties of the formative that are relevant for semantic interpretation (that is. Any attempt to construct a careful grammar will quickly reveal that many formatives have unique or almost unique grammatical characteristics. and frighten. highly idiosyncratic. will differ in specification for the particular syntactic feature of Object deletion. membership of lexical items in derivational classes (declensional classes.). tense. Since many such properties are entirely irrelevant to the functioning of the rules of the base and are. on the one hand. and many others) . but it will not specify the degree of aspiration of the stop or the fact that the vowel is voiced. the phonological matrix of the lexical entry will specify that the first segment is a voiced labial stop and the second an acute vowel. the lexical entry must specify: (a ) aspects of phonetic structure that are not predictable by general rule (for example. This may be im portant for the theory of semantic interpretation.88 CATEGORIES AND RELATIONS IN SYNTACTIC THEORY preterminal string. and Chomsky. Notice. incidentally. Furthermore. where each label stands for a category of strings. This theory formalizes a conception of linguistic structure that is substantive and interest ing. 1 and Chapter 4. Postal. it contains information that is re quired by the phonological and semantic components of the grammar and by the transformational part of the syntactic com ponent of the grammar. the degree and manner of deviation of strings that are not directly generated (see § 2 . a phrase structure (constituent structure) gram mar. 1 964a. 3 . by implication. But introduction of complex symbols constitutes another radical departure from this theory. It is no longer true that a Phrase marker can be represented as a labeled tree-diagram. As described informally in § 2. that the purely semantic lexical features constitute a well-defined set. and hence. a phrase structure grammar consists of an unordered set of rewriting rules. 1 ). rather than freely to generate the full set of actual sentences. 3 . we have already departed from this theory by assuming that the rewriting rules apply in a prescribed sequence to generate a restricted set of (base) strings.3. § 1 . as well as information that determines the proper placement of lexical entries in sentences. 1 . namely the "taxonomic" view that syntactic structure is deter mined exclusively by operations of segmentations and classifica tion (see § 2. the con- . 1 964). In short. These modifications affect the strong gen erative capacity of the theory. A feature belongs to this set just in case it is not referred to by any rule of the phonological or syntactic component. and the separate treatment of the lexicon just suggested is again an essential revision. strictly speaking. Of course. See Katz (1 964b) . It is important to observe that the base system no longer is. and that has been quite influential for at least half a century. This modification restricted the role of the phrase structure grammar. 1 and more carefully in the references cited there. in a given grammar. and assigns a structural de scription that can be represented as a tree-diagram with nodes labeled by symbols of the vocabulary. just as in the case of the notation for transformational rules that involves carrying along labeled brackets in lines of a derivation. where each a. rules applying . eTn = flal . Thus application of such a rule to the last line of a derivation depends in part on earlier lines. . . is a terminal or non terminal symbol of the vocabulary on which the grammar is based). which are not "local transformations" of the sort that appear. notice that a derivation involving only phrase structure rules (rewriting rules) has a strict "Markovian" character. 1 960. . eTn (eTl = IISII. § 6). In fact. A grammatical transformation is. We should now have a kind of transformational grammar. typically applies to a string with a particular struc tural description. in other words. . and we should have entirely lost the intuition about language structure that motivated the development of phrase structure grammar. 1 8 But with the availability of complex symbols. forthcoming. . aspects of the earlier steps of a derivation can also be carried along to later steps. eTn-l and depend com pletely on the string eTn• A grammatical transformation. ASPECTS OF DEEP STRUCTURE 89 ventions for the use of complex symbols in effect allow the use of quasi-transformational rules in the base component . the rules that can be applied to form the next line eTn + l are independent of eTl' .§ 2. 1 963. . . on the other hand. to some extent. Chomsky and Miller. incorporation of brackets into strings provides the most appropriate notation for the transformational rules of the phonological component (see Halle and Chomsky. akll. . in the transformational cycle in phonology. To see why this is so. Suppose. That is. exclusively. that we were to include labeled brackets in the strings that constitute a derivation and were to allow the "rewriting rules" to refer to these symbols. and. however. global operations on strings can be coded into complex category symbols and carried along in derivations until the point of "application" of these operations. a rule that applies to Phrase markers rather than to strings in the terminal and non terminal vocabularly of the grammar. Consequently. in a derivation consisting of the successive lines eT1' . though not for the transformational rules of the syntactic component. the features [Transitive]. We must still give rules to determine whether a V may or may not be transi t ive and so on. the Verb can be positively specified for the feature [Abstract-Subject] just in case it is the environment [+Abstract] ' " .NP. and a grammar using complex symbols is a kind of transforma t ion al grammar rather than a phrase structure grammar. as distinct from the context-free rules (u) that sub categorize Nouns. It would not do simply to add to the grammar the rule (28). incidentally. . that the conventions established for the use of complex symbols do not provide systems with greater weak generative capacity than phrase structure grammars (even if appropriate con ven tions are established to permit complex symbols to appear at any point in a derivation. and it can be positively specified for the feature [Animate Object] just in case it is in the environment . [Abstract-Subject]. has no bearing on the observation that such a theory is no longer a version of the theory of phrase structure grammar.1D As a first approximation. (28) V � [+V. for the analysis of V: . ±Animate-Object] The problem is that an occurrence of the category symbol V can be replaced by a complex symbol containing the feature [+Transitive] just in case it is in the environment .3+ Context-sensitive subcategorization rules.go CATEGORIES AND RELATIONS IN SYNTACTIC TIlEORY to complex symbols are. [Animate-Object] must be introduced by rewriting rules that are res tri cted with respect to context. rather than only in lexical categories . ±Transitive. Similarly. of course.. This fact. analogous to (24) : . We have not yet considered how the category V is analyzed into a complex symbol. in effect.see note 4). [+Animate] . in the case of all ()f those lexical features that are in volved in the statement of �ontextual restrictions . Thus suppose that we have the grammar (2 3)-(25). ±Progressive. . and so on. ±Abstract-Subject. § 2. transformational rules. and must add to the lexicon ap propriate entries for individual verbal formatives. we might consider rules of the following sort.. Notice. Hence. the convention that X. is an abbreviation for the sequence of rules (32) (i) A -+ ZIX1 .. +Animate] [+ [-Animate]-Object]l .D et [+N. +Abstract] Aux [+ [-AbstractJ-Subject]1 [+N. . cf. -Abstract] Aux [+ [+Animate]-Object]1 . -Animate] We can now introduce the standard conventions for ex pressing generalizations in the case of context-sensitive rewriting rules such as (4) ..# J (ii) -+ (n) A } { (34) (i) (ii) (iii) [+V] -+ (iv) [+[+Abstract]-Subject]1 [+N.. -Abstract] Aux [+[+Animate] -Object]/ . for example.Det [+N. +Animate] [+ [-Animate] -Obj ect]I . + Transitive]/ . for discussion of the role of these conventions in linguistic theory). +Transitive] / . in particular.NIL (33) (i) V [ V. Chomsky.§ 2. and other familiar related conventions.Subject]/[+N. -+ + -TransitiveJ/ .Y1 Z /X. Chapter I .Y. (30) (cf. These allow us to restate (29) and (30) as (3 3) and (34). (29).Y.11 -+ -+ -+ [+V] [+V] [+V] [+V] -+ [+ [+Abstract]-SubjectJ/[+N.Det [+N..NP [+V. . -TransitiveJ / . Appendix. +Abstract] Aux [+ [-Abstract]. ASPECTS OF DEEP STRUCTURE (29) (i) V (H) V (30) (i) (ii) (iii) (iv) -+ -+ [+V. -Animate] 1 . respectively.Det [+N. § 7. 1 957.. the schema for grammati cal description that we have so far developed still does not permit us to state the actual processes at work in determining the form of sentences. +Abstract] AUX [+F2] / [+ N . believe (it to be unlikely)].S' [t h in k ( t h a t he will come).Det [+N. somewhat in the manner of ordinary rules of agreement in many languages. though formally adequate. + Animate] [-F2]/ . and that alongside of (33) there are rules specifying various other choices of subcategories of Verbs. where S' is a variant of a sen tence. . feel (sad)]. -An im ate] l . assign features of the Subject and Object to the Verb. .NP'""' S' [persuade (John that it is unlikely» (omitting certain refinements) . an inadequacy that reveals itself in the redu ndancy and clumsi ness of the systems of rules of which (33) and (34) are samples. . This becomes still more obvious when we observe that alongside of (34) there are many other rules of the same kind.Predicate-Nominal [ be com e (presiden t)].like"" Predicate-Nominal [ lo o k (like a n ice person). In the present case. act (like a foo 1) ] . and there are also many rules (of which (33) presents j u st two) that impose a subclassification on the category Verb in terms of the set of frames in which this category appears at the stage of a derivation where it is to be subcategorized. -Abstract] Aux [+F1]/ . These generalizations are not expressible in terms of the schema for grammatical des crip t ion so far developed. grow (o ld).Det [+N. . whereas the rules (35) determine the sub categorization of Verbs in some essentially haphazard way in I [+V] 7 j [+F1] / [+N. . for example.CATEGORIES AND RELATIONS IN SYNTACTIC THEORY It is immediately apparent that the rules (33) and (34). are extremely clumsy and leave important generalizations unexpressed.Adjective [e. in such environments as : . Rules such as (34) systematically select the Verb in terms of the choice of Sub ject and O bject. there is a large set of rules (of which (34) men tions j ust four) that. In other words. Our present difficulty can be seen clearly by comparing the rules (34) with the hypothetical set (35) : (35) (i) (ii) (iii) (iv) where Fl and F2 are certain syn tactic features.g. in effect. 20 Generalizing. A more ex pressive notation would be simply the symbol .W]""" Y/Z . we can now have Z" . let us allow certain features to be designated in the form [X . in our present terms. where X and Y are strings (perhaps null) of symbols.W. We thus introduce the notation .Adjective]. We shall henceforth call these con tex tual features.Y]. We then have a general rule of subcategorization to the effect that a Verb is positively specified with respect to the contex tual feature associated with the context in which it occurs. Let us consider how a more natural and revealing expression of these processes can be developed.W as an abbreviation for the rewriting rule (37) A � X""" [+A . pre-Adjectival Verbs such as grow.NP]. (3 6) A � X-CS-Y/Z . the opposite would be true in this case if the familiar notational conventions are applied to evaluate these systems.§ 2. feel. at the level of explanatory adequacy) .NP. which is therefore shown to be inadequate (in this case. the system (34) is not. where " C S" stands for "complex symbol.w. +Z . ASPECTS OF DEEP STRUCTURE 93 terms of choice of Subject and Object. However. more highly valued than (35) . as an abbreviation for the sequence of rules . Let us regard Transitive Verbs as positively specified for the con textual feature [ . the linguistically significant generalization underlying (34) is not expressible within our present framework.. In other words." Utilizing the bracket conventions. in fact.NP" itself.. as positively specified for the contextual feature [ . and so on. Observe that the feature specification [+Transitive] can be regarded as merely a notation indicating occurrence in the environment . . + ... + that-S1 persuade.NP. y/Z...... we shall have. + . [+ v.NP (o/'-'Det N) S1 . + .. [+V.... the rule (40). + . + .. .. Predicate-Nominal] Prepositional-Phrase that S' NP (of.. N) S' etc... [+V..like. Det. John seems like a nice fellow. + .like Predicate-Nominal] believe. John grew a beard.. + ... a week elapsed. John became sad. + .94 CATEGORIES AND RELATIONS IN SYNTACTIC THEORY A � X .. [+V.. + ... [+A .. + . instead of (33).. +Z" ... John became presiden t.] .W. John grew. [+V... 21 The lexicon might now contain the items (4 1 ) eat....NP... as a better approximation: NP # (40) Adjective Predicate-Nominal V � CS/ . + ... [+V.. John seems sad. + .W... The notation introduced in (35) allows us to express the fact that a set of frames in which the symbol A occurs imposes a cor responding subclassification on A ..like Predicate-Nominal] look.. with one s ubdivision cor responding to each listed context. [+V. and so on .. + .. .22 The rules (40) supplemented by the lexicon (4 1) will permi t such expressions as John eats food.NP] elapse.Adjective..(Prepositional-Phrase) #. . + .Adjective. Thus in the case of Verb subclassification.. .#... .. [+V.Adjective] become....#] grow..Adjective.Predicate-Nominal] seem.. John grew sad... We can use the same notational device to express the kinds of selectional restriction expressed in such rules as (34) . Rules such as (42). I shall call selectional rules .Det [-Animate] 1 -I . John believes that it is un likely. which analyze a symbol (generally. John looks sad. etc.Det [+ Animate] . a complex symbol) in terms of syntactic features of the frames in which it appears. which assign features of the Subject and Object to the Verb . The rules differ in that in the case of (40) the frame is stated in terms of category symbols. whereas in the case of (42) the frame is stated in terms of syntactic features . The rules of (40) and (42) analyze a category into a complex symbol in terms of the frame in which this category appears. Thus we can replace (34) by the rules (4 2 ) (11) (iii) (iv) where now [[+Abstract] Aux ] is the feature denoted in (34) as [[ +Abstract]-Subject]. but not (35)." We shall see later that there are important syntactic and semantic differences between strict subcategorization rules and selectional rules with respect to both their form and function. Rules such as (40). which analyze a symbol in terms of its categorial context. and that consequently this distinction may be an important one. John looks like a nice fellow. We see that with a slight extension of conventional notations the systematic use of complex symbols permits a fairly simple and informative statement of one of the basic processes of sub classifica tion. John looked at Bill. The notational convention (36)-(37) shows in what respect a system of rules such as (34). The latter express what are usually called "selectional restrictions" or "restrictions of cooccurrence.§ 2. In the case of both the strict subcategorization rules (40) and - �) 1 [ V ] + � [+ Abstract] Aux [-Abstract] Aux CS! . John persuaded Bill of the necessity for us to leave. John persuaded Bill that we should leave. John believes me. I shall henceforth call strict subcategorization rules. ASPECTS OF DEEP STRUCTURE 95 John looked. expresses a linguistically significant generalization. and the further fact that no frame which is not part of the VP is relevant to the strict subcategorization of V.2lI This observa tion suggests that at a certain point in the sequence of base rewriting rules. Verbs are not strictly subcategorized in tenns of types of Subject N P s or type of Auxiliary apparen t ly. is relevant to the strict subcat egorization of V. (44) V � CS! - a. This set of rules imposes a categorization on the symbol V in terms of a certain set of frames in which V occurs. Thus the symbol VP will dominate such strings as the following. On the other hand. we introduce the rule that s trictly subcategorizes Verbs in the following fonn: ' . It fails to express the fact that every frame in which V appears. in the VP. . there are still de eper general izations that are not yet expressed.96 CATEGORIES AND RELATIONS IN SYNTACTIC THEORY the selectional rules (42). where a is a string such that Va is a VP The rule schema (44) expresses the actual generalization that detennines strict categorization of Verbs in terms of the set of syntactic frames in which V appears. in derivations generated by rewriting rules of the base : (43) (i) V (ii) V NP (iii) V NP that-S (iv) V (v) V (vi) (vii) (viii) (ix) V V V V (elapse) (bring the book) (p ersuade John that there was no hope) Prep-Phrase (decide on a new course of action) Prep-Phrase Prep-Phrase (argue with John about the plan) Adj (grow sad) like Predicate-Nominal (feel like a new man) NP Prep-Phrase (save the book for John ) NP Prep Phrase Prep-Phrase (trade the b icycle to John for a tennis racket) - and so on. Consider first the case of (40). Corresponding to each such string dominated by VP. there is a strict subcategorization of Verbs. that each feature of the preceding and follow ing Noun is assigned to the Verb and determines an appropriate selectional subclassification of it. then each of the lexical features that was introduced by the rules of (20) would determine a corresponding subclassification of the complex sym bol [+V] . The rules abbreviated by the schemata (45) assert. Here too it is evident that there are linguistically significant generalizations that are not expressed in the rules as given in this form.§ 2. ASPECTS OF DEEP STRUCTURE 97 We have now discussed the problem of formulating the gen eralizations that actually underlie the strict subcategorization rules (40). the Verb) is subcategorized in terms of the contexts in which this element appears. the most natural way to formulate the underlying generalization would be by such rule schemata as (45) [+V] � CS! a {O('AUX. where these contexts all meet some syntactic condition. a a is an N. of which (42) presents a sample. where . Thus the rules (42) do not make use of the fact that every syntactic feature of the S ubject and Object imposes a corresponding classification on the Verb. being a variable ranging over specified features. a certain extension of the notational devices for formulating rules is called for so that the evaluation measure will operate correctly.2 4 not just certain arbitrarily chosen features.D et ---. In all cases. The theory of grammar would fail to express the fact that a grammar is obviously more highly valued if sub categorization is determined by a set of contexts that is syntacti- . Once again. simply. In this case. an important generalization would be missed if the relevant contexts were merely listed. Thus if the rule (45) appears in the sequence of base rules after the rules (20). and have presented informally a device that would accomplish this result. namely dominance by N (with some ordering that is ap parently inconsequential) . It remains to consider the selectional rules.} . The rule schemata (44) and (45) deal with a situation in which an element (in this case. We interpret these schemata as abbreviating the sequence of all rules derived from (45) by replacing a by a symbol meeting the stated condi tion. we can often restate the rules in question with the use of labeled brackets (regarding these as carried along in the course of a derivation). for example. Analyzability. except in the statement of context. The rule schemata (44) and (45) take on the character of transforma tional rules even more clearly. suggest that even the rules (i) must be replaced. and that utilize complex symbols. . Thus. and many other references). Rules of this type are essentially of the form where the expression "X is analyzable as Yl' . . One might. in fact. is dominated by Y" in the Phrase-marker of the derivation under construction. but the rules (ii) are transformational rules of an elementary sort . where XA Y is analyzable as Zl. The appropriate sense of "syntactically definable" is suggested by the examples just given. with certain features being carried over to certain of the "de scendants" of a particular category symbol in the manner of Matthews's system referred to in note 1 3.3. The rules (i) are ordinary phrase structure rules. in such a way that X. by • • • (46) A � CS/X - Y.. A precise account of "syntactically definable" can be given quite readily within the framework of transformational grammar.2 6 Along with a lexicon. or by allowing complex symbols to appear at arbitrary points of a derivation. . in part. then. .3 we pointed out that a system of rewriting rules that makes use of complex symbols is no longer a phrase structure grammar (though it does not differ from such a grammar in weak generative capacity). 1 955. . in this sense. At the conclusion of § 2. 1 956.:.98 CATEGORIES AND RELATIONS IN SYNTACI1C THEORY cally definable. means that X can be segmented into X ::: Xl · · · X. Y. is the basic predicate in terms of which the theory of transformational gram mar is developed (see Chomsky. Z". but rather is more properly regarded as a kind of transformational grammar. or in various other similar ways . the base component of the gram mar contains: (i) rewriting rules that typically involve branching and that utilize only categorial (noncomplex) symbols and (ii) rule schemata that involve on1y lexical categories. would be what we have called a "local trans formation. although such a system does play a fundamental role as a subpart of the base component. for example. furthermore. or by local transformations (cf. where such operations as conjunction are discussed in terms of a framework of this sort). Putting these aside for the moment. ASPECTS OF DEEP STRUCTURE 99 rule schemata that go beyond the range of phrase structure rules in strong generative capacity (cf. where. Chomsky and Schiitzenberger. consequently.§ 2. p. ." Cf. Chomsky and Miller. Thus the strict subcategorization of V involves only frames that are dominated by the symbol VP. to see to what extent freedom to use such devices is actually empirically motivated.. its role is that of defining the grammatical relations that are expressed in the deep structure and that therefore determine the semantic inter pretation of a sentence. The descriptive power of the base component is greatly en riched by permitting transformational rules. This suggests that we impose the follow ing general condition on strict subcategorization rules: each such rule must be of the form (47) A � CS/a - p. p. to suppose that the base component of a transformational grammar should be strictly limited to a system of phrase struc ture rules. and it is frames dominated by VP that determine strict sub categorization of Verbs.. where aAP is (T � A a (T. In short. From the examples just given. note 18. 133. let us continue with the investigation of strict subcategorization rules.. Thus (47).. and there are also obvious restrictions (to which we return in § 4. The symbol V is introduced by rules of the form: VP � V . note 18). The italicized condition guarantees that . 298. (T is the category symbol the left in the rule • • • . . reformulated within the framework of the theory of grammatical transformations. In fact. in the first place. it is important to see what limitations can be imposed on their use . that appears on that introduces A. 1963.2) involved in the use of selectional rules. it seems that there are indeed heavy restrictions.that is. 1963. it has become dear that it was a mistake. If this condition of strict local subcategorization is adopted as a general condition on the form of grammar.]. that is." "The Nile." "the fact that he was guilty. as in the case of "The Hague.] is the category of Proper Nouns." "the habit of working hard" . The category [ . Suppose that the rule that introduces Nouns into the grammar is. This position fixes the set of frames that determine sub categorization. the only characteristic of these rules that must be explicitly indicated in the grammar is their position in the sequence of rules.26 If this is cor rect. The category [Det .1 00 CATEGORIES AND RELATIONS IN SYNTACTIC THEORY the transformation is." "the opportunity for him to leave. then the Proper/Common distinction is strict subcategorial.] (con tinuing with the notational conventions for features introduced earlier). for the impersonal it of such sentences as "it strikes me that he had no choice.S']. and does not fall together with the other features introduced in (20)." "it is obvious that the attempt must . more importantly. rather than as part of a freely and independently selected Determiner system) . Perhaps one should utilize this category to account for "quotes contexts" and." with a fixed Determiner that may just as well be taken as part of the Noun itself. The category [Det .S'].the latter involving a sentential Complement with an obligatorily deleted Subject). then the strict subcategorization rules can simply be given in the form the rest being supplied automatically by a convention. The category [ . the following: (49) NP ") (Det) N(S') In this case. In other words.S'] is not realized in so obvious a way as the others. "strictly local" in the sense of note 1 8. we should expect strict subcategorization of Nouns into the categories [Det . furthermore. essentially. and [ .] is simply the category of Common Nouns." "it surprised me that he left. Nouns with no Determiner (or.S'] is the category of Nouns with sentential Complements (such as "the idea that he might suc ceed. [ . [Det . the sen tence "on the train. "he worked at the office" versus "he worked at the job " . be optionally preposed to the sentence. whereas many types of Prepositional. This observation suggests that we modify slightly the first several rules of the base. It is well known that in Verb Prepositional-Phrase constructions one can distinguish various degrees of "cohesion" between the Verb and the accompanying Prepositional· Phrase. or substituting for it by a strictly local transformation in the man ner described in note 1 8). "he ran after dinner" versus "he ran after John") . we note one further consequence of accepting the general condition sug gested in connection with (47) . Clearly. Returning." Both kinds of phrase appear in (5 1) he decided on the boat on the train that is.that is." Clearly. like a Time Adverbial. in fact. has no particular connection with the Verb.Phrase appear in much closer construction to Verbs. "he laughed at ten o'clock" versus "he laughed at the clown" . although the first Preposi t ional-Phrase of (5 1). The poin t can be illustrated clearly by such ambiguous constructions as (50) he decided on the boat which may mean "he chose the boat" or "he made his decision while on the boat. once again. he decided" is unambiguous. replac ing them by . the second Prepositional-Phrase in (5 1 ) is simply a Place Adverbial . It can. as in the e xamples cited. but in fact modifies the entire Verb Phrase or perhaps the entire sentence. ASPECTS OF DEEP STRUCTURE 101 fail. "he chose the boat while on the train. to Verb subcategorization. cannot . which is in close construction to the Verb. Time and Place Adverbials can occur quite freely with various types of Verb Phrase.§ 2. There are many other examples of the same kind (for example . on the one hand. which." which derive from underlying strings with NP's o f the form: it'"'Sentence (the Sentence Complement either being sep arated from it by a transformation. Duration) remain .three times a week remain . the Place and Time Adverbials that are associated with the full Predicate-Phrase. once again. To illustrate. and that might.Place) win . that Verbs are subcategorized with respect to the Prepositional-Phrases introduced by (50iii) but not with respect to those introduced by (50ii) .namely. in fact. It will follow. be in part more closely associated with the Auxiliary (cf.Frequency) (54) dash . but not with respect to Verb Phrase Complements.into the room (V . then.for three hours (Vi. j l (iv) Prep-Phrase -+ � (v) V CS The conventions governing complex symbols will interpret (v) as strictly subcategorizing Verbs with respect to all contexts intro duced in the second part of rule (iii) and in rule (iv). note 23) or with Sentence Adverbials which form a "pre-Sentence" unit in the underlying structure.in England (V .three times a week (V . but not (54) :27 (53) dash . we have such phrases as (53).Direction) last . Thus Verbs will be subcategorized with respect to Verbal Complements.in England last . in connection with the four types of Adverbials listed in (52iv).into the room win .102 CATEGORIES AND RELATIONS IN SYNTACTIC THEORY (52) (i) S � NP"""' Predicate-Phrase (ii) Predicate-Phrase � Aux"""'VP (Place) (Time) be Predicate (NP) (Prep-Phrase) (Prep-Phrase) (Manner) Adj (iii) VP � V S' (like) Predicate-Nominal Direction Duration Place Frequency etc.for three hours . That this is essentially the case is clear from the examples given. " not "the preacher married John and Mary." and so on. fit (in the sense of "the suit fits me.for example : resemble." "I am fitted by the suit. the other contexts introduced in (5 2 iii) will play a role in strict sub categorization of Verbs . the italicized phrases in "he argued w ith John (a bout politics)." "the letter was weighed by John." "two tons is weighed by this car. marry (in the sense of "John married Mary." which. p. Thus we do not have "John is resembled by Bill. Similarly." not "the tailor fitted me." and so on.§ 2 . and he has also observed that these are. since they are introduced by a rule (namely (5 2 ii» the left-hand symbol of which does not directly dominate V." "he talked about Greece. does take Manner Adverbials freely)." which does take Manner Adverbials freely) . the Verbs with following NP's that do not undergo the passive transformation. do not play a role in Verb sub categorization." "John played Othello in England." "John always runs after dinner. whereas the italicized phrases in "John died in Eng land. 8)." "ten dollars is cost by this book. we may have the rule (55) as a rewriting rule of the base and may formulate the passive transformation so as to apply to strings of the form (56)." not "John weighed the letter. Thus Verbs generally take Manner Adverbials freely." which does take Manner Adverbials freely) ." "he decided on a new course of action. That is." and we can have " I was fitted by the tailor. are of types that induce a subcategorization of Verbs. characteristically." etc. and so on." "he aimed (the gun) at John. 1 960a.2 8 These observations suggest that the Manner Adverbial should have as one of its realizations a "dummy element" signifing that the passive transformation must obligatorily apply. cost." "John was married by Mary. the Manner Adverbial participates in Verb subcategorization .). have. but there are some that do not . The Verbs that do not take Manner Adverbials freely Lees has called "middle Verbs" (Lees." "a good book is had by John." and so on (although of course "John was married by Mary" is accept able in the sense of "John was married by the preacher. with an elementary transformation that sub- . In particular. weigh (in the sense of "the car weighed two tons. ASPECTS OF DEEP STRUCTURE 103 Similarly." "he ran after John. " "John is looked up to by everyone. in which case it will also take Manner Adverbials freely. in fact. 1 955. (where the leftmost . .V . was motivated solely by the passive construction (cf. .by -passive ..NP-Manner] . Third.. in (56) requires further specification e . . . a Verb will appear in the frame (56) and thus undergo the passive transformation only if it is positively specified. 7 3-74) . Chomsky. The reason was that V of (56) had to be limited to trans itive Verbs. it cannot contain an NP) . This formulation has several advantages over that presented in earlier work on transformational grammar (such as Chomsky. for the ordinary passive transformation. Chomsky. - NP . it is now possible to account for "pseudo passives. it was necessary to treat pseudo passives by a new transformation...passive . . 1 957) . In fact. it accounts automatically for the restriction of passivization to Verbs that take Manner Adverbials freely. First of all. 1 957. But if passivization is determined by a Manner Adverbial.. ." In the earlier formulation (cf. the schema (56) already permits these passives. for the strict subcategorization feature [.NP - . .g. . with this formulation it is possible to account for the derived Phrase-marker of the passive by the rules for substitution trans formations . and it will be converted into "john is looked up to by everyone" by the same elementary transformation that forms "john was seen by every one" from "everyone saw jo�n. Thus "everyone looks up to John by passive" meets the condition (56). This makes it possible to dispense entirely with an ad h oc rule of derived constituent structure that. so as to exclude the "middle" Verbs have." such as "the proposal was vehemently argued against. . . then V in (56) can be quite free. pp . etc . That is. with Jo hn as the second NP.. Second.Aux . in the lexicon." "the new course of action was agreed on. as just suggested. and ." by a slight generalization of the ordinary passive transformation ..CATEGORIES AND RELATIONS IN SYNTACTIC THEORY stitutes the first NP for the dummy element passive and places the second N P in the position of the first NP: (55) (56) Manner � by . Chapter IX). resem ble. It is perhaps worth while to trace through the argument again to see why this is so. In particular. Furthermore. provides an indirect j ustification for the proposal (cf. in this way. One of the elements that must be internal to the VP. which is internal to the VP by the strictly local subcategorization principle. Similarly. that the Adverbial introduced by (52ii) is not subject to the passive transformation as defined by (56).§ 1I. 99) that strict subcategorization rules be limited to strictly local transformations. we can have "the boat was decided on" in the sense of "he chose the boat. in accordance with this principle." where the phrase "at the office" is a VP-Complement introduced by (52ii) and therefore follows the Manner Adverbial. Thus "John is looked up to" and "John was seen" are formed by the same rule despite the fact that only in the latter case is John the Direct Object of the deep structure . however." Thus the passive sentence corresponding to (50) is unambiguous. The fact that we are able. Notice. Many other facts can be explained in the same way. where "on the boat" is a . is the marker for " passiviza tion. since it plays a role in strict subcategorization of the Verb. ASPECTS OF DEEP STRUcruRE can be an intransitive as well as a transitive Verb. p." where "at this job" is a Verb-Complement introduced by (52iii). the marker for passivization is associated with the presence of the Manner Adverbial. This accounts for the fact that we can have "this job is being worked at quite seriously" from "Unspecified-Subject is working at this job quite seriously. to account for the nonambiguity of "the boat was decided on by John" as con trasted with the ambiguity of "John decided on the boat. it follows that NP's in VP-Complements are not subject to "pseudopassivization" while NP's in V-Complements may be subject to this operation. since it will follow the Adverbial by'-"passive." along with many similar examples. but not " the office is being worked at" from "Unspecified-Subject is working at the office. By the "strictly local subcategorization" principle we know that certain categories must be internal to the VP and others must be external to it ." but not in the sense of "he decided while on the boat. though (50) itself is ambiguous. Since the passive transformation must be formulated with the structure index (56). . among many others. considering the relational character of these notions. . B · . . ." Further more.namely the assumption that strict subcategoriza tion is limited to strictly local transformations . observing that "the boat was decided on by John" is unambiguous and means only that John chose the boat. of presenting structural informa tion of the sort illustrated in ( 2 ) of § 1 in a set of rules that are designed to express precisely the basic linguistic processes involed. it is subject to pseudopassivization by the passive transformation.has empirical support. § 3. The reanalysis (52) requires that the definitions of functional notions proposed in § 2. ." equivalently.. it seems that they are defined only for those major cate gories A that appear in rules of the form X � .. "on the boat.. We observed in § 2. S] rather than as [VP. ..1 06 CATEGORIES AND RELATIONS IN SYNTACTIC TIlEORY V-Complement in "John decided on the boat" (meaning "John chose the boat"). ( 1 1 » be slightly altered.2 (cf. Therefore. This revised formulation of the rules... or X � .. . posed in § 1.Predicate-Phrase (ii) Predicate-Phrase � Aux.2 that these notions are defined only for what we caIled "major categories. but where "on the boat" is a VP Complement in "John decided on the boat" (meaning "John decided while he was on the boat. . where B is also a major category. A . incidentaIly. S].VP (Place) (Time) . We may now consider a generative grammar with a base component containing. AN ILLUSTRA TIVE FRA GMENT OF THE BA SE COMPONENT Let us now summarize this discussion by returning to the original problem. it is not subject to pseudopassivization since it does not meet the condition (56). illustrates another property of the tradi tional functional notions.. the rules and rule schemata (57) and the lexicon (58) : (57) (i) S � NP . This seems quite natural. John decided"). · B · · ·A · · · . we conclude that the premise of this argument . Thus we might perhaps define the notion "Predicate-of" as [Predicate Phrase. we derive the sen tence "sincerity may frighten the boy" of § 1 .. +Abstract. where a is an N and p is an N Adjectiv� � CS/a . (iv) Pre d icate � � (xiii) (xiv) (xv) (xvi) (xvii) (xviii) (58) (xii) re Icate-Nomma I . Prep-Phrase � Direction. AN ILLUSTRATIVE FRAGMENT OF THE BASE (iii) VP (v) (vi) (vii) (viii) (ix) (x) (xi) .' k e) P ) } ) .Count.. Predicate � { AdjeCtived ' (l . . Adding the rules that realize Definite as the and non-Definite as null before a following non-Count Noun.§ 3. predicate P) (Prep-Phrase) (Prep-Phrase) (Manner» V . Place. [+V. ]) (may. .p) } This system of rules will generate the Phrase-marker (5 9). .. . + [+Abstract] Aux . +Det .) N � CS [+Det . +Count.. +Det . etc. . [+M .. However.AUX . +Human..] � [±Animate] [+Ani mate] � [±Human] [. Freq u en cy. + . . . + Object-deletion. . ..] � [±Count] [+C oun t] � [±Animate] [+ N . .. with the Phrase marker (59).... [+N ... ... . . 1 . Notice that this fragment of the base is "sequential" in the sense of § 2 . of) Article (post-Article) Article � [±Defin ite] j l� COMPONENT 1 07 COpula. Aux � Tense (M) (Aspect) Det � (pre-Article ." holding (sincerity.Det [ + Animate]..(De t . Duration. + .. . . this is a relatively minor matter of appropriate formalization and involves nothing of principle .N P. +Animate.Coun t] � [± Abstract] [+V] � CS/a. ]) . ]) (boy. We have only sketched the procedure for constructing a Phrase-marker of the req uired sort from a derivation. ... In particular. ]) (frighten.. (59) represents not only all information involvi ng the relation "is a.. [+N. V � CS NP � (Det) N (5. # o 00 ..- � I I [+Defini te] N H � I sincerity frighten H � H" o z � I 2 • • • )Jf '" Z '" >< Z B boy � C'l � toJ � o . NP Predicate-Phrase Det Aux VP � M N Article I I V G I may Cl' C2' - F I � Det [-Definite] F� "JV I I NP I'" Article toJ �.(59) 11-5.. .. lZl .= = = = F - COl F F1' = [+Det-] = Common [-Count] G G1' = [+-NP] = Transitive [+[+Abstract]Aux-Det[+Animate]] H H1' = [+Det-] [+Count] Fl = � G2' = [+Object-deletion] � F" H'" = = [+Abstract] H" = [+Animate] [+Human] ." o Z till � C le .l i � � lZl i I i � . where D is a phonological feature matrix and C a complex symbol. The Phrase-marker (59) provides directly all information of the sort specified in (2i) and (2iii). with one exception. Consequently. to suppose that this is in general merely a question of added detail. and we have given no semantic features in either (58) or (5 9). boy must be specified as [-V]. If the analysis that we have given is correct. § 2. and.2). p. but it would be a serious mistake." and not only must frighten be specified as [-N]. First of all. as we have observed. the lexical rule (d. lexical entries must be specified negatively for features corresponding to contexts in which they may not occur. so as to exclude it from the position of turn in "his hair turned gray. Given a lexical entry (D.3. for example. then it is devices of the sort just exhibited that are implicit in the informal statements of traditional grammar summarized in (2). We can deal with this matter by adopting several additional conventions governing the base component. 84) permits substitution of D for any complex symbol K that is not distinct from C. we may assume that a base rule that analyzes the lexical category A into a complex symbol automatically includes the feature [+A] as one of the elements of this complex symbol (see (20). to exclude it from the position of boy in this sentence. but it must also be specified negatively for the feature [ ." and so on. There are clearly other syntactic features that must be indicated. These negative specifications were not actually given in (58) . Thus in (58). so as to exclude it from the position of frighten in "sincerity may frighten the boy.Adjective]. . Notice that neither the lexicon (58) nor the Phrase·marker (59) is fully specified.110 CATEGORIES AND RELATIONS IN SYNTAcrIC THEORY between strings and the categories (many of them now repre sented by features) to which they belong but also the hierarchic relation among these categories that is provided by the rules and mirrored precisely in the derivation. in this case. In part.C). it is clear how these gaps can be filled. functional information of the sort specified in (2ii) is derivable from this Phrase·marker as wen . One final comment is necessary in connection with the lexicon (58). to which we shall turn in the next section. [-M] (cf. For the time being. the distinctness require ment of the lexical rule will now exclude items from certain contexts. those corresponding to features in which it can appear) (ii) list only the features correspon ding to frames in which the item can appear. we may assume that each lexi cal entry automatically. the entry for b oy contains [-V]. we may adopt one of the following conventions : (i) list in the lexicon only the fea tures correspon ding to frames in which the item i n quest ion cannot appear (rather than . There is a fundamental distinction between the rewriting ru les (57) and . and thus pre fer to leave the question open.I. by convention. and permit them in others. [-Adjective]. I have no strong examples to support one or another of these assumptions.29 Third. In any case. We shall return brie fly to the problem in Chapter 4. (iii) and (iv». note 9). and (i i) is correct if the most highly valued grammar is that in which the distribution of items is most constrained (similarly. in the case of features introduced by strict subca tegoriza tion or selectional rules (wh a t we have called the "contextual features"). Th us in (58). These conven tions embody alternative empirical hypotheses concerning val uation of grammar. Summary The fragmen t presente d in § 3 il lus trates the kinds of rules that apparen tly are to be found in the base component. as in (58) (in case (i) or case (ii) we add the further convention that an item is specified in the opposi te way for every contextual feature not mentioned in its lexical entry) (iii) adop t (i) for the strict subcategorization features and (ii) for the selectional features (iv) adopt (i i) for the stric t subcategorization features and (i) for the se le c tional features.§ 4. contains the fe a ture [-A] for every lexical cate gory A. as in (5 8) . unless it is explicitly provi ded with the feature [+A]. § 4. Thus (i) is correct if the most highly valued grammar is that in which the dis tribution of i tems is least constrained. TYPES OF BASE RULES I I I Second. TYPES OF BASE R ULES § 4. (viii). Among the rewriting rules of the base component we can distinguish branching rules. so far. (xiii). § 2 . (iv). for example . The lexical rule need not be stated in the grammar since it is universal and hence part of the theory of gram mar. (xvii). (v). (xv». such as (i). Moreover. although once a subcategorization rule has been applied to a certain category symbol 0' no branching rule can be applied to any of the symbols that are derived from 0'. language-independent definition of the function f of (1 4iv). Thus a branching rule analyzes a category symbol A into a string of (one or more) symbols each of which is either a terminal symbol or a non terminal category symbol. See Chapter 4. It thus has the status of a convention determining the interpreta tion of the grammar. § 6. Its status is just like that of the principles that define "derivation" in terms of a system of rewriting rules. (xviii» or con text-sensitive (such as (vi). Chapter 1 . we may say that the lexical rule in fact constitutes part of the general. (xii). In terms of the framework of § 6. In particular. rather than the status of a rule of the grammar. Notice that these two sets of rules (branching and subcategorization) are not ordered with respect to one another. on the other hand. We have. A subcategorization rule. Branching rules and subcategorization rules may be con tex t free (such as all of the branching rules of (57) and (x). and we must apparently weaken it slightly. restricted the subcategorization rules to lexical categories. introduces syntactic features. and thus forms or extends a complex symbol. the subcategorization rules that are context-sensitive (60) A � Z/X . (iii). (vii). Chapter 1 . we have not permitted rules of the form (60) in wh i ch A is a complex symbol and Z a terminal or category symbol or a string of more than one symbol.W . (xiv). This restriction may be a bit too severe. such as all others of (57).1 12 CATEGORIES AND RELATIONS IN SYNTACTIC TIlEORY the lexicon (58) . Notice that (57) contains no context-sensitive branching rules. All rewriting rules are of the form The branching rules are those rules of the form (60) in which neither A nor Z involves any complex symbols . (xvi). from subcategorization rules. (ii). (xi). in fact. That is. § 4 . We observed earlier that the notion of grammatical function defined in § 2. additional condition on the base. but that once a subcategorization rule has applied to form a complex symbol I. and selectional rules (such as (57xiv).§ 4. furthermore.2 did not yet account for the assignment of the Subject-Verb rela tion to the pair sincerity. but the defini- . no branching rule can later apply to I (but cf. defines a selectional relation between two positions in a sentence . frighten and the Verb-Object relation to frighten. § 2 . Selectional rules and grammatical relations We shall say that a selectional rule. (57xv». have been defined in terms of the heads of major categorie s (cf. p. 99). these may be interspersed in the base. Perhaps this should be imposed as a general. in effect. given the gramma r (57). no strict subcategoriza tion rule applies later to develop I further. Such selectional relations determine grammatical rela tions. which sub categorize a lexical category in terms of the frame of category symbols in which it appears. Chapter 4. it appears from the examples that I have considered. TYPES OF BASE RULES are. but once a selectional rule has applied to form the complex symbol I. namely strict subcategorization rules (such as (57vi) and (57viii». distinguished two important sub types. in the case of (57xiv). at least. in one of the senses of this traditional term.2. The suggested definition of grammatical relation would account for these assertions. (57xv). These are important facts. (58) . such as (57xiv). The same notion of gramm atical relation could. strictly local tr ans formati onal rules (cf. The same relation apparently holds between strict subcategorization rules and selectional rules. 2). which subcategorize a lexical category in terms of syntactic features that appear in specified positions in the sentence. Among the context-sensitive subcategorization rules we have. boy in sincerity may frighten the boy (=( 1 » . to which we return in Chapter 3. the position of the Verb and that of the immediately preceding or immediately following Noun. § 2).for example. So. We noted that subcategorization rules may follow branching rules in the sequence of rules constituting the base. that we were to subcategorize the Verb by a context-free rule. for the Verb. post-+ [+Animate]-Object] .Aux + a + Det - These rules would assign features of the Verb to the Subject and Object. instead. +[+Abstract]-Subject. 7 3-74. and the feature [post-+[+Abstract]- (65) [post-+ [+Abstract]-Subject. and then to use a selectional rule to determine the subcategorization of the Subject and Object. lexically marked as allowing an Abstract Subject and an Animate Object. just as in (S7) we gave such a rule to determine the choice of Verb in terms of Subject and Object.30 Consider now the selectional rules (S7xiv). clearly. + [+Animate]-Object]S1 [+V. Thus we would have such rules as N -') CS/ j . just as (5 7xiv) assigned features of the Subject and Ob ject to the Verb. We might have. For example. the Object would have the features (64) [pre-+[ +Abstract]-Subject. + [+Animate]-Object] 1 14 CATEGORIES AND RELATIONS IN SYNTACTIC THEORY which can be replaced by a lexical item such as frighten. (S7xv). the feature [pre+[+Animate]-Object] is irrelevant to choice of Subject Noun. the Subject would be specified as having the features a l ' where a is a V Similarly. With this notion now defined. We must now give a context-sensitive selectional rule to deter mine the choice of Subject and Object. we have completed the analysis of the in formal grammatical statement ( 2) of § 1 . if the Verb is (62). such a rule as Thus we might in particular form the complex symbol (62) (6 1) V -') [+V.tion in terms of selectional relations seems somewhat more natural and avoids the problem noted on pp. the Subject and Object). + [+Abstract]-Subject. which con strain the choice of Verb and Adjective in terms of a free choice of certain features of the Noun (in this case. pre-+ [+Animate]-Object] But. Suppose. 1 . In short. rather than conversely. for each feature X. the choice of elements for the posi tion "Subject of a Verb with Animate Subject" is the same as the choice of elements for the position "Object of a Verb with Animate Object. Much more serious than this. In § 2. and conversely. Adjective.Subject] is irrelevant to choice of Object Noun. The problems are magnified when we bring into account the independent Noun·Adjective selectional rules. in that it involves no more mechanism than is actually forced by the linguistic facts. If this is true. where X is any feature. and the fact that X occurs in both of them cannot be referred to by a rule of the grammar (unless we complicate the mechanism further by allowing features to have a feature composition them selves). it is possible to take another significant step toward a general characterization of the categories Noun. that within the framework so far · developed. 2 .X-Object]. is the fact that a Noun must be marked in the lexicon for the feature [pre-X-Subject] if and only if it is marked for the feature [post. etc. however. One would imagine that a similar argument can be given for any language. [post-X-Object]. I defined "lexical category" and "major category. Consequently. there is no alternative to selecting Verbs in terms of Nouns (and." § 4. Verb. That is.2).2. the decision to choose the complex symbol analysis of Verbs independently and to select Nouns by a selectional rule in terms of Verbs leads to a quite considerable complication of the grammar. this framework seems to be optimal. Moreover. Furthermore. are single symbols. the features [pre-X-Subject]. by a similar argument. Adjectives in terms of Nouns). (see §§ 2 . In much the same way we can rule out the possibility of allowing Subject to select Verb but Verb to select Object. for each X. only the features [pre-+ [+Animate]-Subject] and [post + [+Animate]-Object] . We see." Animate Nouns appear in both posi tions. But the feature [Animate] is no longer available for Nouns. a mass of perfectly ad hoc rules must be added to the grammar to assign to Nouns with the feature [pre-X-Subject] also the feature [post-X-Object]. TYPES OF BASE RULES . then. Suppose that among the lexical categories. A major category that directly dominates . Among the major categories introduced in the analysis of Sentence. we shall have succeeded also in giving a substantive specification to the functional notions discussed in § 2 . . . and nonmajor categories in general terms . and there are many formal features of the grammar that can be brought into consideration in doing so . so few generative grammars). . NP . There is no problem in principle of sharpening or generalizing these definitions in one way or an other. One might now go on to attempt to characterize other lexical. . major. . It will be obvious to the reader that this characterization is not intended as definitive in any sense. and one that directly dominates VP. we label as Noun the one that is selectionally dominan t in the sense that its feature composition is determined by a context·free subcategorization rule. 2 . . .1 I6 CATEGORIES AND RELATIONS IN SYNTACTIC THEORY the latter being a lexical category or a category dominating a string containing a lexical category . This is a consequence of the fact that there are so few grammars that attempt to give an explicit characterization of the range of sentences and structural descriptions (that is. N . The problem is merely that for the moment there is no strong empirical motiva tion for one or another suggestion that might be made in these directions. and perhaps to give a substantive characterization to the . even in a partial sketch . it will no doubt be possible to give empirical justification for various re finements and revisions of such loosely sketched proposals as these. we can des ignate VP. . The reason has already been indicated in note 2. we can des ignate Predicate· Phrase . . we now designate as NP the one that is analyzed as . as the lexical category X that appears in a string · · · X · · · N p · · · or · · · N p · · · X · · · directly dominated by VP (assum ing that there can be only one such X) or as the lexical category that may obtain its features from selectional rules involving two or more N's (if transitivity is a category that is universally realized) . To the extent that we can do this. its features being car ried over by selectional rules to other lexical categories . As explicit grammatical descriptions with this goal accumulate. We can define V in various ways for example. and only in these. a priori. they are to be stated only in general linguistic theory. To say that formal properties of the base will provide the framework for the characterization of universal categories is to assume that much of the structure of the base is common to all languages. However.§ 4. but only a h ighly restricted set of elementary structures from which actual sentences are constructed by transformational rules. they pertain to the form of language in general rather than to the form of particular languages. 1-2 . To the extent that relevant evidence is available today.82 The base Phrase-markers may be regarded as the elementary content ele ments from which the semantic interpretations of actual sen tences are constructed. whose origins can again be traced back at least to the Grammaire genera le et raisonnee (Lancelot et al. explicitly characterize the full range of sentences..ss Therefore the observation that the semantically significant functional notions (grammatical rela tions) are directly represented in base structures. it seems not unlikely that it is true. 1 660) . should come as no surprise. This is a way of stating a traditional view. and thus presumably reflect what the mind brings to the task of language acquisition rather than what it discovers (or inven ts) in the course of carrying out this task. as part of the definition of the notion "human language" itself. the traditional view that such substan tive characterizations mus t ultimately refer to semantic concepts of one sort or another. Once agai n. Insofar as aspects of the base structure are not specific to a particular language. TYPES OF BASE RULES universal vocabulary from which grammatical descriptions are constructed. as in §§ 2 . Instead. In traditional terms. furthermore. quite natural to suppose that formal properties of the base will pro vide the framework for the characterization of universal cate gories . they need not be stated in the grammar of this language . Thus to some extent the account of the base rules suggested here may not be long to the grammar of English any more than the definition of . in itself. there is no reason to rule out. and i t is. it is clear that this attempt to characterize universal categories depends essentially on the fact that the base of the syntactic component does not. 2. shown a great diversity in the surface structures of languages. Chapter 1 . Since the origins of this work in the Gram maire generale et raisonnee. In connection with the selectional rule (57xiv). Modem work has. the evidence that has been accumulated in modem study of language does not appear to suggest anything of this sort. there is no reason to expect unifonnity of surface structures.. Insofar as attention is restricted to surface structures. namely that the Subject or Object may be selected in terms of an independent.'"'f3 } a'"'Aux CS/ {a'"'AuX . Cf.} Det'J3 CS/ - In terms of evaluation measures that have so far been proposed . the most that can be expected is the discovery of statistical tendencies. should be preferred to the alternative (67) . and. which I now repeat in less abbreviated fonn as (66). choice of Verb. However. but this claim seems to me very much exaggerated. (66) ��)} [+V] (67) ��)} [+V] (11) � (11) � { Q("AUX . or partially independent. indeed. and the findings of modem linguistics are thus not inconsistent with the hypotheses of universal grammarians. we have now conclusively ruled out one possibility. since the study of deep structure has not been its concern. Consequently. Not quite so simple is the question of whether this rule. in fact. it has not attempted to show a corresponding diversity of underlying structures. it has been emphasized that the deep structures for which universality is claimed may be quite distinct from the surface structures of sentences as they actually appear. §§ 6 and 8. such as those presented by Greenberg ( 1 963) . It is commonly held that modem linguistic and anthro pological investigations have conclusively refuted the doctrines of classical universal grammar.1 I8 CATEGOIUES AND RELATIONS IN SYNTACTIC THEORY "derivation" or of "transformation" belongs to the grammar of English. The fact that languages may differ from one an other quite significantly in surface structure would hardly have come as a surprise to the scholars who developed traditional universal grammar. there is no way of choosing between these. however. Thus command would be excluded from the context (66ii). It may appear at fi rs t that this is little more than a terminological question.Det [+Animate]] . for simplicity In (68iii) we can also have command.the platoon our respect (iii) his decision to resign his commission In (68i) we can have the Verb com'TlUlnd (I neglect. when command appears in this context. wh ich can also appear in (68i) but not (68iii) . and [ . That is. questions of choice of Auxiliary). but in a different though not totally un related sense. since a crucial question of exposition. [[+Abstract] Aux . the lexical entry for frigh ten will be positively specified for the feature [[+Abstract] Aux . it will be marked in such a way as to permit it to have either a n Animate or an Abstract Noun as Subject or Object. and (67ii) assigns a feature of Object selection to Transitive Verbs.Det [+Animate]] and [[+Abstract] Aux Det [+Abstract]]. the Verb command will be positively marked for the features [[+Animate] Aux . It should be noted. On the other hand.Det [+Animate]]. as required . that the grounds for this decision are very weak. (66i) assigns certain features to Transitive Verbs and (66ii) to In transitive Verbs. If we select the alternative (67). for example. It is for such reasons that I selected the alternative (66) in the grammatical sketch . (67i) assigns a feature of Subject selection to all Verbs. TYPES OF BASE RULES 1 19 (see. for example. [ .Det [+Animate]]. In (68ii) we cannot have command.]. if we select (67). but we can have. i bid. Thus consider the following contexts: (68) (i) he the platoon (ii) his decision to resign his commission --. If we choose (66).] and [ . If we select the alter native (66). this is not at all obvious. . as in many such cases. command will be positively marked for the features [[+Animate] Aux . baffle.).Det [+Ab stract]]. Chapter 3). 1 9 55 . but not [[+Abstract] Aux . it will be positively specified for the two features [[+Abstrac t] Aux .Det [+Animate]] .§ 4.]. But this specification fails to indicate the dependency between Subject and Object illustrated by the deviance 'of (68ii). Chomsky. In accordance with the usual conventions for obligatory application of rewriting rules (cf. but. Clearly.namely.IES AND RELATIONS IN SYNTACTIC THEOR. However. Recall that the notation [+Animate] Aux . The context-sensitive subcategorization rules are further subdivided into strict subcategorization rules and selectional rules.Det [+Abstract].3. For some further related discussion. One might propose. the same number of features must be indicated in the lexicon. With the choice of (66). Suppose. Further remarks on subcategorization rules We have distinguished. this comment does not exhaust the question. The categorial component consists solely of branching rules. I have so far not been able to find stronger examples. In particular. § 4. how to enter lexical items with a range of distinct but related syntactic and semantic features . which are possibly all context free (see Chapter 3). is a single symbol designating a particular lexical feature. between branching rules and subcategorization rules and between context-free and context sensitive rules. in the case of Verbs for which choice of Subject and Object is independent. It may well be that to a large extent the form of the categorial component . the branching rules of (57) would constitute the categorial component of the base of this fragment of English grammar. a categorial component and a lexicon.1 20 CATEGOR. but this is a misinterpretation of the notational system. for example. then. whereas the context-free subcategorization rules introduce in herent features. These rules introduce contextual features. It seems at first as though a certain redundancy results from the decision to select (66) over (67). that the sub categorization rules be eliminated from the system of rewriting rules entirely and be assigned. the features seem more "complicated. In fact. this is a perfectly feasible suggestion. see Chapters 3 and 4. to the lexicon.is far from settled. in the base. that the base is divided into two parts.Y . in our framework. even in this case. in effect. alternatively. The primary role of the categorial component is to define implicitly the basic grammatical relations that function in the deep structures of the language. by any means." in some sense. Suppose that we have a lexical entry (D. We stipu lated previously that the lexical rule permits D to replace the symbol Q of the preterminal string ((>Q1/1 provided that Q is not distinct from C.YJ. Consider." The subcategorization rules can be assigned to the lexical component of the base in the following way. the rules that introduce contextual features. can be regarded as syntactic redundancy rules. 84). § 2. and hence assigned to the lexicon.§ 4. Suppose that we now require. Obviously. the context·free subcategorization rules. But the rules that introduce contexual features into complex symbols can be eliminated by an appropriate reformulation of the lexical rule. and they assign corresponding contextual features. together with the principle j ust stated. together - . Thus strict subcategorization features for an item of the category A must involve frames that. we require that ((>QI/I equal ((>1 ((>2Q1/11 1/12' where ((>2 is dominated by X an d 1/11 by Y in the Phrase-marker of ((>Q1/1. First of all. the contextual features must appear in lexical items. we can convert it to a context-sensitive rule by con ventions of the following sort. A lexical entry may be substituted in these positions if its con textual features match those of the symbol for which it is sub stituted. then. in addition. that this occurrence of Q actually appear in the frame X Y. that is. This convention can be formulated precisely in terms of the notion "Analyz ability" on which the theory of transformations is based. Our earlier conditions on subcategorization rules (cf.tH) become conditions on the kinds of contextual features that may appear in lexical entries . These rules select certain frames in which a symbol appears. p. That is. Instead of formulating this as a context-free rule that operates by matching of complex symbols. TYPES OF BASE RULES 1 21 is determined by the universal conditions that define "human language. C) where D is a phonological feature matrix and C is a complex symbol containing the feature [+X . We now eliminate all context-sensitive subcategorization rules from the grammar and rely on the formulation of lexical features. to achieve their effect. such as {57 ix-xiii). the rule that introduces lexical items into derivations (cf. and the selectional features must involve the lexical categories that are the heads of grammatically related phrases. C). for each lexical category A .III 2 CATEGORIES AND RELATIONS IN SYNTACTIC 'IHEORY with A . A preterminal string is generated by the branching rules of the categorial component . The complex symbol C contains inherent features and contextual features. there is a rule A � !:t. We now have no subcategorization rules in the categorial com ponent of the base. form the single constituent B that immediately dominates A . strictly local in the sense of note 1 8. The lexicon cons ists of entries associated with certain substitution transformations that introduce lexical items into strings generated by the categorial component. This formulation brings out very clearly the sense in which our utilization of complex symbols was a device for introducing transformational rules into the base com ponent. This transformation substitutes (D. All contextual restrictions in the base are provided by these transformational rules of the lexicon . is a fixed "dummy symbol. in the sense outlined earlier. furthermore. A lexical entry is of the form (D. the substitution transforma tion is. suppose that (for uniformity of specification of transformational rules) we add the convention that in the cate gorial component." The rules of the cate gorial component will now generate Phrase-markers of strings consisting of various occurrences of !:t.) . Thus the categorial component may very well be a context-free constituent structure grammar (simple phrase structure gram mar) with a reduced terminal vocabulary (that is.see note 1 5) for a certain occurrence of !:t. where D is a phonological matrix and C a complex symbol . in the Phrase·marker K if K meets the condition I. (marking the positions of lexical categories) and grammatical formatives . In fact. where !:t. C) (now regarded as a complex terminal symbol . Lexical entries sub stitute for the lexical categories of a preterminal string by the principle just stated . which is a Boolean condition in terms of Ana lyzabiIity in the usual sense of transformational grammar. with all lexical items mapped into the single symbol !:t. We can restate this system of features C directly as the structure index I for a certain sub stitution transformation. Where strict subcategorization is involved. The function of . for the present. the contextual features (structure indices of substitution transformations) that may appear in the lexicon are limited by the conditions on strict subcategorization and selectional rules previously discussed. But in the earlier formulation. some in terms of Subject selection.§ 4. Similarly. The trans formational rules map deep s tructures into surface structures. The ordering of the rewriting rule A � CS places an additional limitation on the class of contextual features that may be used. In both formula tions. The earlier proposal was somewhat more restrictive in certain respects. If not. At least the first of these functions appears to be carried out in a very general and perhaps universal way by these rules. in terms of a lexical rule based on the distinctness condition. . and they determine the ordering of elements in deep structures.that is. This way of developing the base component is not quite equivalent to that presented earlier. the issue discussed in § 4. the system of base rules exclusive of the lexicon and the subcategorization rules that we. Because of the greater flexibility that it allows. this must be the preferable formulation of the theory of the base. If so. It is an interesting question whether the greater flexibility permitted by the approach of this subsection is ever needed. then the other formulation. there is a further restriction. § 4+ The role of categorial rules We have defined the categorial component as the system of rewriting rules of the base . is to be preferred. and some in terms of Object selection. TYPES OF BASE RULES the categorial component is to define the system of grammatical relations and to determine the ordering of elements in deep structures. The rules of the categorial component carry out two quite separate func tions : they define the system of grammatical relations. certain Verbs can be restricted in terms of Subject and Object selection. We shall return to this question in Chapter 4. with sub categorization rules given as rewriting rules. regard as belonging to the lexicon. 2 regarding examples (66)-(68) does not arise in the new formulation. although NP""""VP oF VP""NP. It has been suggested several times that these two functions of the categorial component be more sharply separated. the categorial component should contain the corresponding rules (70). and that the second. since they not only define an abstract system of gram matical relations but also assign an abstract underlying order to the elements. no order is assigned to the elements on the right-hand side of the rule. the Phrase-marker generated by such rules as (70) will be representable as a tree-diagram with labeled nodes and unlabeled lines.CATEGORIES AND RELATIONS IN SYNTACTIC THEORY perhaps reordering elements in various ways in the course of this operation. The greater abstractness of set-systems. Thus the grammatical relations defined by (70) are neither more nor less "abstract" or "order-independent" than those defined by (69) . so far as grammatical relations are concerned. is a myth. in essence. VP} = {VP. The rules (70) can be used to define grammatical relations in exactly the way indicated for the rules (69). and can lead to a study of grammatical relations that is independent of order. Such is the import of the proposals regarding the nature of syntactic structure to be found in Curry ( 1 96 1) and Saumjan and Soboleva ( 1 963 ). that in place of such rules as (69). . thus {NP. VP} VP -+ {V. perhaps.34 They propose. The Phrase-marker generated by such rules as (69) will be representable as a tree-diagram with labeled nodes and labeled lines. be eliminated completely. NP} (70) In (70). Proponents of set-systems such as (70) have argued that such systems are more "abstract" than concatenation-systems such as (69). this being a phenomenon that belongs only to surface structure. where the element on the right is a set rather than a string: (69) S -+ NP""""VP VP -+ V"""" N P S -+ {NP. The rules (69) convey more information than the corresponding rules (70) . NP}. The first set will assign an intrinsic order to the elements of the underlying unordered Phrase markers (that is. the problem of giving empirical support to this theory has not yet been faced. whenever an attempt to account for such structures has actually been undertaken. the systems of grammatical relations defined in the two cases are identical. no proponent of a set-system has given any in dication of how the abstract underlying unordered structures are converted into actual strings with surface structures. and there is no internal organization . in fact. it has invariably been found that different sets in a single language lead to the same decision as to the abstract underlying order of elements.that is. it has invariably been found that there are strong reasons to assign an internal organization and an inherent order of derivation among the items constituting such a set. Con- . It provides the base Phrase-markers required for the application of the sequences of transformations that ultimately form surface structures.within the set of structures. it is an entirely empirical question." "to please John is difficult for us. there is no way of determining which theory is correct. "for us to please John is difficult. The first set of rules simply converts a set-system into a concatenation-system. for the theory of the categorial component.§ 4. There is no evidence at all to suggest that either of these steps can be omitted in the case of natural languages." "it is difficult for us to please John. A priori. But. Furthermore. Presumably. Hence. each member is directly related to the underlying abstract representation. In fact. The second set of rules will be grammatical transformations applying in sequence to generate surface structures in the familiar way. and the evidence presently available is overwhelmingly in favor of con catenation-systems over set-systems. the proposal that." or "John is difficult for us to please"). no order of derivation . it seems that a set-system such as (70) must be supple mented by two sets of rules. TYPES OF BASE RULES in fact. Hence. it will label the lines of the tree-diagrams representing these structures). the categorial component should be a set-system entails that in a set of syntactically related struc tures with a single network of grammatical relations «(or ex ample. is a paraphrase of the original. and what its semantic functions are.1 26 CATEGORIES AND RELATIONS IN SYNTACTIC THEORY sequently. 4 2) and Mohawk. Second. it has no bearing on it at all. even richly inflected languages do not seem to tolerate reordering when it leads to ambiguity. In this case. 1 956. and the rule for realizing underlying abstrac t representations would be extremely simple But there is no J<. that there are several underlying generalizations that determine when such reordering is permissible. p. Nevertheless. No gram matical transformations would be needed. but. It seems. The phenomenon of so-called "free word order" is sometimes mentioned as relevant to this issue. there is no reason to consider the set-system. Suppose that for some language each permutation of the words of each sentence were to give a grammatical sentence that. For one thing. the set-system simply cannot be considered seriously as a theory of grammar. the free word order phenomenon is an interesting and important one. . it seems that the interpretation will invariably be that "Die Mutter" is the Subject (unless it has contrastive Stress. First of all.nown language that remotely resembles this description. In every known language the restrictions on order are quite severe. for obvious reasons. and therefore rules of realization of ' abstract structures are necessary." in which the inflections do not suffice to indicate grammatical function. so far as I can see. the set-system would be much superior for the categorial component of the grammar of this language. in which case it may be taken to be the Subject or the Object) . Thus in a German sentence such as "Die Mutter sieh t die Toch ter. for the time being. the Verb contains affixes designating the Subject and Object. The same seems to be true in other languages as diverse as Russian (cf. it should be emphasized that grammatical transformations do not seem to be an appropriate device for expressing the full range of possibilities for stylistic inversion. in faci. Peshkovskii. as a possible theory of grammatical structure. In the latter. but where the reference is . rather. richly inflected languages tolerate stylistic reordering much more extensively than languages that are poor in inflection. Until some account of such rules is suggested. and much too little attention has been given to it. If this is universal. on the theory of grammatical structure. phenomenon. under normal intonation (I am indebted to Paul Postal for this in formation).35 It might. and it is not statable in terms of the theory of transformations. it suggests the generalization that in any language.that is. (As a special case of this.) Something of this sort seems to be true. TYPES OF BASE RULES ambiguous. it is one that has no apparent bearing. then. . for the moment. it will follow that inflected languages will tolerate reordering much more freely than uninflected ones. the rules of stylistic reordering are very different from the grammatical transformations. In general. of Chapter I ) . which are much more deeply embedded in the grammatical system . in fact. be argued that the former are not so much rules of grammar as rules of performance (cf. though this is surely an interesting. In any event. §§ I . up to the point where a structure is produced that might have been generated independently by the grammatical rules.§ 4. the initial NP is taken to be the Subject. 2 . stylistic inversion of "major constituents" (in some sense to be defined) is tolerated up to ambiguity . consider a base component which generates the Phrase-markers ( 1 )-(3). In § 1 . and continue to use the fragment of § 3.) As it stands. with a different choice of Auxiliary." were we to modify it by deleting S' from the Determiner associated with man . a specialist. the theory of the base component sketched in § 4. would be the basis for the sentence "John was examined by. Similarly. as an illustrative example of a grammar. the base Phrase-marker ( I ) must be supplemented by another Phrase-marker. a surface structure) in the process. which. Chapter 1 .3 of Chapter 2. automatically assign to the sentence a derived Phrase-marker (ultimately. we defined the basis o f a sentence as the sequence of base Phrase·markers that underlies it. however.3 Deep Structures and Grammatical Transformations LET us adopt. (In this case. tentatively. For concreteness.1 The base Phrase-marker (3). Chapter 2. to form the basis for some sentence. furthermore. the passive transformation is followed by the deletion of unspecified agent." The Phrase-marker ( I ) would be the basis for the sentence "the man was fired. appropriately modified to exclude sub categorization rules from the categorial component of the base. The basis of a sentence is mapped into the sentence by the transformational rules. (2) alone cannot serve as a basis for a sentence because the S' appearing in the Verbal Complement 1 28 . a transform of which will fill the position of S' in ( I) and thus serve as a relative clause qualifying man. The base will now generate base Phrase·markers. s.DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS 129 ( 1) #. .# NP A I Aux past � I � VP V fire Predicate-Phrase I � the NP Manner A Det N S' man I by � passive Det the � � � I I I � NP Predicate-Phrase N #-s.# Aux VP man past V persuade I I I NP N of i\� NP John N A I S' b. (2 ) . the sequence of base Phrase-markers ( 1 ) . informally. (5) (1) _ . by the dia gram (5). (3) is the basis for the well-formed sentence (4) the man who persuaded John to be examined by a specialist was fired The "transformational history" of (4) by which it is derived from its basis might be represented. however.1 30 DEEP STRUGrURES AND GRAMMATICAL TRANSFORMATIONS Cs) Det � � � � I I I NP Predicate-Phrase N Aux VP #-s.# a specialist nom V examine N I I /\ John NP Manner I by passive must be replaced by the transform of some other Phrase-marker_ In fact. and introduction of them changes nothing relevant to this discussion." and then Tto. apply the Passive trans formation Tp to the base Phrase-marker (3). as well as other details. well known." giving a Phrase-marker for "the man persuaded John to be examined by a specialist". specified features. category symbols. The basis of the sentence is the sequence of base Phrase markers that constitute the terminal points of the tree-diagram (the left-hand nodes. It represents the transforma tional structure of the utterance (5) very much in the way a Phrase-marker represents the phrase structure of a terminal string. by a generalized (double-base) substitution transformation TE. In fact. giving a Phrase marker for "the man persuaded John of A John nom be examined by a specialist".2 The deep structure of an utterance is given completely by its Transformation-marker. which contains its basis. to this Phrase-marker apply the passive transformation and agent deletion (TAD) ' giving (4)' I have left out of this description quite a few transformations that are necessary to give the correct form of (4). embed the result in the base Phrase-marker (2). but these are. in (5» . and with the developments of the preceding sections. The surface structure of the sentence is the derived Phrase-marker given as the output of the operations represented in the Transformation marker. in place of S'. by and large. which deletes the repeated NP "John. The diagram (5) is an informal representation of what we may call a Transformation-marker. When Transformation-markers are . next embed this in the position of S' in (I). which replaces "of A nom" by "to.DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS 131 We interpret this a s follows: First. to this apply the relative transformation TR' which permutes the embedded sentence with the following N and replaces the repeated phrase "the man" by "who. a Transformation-marker may be formally rep resented as a set of strings in an alphabet consisting of base Phrase-markers and transformations as its elements. by TE. to this apply first TD. just as a Phrase-marker may be formally represented as a set of strings in an alphabet consisting of terminal symbols." giving a Phrase-marker for "A fired the man who persuaded John to be examined by a specialist by passive" . 4 of Chapter 2 . we have used S' as the dummy symbol . They have also succeeded in .8 It follows.this assumption is also implicit in Fillmore. then. since semantic interpretation will now be independent of all aspects of the Transformation-marker except insofar as this indicates how base structures are interrelated. combine semantic interpretations of already interpreted Phrase-markers in a fixed way). 1 957. is what underlies the work in transformational generative grammar that has appeared in the last ten years.) Katz and Postal point out that the principle just stated greatly simplifies the theory of the semantic component. they conclude also that a sentence transform embedded in a matrix sentence I must replace a dummy symbol of I.1 3.. in the course of this work. and by Klima (personal communica tion) for the question transformation. This was pointed out by Lees ( 1 96oa) for the negation transformation. Generalizing these remarks to embedding transformations. by the condition men tioned in note 1 ) . First. 1 962) must be re formulated as obligatory transformations.2 DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS represented as in (5). the branching points correspond to gen eralized transformations that embed a constituent sentence (the lower branch) in a designated position in a matrix sentence (the upper branch). A theoretical apparatus of this sort. (In the foregoing discussion. whose applicability to a string is determined by presence or absence of a certain marker in the string. 1 963. as noted in § 2 . several important points have gradually emerged which suggest that a somewhat more restricted and conceptually simpler theory of transformations may be adequate. that trans formations cannot introduce meaning-bearing elements (nor can they delete lexical items unrecoverably. Katz and Postal ( 1 964) have extended these observations and formulated them in terms of a general principle. at about the same time.e. adopting this suggestion. In fact. it has been shown that many of the optional singulary transformations of Chomsky (1 955. However. namely that the on ly contribution of transformations to semantic in terpretation is that they interrelate Phrase-markers (i. 3 . in its essentials. it is also true for the passive transformation. 4 These rules may state the ordering of transformations relative to one another. there are no really convincing cases of singulary trans formations that must apply to a matrix sentence before a sen tence transform is embedded in it. However. then. In particular. in this view. There is no reason for imposing an extrinsic order on the generalized transformations. In brief. They may apply to a constituent structure before it is embedded. and may designate certain transformations as obligatory. only some of the possibilities permitted by this general theory have been realized convincingly with actual linguistic material. Thus the grammar. and the constituent struc ture embedded in it. The principle. or obliga tory relative to certain contexts.1i On the other hand. must con tain rules generating the possible Transformation-markers by stating conditions that these objects must meet for well-formed ness (what Lees. 1 960a. the descrip tion was in fact incorrect on internal syntactic grounds. Further more. Second.8 . after this constituent structure is embedded. where this general principle has not been met in syntactic description.DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS 1 33 showing that in a large variety of cases. Thus the diagram (5) is typical of the kind of structure that has actually been discovered in Transformation-markers. there are many examples of ordering of singulary transformations. by requiring that they appear in specified positions in Transformation-markers. notice that the theory of Transformation-markers permits a great deal of latitude so far as ordering of transforma tions is concerned. seems very plausible. according to the theory. though this too is a possibility. there are no known cases of ordering among gen eralized embedding transformations although such ordering is permitted by the theory of Transformation-markers. or to a matrix structure. and many examples of singulary transformations that must apply to a constituent sentence before it is embedded or that must apply to a matrix sentence after embedding of a constituent structure in it. presently available descriptive studies suggest the following restrictions on ordering of transformations. calls "traffic rules"). The sin gulary transformations are linearly ordered (perhaps only partially ordered). We have thus revised the theory of the base by allowing #S# to appear on the right in certain branching rules. with #S# in place of the occurrence of S' in (2). preserv ing their linear order. the sequence of transformational rules applies to the most deeply embedded base Phrase-marker. the same base rules can reapply to form the derivation represented by (3 ) . they reapply to the new occurrence of #S# in the terminal line of the derivation represented by ( I ) . That is. A generalized Phrase-marker formed in this way contains all of the base Phrase-markers that constitute the basis of a sentenc�. but it contains more information than a basis in the old sense since it also indicates explicitly how these base Phrase-markers are embedded in one another.7 In the rewriting rules of the base (in fact. From the latter occurrence of ISH.8 In addition to the rules of the base. with #S# in place of S'. which initiates derivations. We now allow the rules of the base to apply cyclically. for example. as well as the information provided by the generalized embedding transformations. we fill this position with the string #S#. In this way. First. Suppose that we eliminate the notions "generalized transformation" and "Transformation marker" altogether. 1 34 DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS . the grammar contains a linear sequence of singulary transformations. and by allowing the rules to reapply (preserving their order) to these newly intro duced occurrences of IS H. the generalized Phrase-marker contains all of the information con tained in the basis. the base rules will generate the generalized Phrase-marker formed from (1). Thus. (2 ) . in i ts ca tegori al component) the string #S# is introduced in the positions where in the illustrative example we introduced the symbol S'. (3) by replacing S' in ( 1 ) by (2) and replacing S' in (2) by (3). That is. so modified. From this occurrence of #S# the rules of the base can generate the derivation represented by (2). after having generated ( I ) .These observations suggest a possible simplification of the theory of transformational grammar. wherever a base Phrase-marker contains a position in which a sentence transform is to be introduced. (For example. in the following manner. where pre viously the dummy symbol S' had appeared. These apply to generalized Phrase-markers cyclically. the effect of this convention is precisely what is described in the Transformation-marker (5). and its surface structure enters the phonological component and under goes phonetic interpretation. we may take a generalized Phrase-marker. The embedding itself is now provided by the branching rules of the base rather than by generalized trans formations. . singulary transformations are applied to constituent sentences before they are embedded. in effect. the sequence of rules reapplies to a configuration dominated by S in which these base Phrase-markers are embedded (to (2). We have. The base rules form generalized Phrase-markers that contain just the information contained in the basis and the generalized transformations of the earlier version. But observe that in accordance with the Katz-Postal principle discussed earlier (p. in the same example). to be the deep structure generated by the syntactic component. and to matrix sentences after embed ding has taken place. The final effect of a grammar. Consequently. in the manner just de scribed. That is. Thus the syntactic component consists of a base that generates deep structures and a transformational part that maps them into surface structures. The notion of Transforma tion-marker disappears. The grammar now consists of a base and a linear sequence of singulary transformations. until finally the sequence of rules applies to the configuration dominated by the initial symbol S of the entire generalized Phrase-marker (to ( I ). in the sense j ust defined. Notice that in the case of ( 1 )-(3). ) Having applied to all such base Phrase-markers.DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS 1 35 it applies to (3). as does the notion of generalized transformation. in the generalized Phrase-marker formed by embedding (3) in (2) and the result in ( I ) . The ordering possibilities that are permitted by the theory of Transformation-markers but apparently never put to use are now excluded in principle. and so on. then. it is precisely this information that should be relevant to semantic interpretation. 1 3 2). The deep structure of a sentence is submitted to the semantic component for semantic interpretation. as described earlier . converted the specific properties of the Transformation-marker (5) into general properties of any possible transformational derivation. in our example). These apply. . simple phrase structure grammars in most significant respects. since Transformation-markers and generalized transformations. which constitutes its sole "creative" part.4. § 2 .9 This is the basic idea that has motivated the theory of transformational grammar since its inception (cf. It is interesting to note. it permits a further simplification of the theory of semantic interpretation presented in Katz and Postal (1 964). Chap ter 2 ) . its categorial com ponent) define grammatical functions and grammatical relations and determine an abstract underlying order (cf.1 36 DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS is to relate a semantic interpretation to a phonetic represen tation . in terms of perceptual mechanis ms. to state how a sentence is interpreted. 2 ) . are suggested in Miller and Chomsky ( 1 963. the lexicon characterizes the individual properties o f particular lexical items that are inserted in specified positions in base Phrase-markers. This formula tion seems to be a natural extension and summary of the develop ments of the past few years that have just been summarized. Chapter 2 ) .that is. In fact. Thus when we define "deep structures" as "structures generated by the base component. apparently without exception. The branching rules of the base (that is. and an improved version is given in Katz and Postal (1 964).lo Some possible reasons for such an organization of grammar. in terms of the modification of syntactic theory proposed there and briefly discussed earlier. § 4. in effect. This relation is mediated by the syntactic component of the grammar. that the grammars of the "artificial languages" of logic or theory of programming are. note 33. need no longer be considered at all. Its first relatively clear formulation is in Katz and Fodor (1 963). The formulation just sug gested sharpens this idea still further. as well as "projection rules" to deal with them. Notice that in this view one major function of the transforma tional rules is to convert an abstract deep structure that expresses the content of a sentence into a fairly concrete surface structure that indicates its form." we are. assuming that the semantic interpretation of a sentence depends only on its lexical items and the grammatical functions and rela tions represented in the underlying structures in which they appear. in this connection. we have returned to a con ception of linguistic structure that marked the origins of modern syntactic theory. where K is a Phrase-marker differing from (2) only in that man in (2 ) is replaced by boy in K . namely that presented in the Grammaire ge nerale et raisonnee. K.12 One additional point must be emphasized in connection with the notion "deep structure. we might just as well have constructed the generalized Phrase-marker M' formed from (I).. they cannot take account of the context in which this occurrence of S appears. of the rules that introduce the initial symbol S in designated positions in strings of category symbols. For example. In the earlier version of the theory. Now the recursive property is a feature of the base component.ll The transformational component is solely interpretive. to the generalized transformations and the rules for forming Transformation-markers. fired the man (I the boy persuaded John to be examined by a specialist I) by passive The string (6) (with its Phrase-marker) is of the form that per mits the rel ative clause transformation to apply. we shall have not the string (6) but rather (7) : (6) !J.DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS 137 Looking more closely at the recursive property o f the gram mar. apparently. at the stage of derivation at which the relative clause transformation ( TB of (5» is applied to K with (3) embedded within it.. fired the man (I the man persuaded John t o be examined by a specialist I) by passive (7) !J." since the condition of identity of the two Nouns . in particular. There are." When the base rules generate a Phrase-marker from an occurrence of S that is embedded in an already generated Phrase-marker. It is worth mentioning that with this formulation of the theory of transformational grammar. and (3). in particular. the recursive property was assigned to the transformational com ponent. But now. replacing "the man" by "who. instead of the generalized Phrase-marker M consisting of ( I )-(3) (with (3) embedded in (2) and the result embedded in ( I » . we have now suggested the following modification of trans formational theory. no other recursive rules in the base. for obviously the generalized Phrase-marker formed from (1). A deep structure is a generalized Phrase-marker underlying some well-formed surface structure. Putting aside questions of formalization. then. The surface structure Ms of S is well formed just in case S contains no symbols indicating the blocking of obligatory transformations. and there is. What. we can see that not all generalized Phrase-markers generated by the base will under lie actual sentences and thus qualify as deep structures. is not the deep structure underlying any surface structure. The transformational rules provide exactly such a test. the transformation will block. just in case the transformational rules generate Ms from MD. note 1 ) . Such occurrences will indicate that certain transformations that should have applied were blocked. The . and (3). In fact. no simpler test. Thus the basic notion defined by a transformational grammar is: deep structure MD underlies well-formed surface structure Ms. The same (or similar) formal devices can be used in a variety of other cases.1 38 DEEP STRUcrURES AND GRAMMATICAL TRANSFORMATIONS is met and we thus have a recoverable deletion (cf. the identity condition of the transformation is not satisfied. in general. as it would if application of the relative clause transformation were permitted in this case. K. Thus if its application is blocked. in this case. (3) does not provide the semantic interpretation of (4) . We can then establish the convention that a well-formed surface structure cannot contain internal occurrences of iF. We can make this observation precise. with the surface structure Ms. A generalized Phrase-marker MD is the deep structure underlying the sentence S. But in the case of (7). The notion "deep structure" itself is derivative from this. although generated by the base rules.that is. the generalized Phrase-marker formed from ( 1 ) . by defining the relative clause transformation in such a way that it deletes the boundary symbol iF when it applies. this symbol will remain in the string. is the test that determines whether a generalized Phrase marker is the deep structure of some sentence? The answer is very simple. ls This is precisely what we want. K. Thus the phrase "the boy" cannot be deleted from (7) because of the general condition that only recoverable deletions are permitted . a well-formed sentence. Thus the categorial rules that generate the infinite set of generalized Phrase-markers can apparently be context-free. whether of base Phrase-markers or lexical entries. Thus it seems absurd to suppose that the speaker first forms a generalized Phrase-marker by base rules and then tests it for well-formedness by applying transformational rules to see if it gives. with all distribu tional restrictions. In the present version this filtering function is simply brought out more clearly. furthermore. But this absurdity is simply a corollary to the deeper absurdity of regarding the system of generative rules as a point-by-point model for the actual construction of a sentence by a speaker.DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS 1 39 transformational rules act as a "filter" that permits only certain generalized Phrase-markers to qualify as deep structures. though this fact was never discussed in exposition . Such a description of the form of the syntactic component may seem strange if one considers the generative rules as a model for the actual construction of a sentence by a speaker. and (b) that these contextual fea tures be regarded as defining certain substitution transforma tions. being determined by the (singulary) transformations. Notice that this filtering function of the transformational component is not an entirely new feature specific to the version of transformational grammar that we are developing now. it was also true of the earlier version. Consider the simpler case of a phrase . Thus a sequence of base Phrase markers might have been selected that could not serve as the basis of any sentence. In § 4. finally. In fact. Thus strict subcategorial and selectional restrictions of lexical items are defined by transformational rules associated with these items.3 of Chapter 2 we suggested: (a) that the distributional restrictions of lexical items be determined by contextual fea tures listed in lexical entries. any system of rules for generating Transformation-markers would certainly permit cer tain structures that do not qualify as Transformation-markers because of inconsistencies and blocks arising in the course of carrying out the instructions that they represent . We have now observed that the transforma tional rules must also carry the burden of determining the distributional restrictions on base Phrase-markers. then M was a deep structure." "M' is a well-formed sur- . If M' is well formed. derivatively. just as they can all be enumerated in many other ways.14 But a generative grammar as it stands is no more a model of the speaker than it is a model of the hearer. The base rules and the transformational rules set certain conditions that must be met for a structure to qualify as the deep structure expressing the semantic content of some well formed sentence. one can develop innumerable procedures for actually constructing deep structures. select ing the words or symbols that he is going to use (deciding what he is going to talk about). or elementary arithmetic. One can study models of performance that incorporate genera tive gra mmars. These will vary in exhaustiveness and efficiency. As noted earlier." first selects the major categories. or some small part of English that might be described in these terms). the grammar of a programming language. One such constructive procedure is to run through the base rules (observing order) so as to form a gen eralized Phrase-marker M. given the grammar. To think of a generative grammar in these terms is to take it to be a model of performance rather than a model of competence. finally. the grammar defines the relation "the deep structure M underlies the well-formed sur face structure M' of the sentence S" and. and in the extent to which they can be adapted to the problems of producing or under standing speech . All deep structures can be enumerated in this way. in formulating an "utterance. and then through the transforma tional rules (observing order) so as to form a surface structure M' from M. Rather. it defines the notions " M is a deep structure. and so forth.1 40 DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS structure grammar with no transformations (for example. then the categories into which these are analyzed. at the end of the process. thus totally misconceiving its nature. otherwise. it was not. and some results have been achieved in such studies. It would clearly be absurd to suppose that the "speaker" of such a language. Given a grammar containing a base component and a transformational component. as has been repeatedly emphasized. it can be regarded only as a characterization of the intrinsic tacit knowledge or competence that underlies actual performance. part of particular grammars. Similarly. and a phonological component. the category symbols appearing in base rules are selected from a . The grammar does not. It merely defines these tasks in a precise way. A grammar contains a syntactic com ponent." "s is a deviant sentence formed by violating rule R or condition C"). provide any sensible procedure for finding the deep structure of a given sentence. a semantic component. Once this point is clear. strictly speaking. the fact that transformations act as a kind of filter will occasion no sur prise or uneasiness. and to specify an abstract underlying order of elements that makes possible the functioning of the transformational rules. they play no part in the recursive generation of sentence structures. The base generates deep structures. the choice of base rules is constrained by a universal condition on the grammatical functions that are defined. which is then given a phonetic interpretation by the rules of the phonological component. we have now suggested that the form of gram mar may be as follows. it is not to be confused with a grammar. the rules of the base may be universal. The syntactic com ponent consists of a base and a transformational component." and many others (such as "s is structurally ambiguous. or for producing a given sentence. or it may be that. To a large extent. to define a certain system of grammatical relations that determine semantic interpretation. although free in part. in essence. in turn." "s and S' are para phrases. The base. it is mapped by the transformational rules into a surface structure. and thus not. A per formance model must certainly incorporate a grammar. in itself. The categorial subcomponent of the base consists of a sequence of context-free rewriting rules. just as it provides no sensible procedure for finding a paraphrase to a given sentence. To summarize. A deep structure enters the semantic component and receives a semantic interpre tation. The latter two are purely interpretive. The function of these rules is." "s is a well-formed sentence.DEEP STRUCTURES AND GRAMMATI CAL TRANSFORMATIONS 141 face structure. this association being me diated by the recursive rules of the syntactic component. Thus the grammar assigns semantic interpretations to signals. consists of a categorial subcomponent and a lexicon. which becomes a generalized Phrase-marker when lexical entries are inserted in accordance with the trans formational rules specified by the contextual features that belong to these lexical entries. we derive a pre tenninal string. reapplying them to each new occurrence of S introduced in the course of the derivation. insert base Phrase-markers into other base Phrase-markers. The transformational subcomponent consists of a sequence of singulary transformations. are presumably drawn from a universal "alphabet. the rewriting rules can. namely that they may intro duce the initial symbol S into a line of a derivation . Thus the lexical entries constitute the full set of irregularities of the language. We call a fea ture "semantic" if it is not mentioned in any syntactic rule. in fact. beginning with S. the choice of symbol may be largely or perhaps completely determined by the formal role the symbol plays in the system of base rules. The lexicon consists of an unordered set of lexical entries and certain redundancy. The base of the syntactic component thus generates an infinite set of generalized Phrase-markers. In this way. Each transformation is fully defined . thus begging the question of whether semantics is involved in syn tax. Some of these are phonological fea tures. this process being iterable without limit. too. in effect.142 DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS fixed universal alphabet. The infinite genera tive capacity of the grammar arises from a particular formal property of these categorial rules. Each lexical entry is a set of features (but see note 15 of Chapter 2). drawn from a particular universal set of phonological features (the distinctive-feature system). We may construct a derivation of a generalized Phrase-marker by applying the categorial rules in the specified order. rules. In this way. The set of phonological features in a lexical entry can be extracted and represented as a phonological matrix that bears the relation "is a" to each of the specified syntactic features belonging to the lexical entry. Some of the features are semantic features ." but little is known about this today. and nothing has been said about it here. These.15 The redundancy rules of the lexicon add and specify fea tures wherever this can be predicted by general rule. applying the sequence of rules to a given configuration only if we have already applied it to all base Phrase-markers embedded in this configuration. these rules apply in a cycle. Thus transformations may refer to specified syntactic features as if they were categories. which. forthcoming. 1 962 b . The interpretive components of a grammar have not been our concern here. whereas the surface structure of S determines its phonetic form. they seem to function in parallel ways. § 2).that is. Halle and Chomsky. (See Chomsky. This deep structure expresses the semantic content of S. first to the minimal elements (formatives).) In this way a phonetic representation of the entire sentence is formed on the basis of the intrinsic abstract phono- . Insofar as details of their structure have been worked out. If none of the transformations blocks. which is a Boolean condition on Analyza bility. C homsky.DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS 1 43 by a structure index. is defined by the rewriting rules of the base and by the lexicon. which is the terminal string of the derived surface structure. the generalized Phrase-marker to which the transforma tions were originally applied constitutes a deep structure. namely the deep structure of the sentence S. and so on. 1 960. we derive in this way a well· formed surface structure. then to the constituents of which these are parts. then to the constituents of which they are parts (a constituent of a Phrase·marker being a sub string of its terminal string dominated by a single category symbol). we construct a transforma tional derivation by applying the sequence of transformational rules sequentially. and a sequence of elementary transformations. Chomsky and Miller. The phono logical component consists of a sequence of rules that apply to a surface structure "from the bottom up" in the tree-diagram representing it. Halle. Given a generalized Phrase·marker. and Lukoff. 1 956. In this and only this case. The notion "Analyzable" is determined in terms of the "is a" relation. until the maximal domain of phonological processes is reached. but we shall not go into this modification of the theory of transformational grammar here (see Chapter 4. That is. "from the bottom up" . 1 963 . In fact. transforma tions must also be designed so that they can specify and add syntactic features. in turn. it appears that permutations can be eliminated from the set of elementary transformations in favor of substitutions. 1 6 Second. be stated as part of a specific grammar. you in imperatives). We shall discuss it briefly here and then return to the question in Chapter 4. the projection rules of the semantic component operate on the deep structure generated by the base. the intrinsic semantic properties of the formatives) and the categories and grammatical relations rep resented in the deep structure. on the basis of the readings assigned to its parts (ultimately. deletions. too.2 .) To the extent that grammatical categories and rela tions can be described in language-independent terms. 1 963. as mentioned in note 1 3 . quantifiers can be eliminated from the formulation of transformations in favor of a general convention on deletion. or the . assigning a semantic interpretation (a "reading") to each constituent.1 44 DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS logical properties of its formatives and the categories represented in the surface structure. § 2 . and other papers by Katz listed in the bibliog raphy. We are proposing the following convention to guarantee recoverability of deletion : a deletion operation can eliminate only a dummy element. can apparently be simplified in various ways. Katz and Postal. In a somewhat similar way. If so. Throughout this discussion. it seems that the structural analyses that determine the domain of transformations can be limited to Boolean conditions on Analyzability. First. or a formative explicitly mentioned in the structure index (for example. The latter point deserves some further clarification. we have simply been presuppos ing the theory of grammatical transformations as presented in the references cited. 1 964. but it is perhaps worth mentioning that this theory. there fore. That is. and adjunctions. which need not. the derived Phrase-markers that would be provided by permutations may not be necessary in addition to those provided by the other elementary transformations. That is. this places a severe additional restriction on the theory of transformations. (See Katz and Fodor. one may hope to find universal projection rules. Elimination of permutations from the base set would greatly simplify the theory of derived constituent structure. the man . we may say briefly that the erasure operation uses the term X to delete Y in such a case . As an additional illustration. sameness of reference requires reflexivization of the second Noun Phrase (this is also true of pronominalization). Chomsky. We say. which are straightforward within the general theory of transformations. and then deletes this new occurrence of X which re placed Y. ��--����----�----� ' 4 the relative transformation can be formulated as an erasure operation that substitutes the first term X of the proper analysis for the third term Y. each occurrence of a referential item is assigned a marker. consider the reflexivization operation (see Lees and Klima. I n£. let us define an erasure transformation as one that substitutes a term X of its proper analysis for a term Y of its proper analysis (leaving X intact). 2 ) . .)." the two phonetically identical Noun Phrases are necessarily interpreted as differing in reference. that an erasure operation can use the term X to delete Y just in case X and Y are identical.[#wh. Various attempts have been made to build an account of this into the syntactic component. say.18 The reflexivization rule can be formulated . pp. In the example of relativization discussed earlier (pp. § 2 . 1 28 f. The availability of lexical features suggests a new approach that might be explored. To clarify the latter point further.DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS 14 5 designated representative of a category (for example. We shall investigate the exact nature of the required relation between X and Y some what more fully in Chapter 4. then.had been fired #] returned to work . but none has been very convincing. 1 963. or an element that is otherwise represented in the sentence in a fixed position. an integer.cf. for a detailed discussion) . Avoiding details of formalization. erasing the latter1 7 in the process. � 11 3 . Suppose that certain lexical i tems are designated as "referential" and that by a general convention. if we have the string (8) the man . 1 964 . the wh question transformations that delete Noun Phrases are in fact limited to indefinite Pronouns -. It has frequently been observed that in a sentence such as "John hurt John" or "the boy hurt the boy. as a feature. a consequence of a more general condition on transformations. and it in troduces the new phonetic element self. Thus "I kept it near me" has a deep structure of the form "I kept . the reflexivization rule (similarly. incidentally. the repeated Noun is in a Sentence-Complement to the Verb . in case they have been assigned the same integer in the deep structure. and of course there are problems in specifying the notion "referential" properly. but not the second. Notice. the erasure leaves a residue. This particular remark about English is. This gives the right answer in many cases. finally giving.it I S I. self) can be introduced into a configuration dominated by S once the cycle of transformational rules has already com pleted its application to this configuration (though items can still be extracted from this constituent of a larger "matrix structure.in particular. the feature [±Human]. There are a few examples that seem to conflict with this analysis (such as "I - ." where S dominates "it is near me. Thus we have "I kept it near me" alongside of "I aimed it at myself." and so on. apparently. that the reflexivization rule does not al ways apply (though pronominalization does) even when the two Nouns are strictly identical and hence coreferential ." the first Noun Phrase is used to delete the second." in the next cycle of transformational rules).it . but there are interesting problems that arise when the referential items are plural. The difference is that in the first. in particular. Thus when applied to "I hurt I. As in the case of relativization (cf. The semantic component will then interpret two referential items as having the same reference just in case they are strictly identical .1 46 DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS as an erasure operation that uses one Noun Phrase to delete another. "I hurt myself. the pronominalization rule) will apply only when the integers assigned to the two items are the same. note 1 7)." But by the recoverability condition on deletion. The reflexivization rule does not apply to a repeated N dominated by an occurrence of S that does not dominate the "antecedent" occurrence of N." But "I aimed it at myself" has a deep structure of the form "I aimed .at me" (there is no underlying sentence "it is at me") . namely that no morphological material (in this case. for reasons that I do not understand. and it may be possible to undertake abstract study of them of a sort that has not been feasible in the past. deletions. it provides an interesting confirmation of the theory of trans formational grammar. erasures) and that the limitations on their application can be given by general conventions of the sort just mentioned. . and. If this is correct. then the formal properties of the theory of transforma tions become fairly clear and reasonably simple." "I drew it toward me"). but it covers a l arge number of convincing cases. Returning to the main theme. in the distinction it makes between super ficially analogous cases that differ only in that one but not the other is based on an independently existing embedded sentence. we can apparently define a grammatical transformation in terms of a "structure index" that is a Boolean condition on Analyzability and a sequence of elementary transformations drawn from a base set including substitutions. It seems also that these form larger repeated units (for example. substitution-deletions.DEEP STRUCTURES AND GRAMMATICAL TRANSFORMATIONS 1 47 pushed it away from me. and adjunctions. Consequently. Transitives. for any language. for the present.I. the problem suggested by the title of this section can. construct a deviant sentence by breaking the rule. only very rudimentary grammatical descriptions are available. The distinction between strict subcategorization features and selectional features. which is formally well defined. and that they involve open questions of a fundamental nature. in each case. Degrees of grammaticalness IT is quite apparent that current theories of syntax and semantics are highly fragmentary and tentative. be at best a source for speculation. violation of the rules will give such strings as : (1) (i) John found sad (ii) John elapsed that Bill will come 148 . pre-Adjectival. Nevertheless.4 Some Residual Problems § I. Thus in § 3 of Chapter 2. etc. Verbs are strictly subcategorized into Intransitives. some of the topics of the preced ing chapters relate to the question of the proper balance between syntax and semantics in a way that deserves at least some further comment. pre-Sentence. so that no satisfactory answers can be given for many factual questions. 1 We can. Each such contextual feature is associated with a certain rule that limits lexical entries containing this feature to certain contexts . In these cases. THE BO UNDARIES OF SYNTAX AND SEMANTICS § I. appears to correlate rather closely with an important distinction in language use. Furthermore. 1 of Chapter 2). That is. Clearly. Sentences that break selectional rules can often be interpreted metaphorically (particularly.this being a task that varies in difficulty or challenge from case to case .I !I (iii) John compelled (iv) John became Bill to leave (v) John persuaded great authority to Bill On the other hand. if an appropriate context of greater or less com plexity is supplied. Bloomfield. These examples are. I think.whereas there is no question of imposing an interpretation in the case of such strictly well-formed sentences as (3) (i) (ii) (iii) (iv) (v) revolutionary new ideas appear infrequently John plays golf sincerity may frighten the boy John loves company they perform their duty with diligence Nevertheless. Clearly. failure to observe a selectional rule will give such typical examples as (2) (i) (ii) (iii) (iv) (v) colorless green ideas sleep furiously golf plays John the boy may frighten sincerity misery loves company they perform their leisure with diligence strict subcategorization rules and strings such as (2) that break (cf. the manner of deviation illustrated in (2) is rather different from that in ( 1 ) .3 . THE BOUNDARIES OF SYNTAX AND SEMANTICS l . these sentences are apparently interpreted by a direct analogy to well-formed sentences that observe the selectional rules in question. the sentences of ( 1 ) . typical of a fairly wide class of . as personification . § 2. for example.cf. one would proceed in quite a different way if forced to assign an inter pretation to sentences that break strict subcategorization rules.§ 1. It is necessary to impose an inter pretation on them somehow . 1 963) or allusively in one way or another. strings such as ( 1 ) that break selectional rules are deviant . It distinguishes perfectly well-formed sentences such as (3) from the sentences of (1) and (2 ) . be imposed on them. This feature is involved in rules that are as inviolable as those that give the book which you read was a best seller and what you found was my book. Similarly. a m use charm. it is important to bear in mind that not all rules involving low-level syntactic features tolerate deviation as readily as do selectional rules involving these features. consider the selectional feature [[+Abstract] ' " . . . . and a grammar of the type just described seems to make them in some measure. Thus no matter how selectional rules are treated. A descriptively adequate grammar should make all of these distinctions on some formal grounds. while excluding (4). . .although of course an interpretation can easily.8 Thus both of the sentences (4) (i) the book who you read was a best seller (ii) who you met is John result from failure to observe rules involving the feature [Human]. and no doubt uniformly. at least."2 It seems that sentences deviating from selectional rules that involve "higher-level" lexical features such as [Count] are much less acceptable and are more difficult to interpret than those that involve such "lower-level" features as [Human] . It further separates the sentences of ( 1). there is no doubt that such features as [Human] play a role in purely syn tactic rules (since surely the examples of (4) are ruled out on purely syntactic grounds). they differ completely from sentences that result from a failure to observe selectional rules involving the feature [Human] . Thus items . which are not directly generated by the system of grammatical rules. from sentences such as ( 2 ) . Both in degree of acceptability and manner of interpreta tion. but are totally unacceptable . Thus it takes several steps toward the development of a significant theory of "degree of gram maticalness.SOME RESIDUAL PROBLEMS cases. which are generated when selectional rules are relaxed. . generated by relaxing strict subcategorization rules. [+Animate]] assigned to such Verbs as frighten. . At the same time. Thus it seems that this selectionally introduced contextual feature is also involved in rules that cannot be violated without serious departure from grammaticalness . there are also perfectly well-formed sentences that may pose great difficulties for interpretation.. On the other hand." but rather to the fact that the rules that they violate are selectional rules. ) person suddenly appeared. . (5) (i) a very walking person appeared (ii) a very hitting person appeared These sentences. charming. . There are sentences such as (4) and (5) that are uniquely. . . and immediately interpretable. so that we have such sentences as a very frigh tening (amusing. uniqueness. like those of (4). than the examples of violation of selectional rules given earlier. are immediately and perhaps uniquely interpretable. although they are paradigm examples of departure from well-formedness. . First. or uniformity of interpretation). . it is clear that features such as [Human] and [[+Abstract] . no matter how narrowly syntax is conceived. but are obviously much more seriously ungrammatical. . as long as it is agreed that (4) and (5) are syntactically deviant. in any simple way. but not. . THE BOUNDARIES OF SYNTAX AND SEMANTICS that are positively specified with respect to this feature can ap pear in the position of pure Adjectives. Second.§ 1. . just as it is clear that various diverse factors determine how and whether a sentence can be interpreted. uni formly. [+Animate]] play a role in the func tioning of the syntactic component. in the intuitive sense that we are now attempting to explicate. it is clear from such examples as (4) and (5) that the notion "grammatical ness" cannot be related to "interpretability" (ease. Examples such as (4) and (5) therefore support two important observations .. at least. More generally. no doubt. for example. and may be subject to a variety of perhaps conflicting interpretations . The special character of the examples of ( 2 ) is not attributable to the fact that these sentences violate rules involving "low-level features. it is clear that the intuitive notion of grammatical well formedness is by no means a simple one and that an adequate explication of it will involve theoretical constructs of a highly abstract nature. [+ . In the example given. referring again to the sample Phrase-marker (59) of Chapter 2 and the formative frighten. Thus we should have the following order of deviance: (6) (i) sincerity may virtue the boy (ii) sincerity may elapse the boy (iii) sincerity may admire the boy This seems to give a natural explication for at least one sense of the term "deviance.. compare the sug gestions of the references of note 2. . . In terms of this order. and others." In this connection. which consider size of category within which substitution takes place in determining .NP]. we might conclude that the only function of the selectional rules is to impose a hierarchy of deviation from grammaticalness on a certain set of sentences. we can define the degree 0/ deviation of a string that results from substituting a lexical item in the position of frighten in this Phrase-marker. . The deviation is greater the higher in the dominance hierarchy is the feature corresponding to the rule that is relaxed. . . following this suggestion. Observe that the rules of the grammar impose a partial ordering in terms of dominance among the features that constitute a com plex symbol in a Phrase-marker. The rules of the grammar impose the dominance order [+ V]. . [+ [+Abstract] . then. [+Animate]]. + .. . we have a complex symbol consisting of the features [+V. deviance would be greatest if the item substituted for frighten is a non-Verb. less great if it is a Verb but a non-Transi tive Verb. .NP. For example. as indicated in (59) . . and still less great if it is a Transitive Verb that does not take an Abstract Subject. . In fact. namely those sentences that can be generated by selectional constraints while otherwise keeping the grammar unchanged . [+Animate]].SOME RESIDUAL PROBLEMS The attempts described in the references of note 2 to give a precise definition to at least one dimension of degree of gram maticalness are much more plausible if limited to the question of deviation from selectional rules than if extended to the full range of examples of deviation from well-formedness. + [+Abstract] . it was pointed out that features introduced by strict subcategorization rules dominate features introduced by selectional rules. that selectional rules be dropped from the syntax and that their function be taken over by the semantic component. (5» . and (iii) conflict with a selectional feature (such as (6iii) and (2» . At the end of § 4. in the lexical entries for these items.the degree of grammaticalness (the extent of syn tactic deviance) of a string. Of course. Further remarks on selectional rules Selectional rules play a rather marginal role in the grammar. we might go on to superimpose on the scale of deviance a split into perhaps three general types. deviation from selectional rules involving high-level features is apparently more serious than deviation from selectional rules involving lower level features. 1 of Chapter 2. That is. although the features that they deal with may be involved in many purely syn tactic processes (cf. etc. subdivisions within at least the third type. furthermore. and in the same section it was further noted that all lexical features are dominated by the symbols for lexical categories.2 . These various observations combine to make the definition of "degree of deviance" just proposed a rather natural one. Such a change would do little violence to the structure of gram mar as described earlier. the features that are utilized and introduced by selectional rules would still appear in lexical entries for strings. Of course.5 This is not surprising. (ii) conflict with a strict sub categorization feature (such as (6ii) and ( I » . § I. boy would be specified as [+Human] and frighten as permitting an Abstract Subject and Animate Object. if we continue to call a feature of the lexical entry a "syntactic § 1. THE BOUNDARIES OF SYNTAX AND SEMANTICS 1 53 . since there are rules of many kinds that can be violated. Furthermore. therefore. . There are. (5» . One might propose. Furthermore. there are also many other types (such as (4) . If the distinction between strict subcategorization rules and selectional rules noted earlier is generally valid. (4) . namely the types that result from: (i) violation of lexical category (such as (6i». then these features of the lexical entry will be syntactic rather than semantic features (cf. Suppose that we were to introduce selectional features by . following Katz. In § 4.1 54 SOME RESIDUAL PROBLEMS feature" when it is involved in a strictly syn tactic rule." can be carried over with little change. of course. as syn tactically well formed. grammatically related constituents of base strings. I shall now mention briefly some considerations that seem relevant to them. Fodor. ( 1 ). more generally. in particular the definition of "degree of deviance. and Postal. two open questions in connection with selectional rules. though not. the discussion of (4). The second was to regard the con textual features of the lexicon as defining certain substitution trans formations that insert lexical items. (5» . The first was to introduce them by rewriting rules and to have lexical items introduced into derivations by matching of nondistinct complex symbols (as in Chapter 2. The projection rules must now be adapted to detect and interpret conflicts in feature composition between grammatically related lexical items and. As noted there. this is not merely a notational question. Let us continue to suppose that the semantic component is an interpretive device based on projection rules of the type discussed earlier. impose a hierarchy of degree of grammaticalness at these lower levels of deviation. the grammar will directly generate even such sentences as (2). We have. This task would now have to be taken over by the semantic component. The earlier discussion of deviance. then. in ac cordance with this proposal. The syntactic component of the grammar would not. § 3). we discussed two alternative proposals for dealing with contextual features. With slight reformulation. Nevertheless.3 of Chapter 2. The same is true of the comments regarding Noun Verb and Noun-Adjective selectional dominance. the same arguments will hold under this revision of the structure of grammar. in particular: (i) Do they belong in the syn tactic or the semantic component? (ii) Should they be rewriting rules introducing complex symbols or substitution transforma tions? Without attempting any exhaustive investigation of these questions. in other words. . . (see pp.§ 1. This is as serious a deficiency as the one noted in the case of a grammar that specifically distinguishes Animate Subjec t from Animate Object. . rewr i t ing (7) (8) the boy is sad the boy grew sad In the case of (7). As these rules are formulated the Adjec tive would actually be assigned different features in these sentences: . it is atypical only in the simplicity of these elements. with its reference to the irrelevant items A ux and Det.8 These features have nothing in common. ] W2 (or is null) and V � V1/fl. whereas in the case of (8) it would be assigned the featu re [[+Human] Aux [+V] ] or some thing of this sort. .. . N o t i ce that the selectional rules differ from the strict subcategorization rules in that they typically involve irrelevant symbols standing between the items that they relate. ]. . ..8 W � W1 [a. . ... (1 0) X WA VY . Y = /fl. . .. 1 1 4-1 1 5). ]. ] V2 (or is null). We may remedy it and at the same time.. the Adje c tive would be assigned the feature [[+H uman] Aux be . /fl] . The rule (57xiv) of Chapter 2 is characteristic of selectional rules in this res pect. in accordance with § 3 of Chap te r 2 . . eliminate the reference to irrelevant intervening contexts in selectional rules by e s tablishing the following convention for these rules .7 We take (9) to be app licable to any string where X = [a. . The result of appl y ing (9) to ( 1 0) is the string .] by rule (57xv) of C hapter 2. where [a] and /fl] are specified features or are null (but either one or the other is nonnull). . in our terms. etc. . Suppose that we have the rule schema . (9) A ? CS/ raJ . .. though they actually identify the same set of lexical items . That this may be more than a purely notational matter is illustrated by (57xv) of C hapter 2. THE BOUNDARIES OF SYNTAX AND SEMANTICS 1 55 rules. . which assigns features of the Subject to a modifying A djective of the Predicate. more importantly. ( 1 3) Adjective � CS/[+N] . however. . (The reader will observe that except for the condition on W. ].( .10 Consider such a typ i cal case of violation of selectional rules as . . in particular. Thus. where [ep] is a lexical feature of the nearest complex symbol containing [a] to the left of A. This condition.I/J].[+Animate]]. and [I/J] is a lexical feature of the nearest complex symbol containing [. ..". . ) What this means is that the rule (9) assigns to A all contextual features [+ep . [+N]) . . . the notion of "applicability" and the conventions for complex sym bols are as before. This fact. in particular. V of ( 1 0). V. might be taken as suggesting that the system involving rewriting rules is preferable. . In this case. respectively: • • .9 More important are certain questions of interpretation that have some bearing on the form of selectional rules and their placement in the grammar. it is necessary only to state the condition on W. and to sad the feature [+ [+Human] -] in the case of both (7) and (8). . is not statable directly in the form of a Boolean structure index for a transformation . Within the alternative framework involving substitution transformations. though stated somewhat differently . In this way we can avoid mention of irrelevant intervening symbols in the statement of contexts and.8] to the right of A .]. can avoid the deficiency of dual-feature assignment noted in the case of (7) and (8). we should now give the rules (57xiv) and (5 7 xv) in the form ( 1 2 ) and (1 3). - These rules would now have the effect of assigning to frighten the feature [+ [+Abstract] . ( 1 2) [+V] � CS/[+N] . the analogous convention must be established. . where X = [ep ] and Y = [". .SOME RESIDUAL PROBLEMS (l l) X WB VY where B is the complex symbol containing the features of A (or [+A ]. though of no great importance. if A is a category symbol) in addition to each contextual feature [+ep . Suppose that the selectional rules are included in the syntax. Since ( 1 4) nevertheless differs from ( 1 5) in "deviance" from the intuitive point of view.§ ( 1 4) John frightened sincerity I. although the constituent frigh ten sincerity of the generalized Phrase-markers of ( 1 5i-iii) is marked as semantically incongruous. Nevertheless. there are frames in which this restriction can be violated with no consequent unnaturalness. and consequently the sentences ( 1 5) (but not ( 14» are finally given a nondeviant interpretation. for example. one might cite the fact that even strict subcategoriza- . as. In further support of this conclusion.ll This seems to me not at all an unnatural or intolerable consequence. the incongruity is removed by the readings assigned to constit uents dominating it. it is presumably a property determined by the joint operation of both the syntactic and the semantic components. Thus the projection rules of the semantic component and the lexical entries for such words as nonsense and speak must be designed in such a way that. in ( 1 5) (i) it is nonsense to speak of (there is no such activity as) frightening sincerity (ii) sincerity is not the sort of thing that can be frightened (iii) one can (not) frighten sincerity Clearly. this intuitive notion does not correspond to grammaticalness. Then ( 1 4) and ( 1 5) are only derivatively generated by the gram mar (in the sense of note 2) . Surely it is not surprising to find that an intuitive concept such as "deviance" can be explicated only in terms of theoretical con structs of various sorts. a descriptively adequate grammar must indicate that ( 1 4 ) is deviant (as in the case of the examples of (2» and that the examples of ( 1 5) are not. they are generated with Phrase markers indicating that they depart in a particular respect from grammaticalness. THE BOUNDARIES OF SYNTAX AND SEMANTICS This is a deviant sentence. Rather. formed by relaxing the restriction of frigh ten to Animate Direct-Objects. which have in themselves no direct and uniform intuitive interpretation. There are various ways to approach this problem. Thus it seems to me that examples such as ( 1 5) do not present a particularly strong argument for removing selectional rules from the syntactic component and assigning their function to the interpretive semantic rules. in any event. and to modify the theory of the semantic component in some way so as to allow it to accommodate these phenomena.. for example. 77). Nevertheless. one might raise the question whether the functions of the semantic component as described earlier should not be taken over. and at least in such cases as these the relation of grammaticalness to intuitive deviance will therefore be much closer. More specifically. coincide with the intuitive notion of "deviance. Alter natively. so that the dLtinction between the . for example. In further support of the argument that grammaticalness cannot. This might be cited as a slight consideration in favor of the decision to eliminate the selectional rules from the syntactic component. by virtue of the semantic properties of certain lexical items and certain constructions. in toto. we may ask whether the cycle of interpretive rules that assign readings to higher nodes (larger constituents) of the underlying generalized Phrase-marker should not be made to apply before some of the syntactic rules. as. too. by the generative syntactic rules. p . We have been considering the possibility of assigning the function of selectional rules to the semantic component. if the latter course is taken." one can cite cases of perfectly grammatical strings that are incongruous on nonsyntactic grounds (cf. in ( 1 6) (i) it is nonsense to speak of (there is no such activity as) elapsing a book (ii) elapsing a book is not an activity that can be performed (iii) one cannot elapse a book Here.SOME RESIDUAL PROBLEMS tion rules can apparently be violated without leading necessarily to semantic incongruity. then ( 1 4 ) and ( 1 5) will be directly generated by the syn tactic rules. one might plausibly maintain that base strings that deviate significantly from grammaticalness are nevertheless con stituents of sentences that receive nondeviant interpretations. Evidently. This notion. but for the present any one that is adopted must be ex tremely tentative. who show that if it is adopted. One can hardly achieve significant understanding of this matter in advance of a deep analysis of systems of semantic rules. Exactly the same can be said about the boundary separating semantic sys tems from systems of knowledge and belief.§ 1. On the contrary. further insight into these questions will await a much more intensive study of semantic interpretive rules than it has yet been possible to undertake. obliterated. Within this framework it is possible to formulate certain reasonably clear questions. which is by no means to be ruled out a priori. THE BOUNDARIES OF SYNTAX AND SEMANTICS 1 59 two components is. is a conservative compromise between the attempt to incorporate the semantic rules strictly within the syntactic component and the attempt to elaborate the semantic component so that it takes over the function of the selectional rules . in effect. § 3 . That these seem to interpenetrate in obscure ways has long been noted. one should not expect to be able to delimit a large and complex domain before it has been thoroughly explored. and that there is quite a range of possi bilities that deserve serious exploration. It is clear from this fragmentary and inconclusive discussion that the interrelation of semantic and syntactic rules is by no means a settled issue. A decision as to the boundary separating syntax and semantics (if there is one) is not a prerequisite for theoretical and descriptive study of syntactic and semantic rules. has laid the groundwork for empirical investiga tion of this sort. Alternative positions can be formu lated. the prob lem of delimitation will clearly remain open until these fields are much better understood than they are today. and it is also fairly clear what kind of empirical evidence would be relevant to deciding them. The work of the last few years. I believe. The approach I have adopted in Chapter 2. There is a general theoretical framework parts of which have received empirical support. In general. the internal organization of the syntactic component must be revised in several essential ways. on the . is explored by Bever and Rosenbaum (forthcoming). discussed in Bever and Rosenbaum (forth coming).] 60 SOME RESIDUAL PROBLEMS one hand. consider Ad jectives that are mutually exclusive in some referential domain. In particular. there is little doubt that the system of "dictionary definitions" is not as atomistic as implied by this account . It is surely our ignorance of the relevant psychological and physiological facts that makes possible the widely held belief that there is little or no a priori structure to the system of "attainable concepts. We have . Concerning dictionary definitions. Or consider the "have a" relation. for example. calor words." Furthermore. quite apart from the question of universal con straints. the system of possible concepts. Such "antonymy sets" (cf.12 Thus. it seems obvious that in any given linguistic system lexical entries enter into intrinsic semantic relations of a much more systematic sort than is suggested by what has been said so far.a system that has no intrinsic structure beyond this. it is important to determine the universal.3. two major problems are open to investigation. and systems of belief. We might use the term "field properties" to refer to these undoubtedly significant though poorly understood aspects of a descriptive semantic theory . Katz. Short of this. Some additional problems of semantic theory One major qualification must be added to this discussion of the relation of syntax to semantics. though it obviously plays a role in semantic interpretation . one can discuss only isolated examples within a theoretical vacuum. on the other.in traditional terms. for example. But such a description is hardly sufficient . universal vocabulary in terms of which these objects are characterized. First. The very notion "lexical entry" presupposes some sort of fixed. It is not surprising that nothing conclusive results from this. 1 964b) provide a simple example of a field property that cannot be described naturally in terms of separate lexical entries. § I. language-independent constraints on semantic features . just as the notion "phonetic representation" presupposes some sort of universal phonetic theory. I have described the semantic component as a system of rules that assign readings to con stituents of Phrase-markers . ). cannot in any natural way be described within the framework of independent lexical entries . though there have been several valuable studies of certain of its aspects. once again. furthermore. Once again. it is important to be aware of the many other problems that face a theory of semantic interpretation of the kind referred to in the preceding discussion ." etc. THE BOUNDARIES OF SYNTAX AND SEMANTICS ( 1 7) (i) the man has an arm (ii) the arm has a finger (iii) the finger has a cut but not ( 1 8) (i) the arm has a man (ii) the finger has an arm (iii) the cut has a finger (except. illustrate relations of meaning rather than relations of fact . and. In this case.§ 1. as possible elliptic variants of entirely different constructions. Finally. These examples. furthermore. Other systems of this sort can easily be found. The consequences of such a decision are not easy to determine. in fact. that an at tempt is made to relate "deviance" in the intuitive sense to "degree of grammaticalness" in the technical sense by excluding such examples as (1 8i-iii) from direct generation (cf. irrelevantly to this point. as Katz and Fodor have .) Suppose. (See note 1 2 . as in "the finger has an ann attached to it." where "the kidney has an ant" is not false or impossible but senseless." "the arm has a man on it. with the irrelevant exception just noted. note 1). This matter is crucial but has been relatively unexplored within any general frame work. It is clear. Thus there is no grammatical objection to "the ant has a kidney. we can do no more here than indicate problems and stress the fact that there are many unanswered questions of principle that might very well affect the formulation of even those parts of the theory of grammar that seem reasonably well established. they suggest that part of the semantic component of a grammar must be a characterization of field properties that is outside the lexicon. we have a hierarchy of terms with systematic relations that. Chapter I . although the deep structures would show that "pompous" modifies "John" in both sentences of the pair.SOME RESIDUAL PROBLEMS emphasized. they would not express the relations of the two Nouns to the Verb that are (in some unclear sense) the semantically significant ones. § 2. for example. the deep structures of the paired sentences are identical in all respects relevant to semantic interpretation of the sort we are considering here. Consider. It is not expressible in transforma tional terms. there are cases that suggest the need for an even more abstract notion of gr amm a tical function and gram matical relation than any that has been developed so far. For example.it is easy for us to please John he came yesterday (ii) it was yesterday that he came - In the case of (20). just those that determine the mean ing of the sentence . This does not seem to be true in the case of ( 1 9). so that the transformational analysis accounts for the (cognitive) synonymy. that the meaning of a sentence is based on the meaning of its elementary parts and the manner of their com bination. for example.2..Bill received a blow at the hands of John Clearly. for example. these sentence pairs : (1 9) (i) John strikes me as pompous . in any systematic way . in the case of ( 1 9i). Thus in some sense the relation of "John" to "strike" is . Cf. However. however. as is possible.Bill sold the book to John (iv) John struck Bill .I regard John as pompous (ii) I liked the play . and Chapter 2. approaching a variety of paraphrase. It is also clear that the manner of combination provided by the surface (immediate constituent) structure is in general almost totally irrelevant to semantic interpretation. whereas the grammatical relations expressed in the abstract deep structure are.the play pleased me (iii) John bought the book from Bill . in these cases. in these cases: (110) (i) John is easy for us to please . § 4. there is a meaning relation. in many cases. Many related problems have been raised in the extensive dis cussion of the distinction between the "grammatical" Subject and Predicate of a sentence and its "logical" or "psychological" Subject and Predicate (see. . To conclude this highly inconclusive discussion." and. Various formal devices for expressing these facts suggest themselves. . it seems that they lie beyond the scope of any existing theory of language structure or language use. in terms of lexical features or grammatical relations of the deep structure. 1 9 26). 1 886. glass . and that any attempt to delimit the boundaries of these domains must certainly be quite tentative. ." Whatever the force of such observations may be." " 'glass: which has the stress. .lS Consequently. nor even something denoted by words in the sentence. both of fact and of principle. would now be transferred to 'glass: " Thus in the statement "glass is elastic. . pp. "the subject and predicate are not necessarily words in the sentence. . that it is found in glass . 1 924. Jespersen. it seems that beyond the notions of surface structure (such as "grammatical subject" ) and deep structure (such as "logical subject"). hence of accounting for the meaning relation. Cook Wilson main tains (1 926. . for example. Paul. Wilson. but the general problem seems to me non trivial." We have no mechanism for expressing this fact. . [and therefore] . I shall simply point out that the syntactic and semantic structure of natural languages evidently offers many mysteries. . THE BOUNDARIES OF SYNTAX AND SEMANTICS the same as that of "John" to "regard:' and the relation of "strike" to "me" is the same as that of "regard" to "I.§ 1. 'glass' would have to be the predicate. Thus the same form of words should be analyzed differently according as the words are the answer to one question or another. and the kind of stress which fell upon 'elastic' when glass was the subject. 1 1 9 f. would no longer be subject.) that "in the statement 'glass is elastic: if the matter of inquiry was elasticity and the question was what substances possessed the property of elasticity. is the only word which refers to the supposed new fact in the nature of elasticity. in gen eral. . there is some still more abstract notion of "semantic function" still unexplained. To mention just one. now adopted the following conventions: (2 1) (i) only positively specified strict subcategorization features and only negatively specified selectional features appear explicitly in lexical entries. THE STR UCTURE OF THE LEXICON § 2. let us assume that the proper method for inserting lexical items is by a general rule that inserts the lexical entry (D. § 3. make a specific choice at each point where. Redundancy The lexicon was described earlier simply as a set of lexical entries. in effect. further structure being necessary in the lexicon to account for field properties .15 These choices do affect the following discussion. Q . . the others being introduced by the auxiliary convention (ii) . in the discussion. in Chapter 2 . . we pointed out that various general conventions can be given that permit significant simplification of such lexical entries. features that exempt items from certain phonological rules. In particular.SOME RESIDUAL PROBLEMS § 2. the latter being a set of features of various sorts (syntactic and semantic features. let us make the empirical assump tion that a grammar is more highly valued if the lexical entries contain few positively specified strict subcategorization features and many positively specified selectional features.C) in a position . in a Phrase-marker (Q being a complex symbol developed by rewriting rules). each consisting of a distinctive feature matrix D and a complex symbol C. but analogous problems arise no matter which of the proposed alternatives is selected. rather than that suggested in Chapter 2. Thus we tenta tively accept alternative (iv) of p . Furthermore. we listed alternative possibilities that seemed to deserve consideration. features that specify which morphological or transformational processes apply to strings con taining the items in question. 1 1 1 . Thus we tentatively accept the method of § 3 of Chapter 2. . § 4. where C is not distinct from Q in the technical sense of feature theory. We have. To explore the question of simplification of lexical entries somewhat further. .3. and so on)_ 1 4 We have just seen that this account is oversimplified in the case of semantic features. let us.I. for concreteness. Furthermore. justice.]) (a. ."'] (where a = + in the case of a strict subcategorization feature and a = ... [+[+N] . . then assign it the specified feature [..) is a hierarchic sequence with respect to the grammar G if [a.. .[+Human]] . .. [±Human] ) (ii) ([+N]. . [-M] . for each i < n. [a"F. [±Common])18 Where such relationships obtain.. with respect to the illustrative grammar (57) of Chap ter 2 we have the hierarchic sequences .. we might give the lexical entry for frighten (cf. ]) The conventions will introduce: the category features [-N] . Adjective. ..alP . in G.in the case of a selectional feature)."'] We also pointed out (in Chapter 2. . (23) (i) ([+Animate]. /I).] is the only specified feature directly dominating [a'+ lFH l]. [+V.]) is a maximal hierarchic .. /I.. .§ 2.. as insertable in the context sincerity . . . [-Coun t]. § 3) that a convention anal ogous to (lIl ii) can be established in the case of features cor responding to lexical categories. . for example. . Thus frighten will be specified (by (u ) plus conventions) as a Verb. we can utilize them to simplify lexical entries by the following rather natural convention: 19 (2 4) suppose that ([a1Fl].. [-Adjective]..]. (58) of Chapter 2) simply as: (22) (frighten.. (. [+ Common] .. . [a"F.. ."'] for the con textual feature [lP . In accordance with these conventions. Thus... C) is not explicitly provided with the feature specification [alP ..F. but not a Noun.. [±Abstract]) (Hi) ([+N]. = + or .S.John _ but not sincerity 16 or sincerity . the strict subcategorization features [.[-Animate] . the selectional features [+ [+N] [+ Animate]]. . Let us say that the sequence of specified features ([alFl]. or Modal..[+N] . + . THE STRUCTURE OF THE LEXICON (ii) if the lexical entry (D.NP..11 We can proceed to develop an appropriate convention to simplify lexical representation of items with inherent features in the case where these are hierarchic rather than cross classifying. .Np . • • . . cost will be specified positively for [ ... Manner] for Verbs.NP..F. and it is unnecessary (by virtue of convention (2 4» to list [+K] in C. to designate the lexical category for any item.NP]. we see that if a lexical item is specified in the lexicon as [+ .... If every lexical entry contains lexically determined features.Cl. it is unnecessary. since we can have "he read the book (carefully. .. where C contains [a. +Human. with great enthusiasm). . as seems plausible. need never receive explicit mention in the lexicon.. We have thus far considered only universal notational con ventions underlying lexical representation. introduced the features [ .. ]) the features [+N].. . every Verb in English that can occur with a Direct-Object and a following Manner Adverbial can occur as well with just a Direct-Object.NP. In the sample grammar (57). [+Animate] now being predictable. where C is a lexical category. .NP.20 Let us say that the feature [aF] is lexically determined in the grammar G if there is a hierarchic sequence ([+K].C) is a lexical entry of G. then the features [+C] and [.. Using this convention. 1 :. with respect to this entry.. for each i..1 66 SOME RESIDUAL PROBLEMS sequence with respect to the grammar G.. then it must also be specified as [+ . but resemble..N P] and [ . among others.. ." "John . Manner].. This is · to say that if (D. for example. 2 1 The strict subcategoriza tion rules of the grammatical sketch of § 3. For example. and that (D. (58) of § 3. i < n. where K is a lexical category (a = + or .]. each lexical item contains lexically deter mined features. [+Common. [aF]) with respect to G. In accordance with the observation just made. Chapter 2...) .C) is a lexical entry and C contains [aF]. there are also many language-specific redundancies. Manner]. then (D. Hence. +Count.. . Thus. Chapter 2. Then C is ex tended automatically to C' containing C along with all of the specified features CalF. though not conversely. However. read will be specified positively for both features.NP] and negatively for [ ...C) is necessarily a member of the lexical category K. we can simplify the lexical entry in (58) of Chapter 2 for boy to the following: (25) (boy. though not necessarily conversely. in the lexicon of (58).]. ...Manner. ]) By the rule (27) followed by the convention (2 1)..[ . ... Manner] � [+ ..NP . Manner]. where F =F." but not "John resembled his father care fully (with great enthusiasm). these will be automatically extended to (29) (i) (walk.NP] to be interpreted in the following manner: if (D.... (24). ..N P].. thus. [+ .§ 2. what is needed is the following rule : (26) [+ .. second. allowing us to use part of the internal structure of the notations for lexical features. hit will be entered in the lexicon in this form: (28) (i) (walk. the rule (26) can be further generalized. Actually. ." etc. [+ ... . .NP . Then such words as walk. . and also the specified feature [+ . .C) is a lexical entry with distinctive feature matrix D and complex symbol C containing [+ .. . Manner.Manner.. What is needed is a convention permitting a variable over strings to appear in the rule gen eralizing (26).. then they can occur without one. or to allow a condition on ({' to be added. Manner.. which contains each specified feature [aF] of C. then C is replaced by C'.. Clearly.NP .. Using cp as a string variable. . + . [+ ... .. ]) (ii) (hit. in effect. It is also true of Intransitive Verbs that if they can take a Manner Adverbial.NP . select any constant string as ({'. THE STRUcruRE OF THE LEXICON resembled his father.NP.. where this is well defined in terms of base rules.. Let us suppose that the rule (2 7) applies before the conventions (2 1 ). we can give the rule in this form: This is to be interpreted as follows: first. interpret the result in the manner described in connection with (26) .NP].. Here again we have a redundancy in the lexicon and a significant generalization still not expressed in the grammar. . It might also be expedient to develop the obvious convention that allows (27) to be stated as a con text-sensitive rule. ]) . . if the second C is a true consonant (that is. both phonological and syntactic.. and. on the other hand. (27) as syntactic redundancy rules. state general properties of all lexical entries. Observe that a distinction must be made between the con ventions (2 1 ). ( 27 ). + .. They are part of the procedure for interpreting grammars (the function f of ( uiv) ( 1 4 iv). except that the features in question are phono logical rather than syntactic. Thus in an initial sequence ICC in English. and therefore must be given in the grammar.. the phonological redundancy rules give a fuller specification to D." and the latter. The phonological redundancy rules that state these facts are precisely of the form (26) and are interpreted in the same way. but only with a Direct-Object. and that I have been referring to here (following a suggestion of his) as phonological redundancy rules. etc. the first must be [s] . though both play the role of eliminating redundant specifications from the lexicon. and the syntactic redundancy . The former are universal. (28) are closely analogous to the phono logical rules that Halle has called "morpheme structure rules" (Halle. and therefore need no specific statement in the grammar.22 I have tried to emphasize this by calling the former "conventions. • Thus walk can appear with or without a Manner Adverbial. consequently.NP. . and hit can appear with or without a Manner Adverbial. . We shall refer to the analogous syntactic rules (26).SOME RESIDUAL PROBLEMS (ii) (h it. and therefore make it unnecessary to provide feature specifications in lexical entries where these are not idiosyncratic. ]) . [+ . I 959b). C). Rules such as (27) . The latter... 1959a." Given a lexical entry (D. These rules are designed to deal with the fact that certain phonological feature specifications are predictable. the first must be an obstruent. "rules. not a liquid or a glide). are partic ular to a given language. but with no Direct-Object. Chapter I .Manner . given others. . The redundancy rules. the generalization to (27) has no analogue. (24) and the syntactic redundancy rules (26).NP Manner. § 6). if the second consonant is a liquid. to give a general. the two systems are analogous. it is an open question whether these have the significance of the phonological redundancy rules. THE STRUCTURE OF THE LEXICON 169 rules give a fuller specification to C . The real function of the phonological redundancy rules is to determine the class of phonologically admissible (though perhaps nonoccurring) se quences in a principled way. in English. To the extent that they succeed in doing this. /blik/) and "systematic gap" (such as. it is necessary to consider an aspect of the system of phonological redundancy rules that has not always been fully appreciated. Consequently. in English. so far as the role that they play is con cerned. and there are many descriptive studies that give charts or rules of one sort or another to specify the set of "phonologically admissible sequences. Halle's achievemen t was not merely to reiterate the fact that such constraints exist but to present a principled basis for selection of one set of rules rather than another to determine them. To this extent. they provide empirical support for the l inguisti c theory that contains the evaluation procedure that Halle proposes. there is still an important difference between them. The fact tha t there are rules for predicting certain phonological feature specifications in te rms of others has long been known. /bnik/). as well as the system of constraints on phonological rules that this procedure presupposes. language-in dependent definition of the notions "accidental gap" (such as. To see this.§ 2." "possi b le syllables " and so on. However. in many crucial cases. He showed that a very general and independently moti vated evaluation procedure for phonology (namely. That is. - . in place of an ad hoc chart or list. This observation suggests that we seek an analogue to the . conforms to the known facts.2 3 He thus was able to propose an explanation for the facts of phonological admissibility. in place of a mere description in other words. minimization of feature specification) seems to provide such a basis. application of this criterion selects a system of phonological redundancy rules that defines the notion "phonologically ad missible" in a way that. But there is no really con vincing analogue to the notion of "phonological admissibility" in the case of the syntactic redundancy rules. Since English is too poor in inflection to illustrate this difference. In a traditional grammar. What must be shown is that this formal tripartite distinction of occurring. thus distinguishing possible from impossible lexical entries (possibility with respect to a particular language. possible but nonoccurring. but instead of continuing with additional examples.24 The problem is reasonably clear. case. from a purely formal point of view. the redundancy rules provide general constraints on all lexical entries. that is. and im possible has the significance in the syntactic case that it clearly does in the phonological case. insofar as the redundancy rules are not universal conventions). but nonoccurring lexical entry" and "impossible lexical entry. § 2 . Thus what must be shown is that the possible but nonoccurring lexical entries have the status of "accidental semantic gaps" in the sense that they correspond to lexical items that the language does not provide for specifically but could in principle incorporate with no alteration of the general semantic system within which it functions. on the syn tactic level. But in general not all of the possibilities will be actually realized in the lexicon. In fact. a particular occurrence of a Noun would be described in terms of its place in a system of paradigms defined by certain inflectional categories. the syn tactic redundancy rules do make a distinction between "possible. The study of syntactic redundancy rules is a large topic in itself. namely the categories of gender. In both cases. we shall turn to German for examples.2. number. Inflectional processes It is useful to compare two ways of dealing with questions of inflectional morphology.SOME RESIDUAL PROBLEMS distinction between accidental and systematic gaps. namely the traditional method of para digms and the descriptivist method of morphemic analysis. I should like to consider briefly some of the problems that arise in the attempt to deal with morphological processes within a frame work of the sort that has been outlined earlier. however. I have no very satisfying examples at present. and merits investigation. and ." precisely as the phonological redundancy rules do. 28 we can represent the Phrase-marker of the sentence . In fact.§ 2. as containing the subconfiguration (30) . [2 Case]. THE STRUCTURE OF THE LEXICON declensional type. let us say. with the specifications being not + and . C» and that [2 N umber] and [2 Case] are in troduced by grammatical rules . associated with this Article N definite B. the specified feature " ' . etc . Regarding each of the dimensions of the system of paradigms as a multivalued feature. we can restate the paradigmatic description directly in terms of syntactic features. . Plural. they are part of the complex symbol C of the lexical entry (Bruder.but. Thus. Mutter. integers con ventionally associated with the traditional designations. . and each word has a particular "value" along each of these independent dimensions. der Brudef' .iider occurrence of Brudef'. and belonging to a certain declensional class along with Vater. Genitive. there will be a feature matrix indicating that this formative is assigned to the categories [1 Gender]. . Each of these categories constitutes an in dependent "dimension" of the paradigm. Presumably. [2 Number]. and [1 DC] (as well as to many others represented in (30) simply by ) Notice that the specified features [1 Gender] and [ 1 DC] are inherent to this formative (that is. . . 2 6 Thus the word Bruder in the phrase der Bruder would be characterized as Masculine. SOME RESIDUAL PROBLEMS [2 Number] is introduced by a context-free rule of the base apply ing to Nouns. Chapter 2) . some of considerable generality.2T and specified feature [2 Case] is introduced by a rule that does not belong to the base subcomponent of the syntax at all but rather to its transformational part (cf. this approach would substitute morphemes. note 35. if declensional type depends on Gender). or where they are partially determined by other aspects of a formative. The characteristic method of analysis of modem linguistics is rather different from the traditional approach that we have j ust restated in our terms. it belongs to the category [3 Case] . giving the form Bruder. [1 DC]. incidentally. and all but [2 Case] will appear in the terminal s tring generated by the base rules. re dundancy rules of the kind discussed earlier will apply. that the specification [1 DC] might be introduced by a re dundancy rule that. Interpretive phonological rules. Where these features are not completely independent (as. finally. for example. then operate on the phonological matrix of the lexical entry. Notice. then of these features only [2 Number] will be a feature of the preterminal symbol for which Bruder is substi tuted by the lexical rule. (A separate rule that is quite general would specify that /(V)n/ is suffixed if. in this case. one (or perhaps some hierarchic configuration) cor responding to each of the dimensions that define the system of paradigms. The system of paradigms is simply described as a system of features. A rule of the (interpretive) phonological component will operate on (30). a phonetic matrix. furthermore. In place of the traditional categories (our features). the theory of syntactic features developed earlier can incorporate the traditional paradigmatic treatment directly. This rule will assert that a Vowel is fronted in a formative that is simultaneously of the categories [2 Number]. in a completely consistent "item-and-arrangement" gram mar: . Thus Bruder in (30) would perhaps be represented in the manner of (3 1). ) In short. some quite specific. giving. takes into account both pho nological and other lexical features. If so. 29 More generally.. where each of these elements is regarded as a single morpheme. For one thing." 28 Rules would then be given that would convert (3 1) into a sequence of phonemes. we simply give no rule at all expressing the fact that certain features are phonetically unrealized. just as we give no rule expressing the fact that [+N]. or. How ever.§ 2. But this extensive set of rules is entirely superfluous and can simply be omitted under the alternative paradigmatic analysis. With the feature analysis (30).. causes cumbersome and inelegant formulation of rules when the representations to which they apply are in the form (3 1). suppletion and internal modification cause no special difficulty at all in the paradigmatic formulation. In the case of (30)' we have the corresponding rule that the Vowel is fronted when the item in question has the features [DC I] and [2 Num ber] . the often suppletive character of inflectional systems.DCI . Plural ' " . (3 1) THE STRUCTURE OF THE LEXICON Bruder'DC{·. is phonetically unrealized.. where the item in question is a Noun. But in the case of the morphemic analy sis we now have the additional rules stating that in such contexts as (3 1 ). There are several reasons for this.. in the .Plural. NP.Masculine . it is necessary to refer to irrelevant morphemes in many of the grammatical rules.. . in par ticular contexts. In the case of (3 1 ) . Gen itive . In each such case a specific context-sensitive rule must be given stating that the morpheme in question is phonetically null. For example.... for that matter. we first apply a rule stating that the Vowel is fronted in the context: . .. as zero elements..... with morphemic representations. Thus compare the rules that must be provided for the paradigmatic analysis (30) and for the morphemic analysis (3 1). Respresentations such as (3 1 ) are clumsy for a grammar based on rewriting rules or transformations. Similarly. DCI being a kind of "class marker. all four inflectional morphemes are phonetically null. many of these "morphemes" are not phonetically realized and must therefore be regarded.. as well as the fact that (as in the example) the effect of the inflectional categories may be partially or even totally internal. . the paradigmatic analysis has many ad vantages and is to be preferred. For example. and this is the usual situation in the case of agreement rules . Postal. it follows that we must allow the transformational com ponent to contain rules that alter and expand the matrix of features constituting a lexical item . in this connection. 1 964a. notice that the order of morphemes is often quite arbitrary. the grammar must contain agreement rules that assign to the Article all of the feature spe cifications for [Gender].SOME RESIDUAL PROBLEMS case of (3 1 ). to be an ill-advised theoretical innovation . and these rules add to Phrase markers specified features that enter into particular formatives. It seems that in in8ectional systems. rules of agreement dearly belong to the transformational component (cf.oo It is difficult to say anything more definite. Finally. not being part of the terminal string. pp. 43f. But in the paradigmatic representation. the rule for fronting of the Vowel must refer to the morpheme Masculine. since there have been so few attempts to give precise and principled descriptions of in8ectional systems in a way that would have some bearing on the theoretical issues involved here . though there may be cases where some compromise should be made.s l If we assume now that the paradigmatic solution is the correct one. need not be referred to at all in the rules to which they are not relevant . these elements. therefore. whereas this arbitrariness is avoided in the para digmatic treatment.) Similarly. which ever permits the optimal and most general statement of some aspect of the syntactic or phonological system. (See note 35 of Chapter 2.). I know of no compensating advantage for the modern de scriptivist reanalysis of traditional paradigmatic formulations in terms of morpheme sequences . the feature (or features) of Case must in general be specified by rules that apply after many transformational rules have already taken effect . for example. dominating their phonological matrices . Within our framework. either paradigmatic analysis in terms of features or sequential morphemic analysis is available. the features being unordered . This seems. In the case of (30). link. send. [2 Number].. would be converted to /der/ by rules of the phonology. /seNd/. [+N ] ] / - . and ')I being variables that range over integers . . N. and Case . P N umber ')I C ase a Gen der ' . and [')I Case] if the fourth element is of these categories. Z). NP. which provide an appropriate formalization for traditional rules of agreement . where /N/ = [+nasal] and the other symbols NP. as in many other languages.§ 2. Y. THE STRUCTURE OF THE LEXICON [ N umber]. N is an This rule is interpreted as asserting that in a string analyzable as (X. [. [2 Case] . For example. This formative. /liNk/. p. Regarding features as con stituent elements of formatives. rule (32) assigns to the formative definite32 in (30) the features [I Gender]. . in effect. This rule thus asserts that the Article agrees with its Noun in Gender. Thus we must have a rule that might be given in the form: . so that the words limp. IliNt/. rules of agreement such as (32) are quite analogous to the rules of assimilation of the phonological component . and [Case] of the Noun it modifies. in English.8 Number]. Article � [ aGender P Nu mbe r ')I Case where Article . /riNg/ in lexical entries. There is no particular difficulty in extending the theory of transformations to allow for the formulation of rules such as (32). so categorized. Article. ring would be re presented / liN p /. lint. where the second plus third plus fourth elements constitute an the second element is to be assigned to the categories [a Gender]. rewrite terminal symbols in certain restricted ways . The rule (3 2) is a transformational rule of the usual kind except that it introduces specified features instead of only non lexical formatives. these transformational rules will. Formally.. nasals are neutralized before stops. as is quite natural . Thus the features play a role which is intermediate between that of formatives and that of true cate gories with respect to the operation of transformational rules. Number. a. In particular. The Adjective. note 32) Other agreement rules expand an already present matrix of features . giving Isi1J #I. that the traditional approach to the de scription of inflectional systems can be fonnalized quite readily within the framework that we have established. semantic.33 Thus (33) asserts that the features [a grave] and [{3 compact] are added to a [+ nasal] that precedes an [a grave]. so that we have the rule (33) [+ nasal] � [ a grave f3 compact ] [ : . Before turning to the much more perplexing problems of derivational morphology. . the only features associated with the non lexical item definite (but cf. will have an independent feature matrix of its own. in other words. 1nl before dentals.} . then. I should like to mention a few addi tional problems that arise when inflectional features are con sidered in further detail. a lexical item may acquire other . It seems. In the case of rule (32). the features added are.for example. . this appears to be the most natural way to deal with inflectional systems. The latter. and IrJl before velars (where the voiced velar then drops in certain positions. f3 range over {+. and syntactic features. apparently. The nasal assimilates to the following consonant with respect to the features of gravity and compactness. which is expanded by the agreement rule. When inserted into a Phrase-marker. Furthennore. We have been regarding a lexical item as a set of phonological. and its features will include its inherent features (those given in its lexical entry) and those associated with the complex symbol that it replaces by the lexical rule. where a. in this case. ] I _ nantal f3 compact interpreted in the manner of rule (3 2).. etc. [{3 compact] consonant.. It asserts. is introduced into the prenominal position by a transformational rule. being a lexical item. I have not given the full statement of required context in (33 » .SOME RESIDUAL PROBLEMS are also abbreviations for certain sets of phonological features. that the nasal is Iml before labials. the rule assigning fe atures of a Noun to a modifying Adjective. quite apart from this. pp. such features as [a Number] are inherent to the Phrase-marker rather than the lexical item.but see note 35. This is so because the element boys (containing the feature [+Plural]) is not identical with the element boy (contain ing the feature [-Plural]) j ust as the element man is not identical with boy. In some cases this decision has just the right consequences. and pp.) that deletions must be recoverable. We have mentioned in several places (Chapter 3. and become part of the formative only after it is inserted into a Phrase-marker. and. for example. Consider. in neither case is relativization permitted. 1 44f. But matters do not always work out quite this smoothly. so the generalized Phrase-marker for "I saw the [11 the boys were clever 11] boy" does not underlie a sentence.§ 2. then contextual features may be added to the lexical entry beyond those that it already contains. In the case of lexical items. Chapter 2 ). THE STRUCTURE OF THE LEXICON features beyond those inherent to it in the lexicon. 1 37-1 38). the features involved in the case dimension are certainly added to a formative by rather late transformations (since case often depends on aspects of surface rather than deep structure . But various problems arise if we follow this assumption consistently. and certain features that are inherent to Nouns (such as Gender) are assigned to Verbs and Adjectives only by transformations. Hence. 145). notes 1 and 13. as we have just observed. Just as the generalized Phrase marker for the string "I saw the [11 the man was clever 11] boy" is not the deep structure underlying any well-formed surface struc ture and hence does not provide the semantic interpretation for any sentence (cf. We have been as suming that these various operations simply extend the set of features constituting the formative. Thus if we adopt the method of lexical insertion described in § 3 of Chapter 2. and have sug gested that this condition can be formalized by the following convention relating to what we called "erasure transformations" : an erasure transformation can use a term X of its proper analysis to erase a term Y of the proper analysis only if X and Y are identical. Furthermore. "identity" might be taken to mean strict identity of feature composition. . the case of the relativization transformation discussed previously (p. this terminal string: (36) John is more than [# Bill is clever #] clever The comparative transformation. To derive (34) from (35) in the manner described earlier.. but. The fea- NP N ------Aux Present I-S. in particular. they reapply to the full configura tion (35). . at this point (omitting various refinements). can be . which has." Next. .. rather. . indicated by .I I+e> John I ------- Predicate· Phrase I Copula ------� Compar VP ---- P red i ca te Adjective NP N ~ Aux Presen t c(ever o l+ A e> Bill I � Predicate· Phrase ------. the transformational rules first apply to the most deeply embedded base Phrase-marker.SOME RESIDUAL PROBLEMS Consider the rules that provide for comparative constructions of various sorts. following previous conventions. for such sentences as (34) John is more clever than Bill In this case. which applies next. namely that of "Bill is clever. Predicate Copula Adjective I VP I clever /\ V tures constituting the lexical formatives of (35) are not given explicitly. the sentence is formed from the underlying deep structure given as (35).. . Similarly. this causes no difficulty . In particular... This gives (38) John is more clever than Bill is A final option is to delete the repeated copula. # NP is . or the perfectly analogous French example (40) : ( 39) (40) these men are more clever than Mary ces hommes sont plus intelligents que Marie In the case of (39). is as-as. derived from (35). . since it differs from the Adjective of the matrix sentence in gender and number. 4 � 5 - ... it permutes 4 and 6 (technically. But consider the example (39).. 3 --.. deleting 5 and #.... deleting 4) . since it has the feature [-Plural] in the embedded sentence and [+Plural] in the matrix sentence . THE STRUcruRE OF THE LEXICON 1 79 formulated as an erasure operation that uses the Adjective of the matrix sentence to delete the corresponding Adjective of the embedded sentence .Adjective # . . some inherent and some added by transformation and as a consequence of insertion into a Phrase-marker. NP .§ 2.. it seems from such examples as these that the features added to a formative by agreement trans formations are not part of the formative in the same sense as .. but our deletion conventions should prevent the deletion of the copula. . But recall that the deletion of the Adjective in the fifth posi tion of (37) by the comparative transformation is possible only when the two Adjectives are identical . Furthermore.. the deletion of the Adjective of the embedded sentence should be blocked. .. Finally. 6 Adjective (where · · · . These observations suggest that it may not be correct to regard a formative simply as a set of features. giving (3 4) . . the deletion of the final copula in (38) requires identity of the two copulas.. in the case of (40). In the case of (34). it places 4 to the right of 6. . 34 Thus it applies to a string of this form : (3 7) 2 � --.is .. . etc. more-than. deletion of the Adjective is straightforward.).. The c ru cial difference apparently is that in this case.s6 Sentences (4 I i) and (4 I ii) pose no problems. which take part in similar rules) the inflectional features that are added by agreement transforma tions are apparently not considered in determining whether the item in question is strictly identical with some other item. similarly." in particular. where a distinction in the feature of plurality was sufficient to block deletion. the Noun Phrase in question is in Predicate position and therefore receives its num ber not inherently (as in the example discussed in connection .1 80 SOME RESIDUAL PROBLEMS those which are inherent to it or as those which it assumes as it enters a Phrase-marker. then. However. as contrasted with the case of relativization." " Tristram Shandy is a novel. for the purposes of a deletion operation. plurality of the Noun (which is a feature that the Noun assumes as it enters a Phrase-marker) is a feature that must be considered in determining whether it is identical to another Noun." respectively. Thus (4 1 iii) implies that Bill is a lawyer. discussed earlier. is a case where plurality is not considered a distinguishing property of Nouns." Here. But consider (4 I iii)." "Bill is a lawyer. one cannot replace "Bill" by "Mary" in (4 I i). But the deletion of "a lawyer.SI! Some further support for this conclusion is given by such examples as the following: (4 1) (i) John is a more clever man than Bill (ii) The Golden Notebook is as intricate a novel as Tristram Shandy (iii) I know several more successful lawyers than Bill It is clear that the deep structures for these three sentences must contain the base Phrase-markers underlying "Bill is a man. Thus in the case of the relative trans formation. and the string with which it is compared is not "a lawyer" but rather its pluralized form. in the case of Adjectives and the copula (also Verbs. is permitted only under the identity condition discussed earlier. "lawyers. However the transforma tional rules are actually formulated. it is clear that we are deleting "successful" and "a lawyer" as Predicates of "Bill" in the under lying structure."37 from the base string "I know several [# S #] lawyers. as we have just seen. " The Predicate-Nominal of the latter is not singular. features intro duced by transformation into lexical formatives are not to be considered in determining when deletion is permitted.a a Notice that this analysis of a formative as a pair of sets of features need not actually be described or mentioned in any way in the rules of the grammar. In both cases. lamp. Thus we cannot have "They are a lawyer. rather. In other words. admittedly. rather. one member of the pair consisting of features that are inherent to the lexical entry or the position of lexical insertion. a forma tive. it is. apparently. in other words. (40». etc. since it is. First. THE STRUCTURE OF THE LEXICON 181 with relativization) but rather by an agreement transformation. on rather slender evidence (but . Only the first set is considered in determining legitimacy of deletion in the manner previously described. find. These examples suggest two conclusions.§ 2. nondistinct from it."Bill is a lawyer." etc.2). and such facts as these show that Predicate Nominals must be neutral with regard to number. the conflict in number between the italicized Noun Phrases of "I know several lawyers" and "Bill is a lawyer" is on a par with the conflict in number and gender between the italicized Adjectives of "ces hommes sont in telligents" and "Marie est in telligente" (cf. determined by a general convention regarding the form of grammar. the conflicting features are introduced by agreement transformations. Thus consider once again the case of "I know several lawyers" . § 2. it is unspecified with respect to number exactly as the nasal is unspecified with respect to point of articulation in the lexical representations of the formatives king. and the example suggests that this is sufficient to permit deletion. the second member of the pair consisting of features added by transformation. Second. it is not identical with the corresponding nominal ele ment of "I know several lawyers".. in the base structure. what is involved in determining legitimacy of deletion is not identity but rather nondistinctness in the sense of distinc tive feature theory (cf. Hence.3. Hence." "Bill is several lawyers. we are tentatively proposing it for consideration as a linguistic universal. Chapter 2. is to be regarded as a pair of sets of features. it is natural to base the operation on nondistinctness rather than identity. The only apparent alternative to the proposal just advanced is a revision of the general conditions suggested earlier on the order of application of transformational rules. Similarly." Consider now one last set of questions relating to the compara- . the other member consisting of the "noninherent" features introduced by transformation.SOME RESIDUAL PROBLEMS see note 2 of Chapter 2). because the features un specified in underlying structures (such as number. then the analysis of formatives that we have suggested is a general condition on the functioning of erasure transformations . being added by what are. the italicized condition formalizes a very reasonable sense of "recoverability of deletion. Hence. and the noninherent features of the formative are precisely those that are determined by the context. we seem to be led to the conclusion that non distinctness rather than strict identity is what is involved in deletion. I do not know. Summarizing. redundancy rules. is this: a term X of the proper analysis can be used to erase a term Y of the proper analysis just in case the inherent part of the formative X is not distinct from the inherent part of the formative Y. the proposal just dis cussed seems clearly preferable. But notice that this is an entirely natural decision to reach. The original intuition motivating this condi tion was that deletions should. If this proposal is a correct one. one member consisting of the "inherent" features of the lexical entry or the sentence posi tion. in predicate position) also make no independent contribution to sentence interpretation. but in any event. Thus they are recoverable in the sense that the context that determined them is still present in the string after deletion of the item in question. in essence. and that only those features of a formative that are inherent either to its lexical entry or to the position in the sentence where it is inserted are to be considered in determining nondistinctness. hence that are recoverable even if deleted. be recoverable. Whether this may be feasible. Formally. in some sense. then. simply a reflection of context. The general principle for erasure operations. we can say that a formative must be regarded as a pair of sets of features. and are. in fact. and in this case. however. Consider such sentences as (42 ) (i) John is as sad as the book he read yesterday (ii) he exploits his employees more than the opportunity to please (iii) is Brazil as independent as the continuum hypothesis? Clearly. then. Furthermore.3 of Chapter 2. and Bill. (57xv) of Chapter 2 . Suppose that we adopt the method of lexical insertion proposed in § 3 of Chapter 2 and based on non distinctness. one class of examples that suggests that in certain cases a difference in the composition of two formatives with respect to such features as [post-Animate] should suffice to block deletion. this difference of feature composi tion does not make the two items distinct from one another. [+[+Animate] . in the sense of the preceding paragraph. There is. Hence. it is not the case that one of them is marked [+F] and the other [-F]. As matters now stand. for some fea ture [F]. in the technical sense of feature theory (that is. rather than that proposed in § 4 . " ' . heavy will have the feature [post-Animate] in the matrix sentence and the fea ture [post-Inanimate] in the embedded sentence of the Phrase marker corresponding to (35) (this Phrase-marker will be iden tical with (35) except that each occurrence of clever in (35) will be replaced by heavy. each occurrence of the Adjective clever will have such features as [post-Animate] (that is. is replaced by the rock. with the features [-Animate]. these are deviant and must be marked as such in a descriptively adequate grammar. THE STRUCI'URE OF THE LEXICON tive transformations. therefore deletion is permitted. In the Phrase-marker (35). with the features [+Ani mate]. In each case. it would be natural to regard these con textual features of the Adjective as noninherent.]) added to it by selectional rules of the base component (in this case.§ 2. associated with rock) . the deleted items . now revised as ( 1 3) of this chapter). one containing the feature [post-Animate] and the other the feature [post-Inani mate] . But we clearly must allow such sentences as "John is heavier than this rock". the two occurrences of heavy that are compared when we attempt to apply the comparative transformation differ in feature composition. " ' . . nominalization transformations of the sort that form the sentences "their de struction of the property .. Fm." etc. we touch on problems of homonymity and range of meaning that are cloaked in such obscurity." "their refusal to come . Derivational processes Derivational processes create much more of a problem for any sort of generative (that is. . Consider. that no conclusions at all can be drawn from them. A nominalization transformation will apply at the appropriate stage of derivation to the generalized Phrase-marker containing the configuration "they destroy the property" dominated by S. Rather. . for the moment. . they in fact raise no serious difficulties. will not be en tered in the lexicon as such. Where derivational processes are productive." "their refusal to participate .41 and where Fb . however.40 forming ultimately the Phrase-marker (4 3 ). refusal.SOME RESIDUAL PROBLEMS differ in selectional features from the items with which they are compared. . " ' . destroy and refuse will be entered in the lexicon with a feature specification that deter mines the phonetic form they will assume (by later phono logical rules) when they appear in nominalized sentences. " (although refusal happens " ' . however. in each of the examples of (42) . This results from the fact that they are typically sporadic and only quasi-productive. The only alternative. It is not at all clear that destruction or refusal should be regarded as Nouns in "their destruction of the prop erty . for example. and possibly this might be regarded as the factor that blocks the transformation and prevents deletion.3 . would be to assume that two homonymous lexical entries are involved. Cn stand for specified features . Thus sad is [post-Animate] in the matrix sentence of (42i) and [post-Inanimate] in the embedded sentence. § 2 . without. Cl . We shall consider several examples briefly. 39 In intro ducing examples of this sort. . arriving at any very satis factory way of dealing with the problems that arise. . . etc. the words destruction. explicit) grammar than do inflec tional systems. . where irrelevant details are omitted. in these cases. Clearly. . . the nominalized Predicate-Phrase as a whole might be said to occupy the Noun position. I 1\ � I N NP Prcd katc·Phrase I nom VP AI A \? T A destroy I+Definite] G1 • • • the I V G. take account of inherent features associated with items in lexical entries. namely the features that determine which form of nom these items take . incidentally. property to be a Noun in "their refusal surprised me. Alternatively. of course. these rules must. and so on. that in the light of these remarks we must revise the description of the example (1) of Chapter 2 . phonological rules will determine that nom"destroy becomes destruction and that nom" refuse becomes refusal. the proposed framework is quite adequate for formulating the syntactic generative rules as well as the rules of semantic and phonological interpretation. THE STRUCTURE OF THE LEXICON s Det I+Definite] their �. In such cases as these. Notice.§ 2.42 To have the proper effect." which derives in part from the string underlying "they refuse"). In any event. terror. (·terrid). candor. in a manner which we shall not describe in detail here. however. gramophone. sincerity is not introduced in (1) of Chapter 2 by the lexical rule.(="sincerity may frighten John"). when the underlying sentence "NP-is-sincere" has an Unspecified Subject and the matrix sentence in which it is embedded has a non-Definite Article. Sincerity is formed by a transformation. so that actually even this very simple sentence is the result of a transformational development from a complex basis. the discussion of the transformational cycle for English in SOME RESIDUAL PROBLEMS . phonograph. it seems that these items must be entered in the lexicon directly. such words as frighten. and so on. Hence. such as those that are involved in the formation of such words as horror. contrary to what we assumed earlier. Details aside." where "sincerity (of manner). In fact. and is a "defective Predicate" in j ust the same way as refusal is a defective Pred icate in "their refusal surprised me" or "the refusal surprised me. destruction. there are no rules of any generality that produce the derived items. there is a transformational rule that operates on "NP-is-Adjective" constructions such as "John is sincere (of manner)" and gives such nominalizations as "john's sincerity (of manner)." can be regarded as a Noun. in the example ( 1 ) of Chap ter 2). In these cases. and it is easy to show that internal structure must be assigned to these items if the phonological rules are to apply properly in forming their phonetic representations (cf. This. or.. or telegram. horrifY." like "refusal (to come). is a very un fortunate conclusion. candid. (·candifY). sincerity would surely not be entered into the lexicon. terrify. as there are in the case of sincerity. But consider now the case of quasi-productive processes. since it is clear that from the point of view of both the semantic and the phonological interpretation it is important to have internal structure represented in these words. it is clear that. though sincere would. which served as a basis for discussion throughout that Chapter. The phrase sincerity ap pears as a full NP. etc. horrid. for that matter. Their meaning is clearly to some extent predictable (or at least limited) by the inherent semantic properties of the morphemes that they contain." That is to say. Chomsky and Miller. phone in the con text tele . at least in some such cases. for a detailed statement. ]) (ii) (Stem{"' ifY . frigh ten. Halle and Chomsky. This dilemma is typical of a wide class of examples with vary ing degrees of productivity. . . forthcoming). and incorporate in the grammar overly general rules that allow for nonoccurring as well as actual cases.4 s Perhaps one must regard the gaps as accidental. in fact. it may be necessary to extend the theory of the lexicon to permit some "internal computation. but not by scope or phone in the context phono -. The same is true in the other cases .. Thus Steml can be replaced by graph. and it is not at all clear how it is to be resolved. scope.. " ." in place of simple application of the general lexical rule in the manner already described.. the rule that replaces categories such as Ste� by items of (45) must be formulated with some care. . these exten sions of base derivations within the lexicon must in general de- . Chomsky. Thus telegraph.. [+N. 1 962b. ]) these now being inserted in strings formed by prior insertion in preterminal strings of items selected from (44). There are contextual restrictions on these replacements that must be specified because these processes are only marginally productive.. Alternatively. THE STRUCTURE OF THE LEXICON Halle and Chomsky. horrifY. 1 960. [HI> . + Stems • . [+Steml ' .. whether there is any non-ad hoc solution that can be achieved at all. or. the lexicon would also contain the entries (45) (i) (graph. might be entered in the lexicon as (44) (i) ( te le'""'S tem l ' [Fl ' . .. . ]) (ii) (horr. . ]) these items being entered into strings by means of the general lexical rule. [+Stem2 " ']) (iii) (fright. More seriously. in the case of morphologically complex forms. .§ �.en. ]) (iii) (Stems. There may be several layers of such extension of base derivations within the lexicon. Furthermore. [Gl. and. . However. 1 963 . . with the substantial enrich ment of linguistic theory and corresponding reduction in its interest and importance that this elaboration entails. we shall have a complex symbol dominating a sequence of symbols. . branching must be permitted in a configura tion dominated by a complex symbo1.44 In the light of such examples. The lexicon would. Thus Stems can be rewritten fright in the context en only when the features HI. This restriction seems to • . Alternatively. it may be better to allow the lexi con to contain context-sensitive rewriting rules to effect these extensions of base derivations. Now. whichever method is chosen for extending base derivations. The former alternative is of course preferable. I do not know whether this approach will prove feasible when it is attempted in detail. It may be possible to determine the restrictions completely by feature specifications in (44). within the categorial component. Hz of (44iii) indicate that it is a pure transitive. since it does not affect the structure of the lexicon. 1 1 2-1 1 3) that branching is not tolerated within the dominance scope of a complex symbol. . In other words.1 88 SOME RESIDUAL PROBLEMS pend also on the feature composition of the item being analyzed. and this can only be done by tak ing into account the feature composition of the only partially specified lexical entries of (44) as well as the feature composition of the items of (45) that substitute for the categories appearing in the entries of (44) . and the lexical rule (now reapplicable) would be the only rule involving lexical entries. However. provi sion must be made for the fact that frighten is not a Verb of the same type as redden or soften. There is apparently no empirical motivation for allowing complex symbols to appear above the level of lexical categories. In the examples just discussed. Limitation of complex symbols to lexical categories implies that no complex symbol will dominate a branching configuration. etc. however. we have some evidence that within a word. . under this alternative. be simply a list of entries. we may have to relax the re quirement (pp. relying on a reapplication of the lexical rule to insert the items appropriately. Precisely how such rules should be formulated is not at all clear to me. (45). takes only Animate Object. in fact. Such an analysis could be extended to many other forms . A number of syntactic arguments might be given in favor of a general "causative" operation to accommodate these and other cases. green." etc. chasten as based on a similar analy sis.for example. but is no less true of syntactic processes." and so on. one might seek syntactic j ustification for a transformational analysis from an underlying causative construction so that "it frightens John" would derive from the structure underlying "it makes John afraid. Postal has suggested. this is particularly clear from a consideration of phonological rules. There is no doubt that items must be specified lexically in terms of the operations that apply to them. afraid.§ 2." "corn grows. we might go on to analyze such words as wizen. the underlying form would have to be lexically dis tinguished from the homonymous Adjective that belongs to the class of those that cannot undergo the transformational process in question). soft would be in one category. Much of lexical structure is. simply a classification induced by the system of phonological and syntac tic rules. discussed in note 15 of Chapter 2. whereas happy . THE STRUCfURE OF THE LEXICON hold only above the level of the word. It might even be extended to account for such words as drop. furthermore." where S is the structure underlying "the ball drops. With this modification. with the underlying Adjective designated lexically as one that must undergo this transformational process (in the case of chasten. tender would be in the other . clarify." Adjectives would then have to be divided in the lexicon into two classes depend ing on whether or not they undergo this transformation. such Verbs as enrage. red. that there should be . In the case of such words as frighten. where it was observed that the Intransi tive occurrences cannot be derived from underlying Transitives. Conceivably. " this in turn deriving from the abstract structure "it makes S" where S dominates "John is afraid. Alternative analyses suggest themselves for several of these examples. grow. A general "causative" transformation might permit a derivation of "he dropped the ball. Thus." "he grows corn. we still retain the earlier restriction of complex symbols to lexical categories.. from an under lying structure of the form "he caused S. how ever. such phrases as "take for granted. To some extent.3+ There we noted that certain Verbs were in close construc- . In such a phrase as "take offense at" the problem is more acute. "I didn't think that any offense would be taken at that remark"). In all cases. On the other hand. but certain transforma tions apply to this phrase as if "offense" were a normal Noun Phrase (cf. Problems similar to those of derivational morphology are not lacking beyond the word level as well. § 2. Again. for example. to give a thoroughly satisfactory treatment of this general question. those which may. and those which cannot be subject to R. I mention these possibilities simply to indicate that there remain numerous relatively unexplored ways to deal with the problems that arise when the structure of a lexicon is considered seriously. the syntactic structure is apparently the same. up. we have a lexical item with a rich internal structure. and has investigated some of the consequences of this assumption. for the present." "bring in (the book). From a semantic and distributional point of view." "bring off.SOME RESIDUAL PROBLEMS a general analysis of lexical items with respect to each rule R. however. its behavior with respect to transformations and morphological processes obviously shows that it is some sort of Verb-with-Complement construction." which abound in English. this phrase seems to be a single lexical item. and it therefore must be entered in the lexicon as such. distributional as well as semantic considera tions suggest that this is a lexical item." "look over"). are of course not to be confused with the very different constructions discussed in Chapter 2." and so on. the Verb""'Particle construction is (distribution ally as well as semantically) a unique lexical item (such as "look up. as in "I brought the book (in. down) . Once again. then. with its unique set of syntactic and semantic features. I see no way. the Particle is a fairly free "Adverbial" element. out.45 The Verb""'Particle constructions "look up (the record)." Often. Verb""'Particle constructions also provide a variety of related problems. into those which must. Consider. with respect to the possibility of applying familiar transformational rules. In these close constructions. and the resulting Phrase-marker will now have the structure required for further rules. the contextual feature [ . will be an erasure transformation that will use the freely generated Particles of the Prepositional-Phrases to delete the Particles of the lexical entries.par ticle constructions. however. are also available in Verbr-. that they take certain particles and not others. argue(# about)(# with).3 is used. argue. as. "argue with X about Y") .on'""N P]. "decide on the boat. the Phrase-marker that - . We must therefore indicate in the lexical entry for such words as decide. Associated with these lexical entries. and that these Verb'""Ad verbial constructions are very different from others (such as "decide (while) on the boat") that involve a much looser associa tion of Verb and Adverbial. and. In this case. This information can be presented in various ways. we rely on the filtering effect of transformations to guarantee correct insertion in well-formed deep structures. There are still further possibilities. the feature [ with'""N P'""a bour-'NP]). etc. Thus we would have the entries decide # on. much as in the case of telescope. then distributing the Particles of the lexical entry by substitution transformations . the choice of Particle is often narrowly or even uniquely constrained by the choice of Verb (for example. to decide. Still a third possibility would be to enter the lexical items in the manner j ust proposed. § 4. to argue. the same Phrase-markers result . and so on. incidentally. and to derive the Adverbials with a dummy element in the Preposition position. A second possibility is to develop the Adverbial freely but to give the lexical entry as a sequence of formatives. Again." in the sense of "choose the boat"). One possibility is to develop the Adverbial freely and to assign a contextual feature to the Verb (for example. we derive the correctly formed Phrase-markers when successful lexical insertion takes place.§ ll. take for gran ted. is commonly done in dictionaries. If either of the methods of lexical insertion described in Chapter ll. THE STRUCTURE OF THE LEXICON tion with certain Adverbials (for example. The same choices. Under this alternative. the Verbs in question will now be inserted only in the permitted positions. then. in fact. once again. even if it is eventually justified. . Obviously. Whether these limitations are intrinsic.SOME RESIDUAL PROBLEMS results from lexical insertion and associated operations must be different from that of the Verb"'-'Adverbial constructions. to be expected in a system as complex as a natural language. these can only be regarded as essentially notational variants. until sharper criteria are discovered. where significant systematization is j ust not possible. this discussion by no means exhausts the com plexity or variety of topics that. it is much too early to draw this conclusion with any confidence and. It is possible that we are approaching here the fringe of marginal cases. resist systematic and re vealing grammatical description. remains an open question. one can barely go beyond mere taxonomic arrangement of data. the questions we have touched on here have not yet been illuminated in any serious way by approaching them within the framework of any explicit grammatical theory. or whether a deeper analysis can succeed in unraveling some of these difficulties. For the present. we still must face the problem of extracting whatever subregularities exist in this domain. In any event. so far. since later rules apply very differently in the two cases. I see little basis for choosing among these alternatives at the moment. Still. which used informant responses and comments as evidence bearing on the psychological reality of some abstract system of phonological elements. (b) behaviorism . To accept traditional mentalism. the latter being taken as the primary object of its investigation.Notes NO TES TO CHAPTER I 1. in this way. in TwaddeIJ's critique (1 935) of Sapir's mentalistic phonology.are the rules that underlie performance represented in a nonmaterial medium?. the data provided by introspection) for the determination of competence. One would guess. for example. sense or nonsense." Men· talistic linguistics is simply theoretical linguistics that uses per· formance as data (along with other data. rather. (c) introspectionism . in this traditional sense. or is he also concerned with other facts.do the data of performance exhaust the domain of interest to the linguist. the enterprise has no point because all . The mentalist. that it is the mentalistic studies that will ultimately be of greatest value for the investigation of neurophysiological mechanisms. is not to accept Bloomfield's dichotomy of "mentalism" versus "mechanism. The behaviorist position is not an arguable matter. since they alone are concerned with determining abstractly the properties that such mechanisms must exhibit and the functions they must perform. in particular those pertaining to the deeper systems that underlie behavior?. and not with questions of truth or falsity. In fact. In particular. It is simply an expression of lack of interest in theory and explanation. For Twaddell. the issue of mentalism versus antimentalism in linguistics apparently has to do only with goals and interests. need make no assumptions about the possible physiological basis for the mental reality that he studies.should one make use of introspective data in the attempt to ascertain the properties of these underlying systems? It is the dualistic position against which Bloomfield irrelevantly inveighed. At least three issues are involved in this rather idle controversy: (a) dualism . he need not deny that there is such a basis. for example. This is clear. Uhlenbeck (1964). though in less concentrated form. . so far as they go. however. on grounds of methodological purity. To maintain. question (b) poses no substantive issue. For further discussion. that introspective judgments of the informant (often. What ever evidence is available today seems to me to show that by and large the traditional views are basically correct. This has been denied recently by several European linguists (e. 1 964). As to Dixon's objections to traditional gram mars. 1 963. to condemn the study of language to utter sterility. the linguist himself) should be disregarded is. [the man aimed] [it at John]. for the present. consider Uhlenbeck's proposal that the constituent anal ysis of "the man saw the boy" is [the man saw] [the boy]. "which is already available for the student of language. or. 2. Halle and Stevens ( 1 962). to take a more recent example. a pro posal which presumably also implies that in the sentences [the man put] [it into the box]. 1 963. Furthermore. Dixon. In any event. It is difficult to imagine what possible reason might be given for this. Perhaps this loss of interest in theory. see Katz ( 1 964C). the consti tuen ts are as indicated. since he offers neither any alternative nor any argument (beyond the correct but irrelevant observation that they have been "long condemned by professional linguists"). 1 963: Uhlenbeck.. for which the only argument offered is that it is the result of a "pure linguistic analysis. For example. but rejected forthwith." Cf. Question ( c) arises only if one rejects the behaviorist limita tions of (b). in Dixon.g. strict operationalism or strong verificationism) that were considered briefly in positivist philosophy of science. [the man persuaded] [Bill that it was unlikely]. it seems to me that speech perception is also best studied in this framework. in this case..g. there is nothing fur ther to discuss. in the usual sense.. was fostered by certain ideas (e. although the discussion of "theories" in the latter is sufficiently vague as to allow other interpretations of what he may have in mind). See. for example. and the discussion there. they support the traditional analysis without exception against this proposal. They offer no reasons for their skepticism concerning traditional grammar. and that the suggested innovations are totally unjustifiable. note 7) : to my knowledge.1 94 NOTES TO CHAPTER 1 that interests him is the behavior itself. etc. There are many considerations relevant to the deter mination of constituent structure (cf. this lack of interest in linguistic theory expresses itself in the proposal to limit the term "theory" to "summary of data" (as in Twaddell's paper. in the early nine teen·thirties." Characteristically. 3. We return to this matter later. This remains true if we consider. The notion "likely to be produced" or "probable" is sometimes thought to be more "objective" and antecedently better defined than the others. since almost all highly acceptable sentences (in the intuitive sense) will have probabilities empirically indistinguishable from zero and will belong to sentence types with probabilities empirically indistinguishable from zero. 1 0 1 )." as long as "situations" are specified in terms of observable physical properties and are not mentalistic constructs. quite untested) observations. or else unacceptable and ungrammatical sentences will be as "likely" and acceptable as grammatical ones. though "probability of a sentence (type)" is clear and well defined. 6. Cf. than the others.." (Furthermore. 5. and the concepts in volved are equally obscure. in any objective sense of this word. For example.NOTES TO CHAPTER 1 1 95 Tests that seem to determine a useful notion of this sort have been described in various places . That it may be true is suggested by several (for the moment. invariably describe the "situa tions" in completely mentalistic terms . to regard it as a pattern abstracted from observed behavior. and hence objec tively describable in purely physical terms. these classes must be extremely small and of mutually substitutable elements. see Katz and Fodor ( 1 964).." Actually. a sentence is described as gaining its meaning from the situation "British Culture. betrays a complete mis understanding of what might be expected from anthropological research." To describe British culture as "a situation" is. . furthermore." but "likelihood relative to a given situation. It is noteworthy that linguists who talk of hardheaded objective study of use of sentences in real situations.g. 2 86) the following example is cited: "anyone who feels that if so many more students whom we haven't actually admitted are sitting in on the course than ones we have that the room had 4.) But in this case. For further discussion. in the first place. Dixon ( 1 963. if the notion is to be at all significant. Miller and Isard ( 1963). in the only illustrative example in the book. These characterizations are equally vague. p. when they actually come to citing examples. Thus the acceptable or grammatical sentences (or sentence types) are no more likely. a category mistake. the latter notions are objective and antecedently clear only if probability is based on an estimate of relative frequency and if sentence type means something like "sequence of word or morpheme classes. where. on the assumption that there is some clear meaning to the notion "probability of a sentence" or "probability of a sentence type. in Chomsky and Miller ( 1 963. not "likelihood. p.for example. e. it is an utterly useless notion. support a conclusion about organization of memory which is not entirely obvious. and not ex tremely low on the scale of acceptability. and per ceptual considerations converge in support of the traditional view that these constructions are typically coordinating (multiple- . perhaps [YO UNG [intelligent man]]. Mary." This is entirely counter to the sense. handsome. Thus to assume (with Harman) that the phrase "a tall. since a positive result concerning (4ii) would.g. phonetic. with con trastive stress on "young"). intelligent man" has the structure [[[[tall young] handsome) intelligent) man] seems to me no more justifiable than to assume that it has the structure [tall [young [handsome [intelligent man)]]). 7. that in the phrase "John. The same is true if we consider the very different type of Adjec tive-Noun construction that we find in such phrases as "all the young. it seems fairly comprehensible.. there is no gram matical motivation for any internal structure. old] and middle-aged) nor [young [old and middle-aged)) has any justification. as well as Mary and her child"). 6 1 3. old.g. p. but the issue is rather whether it is always necessary. as I have just noted." the latter being ana lyzed into the coordinated items "Mary" and "their two children. Notice. Similarly. 1 964). In fact.. self-em bedding of degree two or three seems to disturb acceptability much more severely. again. Though hardly a model of felicitous style. The matter is worth studying. In these cases all known syntactic.NOTES TO CHAPTER I to be changed. see Ornan. young. and their two children" the structure is [John) [[Mary) [and their two children)]. These conclusions seem to me equally unacceptable. so that "John" is coordi nated with "Mary and their two children. as noted. In comparison. "John. [intelligent [young man]] or. and middle-aged voters" (for an interesting discussion of these various kinds of modification relations. neither the structure [[young. the assumption that there is no structure is also supported on grounds of acceptability. 1 960) or left recursive (Harman. is likely to agree that the curriculum needs revision. but surely it is false to claim that it must have this structure. Notice that there are cases where further structure might be justified (e. that conjunc tion can have this structure (e. and." This contains six nested dependencies (along with other dependencies that go beyond nesting) with no self-embedding. it is surely impossible to assume. with Yngve. It has sometimes been claimed that the traditional coordinated structures are necessarily right-recursive (Yngve. then probably auditors will have to be excluded. rule 3i). 1 963. too. Here. semantic. with extremely weak and plausible assumptions about organization of memory. red. and (or) yellow" is a constituent (that the structure is right-branching). his theory assumes also that order of generation is identical with order of production .. green. the optimal perceptual device mentioned earlier is no longer constructible. Beyond the obvious condition of finiteness of memory. the phonetic rules explicitly preclude such constituent analysis. Thus no semantic interpretation can be assigned to "old and middle-aged" in "young. in such cases as these. Yngve (1 960. and (b) that left-branching and multiple-branching in fact contribute to un acceptability exactly as do nesting and self·em bedding. old. To support this hypothesis." if one wanted to argue that "blue. that there are perceptual grounds for the analysis. it would be necessary to show (a) that it has some initial plausi bility. or that the hearer should invariably make all higher-level decisions before doing any lower-level analysis. at the last stage. hence to unacceptability. There are various ways of justifying assignment of constituent structure. then the substructures of these. green. Mary. and several other papers) has proposed a different theory to account for certain observations such as those of (4). red.that the speaker and hearer produce sentences "from top·to·bottom" (they first decide on the major structures. then he would have to show that these analyses are required for some grammatical rule. red. there are no gram matical rules that require these analyses. or something of this sort. in such a phrase as "all (none) of the blue. and the other cases mentioned here.e. green. or that "green. As . that the postulated intermediate phrases must receive a semantic interpretation. and middle aged voters" or to "green. For example. I see no plausibility at all to the assumption that the speaker must uniformly select sentence type. contribute to "depth" in Yngve's sense. All of these claims are patently false in this case. then deter· mine subcategories. there are no perceptual or other arguments to support them. as well as nesting and self·embedding. deciding what he is going to talk about. and their two children". or yellow" in "none of the blue.NOTES TO CHAPTER I 1 97 8. red" is a constituent (i. etc. Under this highly restrictive additional assumption. As to (a) . to see any grounds for objecting to the traditional analysis and insisting on additional intermediate categorization. leaving to the very end of the process the choice of lexical items). etc finally. then. that the struc ture is left-branching). branching). Notice also that this is the weakest assumption. The burden of proof rests on one who claims additional structure be yond this. and (or) yellow pennants. and left-branching and multiple-branching. It seems difficult.. that they define a phonetic contour. or yellow pennants" or to "Mary and their two chil dren" in "John. red.• . whatever these may b e . namely. forthcoming). I shall assume throughout that the semantic component is essentially as they describe it and that the phonological component is essentially as described in Chomsk.y. and Lukoff (1 956). Chomsky (1962 b). such that no sentence contains more than N clauses" (that is. 1 959b. see Chomsky (in press).NOTES TO CHAPTER 1 to (b). the hypothesis is supported by no evidence at all. he states that the language is infinite). In particular. such as (2i). Furthermore. as predicted) and that left-branching structures are far more acceptable than nested structures of equal "depth. the hypothesis is contradicted by the observation (4iii) that multiply coordinated structures (cf. Halle and Chomsky (1960. 1 9624). Part 2). I shall return to this matter in the next chapter. However. However. Postal (1 964). it seems that when Dixon denies that a language has infinitely many sentences. he is using the term "infinite" in some special and rather obscure sense. Aside from terminology. Thus on the same page (p. 1 1 . though very low in "depth. since the unaccept ability in this case follows from the assumption of finiteness alone without the additional assumption of "top-to-bottom" production for speaker and hearer. Halle (1 9594." are still unacceptable. I assume throughout that the syntactic component contains a lexicon. 9. The examples given by Yngve all involve nesting and self-embedding and hence are irrelevant to the hypothesis. This suggests an interesting argumen t as to why grammars should contain transformational rules. 1 2. note 7) are the most acceptable (rather than the least acceptable." one might use the corresponding Humboldtian notions "inner form" ." in Yngve's sense. Some additional weight to this argument is given by the dis cussion of performance models involving transformational gram mars in Miller and Chomsky ( 1 963. It also fails to explain why examples of type (4iv). Yngve makes one important point in these papers. For further discussion of his remarks in this con nection. Either this is a blatant self-contradiction. See Dixon (1 963). and that each lexical item is specified in the lexicon in terms of its intrinsic semantic features. I t is astonishing to find that even this truism has recently been challenged. hence to reduce the perceptual load. Chomsky and Miller (1 963). I follow here the exposition in Katz and 1 0. or else he has some new sense of the word "in finite" in mind. 83) on which he objects to the assertion "that there are an infinite number of sentences in a language" he states that "we are clearly unable to say that there is any definite num ber. that some transformations can be used to decrease nesting. N. Halle. In place of the terms "deep structure" and "surface structure. " respec The terms "depth grammar" and "surface grammar" are famil iar in modern philosophy in something roughly like the tively (as used of a sentence). Thus. 1 05). However. sense here lying structure" and "superficial structure" (Postal. Such phrases are meant by those who use them in a sense that is nei ther obscure nor false. in the sense in which these terms are used here. 1 68) . same notions. forthcoming) for some discussion and references." in the sense in which these terms will be used here. here. But the philosopher puts them into his alembic.NOTES TO CHAPTER 1 1 99 seems to me that "deep structure" and "surface s tructure. is drawn quite clearly in the Port-Royal Grammar (Lancelot et al. Hockett uses similar ter 1964b) for the The distinction between deep and surface structure. and so imagines that he has discovered an error of the vulgar [pp.. . It follows that there are two objects of thought: . 1 660). minology in his discussion of the inadequacy of taxonomic li nguis tics ( Hockett. We speak of pain coming and going. I have adopted the more neutral terminology to avoid the question. superficially similar sentences. there may be something in the structure of them that disagrees with the analogy of grammar or with the princi ples of philosophy . or representation as the immediate object of thought." 1 953. there are phrases which have a distinct mean ing. Chapter 29). p . to have an idea. 1 958. . and removing from one place to another. Wittgenstein's distinction of " Tiefengrammatik" and "O berflachengrammatik. 1 67-1 68] . . of textual in terpretation. In phil osophical discussion. picture. 1 5-16. we speak of feeling pain as if pain was something distinct from the feeling of it. But philosophers take an idea to be "the object that the mind contemplates" (p. " in which " to have an idea of anything signifies nothing more than to think of it" (p. reduces them to their first princi ples. More generally. while at the same time. pp. do correspond quite of a sen tence and "outer form" of a sentence. the surface structure of certain expressions being mis takenly considered to be semantically interpretable by means ap propriate only to other. though it closely to Humboldtian "inner form" and "outer form. then. he criticizes the theory of ideas as based on a deviation from the "popular meaning. draws out of them a sense that was never meant. it is often introduced in an attempt to show how certain philosophical positions arise from false grammatical analogies. 1 05). is to possess in the mind such an image. See Chomsky (1964. Postal has used the terms "under intended (cf. Thus Thomas Reid (1 785) holds a common source of philosophical error to lie in the fact that in all languages. imagination. as Peter Rosenbaum has pointed out. Dans cet exemple. j'ai n'est dit que par une imitation. . as Reid regards them. Chapter 3). One of the sources of these absurdities is the failure of the philosopher to attend "to the dis tinction between the operations of the mind and the objects of these operations . comme nous avons dit j'ai un livre. . etc. fai un diamant. 1 3. although this distinction be familiar to the vulgar. 1 38). je pense.1 80) : Ainsi. and there are other modifications relat ing to the marking of the passive transformation. je veux. and. fai un doute. j'ai une montre. c'est-a·dire. in the follow· ing passage (pp. j'ai une montre. . dia mant. c'est-a·dire. be regarded as em bedded in a Prepositional-Phrase (cf. ne sont que des termes meta physiques qui ne designent que des manieres d'etres consideres par des points de vue particuliers de l'esprit. and the thing represented by it. 1 4 . . 1 79. j'ai envie. concept. " (p. rai une idee. fai la volonttf. doute. p. the treatment of the Verbal Auxiliaries in ( 1 0) and ( l l) is incorrect. c'est-a-dire. montre sont autant de noms d'objects reels qui existent independamment de notre maniere de penser. rai envie. envie. Ainsi. which is in the mind. 1 931). idee. fai une idee. et encore moins des etres sensibles que l'on puisse unir run avec l'autre. In fact. j ai est une expression qui doit etre prise dans le sens propre: mais dans fai une idee. From this conclusion follow the absurdities. These descriptions are not fully accurate. ne marquent point d'objets reels. more properly. fai pitie. Mais livre." Furthermore. the sentential complement in ( 1 0) should. nous disons par imitation. the sentential complement of (l l ) should b e regarded as embedded i n the Noun-Phrase Object of "expect. fievre. fai peur. C'est une expression em pruntee. je con�ois de telle ou telle maniere. in fact. je desire. etc. fai la jievre. of the traditional theory of ideas. be strictly limited to "the detection of the sources in linguistic idioms of recurrent misconstructions and absurd theories" (Ryle. to which we shall return in the next chapter.2 00 NOTES TO CHAPTER 1 the idea. in a work published posthumously in 1 769. 1 1 0). it has been widely held that the aims of philosophy should. ' In more recent years. It seems clear that many children acquire first or second languages . Reid's linguistic observation is made considerably earlier by Du Marsais. peur. Notice that these two senses of "having an idea" are distinguished by Descartes in the Preface to the Meditations ( 1 64 1 . and found in the structure of all languages . au lieu que san ttf. Katz (1 964d). there is no a priori reason why a natural language could not con tain a word "HERD. Similarly.similarly. a psychological truth. Thus it seems that a child must have the ability to "invent" a generative grammar that defines well-formedness and assigns interpretations to sentences even though the primary linguistic data that he uses as a basis for this act of theory construction may." similar to "limb" except that it designates the single obj ect cons isting of a dog s four legs. 1 6.' '' For interesting comments on such definitions. so that "its LIMB is brown" (like "its head is brown") would mean that the object consisting of the four legs is brown. it need not be true of somewhat more abstract and functionally defined notions such as "barrier"). 17. and it is a nontrivial fact that they apparently do. By a "reasonable procedure" I mean one that does not involve ." There is no logical necessity for names or other "object words" to meet any condition of spatiotemporal contiguity or to have other Gestalt qualities. 1 667). and timbers. no doubt. 403b). but to extricate his thought from the signs or words which often agree not with it" (Cordemoy. For example. be deficient in various respects. he is giving an unmotivated definition of "proper name. . see Foot (1961). and heat: " though "the physicist would describe it as 'stones. Russell ( 1 940. if we interpret his notion of "logically proper name" as embodying an empirical hypothesis. It also seems apparent that much of the actual speech observed consists of fragments and deviant expressions of a variety of sorts. • ' . Russell is stating what is. 33: "from a logical point of view. p. Interpreted in this way. the "essence of a house is assigned in such a formula as 'a shelter against destruction by wind." etc. rain.NOTES TO CHAPTER 1 201 1 5. a proper name may be assigned to any continuous portion of space time" ) . from the point of view of the theory he constructs. there is an important element of truth in the traditional view that "the pains which everyone finds in conversa tion is not to comprehend what another thinketh. Inter preted otherwise." like the collective "herd" except that it denotes a single scattered object wi th cows as parts. bricks. In general. Thus for Aristotle (De A nima. and the problem this poses for speech perception is magnified many times for the language learner. it is not true of "United States" . Thus there are no logical grounds for the apparent nonexistence in na tural languages of words such as "LIMB. insofar as the designated objects are of the type that can actually be perceived (for example. so that "a cow lost a leg" implies "the HERD lost a leg. quite successfully even though n o special care i s taken to teach them and no special attention is given to their progress. 20. 1 8. with respect to GJ• Thus an unambiguous sentence should receive one structural description. many possibilities that can be considered here. simplification of the technique for finding a compatible hypothesis. There are. to construct an actual theory of language learning. What I am describing is an idealization in which only the moment of acquisition of the correct grammar is con sidered. it might very well be true that a series of successively more detailed and highly structured schemata (corresponding to maturational stages. when supplemented. in ways that depend in part on primary linguistic data and the order and manner in which they are presented. Actually. the gradual development of an appropriate hypothesis. perhaps. by certain kinds of semantic in- . In fact. given Gj• one for each way of interpreting the sen tence s. and grammars and (throughout this is less obvious) for determining the values of f and m in all cases. For example. one that does not incorporate an "encyclopedia. a doubly ambiguous sentence two structural descriptions. Fur thermore. there is little reason to suppose that reasonable procedures of translation are in general possible. but perhaps in part themselves determined in form by earlier steps of language acquisition) are applied to the data at successive stages of language acquisition. for example. the preconditions (i)-(v) them selves might possibly be developed on the basis of deeper in nate structure. The possi bility of a reasonable procedure for translation between arbitrary languages depends on the sufficiency of substantive universals. (and each structural description must be assigned to exactly one sJ. We assume that mappings are effective . a priori.that there is an algorithm for enumerating sentences. Introduction of these additional considerations might affect the general discussion in many ways." See Bar-Hillel ( 1960) for discussion. etc. 19. in some limited but nevertheless real way.202 NOTES TO CHAPTER 1 extralinguistic information . structural descriptions. It assumes that the technique for discovering the correct hypothesis (grammar) must be based on procedures of successive segmentation and classification of the items in the corpus (which constitutes the primary linguistic data.that is. although there is much reason to believe that languages are to a significant extent cast in the same mold. i t would b e necessary to face several other very serious questions in volving. It is instructive to see how modern structural linguistics has at tempted to meet these conditions. a set of structural descriptions should be assigned by f to each s. and the continual accretion of linguistic skill and knowledge and the deepening of the analysis of language structure that may continue long after the basic form of the language has been mastered. Obviously. For discussion. over a wide range of cases. and since language is a tightly interconnected system. formation the exact relevance of which to the problem at hand has never been clarified). in terms of value. An actual acquisition model must have a strategy for finding hypotheses. 22. is that given primary linguistic data D. step-by·step procedures of classifica tion and segmentation. On the . Cf. that the strategy is to consider only grammars that have better than a certain value (in terms of the evaluation measure (v». the latter possibility is not to be discounted. for example. it is not clear that Quine's position should be taken as in any real sense an empiricist one. Actually. Failure of attempts to justify an explanatory theory may be in terpreted in various ways. at each stage in the process of language learning. See references of note 1 0. 24.in particular. It may indicate that the theory is wrong. as has generally been noted by commentators. although this was clear to Descartes (and was later re-emphasized by Leibniz. In fact. See note 19. that the grammar tested for descriptive adequacy was not the most highly valued one. the methodological discussions of modern linguistics pay very little attention to considerations (ii)-(iv) (though they do imply certain conclusions about them) and concentrate almost solely on development of constructive. in his critique of Locke's Essay). justification of linguistic theory does not avoid the problems faced by justification of any sub stantive and nontrivial empirical hypothesis. This point has some historical interest. In fact. Chomsky (1 964). so that the intersection of the class of grammars compatible with D and the class of grammars which are highly valued be reasonably small. Thus he goes on to propose that in the innate quality space a red ball might be less distant from a green ball than from a red kerchief. then. of course. What is required of a significant linguistic theory. § 8. it was necessary to sacrifice descriptive adequacy. Suppose. the class of grammars compatible with D be sufficiently scattered. so that we have not just a pre-experiential characterization of distance but also an innate analysis of this into distance in various respects. In short. or that its consequences were incorrectly de termined . 25. To compensate for this extremely strong demand on the procedure of grammar discovery. Locke's attempt to refute the doc trine of innate ideas is largely vitiated by his failure to observe the distinction we have just been discussing. see Lees ( 1 957). 23. Since any reason able evaluation measure must be a systematic measure. Only then can language learning actually take place.NOTES TO CHAPTER I 21. What is particularly puzzling. thus the empirical claim of "reinforcement theory" will be that learning of language can not proceed in the absence of data. or to analyze or evaluate them. perhaps." etc. hence as adopting an extreme form of nativism. for he does not even require that the "reinforcing stimulus" impinge on the responding organism. An extreme example. what are intended as empiricist views have gen erally been formulated in such an indefinite way that it is next to impossible to interpret them with any certainty. What this proposal comes to is that the only function of "reinforcement" may be to provide the child with information about correct usage. Quine suggests ( 1960. then this account is so grossly and obviously counter to fact that discussion is quite beside the point. is the insistent claim that this paraphrase is some how "scientific" in a way in which traditional mentalism is not." As Quine correctly notes." "conditioning. at least. 1 957). then. one might interpret him as propos ing that such concepts as "ball" are innate ideas. 82-88) that "some basic predilection for con formity" may take the place of "ulterior values. pp. Alternatively." though many contemporary behaviorists use this term in such a loose way that reference to reinforcement adds nothing to the account of acquisition of knowledge that they propose." "reinforce ment. it is sufficient that it be hoped for . For example. we may interpret these terms as metaphoric extensions of the (essentially homonymous) terms used in experimental psychology." and that society's reinforcement of the response may consist "in no more than cor roborative usage. Skinner's concept of "reinforcement" is apparently still weaker than this. in which case what is proposed is a mentalist account differing from traditional ones only in that many distinctions are necessarily obscured because of the poverty of the terminological apparatus available for para phrase of the traditional mentalistic notions. " this again is congenial enough to Skinner's scheme.204 NOTES TO CHAPTER I basis of these few comments. There seem to be only two coherent interpretations that one can give to this account. it is difficult to see wherein the cited proposal differs from this. which appear in it. whose resemblance to the child's effort is the sole reward. 26. is Skinner's ac count of how language is learned and used (Skinner. If we interpret the terms "stimulus. Actually.. as having the meanings given to them in experimental psychology. for he does not enumerate the re wards" (this being one of the contributory factors to the near vacuity of Skinner's scheme). my note 26). In further support of such an anti empiricist interpretation. Unfortunately. This application is perhaps mediated by "reinforcement. one may point to Quine's virtual renunciation of reinforcement theory (cf. • je demeure d'accord que nous apprenons les idees et les veri tees innees. This work has demonstrated that peripheral processing in the receptor system or iq lower cortical centers may provide a complex analysis of stimuli that. I depart here from the Langley translation. et on ne laisse pas de les apprendre. . to the normal use of language. Cf. the central problem of descriptive adequacy is not really raised within structural lin guistics. previously unheard sentences . which was a dominant theme in rationalistic linguistic theory..NOTES TO CHAPTER 1 or imagined (for a collection of examples bearing on this matter. Ainsi je ne saurois admettre cette proposition. accepting Bloomfield's behaviorism under in terpreta tion (b) of note I (as well as Firthians and "neo-Firthians" and many others). soi t en les verifiant par l'experience. . which renders this passage inaccurately. many "neo-Bloomfieldian" linguists. Thus the suggestion that the various theories of immediate constituent analysis might be interpreted as gen erative. The French original is as follows: " . Hubel and Wiesel ( 1 962). Lettvin et al. That this is a fair interpretation of taxonomic linguistics is not at all clear. Secondly. structural linguistics has rarely been concerned with the "creative" aspect of language use. " Cf. 29. 1 959b). Hence. phrase structure grammars (as in Chomsky. 1962a. for example. to organization of the primary linguistic data. Chomsky ( 1 964) for additional discussion and quotations il lustrating Humboldt's views on these questions. or Postal. Frishkopf and Goldstein ( 1 968). seems to be rather specific to the animal's life-space and well correlated with behavior patterns. Les verites des nombres sont en nous. It has. . Others have held that a grammar should at least describe the "habits" or "disposi- . at least in theory. 1956. limiting the task of grammatical description.that is. 2 7. soit en les tirant de leur source lorsqu'on les apprend par raison demonstrative (ce qui fai t voir qu'eIles sont innees) soit en les eprouvant dans les exemples comme font les arithmeticiens vulgaires . 80. . tout ce q u'on apprend n'est pas inne. Thus it seems that not even peripheral processing can be described within the unstruc tured and atomistic framework that has been presupposed in empiricist thinking. as is now known. and very likely beyond their intentions as well. given li ttle attention to the production and interpretation of new. see Chomsky. 1 964a) certainly goes beyond what is explicitly stated by linguists who have developed these theories. have thereby explicitly rejected any concern for descriptive adequacy. 28. ( 1 959). in other words. need not be at all elementary. soit en prenant garde a leur source. For one thing. These mechanisms. further more. Empiricist views leave open precisely the same question. it is not clear that most behaviorist tendencies should be regarded as varieties of empiricism at all. though the sense in which language use might be regarded as a matter of habit or disposition has never been satisfactorily clarified. more specifically. In general. See Lenneberg (forth- . every known species has highly spe cialized cognitive capacities. To be more precise. there is no clear sense of the term "habit" or "disposition" in accordance with which it would be correct to describe language as a "habit structure" or a "system of dispositions. It is important to observe that com parative psychology has not characteristically proceeded on empiricist assumptions about knowledge and behavior. Whether our assumption about the mind is that it contains the schema for transformational grammar or that it contains mechanisms for making arbitrary associations or for carrying out certain kinds of inductive or taxonomic operations. For the moment. not showing how the postulated internal structure arises. Similarly. it is important to bear in mind that an extremely specialized input-output relation does not necessarily presuppose a complex and highly structured device. and lends no support to these assumptions." In general. as innate structure. there is no better account of how the empiricist data-processing operations might have been developed. apart from their relative success in accounting for the given facts of language acquisition.206 NOTES TO CHAPTER 1 tions" of the speaker. 32. that its various maturational stages (see note 1 9) have critical periods. This is the only respect i n which a comparison of such alternatives is relevant. they renounce any interest in mental processes or faculties (that is. in the problems of descrip tive or explanatory adequacy). On the contrary. 11 1 . as distinct from classical empiricism. in a species. there is no justification for the common assumption that there is an asymmetry between rationalist and empiricist views in that the former somehow beg the question. since. There is reason to believe that the language-acquisition system may be fully functional only during a "critical period" of mental development or. But this consideration ap parently offers no information that has any bearing on the choice among alternative theories. Nor does comparison with species other than man help the empiricist argument. there is apparently little knowledge about the brain and little engineering insight into plausible physical systems that can be used to support these hypotheses. than there is of how the rationalist schema may arise through evolutionary processes or other determinants of the structure of organisms. empiricist or otherwise. to be descriptively adequate. and to attempt. the empiricist approach begins its investigation with the stipulation that certain arbi trarily selected data-processing mechanisms (e. There are no grounds for any specific assumptions. I shall . and it is difficult to see what alternative there can be to it if dogmatic presuppositions as to the nature of mental processes are eliminated. if this can be shown. however. attempting to show that the result of this application corresponds to grammars that can be shown. Sutherland. about the internal structure of this device. 31. Many other aspects of the problem of biologically given constraints on the nature of human language are discussed here and in Lenneberg ( 1 960).g. and the problem we face is that of determining the structure of the device that mediates this input-output relation (the same is true of the more general problem of which language acquisition is a special case). 1 959. independently. what we are given is certain information about the primary data that are presented and the grammar that is the resulting product. we would naturally turn to the study of uniformities in the output (formal and substantive universals). It is a curious fact that empir icism is commonly regarded as some how a "scientific" philosophy. 1 964). to develop a richer and more comprehensive theory of mind. Notice that we do not. it is not implied that these are distinct and separate components of the full perceptual system. It then investigates the applica tion of these procedures to data. Continu ing with no preconceptions. M. This. In fact. of course. imply that the functions of language acquisition are carried out by entirely separate com ponents of the abstract mind or the physical brain. which we then must attribute to the structure of the device (or. this alternative rarely being a serious one in the cases tha t are of interest) . to uniformities in the input. That is. has been the rationalist approach. A nondogmatic alternative to empiricism would begin by observ ing that in studying language acquisition. a theory that meets conditions (i)-(iv) of p. in this way.NOTES TO CHAPTER 1 coming) for an important and informative review of data bearing on this question . without. Actually. in effect. I n the particular case o f language acquisition.. taxonomic procedures) are the only ones available to the language-acquisition device. the empiricist approach to acquisition of knowledge has a certain dogmatic and aprioristic character that is largely lacking in its rationalist counterpart. aspects of cognition share properties of language acquisition and 33 . just as when one studies analyzing mechanisms in perception (cf. principles of association. it is an impor tant problem for psychology to determine to what extent other language use. . but it is much more difficult to study. and. 37. been a fair amount of in vestigation into the formal properties of very simple theories of grammar. The latter is. taking them simply as facts to be accounted for by a grammatical theory. particularly in the case of (2ii). 2. There has. 8 5. those referred to in § 2). in fact. see Chomsky ( 1 963). any such grammar must meet formal conditions that restrict it to the enumeration of recursive sets. falls within the framework of context-free grammar. In fact. For surveys of this work. it has been restricted to weak gen erative capacity. However. NO TES TO CHAPTER 2 I. In detail." "Noun Phrase. a special form of the theory of context-free grammar apparently plays a crucial role within the general theory of transformational grammar. Neither the theory of context free grammar nor the theory of finite-state grammar is an artifact invented for mathematical investigation. I shall raise no further question (except of detail) about the adequacy of these observations. been overwhelming accord about most of them. in fact. and further discussion in Chapter 3 and Chapter 4." "transformation. For present purposes. it seems that.2. that any linguistic theory under discussion attempts to meet at least these conditions. For the most part. Chomsky and Schiitzenberger ( 1 96 3 ) 36. but." "Subject") and ultimately must relate this sys tem of concepts to potential empirical phenomena (to primary linguistic data). Cf.g. there is some room for discussion about both termi nology and substance throughout (2). though there are a few results involving strong generative capacity as well (in particular.lw8 NOTES TO CHAPTER 1 henceforth assume. Each is well motivated formally and has independent interest. I think that the central facts are clear enough. of conditions on deletion transformations. as Postal shows ( 1 964a). 1964C). and there has. also note I. See Postal ( 1 962 b. In particular. during the past few years. 1 964a. For reasons discussed in Chomsky (1957) and . A theory of language must state the principles interrelating its theoretical terms (e. obviously. with no further comment.. by far the more interesting notion. As we shall see later. the conditions on base rules. This possibility cannot be ruled out a priori. alternative conventions and decisions have been applied. apart from linguistics. § 2 . insofar as it is clear. and each has in fact been proposed by linguists as a comprehensive theory of language." "morpheme. "phoneme. almost every linguistic theory that has received any substantial attention in recent years. Chapter 3. it seems definitely not to be the case. when the theory of transformational grammar is properly formulated. in reality. itself. transformations. further more. In general. This is not particularly disturbing. the phonemes. it should be expected that only descriptions con- . then. as the depth and range of linguistic study increases. general conditions on the applicability of grammatical rules rather than aspects of the particular language. in short.. morphemes. We must. Study of descriptive or explanatory adequacy· may lead to such a conclusion." where D is represented in terms of primitive notions of the theory. the difficulty or impossibility of formulating certain condi tions within the framework of an otherwise well-supported theory of grammar provides some evidence that these are. then. etc. expressed in a letter of 1 82 2 to Schlegel (Leitzmann. Several cases of this sort will be mentioned later. noch so rohe Sprache selbst. partial generative grammars will provide the only empirical data critical for evaluating a theory of the form of language. of the language are. und wenn man jede. 84) : "dass jede grammatische Discussion nur dann wahrhaften wissenschaftlichen Gewinn bringt.NOTES TO CHAPTER 2 elsewhere. That is. to be expressed within the system of grammatical rules. that the notion "grammar" must be developed and explained in terms of previously defined notions such as "phoneme. gerade mit derselben Sorgfalt behandelt als Griechisch und Lateinisch. If so. it seems to me that all significant structural notions will have to be characterized in terms of the previously defined notion "generative grammar" (whereas structural linguistics has assumed. 1 908. I am assuming that the basic notion to be defined is "G is a most highly valued grammar of the language of which primary linguistic data D constitutes a sample." Study of a wide range of languages is only one of the ways to evaluate the hypothesis that some formal condition is a linguistic universal. als llige in ihr allein der ganze Zweck. such evidence must be drawn from grammatical descriptions of relatively few languages. considera tions internal to a single language may provide significant sup port for the conclusion that some formal property should be attributed not to the theory of the particular language in ques tion (its grammar) but rather to the general linguistic theory on which the particular grammar is based. p. accept Humboldt's conclusion." and "morpheme"). the elements that play a specified role in the derivations and represen tations determined by G. wenn sie so durchgefiihrt wird. What is important is that such assumptions be supported by available evidence and formulated with enough clarity so that new or improved generative grammars will have bearing on their correctness. Paradoxical as this may seem at first glance. in general. For the present. The terminological equivocation has only the effect of suggesting to the casual reader. These demonstrations of the inadequacies of phrase struc ture grammar have not been challenged. Postal. quite erroneously. A stronger but rather well-motivated condition is pro· posed by Postal (1 964a).). The essential adequacy of phrase structure grammar as a model for taxonomic grammatical theory (with the possible but irrelevant exception of problems involving discon tinuous constituents . That is. we have pairs ( a. 4. or anyone else. and many others. in an article with the subtitle "a defense of phrase structure. 1 957. where many of the standard arguments against phrase structure grammar are repeated. to my knowledge. Harman in effect re states the arguments against phrase structure grammar as argu ments against limiting the term "phrase structure grammar" to the particular systems that have previously been defined as "phrase structure grammar. § 4). in the sense of phrase structure grammar." This terminological proposal does not touch on the substantive issue as to the adequacy of the taxonomic theory of grammar for which phrase structure grammar (in the usual sense) is a model. Chapter 6). 1 964a) is demon strated quite convincingly by Postal. 1 6. Chomsky (1 963. The most extreme example of this can be found in Harman ( 1963). and is not challenged by Harman. with approval.2 10 NOTES TO CHAPTER 2 3. is whether the term "phrase structure grammar" should be restricted to taxonomic models or whether the term should be used in some far richer sense as well. that there is some issue about the linguistic adequacy of the theory of phrase structure grammar (in the usual sense). although some confu sions have been introduced by terminological equivocations." This curious situation results simply from the author's redefinition of the term "phrase structure" to refer to a system far richer than that to which the term "phrase structure grammar" has been universally applied in the rather ample literature on this subject (in particular. contextual restrictions. see the references cited on p. Since descriptions of this sort are few. to a system in which in place of category symbols. The only issue that Harman raises. . q» . A weak though sufficient condition is given in Chomsky ( 1955. § 3)· For some discussion. cerned with deep structure will have serious import for proposals concerning linguistic universals. in this connection.see Chomsky. Some aspects of this question are dis cussed in Chomsky and Miller ( 1 963. and this terminological question is of no conceivable importance. etc. any such proposals are hazardous. but are clearly no less in teresting or important for being hazardous. where a is a category symbol and q> is a set of indices used to code transfor· mations. As noted earlier. See the final paragraph of § 2. '1') as a single. or we may interpret it as proposing that the term "phrase struc ture grammar" be used in some entirely new sense. rather than extrinsic.NOTES TO CHAPTER 2 211 A further source of possible confusion. that there is only one occurrence of B in X. and. unanalyzable category symbol. at least). and some substantive disagreements about the usage of these terms. Bill. An analogous assumption with respect to transformational rules is made in Matthews (1 964. in which case it is immediately refuted on grounds of descriptive inadequacy. For some further discussion see Chomsky (in press). these being context-free grammars where the sequential property is. a proposal which is immediately refuted by the fact that under this interpretation. to provide a binary analysis of the major category S into sincerity (NP) and may frigh ten the boy (VP). furthermore. what we have here is a new proposal as to the proper evaluation procedure for a phrase structure grammar. Thus if we were to change the rules of (5). intrinsic (in the sense of note 6. 5. Under this interpretation. Thus we have the following alternatives: we may interpret the paper as proposing a new evaluation measure for phrase structure grammars. Z are unique. that Y.4 for an emendation of these suggested definitions of functional notions. to my knowledge. and. did Tom see you'!. the structural description provided by the Phrase-marker of the now highest-valued grammar is invariably incorrect. 6. there are rather different conventions. in John saw Bill. Appendix A. Chapter 3). This assumption is made explicitly in Chomsky ( 1 955). Let us assume. Formal properties of se q uential grammars have been studied by Ginsburg and Rice (1962) and Shamir ( 196 1). in the discussion of the base of a transformational grammar (Chapter 7). and would have no categorial assignment in common. as presupposed here (for the context-sensitive case. Tom would belong to three distinct and entirely unrelated categories. in which case it has no bearing on the issue of the adequacy of phrase structure grammar. where this and other criticisms of transformational grammar. the Phrase-marker (3). some real. the three elements John. some only apparent. is that there is a way of interpreting the grammar pre sented there as a phrase structure grammar. The definition can be generalized to accommodate the case where this . are taken up. furthermore. § 2). For example. in all subsequent empirical studies of transformational gramma r. in connection with this paper. in this case in other words. namely by regarding each complex element (a.3. correspondingly. 7 . then the latter would be the Predicate-of the sentence in the sense defined in (1 1). may be found in Chomsky ( 1 955). Halle. and for a much more far-reaching study. Z is an A and Z' is an A. In particular. Precise definitions for most of the notions that will be used here. elements of nonlexical formative categories (e. or whether. clearly. What seem to be insuperable difficulties for any such analysis have been raised and frequently reiterated. taking occurrences into account. Halle and Chomsky. 1 963). Notice that accurate definitions require a precise specification of the notions "occurrence. 1956. Cf. consider the phonological rules that assign stress in English by a transformational cycle (see Chomsky. Chomsky and Miller. 9. the rules M -. forthcoming. as in John may try (it is permitted) and John may try (it is possible). The significance of this distinction will be discussed later. 1 1. By and large. Second. should not be included in the set (51). Some have argued that the distinction in question has nothing to do wi th rules of English. then we may form the string X'""Z'""a nd'"'Z''"'Y.. in fact. 1 0. But. one might want M to be a nonlexical category. see • • • . To illustrate the range of possibi1ities that may be relevant. § 2 . may. This is by no means merely a terminological issue. M should be a lexical category. cf. though even here the situation is unclear. 1 960. Thus. with a small amount of supporting evidence.2. but on l y wi th statistics of usage. and I see no point in considering this possibility any further as long as proponents of this most implausible view make no attempt to deal with these objections. A must b e a category of a special type." "dominate. By this criterion. For some discussion o f a possible syntactic basis for such sub categorization. The gen eral rule for conjunction seems to be roughly this: if XZY and XZ'Y are two strings such that for some category A. the well-known contrast of may-may." etc. Gleitman. and throughout the informal discussion here I shall simply avoid these questions. condition is violated. can. Articles) are unstressed. Chapter 1.g. 1 96 1 ). By this criterion. I mention just two considerations. alternatively. This raises no difficulty of principle. the categories in question seem to be the major categories. but it seems to me reasonable to impose this condition of uniqueness on the system of base rules.212 NOTES TO CHAPTER 2 8. we might hope to estab· lish general conventions involving the distinction between lexical and n o n le x ical categories. and Lukoff. we come much closer to charac terizing the actual range of possibilities if we limit A to major categories. § 5. in the sense just de scribed. These rules assign stress in a fixed way in strings belonging to certain categories. for example. One might question whether M should be regarded as a lexical category. where Z""'and""'Z' is an A (see Chomsky 1 957. + grave]. what Harris. Thus [s] is an abbreviation for th e set of features [+ consonantal. . but differs from them in certain respects. . but that they can perhaps be met by narrowing the scope of the proposals to just what is being discussed here. Ill . 14. Yngve. See. The method that I shall elaborate later incorporates various features of these pro posals. Schachter. + voiced. The problem of remedying this defect in phrase structure grammar is clearly very much open. customary on the phono logical level. Chapter 4 ) . I think that Katz's major criticisms are correct. in this connection. + nasal. + continuant. Matthews qevised a techni q ue of indexing category symbols to meet the difficulties that he found. Rule ( 1 8) applies to any segment specified as [+ continu ant] (hence to [s]) in a context which is specified as . + voiced. These devices be come unnecessary once grammatical transformations are intro duced. 1 962). summarized in part in Chomsky (196 1 ) and Miller and Chomsky ( 1 963). the discussion in Postal ( 1 964a). . 1 964). Bach has also dealt with this question. 1 962. and P. + continuant. if we index each specified feature by an i n teger indicating the column it occupies in the matrix. S tockwell. Anderson. and they have suggested a somewhat different way of handling them (see Stockwell and Schachter.m]).[+ voiced] (hence to the context [ . namely the question of subcategorization of lexical categories within the framework of an independently justified generative grammar. indices on category symbols were used to express agreement (and. converting the segment to which it applies to a voiced segment with. T.vocalic. there was no attempt to deal with it in most of the published work of the last several years. + strident. B u t notice th a t a phonological matrix can b e regarded simply as a set of specified phonological features. in a somewhat different way (B ach. E. of enclosing sets of features by square brackets. In the syntactic component of this (pretransformational) grammar. . the same features as before (hence converting Cs] to [z] = [+ consonan tal. I shall henceforth use the convention. Schachter. calls long components) but not subcategorization and selectional restrictions. .grave] and [m] for the set of features [+ consonantal. and he later incorporated this technique as one of the main devices of the COMIT programming system that he developed with the collaboration of V. Similar difficulties were noted independently by R.grave] ) .NOTES TO CHAPTER � Chomsky ( 1 955. 1 3. otherwise. A critique of these and other discussions is given in Katz ( 1 964a). Thus the two·column matrix representing the formative .vocalic. . in general. + strident.vocalic.voiced. 1 95 1 . Although this defect was pointed out quite early. and deserves much further study. 1 5 . The same would be true of "drop" in "the price dropped. 1 89. some syntactic. meaning. The same point 17. regard a lexical entry as a matrix If we regard a lexical entry as a set of features.continuantl• .) We might. some semantic . then items that are similar in sound. • . " "he dropped that silly pre tense". it raises many problems of fact and principle to which I have no answer." Cf.consonantal2' + vocali�. 1 1 9. be related to one another in the lexicon. 3 1 ). A lexical entry can now be regarded simply as a set of features. but it does not seem that the further structure they impose plays any role in their theory. who holds that "grammar deals with the general facts of language. Of course. 2 74)." More generally. . However. • . take a lexical entry to be simply a set of features.grave2• . facts. we shall not follow this course but shall. or syntactic function will not complex symbol pair. p. a lexical entry must also contain a definition. and I therefore continue the exposition without develop 1 6." "he dropped the ball. some syntactic. . the phonological redundancy rules. p. the In transitive "grow" of "the boy grew" or "corn grows. 1 968) that this too consists simply of a set of features . . Alternatively. if the traditional view that syntactic categorization is in pan de- . as can be done in the case of "the window broke. such relationships can be expressed by taking a lexical entry to be a Boolean function of features. lexicology with the special ing it." and the Transitive "grow" of "he grows corn" would hav e to constitute two separate lexical entries. Although it is likely that such a modification of the theory of lexical structure is necessary. which de termine such features as voicing of vowels or unrounding of high front vowels in English. largely for ease of exposition. For example. despite the meaning relation that holds between them. can be supplemented by analogous syn tactic and semantic redundancy rules. rather. For example." "someone broke the window. Recall Bloomfield's characterization of a lexicon as the list of basic irregularities of a language ( 1 9 83. p. and it can be plausibly argued (see Katz and Fodor. some phonological. as in the text.vocalicl' . (Actually the Katz-Fodor definitions are not simply sets. and in innumerable other cases of many different kinds. since there is apparently no way to derive the Intransitive structures from the Transitive ones. is made by Sweet ( 1 9 1 3.2 14 NOTES TO CHAPTER 2 bee can be regarded as consisting of the features [+ consonantall• . redundancy rules may relate features of these various types. in a complete gram mar. some phonological. ] . or of "command" in the example discussed on p. then. Furthermore. Notice. Let us call such a transformation strictly local. one might consider a new extension of the theory of context-free grammar. By a local transformation (with respect to A) I mean one that affects only a substring dominated by the single category symbol A.NOTES TO CHAPTER 2 termined semantically can be substantiated in any serious way. Thus Adverbial Phrases consisting of Preposition""'Determiner""'Noun are in general restricted as to the choice of these elements. incidentally. and these restrictions could be stated by local transformations to the effect that Preposition and Noun can be rewritten in certain restricted ways when dominated by such category symbols as Place Adverbial and Time Adverbial. Thus all rules of the transformational cycle in phonology are local. nominalization transformations that give such forms as "I persuaded John of my seriousness" from an underlying form "I persuaded John of N S. The only motivation for this highly special definition is that many of the examples of local transformations that come to mind meet this restrictive condition as well (for example.e. all rules that establish a partial hierarchy among syntactic features) might be regarded as redundancy rules rather than as rules of the base. which is immediately dominated by the category symbol NP with respect to which the transformation is local). it can be expressed by a redundancy rule determining syntatic features in terms of semantic ones. 18. There is some reason to suspect that it might be appropriate to intersperse certain local transformations among the rewriting rules of the base. quite generally. furthermore. In fact. in fact.. Such a decision would have various consequences. in this case.3. J 9 . being some type of Adverbial). that introduces a string into a position dominated by the lexical category B which is immediately dominated by A . in terms of the dominating category symbol). in this sense . The example of the preceding paragraph involves a transfor mation that is local with respect to a category A (A. that the rules (lW) (and. to which we shall return in § 4. alongside of the fairly widely studied extension of con text-free grammar to con text-sensitive grammars that permit rules that restrict rewriting in terms of contiguous symbols. We shall return to the question of redundancy rules in § 6. and. permitting rules that restrict rewriting by local transformations (i." where S dominates the string under lying "I am serious" and the transformation substitutes a trans form of this string for the dummy symbol occupying the position of the lexical category N. Notice that an importarit question is begged when we assume that Noun subcategorization is independent of context and that the . " "I eat the apple. understand." alongside of "I know the answer"). To maintain the suggested generalization concerning strict sub categorization. must) cooccur with this Auxiliary element.).2. as "generic. (cf. 14. 20. An apparent exception to the last remark is the subcategorization of Verbs in terms of choice of the Progressive form be + Ing. as we have defined "syntactic feature" (cf. lexical items can be inserted in the wrong position by the lexical rule. and the characteristic Adverbials of Progressive do occur with the Verbs own. But. Actually. this is not the case. not introduced by such general rules as (20)-(U ) but simply listed in the lexical entries. " .like""'Predicate Nominal" is a single sy mbol. "I know the answer right now. it is only the features involved in the set of rules of which (2o)-(n) constitute a sample that de termine selectional classification. understand. It will indicate the position of a transform of a sentence. This decision. We shall return to this question in § 3 . as has been pointed out to me by Barbara Hall. know. u. Observe that in (36) such an expression as " . as several of the others." are ruled out (except. The status of the symbol S' in this rule is unexplained at the present state of the exposition. pp. and know occur freely with or without Progressive (along with all other Verbs). 13. Idiosyncratic syntactic features of particular lexical items. Thus each element of the Auxiliary has associated with it certain characteristic Adverbials that may (or. Actually. We shall return to this matter in § 4. etc. play no role in Verb subclassification. we should have to claim that such Verbs as own. in fact. but that the Progressive form is deleted by an obligatory transformation when it precedes these Verbs (this peculiarity would be marked by a feature that consti tutes part of the lexical entries for these forms).216 NOTES TO CHAPTER 11 selectional restrictIons on Subject-Verb-Object are given com pletely by rules determining the subcategorization of Verbs in terms of previously chosen Noun subcategories. The careful reader will notice that as these rules are formulated. avoiding it now only so as not to overburden the exposition. u." which can. will be slightly modified later in the text. 821. although such forms as I eat the apple right now. be treated as involving deletion of a "dummy" Adverbial). as the theory of the syntactic component is extended later on. standing for a particular syntactic feature. in fact. in the latter case. Strictly speaking. in the case of Present tense. a more careful analysis would revise (40) and (4 1 ) in detail. there is good reason to assume this. . Notice that Adjective modifiers can derive from either restrictive or nonrestrictive relatives (consider." analogous structurally to "John stayed in England" with a Verbal Complement introduced by rule (52 iii). the ambiguity of the sentence "the industrious Chinese dominate the economy of Southeast Asia"). others sug gest that a redundancy rule may be needed.. Cf. an ambiguous representative of both constructions. But although restrictive relatives belong to the Determiner system. This matter is discussed in the Port-Royal Logic (Arnauld et al. of the three mentioned only the one we are using permits also the free use of variables. note 4) gives a form of transformational grammar that is richer in some respects and poorer in others than the formulation in terms of Boolean conditions on Analyz· ability. that the relation of the Verb to the Place Adverbial in "John died in England" ( = "in England. or roughly as "in England. rather. for example. for example. On the other hand. Notice. John really lived" or "in England. John died") is very different from that in "John stayed in England" ("John lived in England" is. in the lexicon. Cbom sky (in press) for some discussion. 1 7 . "I found John likable. in fact. Chapter 8). Some such expressions may be derived from Proper Nouns with nonrestrictive relatives by transformation. See the dis cussion of the problem of justification at the outset of § 2. in restricted ways (e.g. 1 .3. "this cannot be the England that 1 know and love. 1 662). Use of complex symbols at arbi trary nodes (as in Harman. there are several rea sons for supposing that nonrestrictive relatives are. Adjective modifiers derived from nonrestrictive rela tives . 2 Notice that these alternatives are not strictly equivalent. in more recent times. 16. as signing certain of the features of Common Nouns to Proper Nouns. Notice also that Proper Nouns can also be used as Common Nouns. by Jesper sen ( 1 924..g. as in most current work on generative grammar. Thus. this is not to deny that an interpretation can some times be imposed on such phrases as those of (54). which surprised me very much"). 1 963 . of a full sentence e." "I once read a novel by a different John Smith"). in particular. "clever Hans" or "old Tom").e. and. the use of labeled brackets is appropriate for the formulation of the transformational rules of the phono· logical component.cf. and the references of footnote 1 1 . as in the case of schema (44)... being interpretable as either "John re sided in England. marginally.g. Once again. Com plements of the full NP (and in some cases.NOTES TO CHAPTER 15. Proper Nouns of course can have nonrestrictive relatives (and. as conflicting with this generalization.2 18 NOTES TO CHAPTER 2 John remained alive. "a good time was had by all" or "recourse was had to a new plan"). as does the distinction between Adverbials that qualify the Verb and those which might more properly be said to qualify the Subject. has so little importance for the study of the grammatical structure of the language in question." Cf. stayed away yesterday" than like that of the Ad verbial modifier of the Verb in "John laid his plans cleverly.cf." See Austin ( 1 956) for some discussion of such cases. if this turns out to be the case." 28." which means. In either the case of past tense or that of passivi zation. consider the Adverbial of "John married Mary with no great enthusiasm. p." with a Place Adverbial that is a Verb Phrase Complement introduced by (5l!ii) . roughly. 26). 1 04 for further d iscu ssi on of such "pseudopassives. in the lexicon. the generalization is invalidated (in the sense of "internal justification" .g. p. (As an example of the latter. like many other Adverbials. in a system of the complexity of a natural language) is generally so unrewarding and. "John will surely die on the Continent. 1 960a. of course. unless. § 4) only jf a more highly valued grammar can be constructed that does not contain it. There are well-known marginal exceptions to this remark (e.in fact. "John was not very enthusiastic about marrying Mary. Chapter I. It must be borne in mind that the general rules of a grammar are not invalidated by the existence of exceptions. and it is also clear that the locution "take Manner Ad verbials freely" requires considerable further analysis and clarifi cation (see Lees. " and therefore seems to play a role more like that of the Adverbial modifier of the Subject in "John. in itself. it leads to the discovery of deeper generalizations.cf. cleverly. It is also worth noting that many of the Manner Adverbials. but he may live in England").) Nevertheless. this remark is true only when "live in" has the sense of "reside in" or "inhabit. It is for this reason that the discovery of peculiarities and exceptions (which are rarely lacking. nor is me generalization that relates Manner Adverbials to passivization invalidated by the fact that certain items must be listed. This differ ence of structure between "live in England" and "die in England" accounts for the fact (noted by Ralph Long) that "England is lived in by many people" is much more natural than "England is died in by many people" . the essential correctness of the comments in the text does not seem to me in doubt. are Sentence transforms with deleted .. Thus one does not eliminate the rule for forming dle past tense of Verbs from the grammar on the grounds mat many Verbs are irregular. as is usual (cf. Adverbials are a rich and as yet relatively unexplored system. . .for example. but also the feature [. There may be some point to allowin g a lexical item to appear in several categorial positions (either by specifying it positively with respect to several lexical ca tegories." in the position s of the Noun and Verb. the feature [-] designates both Proper Nouns and I ntran si tive Verbs. for example. .NOTES TO CHAPTER 2 219 119. we may drop this condition and extend the first convention so that the complex symbol introduced in the a nalysis of a lexical category A contains not only the fea ture [+ A]. since it might be non-Proper as a Noun but Transitive as a Verb. be regarded as Sentence transforms (so that. Subjects." "desire. V that S . Thus. or perhaps always. Thus underlying the sentence "John gave the lecture with great enthusiasm. for example. . . Chapter 11 and Chapter 4. Place Adverbials (at least those which are VP complements) must sometimes. It would have the advantage of overcoming a defect in our notation for features introduced by context-sensitive subcategorization rules. "I read the book in England" derives from an underlying structure very much like the one that underlies "I read the book while (I was) in England").) This might lead to difficulty if a certain lexical item were both a Noun and a Verb. to these categories) . in the case of such words as "proof. and therefore anythi ng we say about them must be regarded as quite tentative. Alternatively. or Transitive as a Verb and Proper as a Noun. (This is why the feature [+ N] had to be mentioned in rul e (57iv) . . .B] for any lexical category B other than A . in the grammar (55). but are permitted to enter either the Noun or Verb position. § 11 . N that S . with the re peated NP "John" deleted. A l tern a ti vel y. in the case of " . . If the proposal of this no te is adopted. 11 ) . as is necessary. Hence it will not be necessary to derive the former by transformation from the latter. Then the lexical insertion rule will place them in either the frame " . . respectively. ." Suppose that these are specified as taking Sentential Complements of various forms. proving that S . " is the base s tring "John has great en thusiasm" (note that "with" is quite generally a transform of "have"). " or the frame " ." with the Adverbial "with great enthusi asm." "belief. it will be necessary to designate such features by a more complex notation i ndicating not only the frame in question but also the symbol that dominates it. the problem cannot arise . . . and it raises unanswered questions about the structure of the lexicon. This convention entails that a word specified a s belonging to two lexical categories must have two separate lexical entries. Similarly. ". or by leaving it totally un specified with respect . . . 220 NOTES T O CHAPTER 2 80. 81. 81. Under such an analysis, "John's proof that S" would derive from the structure underlying "John has a proof that S" by the sequence of transformations that derives "john's book" from the structure underlying "John has a book." One might go on to relate "John has a proof that S" to "John proves that S" (perhaps, ultimately, as "John takes a walk" is related to "John walks"), but this is another matter. In connection with this discussion, it is also necessary to estab· lish a general distinctness condition regarding the idiosyncratic, purely lexical features (e.g., the feature [Object-deletion] in (58), (59» . For discussion of this question, which becomes critical in case these features relate to the phonological component, see Halle and Chomsky (forthcoming). It has been maintained that these relations can be defined in terms of some notion of cooccurrence, but this seems to me duo bious, for reasons presented in various places (e.g., in Bar·Hillel, 1 954; and Chomsky, 1 964). Observe that the definitions of gram matical relation or grammatical function that have been suggested here refer only to the base of the syntax and not to surface structures of actual sentences in other than the simplest cases. The significant grammatical relations of an actual sentence (e.g. (7), p. 70), are those which are defined in the basis (deep structure) of this sentence. I give these informally, instead of using the notation developed earlier. to simplify the reading. There is nothing essential in· volved in this change of notation. For example. if we were to adapt the definitions of universal categories and functions so that they apply to such sentences as "in England is where I met him," which are often cited to show that phrases other than NP's can occur as Subjects, these proposals would fail completely. This sentence, however, is obviously trans formationally derived. It would be perfectly correct to say that "in England" is the Subject of "in England is where I met him," extending the grammatical relation Subject,of. that is, [NP, S], to the derived Phrase·marker (the surface structure). In the basis, however. "in England" is an Adverbial of Place, associated with the VP meet him in the Predicate·Phrase "met him in England," and the sentence is interpreted in accordance with the gram· matical relations defined in this underlying deep structure. This extension to surface structures of such functional notions as Subject-of is not an entirely straightforward matter. Thus in base structures, there is apparently never more than a single occurrence of a category such as NP in any structure immediately dominated by a single category (cf. note 7). and our definitions NOTES TO CHAPTER 2 22 1 83. of these notions relied on this fact. But this is not true of surface structures. In the sentence "this book I really enjoyed," both "this book" and "I" are NP's immediately dominated by S. Ap parently, then, order is significant in determining the grammatical relations defined by surface structures (not surprisingly), though it seems to play no role in the determination of grammatical rela tions in deep structures. Consequently, somewhat different defi nitions are needed for the surface notions. It might be suggested that Topic-Comment is the basic gram matical relation of surface structure corresponding (roughly) to the fundamental Subject-Predicate relation of deep structure. Thus we might define the Topic-of the Sentence as the leftmost NP immediately dominated by S in the surface structure, and the Comment-of the Sentence as the rest of the string. Often, of course, Topic and Subject will coincide, but not in the examples discussed. This proposal, which seems plausible, was suggested to me by Paul Kiparsky. One might refine it in various ways, for example, by defining the Topic-of the Sentence as the leftmost NP that is immediately dominated by S in the surface structure and that is, furthermore, a major category (cf. p. 74 this will make John the Topic in the cleft sentence "it was John who I saw"). Other elaborations also come to mind, but I shall not go into the question any more fully here. This very fruitful and important insight is as old as syntactic theory itself; it is developed quite clearly in the Grammaire generale et raisonnee of Port-Royal (cf. Chomsky, 1 964, § 1 .0; forth coming, for discussion). What is, in essence, the same idea was rein troduced into modern linguistics by Harris, though he has not dis cussed it in quite these terms (cf. Harris, 1 952, 1 9 5 4, 1 957). For further discussion of this notion, within the framework of trans formational generative grammar, see Chomsky ( 1 957), and for steps toward a substantive theory of semantic interpretation based on this assumption, see Katz and Fodor (1 963) and Katz and Postal - 34. ( 1964). Curry's proposals are so sketchy that it is impossible to extract from them more than a general point of view. The position of Saumjan and Soboleva is much more explicitly worked out, but it is defective in crucial respects. Cf. Hall (1 965), for an analysis of this approach. It is possible that "stratificational grammar" also adopts a similar position, but the published references to this theory (e.g., Gleason, 1964) are much too vague for any conclusion to be drawn. Notice, for example, that Case is usually determined by the posi tion of the Noun in surface structure rather than in deep struc- 35. NOTES TO CHAPTER 2 ture, although the surface structures given by stylistic inversions no t affe ct Case. Even in English, poor as it is in in8.ection, this can be observe d For example, the Pronoun in the sentences "he was struck by a bullet," "he is easy to please," he frightens easily" is, in each case, the "logical Object," that is, the Direct-Object of Verbs strike, please, frighten, respectively, in the underlying deep structures. Nevertheless, the form is he ra ther than him. But stylistic inversion of the type we have just been discussing gives such forms as "him I re ally like," "him I would definitely try not to antagonize." Where in8.ections are richer, this phenome non , which illustrates the peripheral character of these processes of inversion, is much more apparent. The relation between inflection, ambigui ty and word order was discussed at some le ngth in traditional linguistic theory. See Chomsky, forthcoming, for some references. . " , do NO TES TO CHAPTER J I . Some details irre le va n t to th e problem under discussion are omit ted in these examples. We here regard each lexical i tem as stand ing for a complex of features, namely those that constitute its lexical entry in addition to those entered by redundancy rules. The use of the dummy symbol a has been extended here to the case of various unsp ecified elements that wi ll be del eted by oblig 2. atory transformations. There is, in fact, good reason to require that only "recoverable deletions" be permitted in the grammar. For discussion of this very important question, see Chomsky, 1 964, § 2.2. We shall return to it at the end of this chapter and in Chapter 4, § 2.2. The formative nom in (!I) is one of several that might be as signed to the Tense....... Modal position of the Auxiliary, and that determine the form of the Nominalization (for-to, possessive-;ng, etc.). The details of this, both for Transformation-markers and Phrase markers, are worked out in Chomsky ( 1 955), within the following general framework. Linguistic theory provides a (universal) system of levels of representation. Each leve l L is a system based on a set of primes (minimal elements - i.e., an alphabet); the operation of concatenation, which forms strings of primes of arbitrary finite length (the terms and notions all being borrowed from the theory of concatenation algebras - cf. e_g., Rosenbloom, 1 950); various re lations; a designated class of s trings (or sets of s trings) of primes called L-markers; a mapping of L-markers onto V-markers, where L' is the next "lower" level (thus levels are arranged in a hier archy) . In particular, on the level P of phrase structure and the NOTES TO CHAPTER 3 223 level T o f transformations w e have P-markers and T-markers in the sense just described informally. A hierarchy of linguistic levels (phonetic, phonological, word, morphological, phrase structure, transformational structure) can be developed within a uniform framework in this way. For details, see Chomsky (1 955). For a discussion of T-markers, see Katz and Postal ( 1 964). 3. For discussion of negation, see Klima ( 1 964), Katz (1 964b). The formation of questions and imperatives and the semantic inter pretation of the question and imperative markers are discussed in Katz and Postal ( 1 964). In Rockett ( 1 96 1 ) the proposal is made that the passive transformation be conditional on a marker in the underlying form, but no supporting argument is given for what, in the context of that paper, is no more than a notational innovation. Notice that the reformulation of the passive transformation as obligatory, relative to choice of an optional marker in the under lying string, is independent of the principle that we have just cited, since the passive marker, as distinct from the question, nega tion, and imperative markers, has no independent semantic inter pretation. Furthermore, we have noted in § 4.4 of Chapter I that there are good reasons to distinguish such transformations as pas sive from purely stylistic i nversion operations. These observations suggest that we attempt to formulate a more general condition of which the principle just cited is itself a consequence, namely that "nonstylistic transformations" are all signaled by optional markers drawn from a fixed, universal, language-independent set. This attempt presupposes a deeper analysis of the notion "nonstylistic transformation" than we have been able to provide here, however. 4. For illuminating discussion of this question, and several others that we are considering here, see Fillmore (1 963) and Fraser (1 963). 5. Both of these observations are due to Fillmore ( 1 963). 6. In connection with ordering of rules, it is necessary to distinguish extrinsic order, imposed by the explicit ordering of rules, from intrinsic order, which is simply a consequence of how rules are formulated. Thus if the rule R I introduces the symbol A and R2 analyzes A, there is an "intrinsic order relating R I and R2, but not necessarily any extrinsic order. Similarly, if a certain transforma tion TI applies to a certain structure that is formed only by application of T2, there is an intrinsic order T1, T2• Taxonomic linguistics disallows extrinsic ordering, but has not been clear about the status of intrinsic ordering. Generative grammars have ordinarily required both. For some discussion of this matter, see Chomsky ( 1964). 2 24 7. NOTES TO CHAPTER 3 We are discussing only embedding transformations here, but should extend the discussion to various generalized transformations that form coordinate constructions (e.g., conjunction). There are certain problems concerning these, but I believe that they can be incorporated quite readily in the present scheme by permitting rule schemata (in the sense of Chomsky and Miller, 1 963, p. 298; Chomsky and Schiitzenberger, 1 963, p. 1 33) introducing coordi· nated elements that are then modified, rearranged, and appropri. ately interrelated by singulary transformations. If the suggestion of note 9, Chapter .11 , is workable, then such rule schemata need not be stated in the grammar at all. Rather, by a general con· vention we can associate such a schema with each major category. This approach to coordination relies heavily on the filtering effect of transformations, discussed later. Thus wherever we have co· ordination, some category is coordinated n times in the matrix sentence, and n occurrences of matched sentences are independ. ently generated by the base rules. 8. Notice, incidentally, that we can now eliminate Complement from the set of category symbols. We could go on, at this point, to define "Complement" as a functional notion (to be more precise, as a cover term for several functional notions), in the manner of pp. 70-7 1 • 9. As it stands, this claim seems to me somewhat too strong, though it is true in one important sense of semantic interpretation. For example, it seems clear that the order of "quantifiers" in surface structures sometimes plays a role in semantic interpretation. Thus for many speakers - in particular, for me - the sentences "every. one in the room knows at least two languages" and "at least two languages are known by everyone in the room" are not synony mous. Still, we might maintain that in such examples both in terpretations are latent (as would be indicated by the identity of the deep structures of the two sentences in all respects relevant to semantic interpretation), and that the reason for the opposing interpretations is an extraneous factor - an overriding considera tion involving order of quantifiers in surface structures - that filters out certain latent interpretations provided by the deep structures. In support of this view, it may be pointed out that other sentences that derive from these (e.g., "there are two lan guages that everyone in the room knows") may switch interpre tations, indicating that these interpretations must have been latent all along. There are other examples tha t suggest something similar. For example, Grice has suggested that the temporal order im· plied in conjunction may be regarded as a feature of discourse rather than as part of the meaning of "and," and Jakobson has pp.B) constitutes a branch of K. incidentally.g. that is. what we are suggesting is this. l :E. This is important. note 7) also provide infinite generative capacity. For some references to remarks in the Port-Royal Logic on the effect of grammatical transformations on meaning. The other function of the transformational component is to express restrictions on distribution for lexical items and for sen tence s tructures. that this identity condition need never be stated in the grammar.C) is a branch. Suppose now that . Notice that the schemata that underlie coordination (cf. then (A. and that Al = A".0. the only way to form new deep structures is to insert elementary "propositions" technically. Cf. if this occurrence of B immediately dominates ZCW (where C is a symbol). 10. etc. At = S. 1 960a). . in fact. Also relevant in this connection is the notion of Topic-Comment mentioned in note 32. u.in other Phrase-markers. Notice. 1 1 .) is a branch of the generalized Phrase-marker K (A I ' formed by base rules.. 1 1 7-1 1 8. . n. For some discussion of Adverbial sequences.. in all cases in which identity conditions appear in transformations. see Matthews (1961). In other words. A. etc. the condition is not really identity of strings but rather total iden ti ty of structures.NOTES TO CHAPTER 3 225 also discussed "iconic" features o f discourse involving relations between temporal order in surface structure and order of im portance. so that (B C) is a branch.B. . § 1 . This is by no means a logically necessary feature of phrase structure grammars. and forthcoming). . see Chomsky (forth coming).. But to define identity of structures in terms of Analyzability it is necessary to use quanti fiers. . base Phrase-markers ." This formulation leaves unexplained some rather marginal phenomena (e. Then it must be that for some i. since (as was pointed out by Lees. . but here too the true recursive property can apparently be limited to the schema S -+ S#S# I S. since it is a general condition on the func tioning of grammars.g. Furthermore. Suppose that the symbol A immediately dominates XBY (where B is a symbol) in the Phrase-marker K. Then (A . very.. the source of such expressions as "very. Chapter 2 . For some discussion. A -+ XBY was one of the cate gorial rules used in generating this Phrase-marker. . the status of which in general is unclear). very Adjective" and some more significant ones (e.. . i :E. 1 3 . Formally speaking. this may be the only case in which quantifiers must . see Chomsky (1964. hence to rules introducing "propositions. • • • . the possibility of iterating Adverbials and various kinds of parenthetic elements. Schlesinger ( 1 964) : Miller and Isard ( 1 964): and the rbume in Chapter 1 . 1 7. See § 2. . 1 8. ex pressed in Chomsky (1957) and elsewhere. and Postal (see bibliography. appear in the structural analyses that define transformations. there is. the only serious work tha t I know of on the relation of th ese domains is that of Katz.3. Ex tracting the identity condition from grammars. This is often true of what we are here calling erasure operations. thus greatly restricting the power of the theory of transformational gram mar. A serious discussion of this question. that is. 1 957. no way to show that semantic considerations play a role in the choice of the syntactic or phonological component of a gram mar or that semantic features (in any significant sense of this term) play a role in the functioning of the syn tacti c or phonological rules. The exten t to which the complexity of the theory of derived constituent structure depends on the presence of per mutations is quite clear. this term is deleted except for the feature [± Human] . at pres ent. Thus no serious proposal has been advanced to show how semantic considerations can contribute to an evaluation procedure for such systems or provide some of the primary linguistic data on the basis of which they are selected.NOTES TO CHAPTER 3 14. as well as the question of dependency of syntax on semantics. 1 of Chapter 2. and Chap ter 4. and § 1 of Chapter 4. I see no reason to modify the view. see Chomsky. 1 6. § 6. A natural notational decision would be to restrict the integers one and two to first and second person. Notice that in this case the third term of the proper analysis is not strictly deleted. for some additional related discussion. awaits a development of the theory of universal semantics. which. 1 5. that although. obviou sly. we are therefore able to formulate the structural analyses that define transforma tions strictly as Boolean conditions on Analyzability. respectively. Rather. Chapter 8). and many other publica tions). for discussion of other claims that have been made. obviously the theory of syntax should be designed so that the syntactic structures exhibited for particular languages will support semantic interpretation). Fodor. Some of the details of this modification are worked out in Fraser (forthcoming). § 1 . semantic considerations are relevant to the cons truc tion of general linguistic theory (that is. See Ch apter 1 . an account of the nature of semantic representa tion. for example. For the moment. which then assumes its phonological shape (giving who. § 2 . Although various positions about these q uestio ns have been stated with great confidence and authority. from the analysis of these notions in Chomsky (1 955. For discussion see Miller and Chomsky ( 1 963). or that) by later rules. they do not illustrate the use of orderl. § 4 . But not all of the rules referring to intrinsic syntactic features of Nouns are selectional rules. . . 1 963. The question as to whether the grammar should generate deviant sentences is purely terminological. . . 1 96 1 . but these seem invariably to have other affixes as variants of ing ( bot heTsome for botheTing. see Chomsky ( 196 1) and many other references. . having to do only with termino logical decisions. we shall assume the latter. 4. for con venience of exposition. as illustrated earlier. For discussion.NOTES TO CHAPTER 4 2 27 NO TES TO CHAP TER 4 Whether the rule is a rewriting rule or a substitution transforma tion . are rules that insert Verbs and Adjectives into generalized Phrase-markers on the basis of the intrinsic syntactic features of the Nouns that appear in various po si tio n s . To avoid what has been a persistent misunderstanding. . 1 955.cf. with no implication that deviant sentences are be i ng "l egi sla ted aga i ns t " as "without a function" or "illegitimate. Chapter 2. with their structural descriptions. This is a substantive issue." Quite the contrary is true. impressive for impressing. but many of the other questions that have been debated concerning these notions seem to me quite empty. For one thing. These examples do not begin to exhaust the range of possibilities that must be considered in a full study of interpretation of deviant sentences. 1 964a). The grammar deTivatively geneTates all other strings (such as ( 1 ) and (2». The principles that deter mine how interpretations can be imposed on deviant sentences may be universal (as suggested in Chomsky.). etc. scary for sca ring. [+ Animate]] do not have Adjectival forms with ing.does not concern us here. Many of the Verbs of th e category [+[ + Abstract] . in particular. it must be emphasized again that "grammaticalness" is being used here as a technical term. and again here) or specific to a given language (as suggested in Katz. in discussions of generative grammar. as has repeatedly been stressed and illustrated. having to do with nothing more than the technical sense of "generate. Recall that selectional rules.3 . the rules violated in the formation of (4) involve such features but are not selectional rules. 3 . Miller and Chomsky. with their structural descrip tions.." A descriptively adequate grammar mUSt ass i gn to each s tring a s tructural descri p tion that indicates the manner of its deviation from strict well-formedness (if any). These structural descriptions will indicate the manner and degree of deviance of the derivatively generated sentences. 5. 2. A natural terminological decision would be to say that the grammar diTectly geneTates the language consisting of just the sentences that do not deviate at all (such as (3». Notice. in effect. in fact. For for derivation 1 0. in the case of "become" this analysis is well motivated.g.4. of certain of these forms.. . The discussion of deviation from grammaticalness that has been carried on here o ffe rs no i nsight into this phenomenon. assuming the co nventi on e = [e. "John seems sad" from an u n derl ying structure con taining the base stri n g "John is sad. Notice that features are unordered in a com plex symbol. consider the following line: "Me up at does / o u t of the Hoor/quietly Stare/a poisoned mouse/still who al i v e / i s asking What/have i done that/You wouldn't have" (E. ]. S. One might therefore try to show that part of th e semantic effect of is to can cel semantic conflicts of certain sorts. this i s surely correct (e. where e 7 . and it would surely be quite beside the point to try to assign it a degree of deviation in terms of the number or kind of rules of the grammar that are violated in generating it. here to present an absolutely precise account or to give these definitions in their s imp l est and most general forms. 6. some notion of "generic" seems to be involved critically (I owe this observation to Barbara Hall) . E. is the null element.similarly. which is resolved only by the convention that we now state. As elsewhere in this discussion. Notice that the formulation given previously left an ambiguity in the latter case. We are. not arise if we were to give a some what different analysis of post-Yerbal Adjectives in E nglish . Y is null if [�] is null. I am i ndeb ted to Thomas Bever and Peter Rosenbaum for many interesting and suggestive commen ts rela t i ng to this q ues t ion. Cummings). that the dee p structure of each of the sentences of ( 1 5) will con tain a s trin g with sincerity as the Direct-Object of the Main Ye rb frigh ten (and with an unspecified Subject). This difficulty would.228 NOTES T O CHAPTER 4 inversion as a s tyl is t ic device (d. It is worth no ti ng that a condition like that impose d on W and V in the discussion of the schema (9) is probably necessary in the theory of transformations. and i t may be correct to extend it to many or all such cases. In many or all such cases. I make no attempt 9. in particular. ." and then "J ohn seems sad" by further transformations . In some cases. some other proposal s Z i erer ( 1 964). This poses not the slightest difficul ty or ambiguity of inte rpreta tion. because it can pro vide a basis for excl udi ng "become" from passivization). " which becomes "Joh n seems to be sad. deriv i ng them from underlying strings with Sentence-Complements to the Yerbs. Thus X is null if [a] is nul l . § 4. . For example. see . al though this p ro ble m has never been discussed ex p l i ci tly. " ge n eri c " incidentally. Ch ap ter 2. for some discus s ion). 1 1. " "he struck an outlandish pose. still it is by no means true that a "cooccurrence relation" of the sort that has been discussed by Harris (1 957). although the sentences "I regard John as pompous. are freely subject to passivization. 24-25) ---. Thus pompous can be replaced quite naturally by a friend in "I regard John as . such sentences as "John strikes me as pompous. It is dear. See Ullmann ( 1959) for discussion of much related descriptive work." etc. The rules involving contextual features.NOTES u. then. note 1 5. If this analysis can be justified on syntactic grounds. may be partially independent of semantic properties. and others holds be tween them. In connection with ( 1 9iii). Katz). For example." and so on. then the deep structures will be somewhat more appropriate for the semantic interpretation than assumed in the text." "his remarks impress me as unintelligible" do not passivize. One alternative would be to take the underlying structure to be it"'"'S . in discussing (1 9i) that the Subject-of strikes in the deep structure is Joh n. such as Luria and Vinogradova ( 1 959). but it should be noted that this is not at all obvious. Such examples must be borne in mind if any attempt is made to give some substance to the widely voiced (but." The lexical item strike of ( 1 9i) would then have very different strict subcate gorization features from the phonetically identical item of "it struck me blind. in other words." An obligatory transformation would gi ve the structure unde rlying i t strikes me that John is pompous." "it struck me blind.strikes me. for representative statements of his. pp. Chapter 2). and much current work in "componential analysis. inversion of the Subject-Verb-Object relations) does not determine a corresponding similarity of distributional restrictions. totally empty) claim that semantic considerations somehow determine syntactic structure or distributional properties. there are other relevant syntactic differences between the paired examples of ( I gi). TO CHAPTER 4 2 29 I ll . (cf. where it"'"'S is an NP and S dominates the structure underlying "John is pompous. for the moment." Although the sentences of ( 1 9i) are near-paraphrases." but hardly in "John strikes me as " (I owe this observation to J." and a further optional trans formation would give "John strikes me as pompous. I have been assuming. see Chomsky ( 1 964). " . Interest in these questions can be traced to Humboldt (1 836) . As several people have observed." while both would differ in strict subcategoriza tion from strike in "he struck me. that the dose meaning relation between regard and strike (involv i ng in particular. Harris has suggested ( 1 952. Also relevant are some psychological studies that have attempted to place a linguistic item in a contex t of somehow related items. Hif ( 1 96 1 ). in these cases. Notice . justice.for example. [± Abstract]}. See note 110. with further structure defined on them in the manner suggested informally in this discussion. 1 6. As pointed out in note 1 5. With respect to se1ectional features. particularly in the case of {[. a distinctive-feature matrix is simply a way of representing a set of abstract phonological features. But there is a subdivision of Inanimate Count Nouns that seems to correspond to this. Chapter 2 .cf. in fact. The latter would be interpretable only as a deviant sentence.that is. Such examples as "sincerity frightens" can be found. dirt. Notice that these or any anal ogous conventions make a distinction amounting to the marked! unmarked distinction that has often been discussed.Count]. but such proliferation of terminology contributes nothing toward clarifying the serious issues raised by such examples. namely the distinction into [+ Concrete]. but his suggestions as to how this might be possible have not yet been developed to the point where their merits can be evaluated. suggests that one might seek a generalization involving also the passive transformation. ) ." various kinds of "sememes. [. 1 7." etc. as in "John frightens easily" (this in fact is much more general . in connection with features and categories." and so on. The possibilities for this are. Notice that words of the category of "frighten" do appear quite naturally as Intransitives in surface structures.Abstract]} characterize the Mass Nouns such as water." The often obligatory Adverbial of Manner. though quite inconclusively.Count].. "the book reads easily.Count]. that the problems mentioned here admit of no merely terminological solution. alternative (iv) is well moti vated. [+ Abstract]} characterize the true Ab stract Nouns such as virtue. so that a lexical entry (a formative) may be regarded simply as a set of features. In such a case.NOTES TO CHAPTER 4 14. One might question the factual correctness of this. the Direct Object of the deep structure "Un specified-Subject frightens John easily. To say that a feature is positively (negatively) specified is to say that it is marked + (respectively.)." "semantic object. but only as (rather mannered) transforms of "sincerity frightens Un specified-Object. Thus we could perfectly well state the facts relating to ( 1 9 ) in terms of such new notions as "semantic subject. But this is irrelevant here. of course. the "grammatical Subject" is the "logical Object" . 1 8. 15. that it may be possible to express the meaning relation on dis tributional grounds. no one would interpret "his sincerity was frightening" as ambiguous. quite limited . while the features {[. I have been assuming that the features { [. such - ." etc. then this convention would have to be extended to selectional features as well.example. these are formulated. no doubt. 1 96 1 ." or.: Abstract] (as p roblem . 24. these constraints go well beyond what has been illustrated here. Cf. "acci dental" versus "sys temat i c gap") suggests what in Chapter 1 was called a "formal" rather than a "substantive" linguistic universal. also substantive constraints to be dis covered here.. 1962a." so that "he X's dogs" is parallel in meaning definition of the notion "phonologically admissible" (i . Notice that if we were explicitly to list positively specified rather than negatively specified selectional features in the lexicon. and in the re ferences given there. h owever. treat a selectional feature as itself being a kind of complex symbol . though there are. As possible examples of "accidental gaps" we might point to the nonexistence of a Verb X taking as Direct-Object expressions designating ' animals and having otherwise the same meaning as the transitive "grow. replaceable. 1 964. See Halle (1 959a. §§ 6. 7. 2 11 .Concrete]. Notice that Halle's . This question is not easy to resolve without much more empirical study. 21. Subject" for "run. perhaps. A Verb that can appear only in the passive would therefore be an exception to this rule (e . of evaluation procedures and explanatory adequacy in Chapter 1 . respectively) should be identified. in fact.e. mountain. subfeatures of [. 19. for all features. in effect. and there is no doubt that. Such a convention would. Thus we should not want to have to l ist both the features corresponding to "takes Human Subject" and "takes Animate it turns out that the features [± Concrete] and [:. they will also play the role of general con ventions (i.e .Animate] and [. Recall that we have presented some reasons for regarding the phrase by"""passive (where passive is a dummy terminal symbol. As rather decent fellow. aspects of the general definition of "human lan guage") that can be relied on to reduce the specificity of particular grammars. "he is said to be a 22. then the feature [Abstract] would be cross-classifying rather than hierarchic with respect to [+ Count].NOTES TO CHAPTER 4 and [. ." for . by the universal dummy symbol A) as a Man ner Adverbial.Count]. also the discussion The phonological redundancy rules are also subject to certain uni versal constraints. As always. I f 20. there are a few exceptions that require separate state ment. such as as table. effort. such forms as "he was shorn of all dignity").g. 1 959b).. The desirability of such a conven ti on was pointed out by Paul Postal. " It is also possible to try to give a language-independent characterization of these categories. Actually.g. in fact. See Chomsky (1964. 27. and further structure must be imposed along these "dimensions.) for discussion. 3 d. Thus we can regard the category of case in German as a four valued. Simply for expository purposes. to "he grows corn" ("raise" appears to cover both senses). let us take the integers in the order of conventional presentations. 118. Halle. These are important matters and have been the subject of much study that. It has been maintained (cf.for other systems). 1 957). Jakobson. in practice. 116.. in descriptivist grammars of the item-and-arrangement type the latter might be omitted. Thus in the case of the feature Stress. in fact. so that [1 Gender] is Mas culine. though it may have more values or a different organiza tion of values d. gender as a three-valued. this is not the optimal analysis. p. designed in such a way as to exclude the possibility of all but the most elementary general rules.for English or German. In phonology. That is. e. Notice that we have assumed all along that features are "binary" . and other phonetic features would also have to be regarded as multivalued in a detailed grammar. though obviously not always in their phonetic func tion. Presumably. There was no logical necessity for this. the categorial rule that develops Nouns will not be N � Il (d. note 115 . goes well beyond the scope of this discussion. a Number] (a = + or . was pointed out to me by Morris Halle. 29. which is quite serious.that they simply partition their domain of applicability into two disjoint classes.23 2 NOTES TO CHAPTER 4 115. best regarded as binary in their phonological function (d. and number as a two-valued dimension. but rather N � [Il. however. 1 936) that such "dimensions" as Case should also be analyzed into a hierarchy of binary features (like phonological distinctive fea tures). but we shall not consider this question here. Bever). This defect of morphemic analysis of inflectional systems. - . pp. G. [11 Number] is Plural. I shall therefore consider only an unstructured description in these illustrative examples. it seems clear that the distinctive features are. and we can consider all No uns as being arrayed in a single multivalued dimension of dedensional classes. U2). and Bruder is assigned to Class 1 along the "dimension" of declensional class. we can easily find five or more degrees that must be marked in English. or the absence of a word that bears to plants the relation that "corpse" bears to animals (this example was suggested by T. [11 Case] is Genitive.. since its only function is to per mit some generality to be introduced into the "morphophonemic" rules and since these grammars are. y Case] . furthermore. p. there has been very intensive and fruitful study of the transformational cycle of Russian and Latvian phonol ogy (for references. 33. This formative might. in fact. note 6. In the last few years. and Lukoff (1 956) and Halle and Chomsky ( 1 960). been no serious investigation of how a transformational cycle applies to a feature system and to Phrase-markers such as (30). and to regard the Ending as belonging to the paradigmatic categories. � Number. See note 38 for some support for this assumption. This is. 3 1 . 34. see Chomsky. Ablaut).NOTES TO CHAPTER 4 30. pp. lP435. The rules that constitute this system apply to Phrase-markers. so far. the transformation is not blocked even when they are dis tinct. it will be possible to bring phonological evidence to bear on the question of morphemic versus paradigma tic represen tation inHectional systems. Bever ( 1 963). not features (though of course certain rules may be restricted in application in terms of syntactic features). Notice that the distinction that is emerging in this discussion is not coincident with that suggested in note 30. hence as a degenerate complex symbol that is expanded by the rule into the full complex symbol [+ Definite. 32. T. 1 964. Harris ( 1 957). When the two Adjectives are paired in a rather special way that is for the present poorly under stood. the empirical evidence sug gests that the ordering of the transformational cycle in phonology is determined completely by categories. Thus an alternative to the analysis presented in (30) would be to regard a lexical item such as Bruder as consisting of a Stem fol lowed by an Ending. When this is clarified. Cf. the most natural assump tion. In one of the earliest descriptive studies of French. G. Variables over feature specifications were used in Chomsky. p." Cr. The idea of using them to deal with assimilation is due to Halle ( 1 962 b). and conse quently their formulation depends very heavily on answers to the questions being considered here. There has. 46 1 -462) maintains that such a fafon de parler cannot be considered either "absolument of . Vaugelas ( 1 647. in developing the transformational stress cycle. Halle. if we regard the features as actually constituting the terminal symbol (the formative). be regarded as consisting of the feature [+ Definite]. a Gender. It is interesting to note that the correctness of such examples as (40) has been questioned. Thus we have such forms as "this is taller than that is wide. 1 4). For the moment. Bever and Langendoen ( 1 963). Bever has pointed out that the same device can be applied to a description of various kinds of alternations that involve feature shift (e..g. See Lees (1961) and Smith ( 1 96 1 ) . we must somehow account for such facts as the following: the i mpossibility of "I know a more clever man than Mary" or "I have never seen a heavier book than this rock. The problem. does not imply that TTistTam Shandy is a poe m . 11 8. Thus from "I have a [# the friend is from England #] friend" we can form. rather." and the problem is the deletion of the Article. a man speaki ng to a woman should not say je suis plus b eau que vous. is taken to be its source. Again. The deletion of the pluralized non-Definite Article is automatic." which." as seems qu i te plausible. The embedded sentence could not be "a friend is from England. The deleted Noun Ph rase is "the friend. Thus." in which case the problem would not arise. by relativization. whereas the sentence "I have never read a poem (which is) more intricate than TTistTam Shandy. in this position. if such sentences as (4 1 iii) are regarded as derived from "I know several lawyers (who are) more successful than Bill" by Noun-Adjective inversion following deletion of "who are. the fact that the sentence "I have never read a more intricate poem than TTistTam Shandy" implies that the latter is a poem. 11 7. al though it would be perfectly all righ t for him to say je suis plus Tiche que 116. Similar considerations may account for another apparent violation of the general condition on recoverability of deletions. But the fact that definiteness is obliga tory suggests that in the underlying Phrase- vous." and suggests that it be avoided when masculine and feminine forms of the Adjective differ. since def initeness of the Article is automatic in this position." although the pre sumed sources of these (namely.2 34 NOTES TO CHAPTER 4 mauvaise" or "fort bonne. but should rather ( "p our parler reguli�rement") resor t to the paraphrase je suis plus beau que vous n'etes belle. the identity condition for relativization involves only the Noun. raises various difficulties for the analysis of comparatives. This fact. is to provide some explanation for such phenomena as those of the preceding paragraph. In parti cul ar. etc. pointed out to me by Brandon Qualls. and not the Determiner of the deleted Noun Phrase . . as throughout this discussion. I should like to emphasize that there is no particular difficulty in formulating an ad ho c sys tem of transformational rules that will have the desired properties. "I know a man (who is) more clever than Mary" and "I have never seen a book (which is) heavier than this rock") are perfectly all right . As has fre quently been observed. "I have a friend (who is) from England" in the usual way. in this view. which differs from the Article that is used to erase it by the relative trans formation. Notice that although sad. might contain a certain nominalization mor pheme in place of the pre-Aspect part of the Auxiliary. i n poorly understood quasi-productive processes. w e might raise the question whether the nominalization element should be represented as a morpheme nom or as one of lexicon for post-Animateness (if we decide that what is involved here is not a matter of homonymity). If this is the correct analysis. in this case. it may very well be assigned contextual features corresponding to various subfeatures of [. Thus the constituent base Phrase marker. since in its underlying form it is nondistinct from the Article of the Noun Phrase of the matrix sentence. Note that this decision requires a feature analysis for Articles. so as to characterize as deviant such sentences as "the pencil is sad. 1 2of. Here. then by the princi ple just S9 . Perhaps it will be possible to rephrase this con vention as part of a general definition of the notion "word. deletion of the Article will be permissible.. See now also Zimmer (1964). Chap ter 4.g. Fm . 41.in this case. and should be further the features Fl' • • . too." That is. We oversimplify somewhat. pp. . 411. one might try to state a general rule determining placement of word boundaries in terms of lexical categories and branching wi thin the scope of complex symbols . 5). 4 5.Ani mate]." Bolinger. Such lists sim ply indicate areas where all presently known theories of language have failed to provide any substantial insight. namely formation of compound no u n s. 4 7f.) . though it ra ises non trivial problems of a different sort. pp. § 4. in various articles (e. See Lees (1960a. This possi bili ty was sug ges ted by some observations of Paul Postal's. for example. 40. Chomsky (1964. an obligatory trans formation) . explored. a feature added by the transformation. at issue. this being added by a "redundancy rule" (in this case. need not be marked in the 4S' 44. A detailed study of one system of essentially this sort. has listed many examples of .). marker the Article be left unspecified for definiteness. and appendices). and Katz and Postal (1964. Bolinger suggests that his discussion of "quasi-transformations. with little difficulty. A related cl ass of problems is examined brie8y by Harris ( 1957 ." This matter has no relevance to the point with [± Definite] taken as a syntactic feature. also note 30. pp. 64f." which cannot receive an in terpreta tion analogous to that of "the book was sad. These constructions are interesting in many respects. 1 961).NOTES TO CHAPTER 4 estabJished.) for discussion. Cf. and they can be ex tended in many ways. is presented in Lees (19600. Bolinger. but this seems to me an entirely unwarranted conclusion. Chomsky.NOTES TO CHAPTER 4 his examples support an alternative theory of grammar. . 1 964. 54). p. for reaSons dis cussed elsewhere (in particular. 1 94 1 . Office of Naval Research. Two Studies on Syntax and Semantics. "A plea for excuses. Alt (ed." Quarterly Progress Report. A. and T. J. Research Laboratory of Electronics. 69.). Joos (ed. New York: Academic Press. Warnock (eds.). Bach. Report for U. T. Austin. 2 6. La Logique. Reprinted in J. pp. A dvances in Computers. "Studies in colloquial Japanese IV: Phonemics. Readings in Linguistics. "Logical syntax and semantics. The Basic Works of Aristotle. Measures of Syntactic Com plexity. I. New York: Random House. J. L. McKeon (ed. ( J 963). Arnauld.). Proceedings of the Ninth International Congress of Linguists." In F. B. (1 964). Rosenbaum (forthcoming). Bar-Hillel. A. 2 30-237. Reprinted in M. New York: Holt. pp.LT. 86-1lI5. 202-203. Bedford. "The reciprocating cycle of the IndO-European e-o Ablaut." Language. Theory of Language. De A nima. "The present status of automatic translation of lan guages. Lunt (ed. Bever.LT. pp." Lan guage. Bloomfield. M. Kasher. 3 0. Philosophical Papers of J. No. Washington. Jerusalem. Mass. ( 1 956). Shamir ( 1 963) . and E. (1933). L." Quarterly Progress Report. A ustin.). The Hague: Mouton Se Co. Vol. Language. ---. Smith. O. Information Sys tems Branch. Translated by J. ( 1 954).Bibliograph y Aristotle. pp. In R. pp. No. ---. Langendoen ( 1 963).. : Mitre ' Corporation Technical Re ports." I n H.. 9 1 1 63. ---. ou l'art d e penser. L. 1 96 1 .S." Proceedings of the Aristote lian Society. Urmson and G. E .). 69. 203-207.( 1960).. G. "Subcategories i n transformational grammars. and P . J. Y. 1 957. and P. 23 7 . ( 1 788). Bloch. (1 950). M. Nicole ( 1 662). --. A. "The e-o Ablaut in Old English. London. London: Oxford University Press. Re search Laboratory of Electronics. L. Beattie. D . 26-58. N. "A gramma tical allegory. pp. 1 965. Bush. Breland ( 1 961).(1955). The Logical Structure of Linguistic Theory. R. Reprinted in R. New York: Wiley. Texas. Library. Vot. with corrections. Bush. "Some methodological remarks on generative grammar:' Word. 366-38 1 • Breland. Luce. 681-684. Reprinted in Fodor and Kau (1964). "Topics in the theory of generative grammar.!. -. and M ." Language. Vot. 323-4 1 8. 1 962.T. 124-1 48. P. Galanter (eds. -. 11. California: Stanford University Press. Mimeo· graphed. pp. 3 5. Syntactic Structures. D . Bush. It appears under the title of the session in H. The Hague: Mouton 8c Co. .). 11." at the Ninth Inter national Congress of Linguists. 1 965. -. IT'2.). Current Issues in Linguistic Theory. Bolinger.(1956). pp. in R. Austin. -. Reprinted in part in Fodor and Katz ( 1964). and E. pp. New York: Wiley. Unpub· lished Master's thesis. Current Trends in Linguistics. Nagel. "Formal properties of grammars:' In R. pp. -. Galanter (eds. Tarski. and E. A slightly earlier version appears in Fodor and Katz ( 1964). Chomsky. (196 1)..).(1957). 1 1 3-1 24. Vol. "A transformational approach to syntax:' In A. Lin guistic Theory. Re printed. Luce. R. New York: Wiley. Suppes. " Modem Philology." American Psychologist. 60. -.(1961). "On certain formal properties of grammars:' Information and Control.BIBLIOGRAPHY Bloomfield. "The misbehavior o f organisms. Review of Skinner ( 1957). 1 37-1 67. R. -. Mass. Vot.R.). Handbook of Mathematical Psychol ogy.(1g64). Language. M. L . Luce. Methodology and Philosophy of Science. and other matters. A. pp.(in press). -.( 1959a).( 1 96ab). (195 1 ) . 2 1g-239. Lunt (ed.E.." I." In T.(1959b). The Hague: Mouton 8c Co. D. Mass. Vol .(1963). "Explanatory models in linguistics:' In E. Ill.(1962a). Sebeok (ed. Cambridge. Stanford. The Hague: Mouton 8c Co. "Syntactic blends 3 7. K. Morphophonemics of Modem Hebrew.). 11. Readings in Mathematical Psychology. This is a revised an d expanded version of a paper presented to the session "The logical basis of linguistic theory. 2. pp. and E. and A. Hill (ed. Readings in Mathematical Psychology. Galanter (eds. 1 964. I7. The Hague: Mouton 8c Co. -. approach to personification 161-1 7 1 . -. Logic. Proceedings of the Congress. Pmceedings of the I958 Conference on Problems of Linguistic Analysis in English. M. D. ( 1963). I6. A.. University of Pennsylvania. "Three models for the description of language. -.). pp. Transactions on Information Theory. Cambridge. Re printed in Fodor and Kau (1 964). pp. Du Marsais. Curry. Descartes. For Roman ]akobson. Handbook of Mathematical Psychology." M. Chandler. and F. and J. J. (1 963).) ( 1 964) . II. A. Vol. ( 1 96 1 ) .). R. "Cartesian Linguistics. On the dating of this manuscript. 208-2 3 1 . "Introduction to the formal analysis of natural languages.: American Mathematical So . de ( 1 66 7). Ch. 1 1 9. J. (196 1 ). and G. The Hague: Mouton &: Co. Linguistic Science and Logic.1 6 1 . Fraser. Lunt.(forthcoming) "On the notion 'derived constituent structure. C. Oeuvres Completes de Diderot. Cudworth. MacLean (eds. -. (1 963). S. B.' " Aspects. A Treatise Concerning Eternal and Immutable Morality.( 1 647). Com puter Programming and Formal Systems. Fodor. H. pp." In R. Bush.. Mass. Halle. Braffort and D. J. "On accent and juncture in English. Studies in Logic Series. I (1 875). G. and H. Paris: Garnier Freres. Miller ( 1 963)." Proceedings of the Aristotelian Society. Asswt (ed. Schiitzenberger ( 1 963). Logique et principes de grammaire. Luce. and E." Word. pp. -. Page references are to J.). Hirschberg (eds. pp. R. ( 1 7 3 1 ). 45-80.). 1 955Diderot." In R. A Philosophicall Discourse Concerning Speech. see Sahlin ( 1 928). The Structure of Language: Readings in the Philosophy of Language. "Notes directed against a certain programme. R_ W. Foot. Fillmore. R. Dixon. B. T. ( 1 75 1). The Hague: Mouton &: Co. Amsterdam: North-Holland. (1 963).). New York : Dover." Mimeographed. LukofE (1956). p. -. 65-80. Galanter (eds. Ross in The Philosophical Works of Descartes. ix. I. Lettre sur les Sourds et Muets. Haldane and G. Vol. P. C. 269-322. "The position of conjoining transformations in a grammar. Vol.). and M. ciety. I. R. Katz (eds. ( 1 64 1 ). Supplementary Volume 3 5." In M. Bedford. pp. Jakobson (ed. "Goodness and choice. Providence. "The position of embedding transformations in a grammar." Both works by Descartes translated by E.: Prentice-Hall. Edited by E. D. Proceedings of the Twelfth Symposium in Applied Mathe matics. 56-68. pp. Meditations. Halle. I9. Cordemoy. --. H. Les vhitables principes de la grammaire. D. Englewood Cliffs.: Mitre Corporation.BIBLIOGRAPHY -- 1I �9 (forthcoming). pp. A. "The algebraic theory of con text-free languages." In P.( 1 76g). P. N. (1 729). Structure of Language and Its Mathematical --. J. "Some logical aspects of grammatical structure. New York: Wiley. --. The English translation is dated 1 668. --. Jakobson (ed. 10. H. M." In R. Frishkopf.I.LT. Reprinted in Fodor and Katz (1g64).. Halle. "Conjunction with and:' Transformations and Dis course Analysis Projects. Gross. "On the bases o( phonology. No. Rinehart Se Winston. S. pp.BIBLIOGRAPHY --- Proceedings of the 1964 Magdeburg Symposium. S." Journal of the Acoustical Society of A merica. pp. Review of Saumjan and Soboleva (1963). pp. "On the equivalence of models of language used in the fields of mechanical translation and information retrieval. Research Labora tory of Electronics. Gleitman. H. 2. 75-95. M. pp. l u g-u28. ( 1 964). Structure of Language and Its Mathematical Aspects. D. "Responses to acoustic stimuli from single units in the eighth nerve of the bullfrog. Gleason. "The organization of language: a stratificational view. "Phonology in generative grammar. Greenberg. Universals of Language. Pulgram (ed. M." Word. (196 1). Language 40.. H ." In J. Zeichen und Sys tem der Sprache." In E. 295-296. (1g62a). "Questions of linguistics. Washington. pp.T. J5. Report of the Fifteenth A nnual Round Table Meeting on Linguistics and Language Studies. C. J. ( 1 963). Hall. Studies Presented to Joshua Whatmough. pp. B..I. "Two families of languages related to ALGO L.: American Mathematical Society. (1957). pp. 397-4 10. pp. 5472.). (1964). L. "On the role of the simplicity in linguistic description.: Georgetown University Press.(1962 b). " ------ . L. Greenberg (ed. Ginsburg. Cambridge: M. Proceedings of the Twelfth Symposium in Applied Mathe matics. --. 350-37 1 ." In C. second edition. pp. ( 1959a). I." Quarterly Progress Report. H. R. pp. IJ. "A descriptive convention (or treating assimilation and dissimilation. Philadelphia: Uni versity of Pennsylvania. 18.). -. ( 1 961).(1964). and M. Journal of the Association for Computing Machinery. (1959b)." Nuovo Cimento. Stuart (ed. The Hague: Mouton Se Co. Rice ( 1 962). Press. Providence. New York: Holt. The Sound Pattern of Russian." Informa tion Storage and Retrieval.).)." In Fodor and Katz ( 1 964). mimeographed. 40. M. 66. and H. "In defense of the number two. 89-94. J. Introduction to Descriptive Linguistics. 4945 1 7. (1964). G.(196 1). "Some universals of grammar with particular reference to the order of meaningful elements. The Hague: Mouton Se Co. 43-57. Goldstein (1963). 58-g0. A. No. Humboldt. (1 8 8 6) Ober die Verschiedenheit des Mensch lichen Sprachbaues. (1924). 3 7." Word. ." In R. Berlin. . 1 8-28. 106-1 54. "Beitrag zur allgemeinen Kasuslehre. 1 55." I. 6. pp. O. 48-50. "Speech recognition: a model and a pro gram for research. 58. Reprinted in Fodor and Katz (1964). M. Principles of Behavior. University of Bristol Studies. Hiz. Harris. (1948). --. Jakobson. H. pp." Language.( 1957)." Travaux du Cercle Linguistique de Prague. Hubel..159." Journal of Comparative and Physiological Psycholor.. London: AlIen Se Unwin. 240-288. "Congrammaticality. S.T. N. Proceedings of the Twelfth Symposium in Applied Mathematics. ( 1 952). pp. Philosophy of Grammar. VoI. "The morphophonemics of English. Jakobson (ed. Harman. C. 1 46-162. and T. An Enquiry Concerning Human Understanding. 33 . D. Methods in Structural Linguistics. Carre ( 1 987). IO. and N. ---.R. 275-28 1 ." Quarterly Progress Report. von. New York: Mac millan. "Co-occurrence and transformation i n linguistic structure." Journal of Physiology.E. 28.). pp. D. -. New York: Appleton-Century Crofts. I12. pp. pp.( 1 96 1 ). : American Mathematical Society. and S. "Plasticity in human sensorimotor control. Wiesel ( 1 962). p p. pp.BIBLIOGRAPHY -- --- . H. Hull. "Linguistic elements and their relations. I T 8. No. pp. Translated by M.I . Herbert of Cherbury (1624). Freedman ( 1 968). pp. C. "Generative grammars without transformational rules: a de£ense of phrase structure. De Veritate. Stevens (1962). H. R. I 60. (1954).(in preparation). 3 9. Research Laboratory of Elec tronics.I. F. Z. 29 8-840. W. "Discourse analysis." Language. 872-876. (1958)." Language. binocular in teraction and functional architecture in the eat's visual cortex. ----. and K. (1961). --." Science. Transactions in Information Theory. 455-462. "Distributional structure. Hume. (1968). (1 986). and A. Chicago: University of Chicago Press. Held. Jesperson. "Movement-produced stimulation in the development of visually guided behavior. "Receptive fields. R. 2 9-5 8 . J. H. Hein (1968). R. (1951).." Language. batteries of transformations and grammatical categories. G. 56. pp. ( 1 748). L. Chomsky (1960). New York: Harper Se Row. Hockett. pp. Providence. The Sound Pattern 0/ English. Structure of Lan guage and Its Mathematical Aspects. No. 597-0 1 6.y. 6. A Course in Modern Linguistics. Fodor ( 1 9 64). and E. 1 70-2 10. Briefwechsel %wischen W. -.R. W. 17. Trans lated by A. G. G. 145. Postal (1964). 1 7 1-185. " --. (1964b). von Humboldt und A. J. An Integrated Theory of Linguistic Descriptions. S. W. pp . Klima. Vinogradova (1959). New York: Columbia U niversity Press. Review of Chomsky (1957). "Rules for English pronominalization. p p . 40.(forthcoming)." In Fodor and Katz (1964). R. 1 940-195 1. ( 1908). "The structure of a semantic theory.. Luria. The H ague : Mouton Bc Co. 1 949. 1 24-1 37. "Language.: Open Court.(1960b). The Hague: -. S. Arnauld. ( 1964a). The Biological Bases of Language. 3 75-" " 407 · -. Mass. H. "Semantic theory and the mea n i ng of 'good. A. J." Lan· guage. p. Fodor. Lancelot. E. pp.." Proceedings of the I. 3 6.(1964d). evolution. and purposive behavior. Patterson (1964). Ill. 835. Lettvin.: M. . R. R. The Grammar of English Nominalizations. B . (1957). H. Matthews. 89-105. "Analyticity and contradiction in na tural language. C. "Mentalism in linguistics. "Grammatical analysis o f th e Engl ish comp ara tive con struction. Hidatsa Syntax. A. " Science. Lees." In Fodor and Katz (1964). pp. A." Language. E." Language. and W. and O . 47. "Depth perception in sheep.I.E. Lenneberg.. "I nnate ideas . New Essays Concerning Human Understanding. multi ply ambiguous adject ival construction in Eng· lish. (1964). A reply to Dixon's 'A trend in seman tics." I n Fodor and Katz (1964). and J. 207-2 2 1 .. --." In S. LaSalle. Lemmon. Semi-sen tences . pp. Leitzmann.BIBUOGRAPHY Katz. pp." Language. Maturana. Cambridge. Reprinted in a re vised and extended version under the title "The capacity for lan guage acquisition" in Fodor and Katz (1964). A. Y.(in pre para ti on) . 50. pp.' '' Journal of Philosophy. (1964). Repri nted in Fodor Bc Katz (1964). "An objective investigation of the dynamics of semantic systems . et al. H. G. 3 9. A . 3 . Halle: Niemeyer. H.(1960a). 33 . Diamond (ed. Leibniz. S. W. Schlegel. Langley. Klima (1963). Language.).T. (1960). 1 7-2 8 . -. Culture in History : Essays in Honor of Paul Radin. B. S .(196 1). Grammaire generale et raisonnee. and G. ''What the frog's eye tells the frog's brain. -. " British Journal of Psychology. McCulloch . -. -. 19-29.(1964C). Pius (1 959) . J. ( 1 660). J9. W. pp. "Negation in English. "A Mouton Bc Co. and P. --. Press. pp.' " Linguistics. and J." Word.. --. Ch. pp. Essays on the Intellectual Powers of Man. "Free recall of self-embedded English sen tences. pp. 10-1 1 . Peshkovskii.I." Quarterly Progress Report No. .. The Hague: Mouton Bc Co. Vol. 64. Russkii Sintaksis v Nauchnom Osveshchenii. A . Some Syntactic Rules in Mohawk. "Transformational grammar . t3 1-238. 7. U npublished doctoral dissertation. W. R. and K. ----- . No. 1-1 7. Jerusalem. Finitary models of language users. 3. M. Mass.I. Miller. Paul. Oman. pp. pp. Strong. IJ. Janu ary-June 1 964. Research Labora tory of Electronics. New York: Henry Halt. Hebrew University. Grammarama. Center for Cognitive Studies. Reichling. and S. A. Stein (1 963).. Press. "Limitations of phrase structure grammars. T. Nominal Compounds in Modem Literary Hebrew.)." A rchivum Lin guisticum. second edition. H. "On the limitations of context-free phrase-structure description. H. Bush. Galanter (eds. ( 1 960). 29t-303' --. . pp. Cambridge: Harvard University. Semi-annual Technical Report Department of Defense. G. Reid. ( 1 964). Pribram (1960).( 1964C). U. ( 1 96 1 ). . and E. New Haven. "Some perceptual consequences of linguistic rules." In R.T. Scientific Report No. A. Plans and the Structure of Behavior. pp. and D. Center for Cognitive Studies. ( 1886)." Information and Control. P. Postal. H. Chomsky ( 1 963). "Principles and methods of syntax: cryptanalyti cal formalism. New York: WHey. CS-t. Luce. ( 1 962b). A. "Underlying and superficial linguistic structure. . Prinzipien der Sprachgeschichte. pp." In Fodor and Katz ( 1 964). Green Bc Co.BIBLIOGRAPHY --- Matthews. Quine. 1 89 1 . M ." Journal of Verbal Learning and Verbal Behavior. Constituent Structure : A Study of Contemporary Models of Syn tactic Description. Isard (1964). t46--2 66.T. 2. 1 3. and S. ( 1956). London: Longmans. Galanter. Unpublished doctoral dissertation. J4. 1 96--209." Har vard Educational Review. ( 1 96 1 ). Cambridge. Page refer" -----. --. 2 1 7-228• . Yale University. Word and O bject.: M. pp. Moscow. Cambridge: Harvard University. Norman (1964). December. Research on the Use of Formal Lan guages in the Behavioral Sciences. 419-492 . IO. M. V.( 1 964b). (1 96ta). and M." Lingua. --. 11. P. A. and New York: Wiley. (1 785). E. Handbook of Mathematical Psychology. D. Advanced Research Projects Agency. Isard (1963). ( 1 964a). and N. Trans lated into English by H. "A class of complex modifiers in English. "Grammar of English nominal izations. 10. . D. "Systematically misleading expressions. Soboleva (1 963) Applikativnaja porozda juJcaja model' i iscislenie trimsformacij v TUSskom jazyke. Flew (ed. The Influence of Sentence Structure on the Reading Process.N. National Physical Laboratory Symposium No. 195 1 . November 19 6 1 To appear in Zeitschrift fur Phonetik. Washington: 1957. (196 1 ). University of California. London. New York: Dover. Reprinted in part in M. and P." Lingua. Oxford: Clarendon Press." Technical Report No. Review: R. 3411-865.R. Language Mono graph No. B. "An appraisal of transformation theory. ( 1 96 11 ) . (1 953). Schachter (19611). ( 1964) . Rosenbloom. Sweet. "Rules for a segment of English syntax. Reading in Linguistics. London: Macmillan and Co. Smith. Stockwell. Woozley. A n Inquiry into Meaning and Truth. Reprinted in A. Moscow: Izdatel'stvo Akademii Nauk SSSR. N. pp. "Visual discrimination in animals. Oxford: Blackwell. (1 985) . I2. arranged by H. 3 7. Information Systems Branch. . Sahlin. ( 1 93 1 ) . P. 7." International Journal of American Linguistics. C. ( 1 96 1 ) . 1 34.). TwaddelI. W. Lees. Los Angeles. Uhlenbeck. N.). P. --. Hebrew University. S. B. (1940). (1 9 1 8) Collected Papers. (1 9118). E. first series. 2 0. G. 194 1 . Sutherland. 54-59." British Medical Bulletin. Schlesinger.. The Elements of Mathematical Logic. Schachter. -.244 BIBUOGRAPHY ences are to the abridged edition by A. Skinner. and P. K." Mimeographed. Joos (ed. Vol. ( 1959). pp. C. Logic and Language. pp. H. London: AlIen Se Unwin. pp. Saumjan. Cesar Chesneau du Marsais et son role dans /'evolu tion de la grammaire generale. ( 1 950). Shamir. pp . "Stimulus analyzing mechanisms. F. Il. M.. Jerusalem." Philosophical Review. On Defining the Phoneme." Proceedings of the Aristotelian Society. O.1 45. Paris: Presses Universitaires. B. New York: Appleton-Century Crofts. E. "On sequential grammars. I. RusselI. 1671 86. 1-1 8. R. F. Sprachwissenschaft and Kommunikations forschung. ( 1 95 7) Verbal Behavior. Unpublished doctoral dissertation. 62. 28. "Ordinary language. S. (1963). A. . Wyld. S. G." Language. I6. G. (1 9 64) ." Mechaniza tion of Though t Processes. Ryle. " Languaje " Ciencias. J. The Hague: Mouton Bc Co. Oxford: Clar endon Press." pp. (1 926). Monograph No. Son Bc Co. Facsimile edition. K. Vol.BIBLIOGRAPHY 245 ( 1 964). Paris: Librairie E. Lunt (ed. ( 1 953). C. Oxford: Black well's. S. IO. I2. Zimmer." In H. 981-983. C. Zierer. L. E. ( 1 960). 1 3-20. Vaugelas. 20. 444-466. pp. Proceedings of the Ninth Congress of Linguists. "Linking verbs and non-linking verbs. pp. ( 1 964). 5. Wittgenstein. Droz. ( 1959). F. Remarques sur la langue Franfaise. V. I. Second edition. Glas gow: J ackson. Discussion in the session "Logical basis of linguistic theory. de ( 1 647). 1 934." Proceedings of the American Philosophical Society. E. Ullmann. Wilson.). Supplement to Word. Affixal Negation in English and Other Languages. Philosophical Investigations. The Principles of Semantics. ( 1964). Ynvge. --- . Statement and Inference. "A model and a hypothesis for language structure. . 3()-lJ8. types of. 12-14. 8-10. 191. L. 95.212 Competence. 12-14. 196-198 mUltiple-branching. 196. 220. 56.225. 234 Adverbials. L. 114.104.• 105. 46• 5g-62. 113. 188. 115. 12-14. 189 comparative. 79 major. 140• 193 and performance. 99.• 113 Base component general structure. 183. 141 transformational. 1 18. 231. 122-126. B. 122. 235 Boundary symbol. 141 illustrated. 116. 116. 66 Case. 221 Deep structure.. 5 Behaviorism.133 Bloomfield. 74. 84. 120. 137. 142 see also Grammar. 199. 197 self·embedded. 136. Ill. 224. 1 10. 220-221. 212.. 119 subcategorization of.211. 3. 97.1 4. H. rules of. 17. 203. 12-14. 15. 198 right· branching. 224. 120. 198. explanatory.227 166. 123.. T. 121. 98. 228. 29. 140. 122. 12. 23.209. 74. 131. 95. 115.133 Bach. 114 Bolinger. 215-219 Analyzability. In3 Assimilation. E. 106-111 transformational rules in. 70. 144 lexical. see Language learning. 98. 206. 64-106 205. 82. 84-gD. 115. 144. acquisition model Adequacy. 27. lIlO. 19. 16. 212. 119 aspects of. 120. 34. 182. 162. 86. lOll. 121. 115 Constructions causative. III Categories grammatical. 233 Component base. G. 40• 41• 44. 32. 101-105.. 25-27.• 49 Curry. 111-127 Basis of a sentence. 41• 5962. 134. 163. 136. 225 left-branching. 64-69. 160. 99. 147. 98. 128. 226 Anderson. 140. 18. 175. R . 6. 4. 157. 122 Base rules.233. 197 nested.178-180. 150. 197 Complex symbols. 139. 143. . James. 24. 194. see Base component categorial. 11. 217.Index Acquisition model. 16-18. 10. 196. 193. 189. 811-gD. 27. 229 Cudworth. 178. 78• 93. 128.206. 203. 234 coordinated. T. 24. D. generative Concatenation·systems. 130 Beattie. 76. 196-198 Cooccurence. empirical descriptive. 106. 123. 78. 196. 30-38• 40. 1 19. 128. 25.206 Bever. 68--14 Functions. 16g. 132 Grammar.181. 214216. Ch. 141. 86. 75.81. 47..132. 89. 137. 105--111 defined. 224 Grammaire genera le et raisonnee. 175. 198 base of.142 sequential. 164.40. 65. 144 syntactic component of. 235 distinctive or phonological.75. 17.89. 97. 5 Grammar and philosophica l error. 135. 171. 208 INDEX lexical.203 Dictionary definitions. 65. 39.211. 88. 122 . 67. 82-87. 135. 88. 142. 160 Diderot. 1I0. 17.208. 222 Empiricis m.85. 6. 144. 51-54. 55.86. generative base component in. 141. 215 selectional rules. 80. 84. 132. 78.88.35. 164.148. 154.61. 65. 141. 231 and notational conventions. llO.2Dg delimitation of. 5.143.214 Formatives. 148. 221 Grammar. 120. 17.164. 198 semantic component of.227. 95""97 subcategorization rules. 81.81. 1119. 117.45. 135.80.57.. 142. Generative capacity strong. 136. 121. 151. 233. 75--19. 61. 113. 136. 235 Field propert ie s.85. 139. 156. 48.215 context-sensitive.98.175. 78. 151. 5-'1 opposed to universal grammar.88. 28. 34-47. 141. 192. philosophical. 79 organization of. 23 Distinctive-feature theory. 199. 148.60-62. 198 projection rules of. R. lll. 145.122. 154. 232 selectional. ll3 finite-state. 144. 16. gD-106..248 Deep structure (continued) defined. 29.229 Firthian linguistics. 113. 128-147 Derivation. 123. 95. 154. 154. 154. 206 Evaluation measures. R. by Lancelot et al. 142.230 semantic.15 phonological component of.67. 16.9°. 203. 165 syntactic. 231 Features contextual.160.4. 15-18.8. 153. 126. 23. 139. 7 Direct-object. 190. 82. 226. 16. 99. 60-62. 160164. particular.98. 65. 146.87 Distinctness of featu re matrices. D. 59. 120. 42-46 Exceptions. 61. 123. 164. 230. 233 grammatical. 97. 157. 162. 28. 68. 85. 58. 143. 181. 161. 154. 28.61. 190. 10. 136. W. 203. 208 J. 81.200. 122. 220.182. 14. 84. A. 214 strict subcategorization. 199 and grammatical transformations. 161. 164. 87 see also Lexical entries Functional notions. 88. 117. 229. 159.200 Dummy symbol. 214. 3. 122 context-free. 198. C. 74. 153. 3.220 Dixon. 79. 157159. 68--14. 6 Grammar. 88.208 weak.. logical. 141 transformational component of. 218. 28. 111. 158. 112. 206. 63 defined. 93. 200 Grammar. 66. 77.. 67 transformational. 205 Fodor. 199. 213.207 Ethology. 118. 16.9.. grammatical. 165. 230. 32. 172. 35 first attempts at. 63 in performance models. 194 Du Marsais. 135. 143 Descartes.165. types of constituent stru cture. 198. 79.110. 112. 2011. 47. G. 205. 59. 28. 6. 11. 142. 211 stratificational. 18-27. 199. 212. 2011 Long components. 8. 148-1511 Grice.. 25. 9. 85. 1111 Herbert of Cherbury. 198. 20g Hu·me. 2211 transformational. 224 Katz. 51. 98. 5. artificial. 174. 112. 164-192. 134. M. 214 structure of. 170-184 systems. 27. 207 and linguistic theory. 6 . gB. 2211. 511. 6 traditional. 217 t heory of. 2111 Grammars justification of. 204 Leibniz. 28. 2111 Main·verb. H. 201 88-go. 47-59 Language use. 1211. 45. 116• 117. 40. 196 Mentalism. 110-115. 19B. 1115. 49. 219.. 198 Inversion. John. 1114 Kernel sentences. 5--'7. 144. 52. 205. 47. 117. 118. 110-1111. 228 Jakobson. 51. 142.55. 54. 55. 1112. 17. 75--'79. 18 phrase structure. 1116. 47.176. 1116. 19. 2112 Harman. 20S structural. 25. 79. J. 8. 224 Hall.INDEX Grammar. 122. 205 Inflection processes. 154.228 degrees of. 6. 45-47. 141. 16. 5. 72 Matthews.2112 Inner form. 197. 8. 62. Richard. 99. 28. 2011. 42-45 Nouns. 199. 5 1 Iconic elements o f discourse. 140 Langue·parole. 4 Learning theory. 71. 1116. 227. 57. universal. 65. 198 Notational conventions. von.. R. 122. 214 see also Formatives Lexical rule. 58. lUll Memory finiteness of. G. 228 Halle. 1117. 64. 46. 122. 154. primary. 611. proper. &" Lancelot. 41 recursive property of. 207 acq uisition model. 4. 117. 2011 Lexical entries. 200. 87. 200. 2011 Immediate constituent analysis. 188 Lexicon. 2111 sequential. 197 organization of. 2oS. J.. 117. 711. 121. 87 Natural order. 1110. 15. 57. 2011. Barbara. 120. W. 201. 1711. 205. 14.. 49-52. 11. P. 6. 208-210. 54. lIO. 110. 40 Grammaticalness. 94. 88. 142.. 216. 226 Linguistic theory. A. 19B. 217 Object words. 208 Grammar. lIO. 2111. 204. 19B see also Perceptual model Morpholo gical properties . 57. 211. David. 115. 9. 222. 1211. 70. 89. 210 Held. 14.. 11-0. 98. 87. 58. 221 structuralist. C. 207. H. 54. 1116. 206 Models acquisition. 1119. 49 Humboldt. 84. 201.2111. 206. 9. types of 2 49 (continued) Language learning. 88. creative aspect of. 205. 107. 225 simplicity of. 172. 202. 7 Nesting. 194. 84-88.. 115118. 141.1111. 2211 Locke. 1411. 202. 140. 172-174. Ill. go. 161. 118. 11. 195. 193. 411. 1117.225 Ideas. G. 16.122. 198. 2011 Linguistics mathematical. 51-58. 1116. 4. 61. stylist ic. 25 performance. 1116. 186. theory of. 17. 222 Linguist ic data. 205 Languages. IIll. 217 Rules of agreement. 144. 141. 120. 119. 7 0. 99. 228 Rule schemata. 235 Proper name. 224. 1 55 . 86-g0. 186. P . 137-140. 164-170 Reflexivization. gB. 122. 128 strict. 78. 1l0. 107. 9. 149-160. P. 71. 1!)-21 Operationalism. 1104 Rationalism. 125. 120. 138. grammatical. 123-125. 74. 65--67. 201 Psychology. 95-100. 85. 18. 13 6. 128. 225 Rules branching. 128. 180 restrictive. 28 Phonological admissibility. 235 derived. G 5 Sapir. 140• 141. M. 131. 1121 Processes derivational. 112. 227 Plato. 158 . 157 redundancy. 112. 12 1. 144. S. 133. 199 Perceptual model.. 141. 177. 142. 48 -5 4. 188 selectional. 3. 124. 113 Relative clause. 24. 142. 142.• 160. 223 Outer or external form. 151. 98. Thomas. Ill. 194 Ordering of rules. 74. 141• 197 Phonetics. 162. 135. 134. 219 excluded from categorial com ponent. 142. I lI5. 216 strict subcategorial. 120. 95""97. 123. 88. F. 220. 207 Performance. 39 . 135. 222 phonological. goff. 149. 1911 . 157. 139. 177. 221 selection aI. P. 132• 1 3 6. 155 context-free. 120. 205-207 Redundancy. U9 75. 112. 17. 136. 136. 174.• SchlegeI. 211. 121.• INDEX Representation. 120. 184-192 quasi-productive. 107 subcategorization. E 9 Postal. 135. 161. 67. 99. 65. 154. 1 17. 13-15. 200 Relations. 71. 143.. 1lI2. 137. 22l1. 84. A. 214. 142 context-free. 145. 127. 209 Semantic interpretation. 220. tested. 84. 132. 186 Phonology. 175. 205 . levels of. 16 Phonetic interpretation. 131. 146 Reid. W_ V. 5 1. physiological. G_. 51. 1 23. 2112 Restriction. loti. 14 5. 123. selectional. 70.. 79. 144. 159. 169. 222 base. 215. 136• 194. performance Phones. 117. 131• 134-136• 139. 120--123 context-free. 127 theory of.• Saumjan. 130. 23 1 Phonological interpretation. 141 context-sensitive. 1 68-17 0 rewriting. go--l06. 1 38. 186. 143. 79-83 Phrase-marker. 123 . 121 projection. 197. 136. 106. lUll . 1 12. 45. 70. 144 generalized. E. 113. 4. 139 Rosenbaum. 231 syntactic. 114. 8 Sahlin. 89. de 4. K . 227 context-sensitive. 224 Quine. 135. 134 . 99. 79. 87. 2111 . 154. 106.25° Operational procedure. . 168-170. 1lI8. 47 Schachter. 1 21 context-sensitive. 2 01. 224. 199. go. Ill. 114 113- Quantifiers. 179. 210 Predicate. 9. 87. 51 Post. 200. 72. 66. 10--15 see also Models. 99. W_. 120. 182. 150 • 1 5l1 . 163. 95. 74. 180 Ryle. 112. 111. 117. 2 2 9 transformational. 188 categorial.. 18. 105. 154. 134. 1 04. 91. 121 context-sensitive. 88. 72. 177. 1113-127. 8. 203. 1 03.• 221 Saussure. 113 -12 0. 40• 67. universal. innate. E. 160 T op ic-comm ent. 179. !la. 160. 204 Soboleva. 1114. 165. 59. 212. 84..• 19'7. 194 . 78. 207. Uhlenbeck. Ill. 2!14 erasure. 1&-18. 84. 2!11 phonetic. 115. 199. 224 local. 182. !l5. 141• 162. C 1 63 Wittgenstein. 141.77. 29. 14!1. 62 Twaddell. 132. 51. 177. 157. 48-54. 179-181 208. 36• 41• 55. 223 relative clause. 70. 80 Turing machines. 73. 144. !I!I. 118. III Surface structure. 6!1. 89 125.• - Verificationism. 138. . !l5. 65. 106. 143. 144.220.201 substantive. 188. 14!1. V . 226 Systems of belief.114 defined. 126 . 55. 215 and meaning. III Transformational history. 108-106. 88. 105. 117. u6 filtering function of. 230. I!l5 Word order. hierarchic. 46• 53. 104. 132. 147 grammatical. 187. 135. 32. 1 7. 124-126 Simplicity measure. P. 156 Structures. 191. 138• 140. 177. 163 logical. 182. log. 134.121 Syntax. 207 Subject granImatical. 13. 224 generalized. 121. !l6. 13!1-1!15. 163. 58. 141. 145. J. 199. 55. 13!1. 130 Transformations elementary. W. 85. 206. 122.INDEX Sentences acceptability of. B. 231 Usage. Grammars. 73. 185 Sequence.• 1!1!1. 146. 89. 111. 144. 23. 194 Wilson. F. 225 Transformation-marker. 70. F. 221 Stockwell. 195. 28.160. 135. statistics of. 211 . 182. ag. 77. 117. linguistic. 75. 13. 127147. 207. 1!11. 10. 220. 1 60 semantic. 46• 5!1. 14. 11. L 51 Word. 132-137.113 . 139. 101. 141 terminal. 166 Set-systems.• 19!1. 222. 99. sim plicity of Skinner. 142 strictly local. 147. 100. 11-14 constituent. 101. 218 Strings basic. 217. 210 formal. 27 80. see Evaluation measures. 14!1 Structure index. 234 singulary. 198. 195.• . 29. !l2-84. 135 matrix. 1 1 8. 55. 198. 191.. 185 deviant. 118. 186. 225 Translation. 189. A. M 194 Universals. 139.. 16!1.203. 10. 175 deletion. 196 and performance. semantic basis for. R. 186• 1!11. Transformational cycle. 1115. 2!18 Yngve. 37. 123. 148. 215 passive. 18 preterminal. 66. 1118. 17-!lo. 25-27. 122. 36. 181. 70.
Report "Aspects of the Theory of Syntax - Noam Chomsky (1965)"