Subcategorization and syntax-based theta-role assignment

J O S E P H E. EMONDS SUBCATEGORIZATION THETA-ROLE AND SYNTAX-BASED ASSIGNMENT* Classic problems of how to generalize over predicate-argument relations (e.g., buy vs. sell; spray paint vs. spray a wall) have led to postulating semantic representations which are structured differently than deep syntax, such as (linked) theta grids and (lexical) conceptual structures. I argue that such autonomous semantics massively violates parsimony, and that theta-roles are better predicted by using only modestly enhanced, independently justified deep structures. In addition, I claim that several recent generalizations (of Rizzi, Levin and Rappaport, and Randall) are better formulated as deep syntactic properties than in terms of theta-roles. This syntactic approach to predicate-argument relations thus reinitiates a line of research implicit in Chomsky's Aspects but never developed. The first section argues that only this approach faithfully applies the syntactic revolution to lexical (headcomplement) semantics. Principles invoked include Chomsky's Full Interpretation and Rule for Agents and Talmy's Figure/Ground separation, along with a new Ground Specification and syntactic counterparts to two formal devices from Jackendoff's Conceptual Structures. The thematic role constellations for many verb classes (mostly but not all from English) are shown to follow from these principles. The conclusion speculates that the theta-roles assigned to a sentence are not its properties at a linguistic level, but rather indicate how that sentence is to modify cognitive representations. 1. P A R S I M O N Y IN S E M A N T I C S H u m a n beings obviously possess a linguistic system which incorporates, but goes far beyond, their endless capacity to name. In contrast to Swift's Laputian authors, humans can connect their names in complex creative ways which are far from random - as is well known, only a small percentage of the logically possible combinations of morphemes are well * I am grateful to Wendy Wilkins and Rastko Mo6nik for encouragement of this work and for organizing fora where it was discussed: respectively the 1985 Winter LSA Symposium on Theta Roles and the 1987 Conference on the Formation of Culture in Ljubljana, Yugoslavia. My intellectual debts obvious here throughout are to the pioneering works of Noam Chomsky and Ray Jackendoff. I am indebted to them both as well as to Carol Georgopoulos, Yuki Kuroda and a referee for careful critical readings, and regret that length has prohibited pursuing every comment that merited consideration. I also remain appreciative of the feedback from my fall 1986 seminar at the University of Washington, and especially of the stern critiques of Koichi Takezawa. Funally, my sincere thanks go to Jan M. Griffith of Wordwright, Seattle, who has efficiently prepared many versions of this work. Natural Language and Linguistic Theory 9: 369-429, 1991. © 1991 Kluwer Academic Publishers. Printed in the Netherlands. 370 JOSEPH E. EMONDS formed. The rules for connecting names (more generally, words) are called syntax. Moreover, when humans carry out mental computations (whether verbalized or not) which, to humans at least, seem to far surpass primate abilities, they are said to be thinking. However, we have no introspective access to the microstructure of this thought, nor, non-reflectively, do we have access to the categories of syntax. In investigations of our own capacities, we in fact can only analyze products of our thought and speech: spoken, whispered, imagined, or written sentences. By studying these products, we have been able to construct a discipline of syntax - of explicit rules for well-formed combinations of morphemes. This syntax has its own theory of fundamental categories and of laws of their combination. Now, a reflective non-linguist might plausibly entertain the hypothesis that these categories of syntax are also the categories of connected thought. Or, since language seems to resist facile solutions, that the categories of non-obvious canonical syntactic representations discovered through linguistic argument might also serve as the principal (though perhaps insufficiently elaborated) elements of thought. Such a language of connected thought can also be called (propositional) semantics. What is the alternative? That there is a language of connected thought which essentially is independent of syntactic theory. In other words, in addition to an ability to name, humans have two further independent mental faculties for combining names which set their expressive/communicative system apart from that of the primates. However, such a multiplicity of faculties does not square well with what seems to have been the most fruitful approach to the study of the human in the last century, the human as an object of science. The approach began with a recognition of many apparently distinct properties unique to humans; all and only humans speak in sentences, pervasively use symbols, create tools, exhibit sexual repression, observe an incest taboo, have a long immaturity, are upright mammals, are continually sexually receptive, are religious, have a moral sense, produce humor, know (some of) what they know, recognize beauty, have a history, show a division of labor independent of biology, regularly kill and enslave within their species, etc. By parsimony, a scientific approach could not accept all of these human properties as primitives. As a result, scientific analyses have been forced to reduce rather than multiply the number of independent properties of the species. The central examples of such unifying approaches to the phenomena of the human-specific are Marx's theory of creative labor in the context of scarcity and Freud's theory of sexual repression in the human uncon- THETA-ROLE ASSIGNMENT 371 scious, t Today, one could not claim for systems derived from Marx and Freud a success based on popularity among academics in the industrialized world. Nonetheless, their present unpopularity seems to be accompanied by a growing intellectual void rather than by some fruitful new model of investigation. The fact remains that by far most of what general knowledge has accrued in the last 150 years about how humans act in and against society is based on results from lines of thought that are at least obliquely Marxist or Freudian. Both of these approaches are thoroughly imbued with the imperative to drastically reduce the number of human-specific primitives. Unfortunately, neither of these productive lines of unifying thought has succeeded in linking social and interpersonal human properties to the particular ability to speak (appropriately) in sentences. 2 Nonetheless, I see no reason not to go on seeking explanations in terms of the interplay of a very few basic human drives and faculties. I therefore privilege the hypothesis, attributed above to the reflective non-linguist, that the categories of syntax are the categories of connected thought. Since connected thought is surely but another name for symbolic reasoning or for knowing (some of) what one knows, this hypothesis has its place in the scientific program of reducing the number of human-specific properties we are obliged to postulate. Since the study of language and syntax has passed out of infancy, the nature and some of the combinatorial principles for the categories of syntax are in many ways reasonably well understood. There exist nonobvious canonical syntactic representations discovered through empirical argument that are natural candidates for representing at least skeletal elements of thought as well; in Chomskyan grammar, these representa- 1 SO, for Marx, the ability to create wealth (tools) yields history and a division of labor, and these in turn give rise to war, enslavement, crime; certain forms of morality, religion, and certain forms of sexual repression. For Freud, the ability to reflectively know (and selectively forget), perhaps itself due to extended human immaturity, is the motor for sexual repression, which in turn engenders phenomena such as morality, religion, beauty, humor, the incest taboo, and aggressive actions such as murder and warfare. The early Wilhelm Reich tries to extend Marxist and Freudian analysis to a full theory of sexual repression, morality, and the ideology of scarcity. The Freudian Jacques Lacan rethinks Freud's schema in terms of the pervasive human use of symbol systems, so that the (awareness of) displacement of sexual desire is the prototypical symbolic operation. 2 Some possible connections between ordinary language use and other human creativity are discussed in Chomsky (1968). Perhaps an insight about how other human properties are linked to language can be found in Langer's (1942) essay, where the differentia specifica of 372 JOSEPH E. EMONDS tions are called deep structures. What deep structures consist of has not changed drastically between Chomsky (1970) and the present day, though the mechanisms which are thought to generate them have evolved greatly (Chomsky 1981; Stowell 1981; Emonds 1985, Introduction and ch. 1; Chomsky 1986). As is well known, the degree of fit between deep structures and semantic representations of propositions has always been a point of heated debate in generative studies (Newmeyer 1980). The question is all the more interesting because (perhaps, only because) syntactic theory, in its most influential variants, has been constructed via argumentation which makes minimal use of our necessarily pre-theoretic notions about meaning and connected thought. That is, to the extent that indirect and/or partial representations of thought that can be justified without recourse to intuitions about meaning actually seem to reveal the structure of thought, we can have some confidence that we have been enlightened, rather than trapped by circularity and/or speculation. It would have been rewarding if these syntactic deep structures typically reflected most of what linguists (or others) think they know about the content of connected thought - to use the term of Jackendoff (1987), about conceptual semantics. For Chomsky (1965), the fit between conceptual semantics and deep structures seemed far from satisfying; many of the problems are laid out with admirable clarity in his chs. 1 and 4. For example, in spite of the syntactic parallelism, the locus of the emotion in (la) is Bill, but in (lb) it is John, while (2a,b) exhibit no such variation in the locus of knowledge. (1)a. b. John worries Bill. John worries about Bill. (2)a. b. John knows Bill. John knows about Bill. Conversely, Chomsky observes that the pairs in (3) and (4) are semantically alike (they apparently have the same truth conditions) even though one cannot find syntactic motivations for deriving them from similar sources. (3)a. b. John bought the book from Bill. Bill sold the book to John. human primates is the abilityto manipulate symbols(for her, symbolsare representations of the potentiallyabsent) in the auditory-acoustic mode; both language and music, the first for symbolizingconceptsand the secondfor symbolizingemotion,are thus unique to humans. THETA-ROLE (4)a. b. ASSIGNMENT 373 I liked the play. The play pleased me. Other typical data of conceptual semantics are grouped together in (5) and (6). In (5), the triple (smear, wall, paint) exhibits a common core of semantic content (i.e., thewall is the location of the smeared paint). (5)a. b. c. d. John smeared the wall with the paint. John smeared the paint on the wall. The paint smeared on the wall. Who smeared the paint on the wall? In (6a,b), the relation of the subject noun phrase (as moving object) to the verb of motion remains constant, although adding a direct object to many English verbs expressing motion, as in (6d), necessarily implies that the direct object is moving and that the subject, no longer necessarily involved in the movement itself, is agentive. (6)a. b. c. d. A A A A (*stationary) UFO was passing by. (*stationary) UFO was passing the (immobile) train. (*stationary) UFO was moving by. (stationary) UFO was moving the (*immobile) train. That is, the locus of the necessarily moving object shifts in (6c,d) but does not in (6a,b). Such semantic facts do not seem to be expressed in the deep syntactic representations, essentially the same as the surface structures for these sentences, that can be justified in non-circular (that is, non-semantic) terms. On the basis of this kind of data, as well as judgments of coreference, certain types of shared inference, etc., all investigators seem to take for granted some capacity of "connected thought" or "(conceptual) semantics" which carries out computations relating to sentences but involves more than their syntax. That is, when normal speaker/hearers with differing sets *Bi of beliefs and predispositions "understand" a wellformed sentence Sj, they are thought to be provided with an invariant core of information on the basis of which they can appropriately modify *Bi independently of variation across *Bis, where this information is not all specified in Sfs syntactic structure. This core of information associated with Sj we can call its semantic representation SRj. One is often asked, how much of semantics can be gotten from the syntax? But this question assumes, independent of the syntax of Sj, that we already have a good idea about what is in SRj, which I flatly deny. Such an assumption is question-begging, and has led to many 374 JOSEPH E. EMONDS ideas about SR that are demonstrably ill-founded, or unenlightening at best. Rather, we should ask, how much semantics is there beyond syntax? This implies that elaboration of syntax precedes semantics and allows questions of semantics to be posed more sharply, as exemplified in the just-cited material of Chomsky (1965). Faced with problems such as those in (1) through (6), as well as others brought out in Chomsky (1965), what research strategies might lead to accounts of at least significant subsets of basically semantic data? In the next section, I outline three approaches to semantics, briefly criticizing what I take to be fundamental drawbacks in two, and explaining why I follow a "syntax-based" approach. In Section 3, I implement the syntax of this approach, and in Section 4, the resulting general principles of theta-role interpretation are formulated and defended. Four subsequent sections address semantic problems often treated in the literature, and two final sections relate my syntax-based approach to recent findings in syntax and cognitive science. 2. T H R E E APPROACHES TO N A T U R A L LANGUAGE SEMANTICS 2.1. Syntactically Based Semantics It seems to me that in the last twenty years three research strategies in semantics have been pursued. The one privileged here, which we can call syntactically based semantics, has been only sporadically followed. It parsimoniously holds that the newly unfolding categories and principles of syntax, if carefully refined and interpreted, are equal to the task of representing and explaining significant domains of semantic facts previously considered recalcitrant or autonomous. At the same time, syntactically based semantics holds constant the mode of empirical justification of syntactic constructs, relegating to at most a peripheral role any justifications based only on semantic intuitions. Moreover, it excludes any semantic formalisms whose validity cannot be amply supported by empirical arguments based on syntactic co-occurrence. In particular, syntactically based semantics accepts no categories or representations of formal logic whose existence cannot be justified through syntactic argumentation. The reason for this is the notorious murkiness of the notion of "semantic fact". Apart from certain well-formedness judgments concerning possible co-reference, and intuitions about gross similarities in head-complement relations, semantic-based reasoning often founders on any empirical extrapolation from the simplest cases. Worse, while intuitions abound about, for example, inference, we have no idea THETA-ROLE ASSIGNMENT 375 of when these inferences should be related (i.e., treated as a paradigm) in the absence of any syntactic correlations. In my view, Gruber's and Jackendoff's earlier works on head-complement semantics are syntactically based. So also are the investigations of Chomsky (1970, 1972a, and 1972b), Kuroda (1979), Milner (1978), Banfield (1982), and Reinhart (1983), citing only books among studies not emphasizing head-complement relations. Although Gruber (1965) refers to his work as "generative semantics", this term is historically and I think notionally inappropriate for his approach. Provided the term "syntax" continues to refer to the types of arguments found in "Chomskyan studies" (broadly construed), then I take the following comment of Chomsky (1972b) as summarizing the research direction of syntactically based semantics: " . . . deep structure is a well-defined level w h i c h . . , provides the appropriate grammatical relations for interpretation in terms of 'semantic relations' or 'conceptual structures'". Many of the important early proposals of this approach are summarized or introduced in Chomsky (1972a), in a collection aptly named Studies on Semantics in Generative Grammar. By definition, the other two common research strategies deny that representations derived solely from syntactic argumentation in its usual sense are equal to thc task of elucidating semantics. Although these strategies occasionally appeal to the methods and results of other fields (e.g., vision psychology), the basic set of tools that are imported as the helpmate syntax needs to represent semantics are inevitably those of formal logic and analytical philosophy. But the logical constructs of analytic philosophy are nothing more than some rigorous and internally consistent ways of representing conscious versions of simple thoughts, invented by modern Anglo-American philosophers who use exclusively the anti-theoretical research method of common sense. There is no reason to think that these logical formulae represent the unconscious structure of even simple thoughts, and much less that they can elucidate anything linguistically complex; nonetheless, analogical or literal adoption of some system of logical representation is usually interpreted as bringing rigor to the search for semantic representations. And additionally, in a strange twist of circularity, recourse to logical systems as a tool for elucidating semantics sometimes leads investigators to consider properties of logic themselves as linguistic facts in need of explanation. In all of psychology or linguistics, there isn't a whit of independent evidence that logical devices such as predicate argument structures or standard rules of inference have psychological reality. For example, it seems that understanding natural language does not entail knowledge of the validity of the contrapositive in anything like its general form; rather, 376 JOSEPH E. EMONDS use of the contrapositive must be learned in logic classes. To see this, consider the semantic commonplace that a proposition with a factive predicate and its negation imply the truth of the complement. Thus Chris doesn't realize that Terry is sick (p) implies Terry is sick (q). By the contrapositive (a simple inference from syntax-independent "logical" semantics), Terry isn't sick ( - q ) implies Chris realizes that Terry is sick ( - p ) . This absurd result can be avoided (perhaps) by denying that natural language negation obeys logic (but changing n't to It isn't so that doesn't suffice), but this step just grants m y point. There is no reason to believe that natural language anything obeys logic, even in the most elementary cases. 3 Similarly, "the well-known discrepancy betwecn the material conditional of logic and the 'if-then' of English (which seems to be true only if there is some kind of connection in content between antecedent and consequent) should be enough to warn anyone not to m a k e a blind leap from mathematical systems to ordinary linguistic behavior" (Chomsky 1955, p. 39). Thus, if the US contains 48 states, then Einstein has proposed a theory of relativity is true by logical standards (since special relativity predates Arizona statehood), but is either false or inappropriate in natural language. In fact, two native speakers who I confronted with this sentence, one a professional linguist, declared it unambiguously false. For a semanticist to declare it true is simply to declare that semantics is not empirical, and hence not part of linguistics. In the cited article, Chomsky gives m a n y other examples that one might imagine would have deterred the facile logical enterprise, but it seems to no avail. In their extreme form, the purveyors of logic treat the syntactic basis of connected thought as simply a stumbling block, to be ignored whenever it seems "illogical." Whatever I say below about two schools of thought which less wholeheartedly depend on logical representations and retain 3 Within generative syntax, it is quite common to accept a system like May (1985) which uses operator-variable notation for the natural language analogs to logical quantification (e,g., for all, every, each, some, no, etc.). Since this study concerns the representation of heads and arguments, I cannot enter into disputes about how quantified arguments obtain "scope" over propositions which contain them, such as the conditions under which elements may move to "operator" positions. My view is that empty operators in complementizer position are well justified, but that a logic-inspired system for representating logical quantifiers in a similar way has ended up being entirely sui generis within formal syntax; and further, it seems to have been adopted without systematic investigation of plausible alternatives. I find the criticisms and alternatives discussed in Baltin (1987), Lappin (1984), van Riemsdijk (1982), Whitney (1984), and Williams (1986 and 1988) to be a more promising approach. In this study, we will leave aside debates over whether logic plays a role in establishing linguistic representations of scope. THETA-ROLE ASSIGNMENT 377 some syntactic basis for their representations applies afortiori to formal semantics elaborated without serious reference to the results of syntax. 2.2. Generative Semantics Historically, the first attempt to use ideas from logic in order to incorporate results from generative syntax into more adequate semantic representations is known as generative semantics. For an overview of this school, see Newmeyer (1980). Not accidentally, its adherents for some time referred to the formal semantic objects they proposed as "natural logic". Influenced and inspired by systems of logic, generative semantics extracted quantifiers of nouns to sentence-external position and introduced predicate-argument structure and terminology for sentences ("propositions"). All syntactic coordination was analyzed as deep semantic propositions connected by and and or. McCawley (1970), using dubious argumentation (cf. Berman 1974), claimed that English, at a deep semantic level, exhibits a kind of "Polish notation", in which the verb (predicate) is initial in propositional structure. The weakness of the generative semantics program was not, however, that it supplemented the categories of syntax with new sets of semantic categories imported from logic. On the contrary, it identified the two types of categories, and then embarked on a reductionist program to eliminate whichever syntactico-semantic categories did not fit into a plausible isomorphism with the categories of logic. To my mind, this identification, less the asyntactic reductionism, was a good idea. Under this approach, when you look at a syntactic deep structure of a relatively simple sentence (e.g., John claimed to me to feel sick), you are looking at its semantic structure as well. For example, the following plausible deep structure (7) is quite removed from a well-formed surface structure. 4 4 According to analysis I have adopted or defended elsewhere, the surface morphemes ed and to of both I are inserted only in surface structure, with ed inserted not under I but under the neighboring V, and in addition, the higher I with features as in (7) counts as syntactically specified while the lower one does not; the preposition to is likewise inserted "late" under GOAL; the lower PP (=S) clause undergoes extraposition; its empty P, the empty NPi and the two empty N, while remaining empty in the surface structure, are all licensed by different principles of grammar. 378 JOSEPH E. EMONDS (7) s NPi P I N I John PAST J [ claim PP N p [ P ] S NPi NP SP(N) I N VP V AP feel sick I Of course, investigators don't yet consciously understand how a speaker uses this syntactic structure and its morphemes in a range of semantic and pragmatic calculations, but this is the investigators' problem, not the structure's. Fatally, generative semantics was so influenced by its logic-inspired program that it often took the weak distributional considerations it used to arrive at logico-semantic structures as strong arguments, precisely because the outputs looked like logic. 5 It thus abandoned the more rigorous practices of syntactic analysis, taking a very benign view of exceptions and of a wide variety of rule operations. The classic attacks on generative semantics are in Chomsky's three-article volume (1970, 1972a, 1972b) and Jackendoff (1972). Nonetheless, this school's determination to reduce the number of syntactico-semantic primitive categories led its Chomskyan 5 For example, if a pronominal NP (a "variable" in an "argument" position) could refer back to some other type of phrase or even (discontinuous) string of worlds, that was evidence that the antecedent was in fact an NP - an "argument", in natural logic. The hidden logical assumption, that every word string which enters into (co)reference is an argument of a predicate, was buried. T H E T A - R O L E ASSIGNMENT 379 adversaries to respond in kind. Chomsky's bar notation, in a preliminary version that is still recognizably current, severely constrained the categories of syntax, but in a way not envisioned in generative semantics. 6 For the present discussion, the most serious and quite intrinsic problem with generative semantics is its facile and linguistically grotesque assimilation of the categories P and PP to various elements in the natural logic system (features on verbs, purely surface features on NPs, predicates). This step is exactly what divides "mainline" generative semantics from Gruber's version, in which a PP system firmly rooted in syntax plays a central semantic role. Two decades of subsequent positive semantic results, obtained in the work of Talmy and Jackendoff by according the category P and its complement structural autonomy and centrality, thus retroactively testify to the invidious influence of logic on generative semantics, in which PP and P have no primitive status. 7 In line with the research program of syntactically based semantics, the extensive studies of the central role of P in syntax (Jackendoff 1973; van Riemsdijk 1978; Emonds 1985) demonstrate that this important semantic category is fully justified on purely syntactic grounds. Thus, while I retain the concern for parsimony in generative semantics, I reverse its verdict on P and PP; these categories are important in "natural logic" because syntactic motivations alone establish their central role, and empirical semantic investigations such as Talmy's have confirmed this. The non-occurrence of P and PP in systems of formal logic simply indicates the irrelevance of these consciously constructed mathematical systems to human psychology.S More generally, generative semantics lost sight of the fact that solid syntactic analyses unexpectedly shed light on recalcitrant semantic prob6 The arguments in Emonds (1972) for identifying the then current syntactic category PRT with P (preposition) were inspired by the generative semantic move toward parsimony in categories; less directly, so is the essay in Emonds (1985, ch. 7) to the effect that COMP must be taken as an instance of P. 7 Talmy, who has often identified himself as working within generative semantics, might object that natural logic can accommodate his central category of Adposition (my P) by adding it to its inventory. Such a move is not typical in generative semantics, but of course I welcome any revision of "natural logic" in the direction of syntactically based semantics. Since Talmy uses two levels of categories in his representations, which he labels "semantic structures" and "underlying syntactic structure", it seems to me that methodologically he is rather situated with Jackendoff's autonomous conceptual semantics, though the structures used by the two authors differ; for example, Talmy does not take the Adposition and the "Ground-specifying" nominal to form a constituent together, at either of his abstract structural levels. 8 A most pertinent if minor example of confusion resulting from ignoring P is the exchange between Hacking (1975) and Jackendoff (1979): the logician, blind to the role of PPs in language, raises and fails to solve a comp!ete non-problem; the linguist subsequently clarifies the issues effortlessly by reference to some simple syntactic facts about P. 380 J O S E P H E. E M O N D S lems. Unfortunately, the tradition of assuming that semantically formulated problcms must forever remain so is still with us, so that now, when investigators clarify or solve problems through syntactic analysis, they often go to pains to recast their solutions in terms of the more murky concepts of present-day semantics; cf. the discussion of lexical semantics in Section 9. 2.3. Autonomous Conceptual Semantics Under the assumption of syntactic inadequacy, a third approach to semantics is to hypothesize, either implicitly or explicitly, some representations whose categories and laws of formation are fundamentally independent of those of syntax, but interact with them. 9 I call such approaches autonomous conceptual semantics. Sometimes, as in Zubizarreta (1989), the independent semantic representations are offered as the format for lexical entries; a reviewer (Roberts 1990) calls them "a rather articulated theory of thematic grids". Since many of their categories (e.g., symbols for variables, names of constants, linking devices) are not found in the syntax, their laws of combination must in any case be independent of the syntax. A more radical proposal (Jackendoff 1987) argues for a component of "conceptual formation rules" that yield "conceptual structures". In Jackendoff's recent work, therefore, the categories of syntax are not, properly speaking, the categories of thought. 1° Rather, his conceptual component is autonomous, and hence is subject to "conceptual formation rules" (his Sections 2 and 3). No arguments are provided in defense of the particular format chosen, but, as is usual when syntax is found wanting, logic-inspired predicate argument structures with initial predicates are utilized, and thus give the impression of independence from syntax. Jackendoff's justification for them is the possibility of calculating inferences, 9 An early influential paper with this perspective is Grimshaw (1979), who argues in examining head-propositional complement relations "that the combinatorial properties of predicates and their complements can be explained only in terms of two independent sets [my emphasis, JE] of cooccurrence restrictions". In a separate work (Emonds 1992), I will counter Grimshaw's arguments for semantic selection, and show that syntactic subcategorization captures regularities that are stipulated or unexpressed in her approach. 10 A purely terminological move would be to say that the "categories of thought" include both syntactic and conceptual structure categories. But since there then remains a set of categories of thought which are independent of syntax and obey a separate set of wellformedness rules, my comment in the text is still appropriate. When these syntax-independent representations are purely lexical, as in Zubizarreta (1989), it can be maintained that connected thought is still syntactic. Nonetheless, her lexical entries autonomously link various predicates and arguments, and only then interact with syntax. THETA-ROLE ASSIGNMENT 381 according to familiar logical tools. But for whatever types of inferences are demonstrably natural language abilities, I see no reason in principle why an empirically based inferential system (with a counterpart, for example, to Jackendoff's 1987 inference rule (15)) cannot as well take syntactic deep and/or surface structures as their input. With the proviso that lexical information associated with heads of phrases is part of interpreted syntactic structures, all relations present in conceptual structure are just as precisely represented in deep structures. 11 Since Jackendoff claims that the conceptual categories and their mode of combination are partly justified on the basis of syntactic argument, my main difference with him is that I deny that his conceptual structures are linked in a grammatical derivation with deep structures. For me, his arguments lead only to the conclusion that the syntactic and semantic categories play some role in his conceptual semantics. The crucial property of his conceptual representations that allows them to be conceived of as plausibly linked to syntactic ones by some restricted type of derivation is his further claim that they have the form of propositions. But this "logical form" of his conceptual structures is not justified independently of syntax and/or logic, so the "derivational link" between his conceptual structures and syntactic deep structure is built in by assumption. Thus, I do not deny that syntactic categories and in fact the many purely semantic categories associated with individual lexical items appear in whatever conceptual structures are needed to articulate *B. However, if the specific forms of postulated conceptual structures in *B are to begin with extrapolations from syntax (a justified move in the absence of alternatives) and from logic (as I have argued, largely unjustified), we can't then deduce anything about syntax from them. Rather, the conceptual structures are infinitely more problematic at this time than syntactic ones and, as Jackendoff would be the first to claim, are very largely dependent on the evidence provided by language. It would thus be a total lapse of logic to say that such hypothesized propositional structures in *B explain the presence of propositions in syntax. In work influenced by Jackendoff's, lexical items inevitably include two types of combinatorial devices: "predicate argument representations" and autonomous "lexical conceptual structures" (Rappaport and Levin 1986; 11 The allure of representing semantic relations in predicate-argument structures is that the "already formulated" predicate calculus can be used to perform operations on them. This move is thoroughly anti-scientific, however, since there is no evidence that the predicate calculus reasonably correlates with a native linguistic ability. Analogously, one could conclude that space is Euclidean because more theorems have already been formulated in Euclidean geometry than in any other. 382 JOSEPH E. EMONDS Hale and Keyser 1987). For example, in the system of Zubizarreta (1989), lexical entries of verbs have two representations, Lexico-semantic and Lexico-syntactic Structure. This system further modifies a framework she earlier accepted in Zubizarreta (1985 pp. 247-252), in which the lexical repesentation of verbs specifies (i) the number of a verb's arguments, (ii) the thematic role of each, (iii) the stipulation of which argument is external (Williams 1981), (iv) the syntactic frames for each argument, (v) the distinction of subjects/objects vs. objects of P, and (vi) referential indices on the arguments, sometimes stipulated as coreferential with the subjects or as constants. It is this set of assumptions, commonly accepted by many syntacticians, which I challenge here. As will be seen, I believe that not only (ii), but also (iii), the "linking" part of (iv), and (v) are all misguided. (Assumptions (i) and (iv) without linking are just subcategorization; the stipulations in (vi) can, as will be discussed in later sections, be incorporated into subcategorization.) These assumptions result from construing lexical entries for verbs as a sort of snapshot of the "core meaning" of a verb in whatever is taken as its "canonical" sense. The particular 'thematic roles' of NPs then result from stipulated item-particular linking with certain syntactic positions, rendering real lexical and interpretive generalizations, to my mind, impossible. In Zubizarreta (1989), thematic roles (ii) are eliminated, as purportedly derivable from inherent features such as CAUSE; but this step is criticized as inadequate in Roberts (1990, Section 5). The other information is distributed to the two different levels, but remains in the lexicon. I welcome her move toward "syntacticization," but feel it is still far too stipulative. Zubizarreta's recent system seems to be a natural development of trying to make the inadequate (i) through (vi) more precise and restrictive. It would be fruitless to briefly criticize aspects of her system in isolation. Arguments against such highly structured lexical entries must consist in showing how the generalizations they claim to express can be as well or better expressed in an even simpler syntactically based system. While my efforts here question the necessity of (ii) through (v) above, I will not take on the ambitious task of offering alternatives to Zubizarreta's careful accounts of constructions such as psychological predicates, Romance causatives, and inherent reflexives. ~2 I would rather use much simpler constructions to re-initiate a line of research I consider was implicit in 12 Many of the properties of derived nominals she analyzes I believe are accounted for adequately in Emonds (1985, ch. 1). In work in progress, I analyze English verbal and adjectival passives as well as prepositionless indirect objects in ways that obviate lexical stipulations about argument structures. THETA-ROLE 383 ASSIGNMENT Chomsky (1965), but never developed. As researchers see the implications of this approach, I believe we will offer ultimately more revealing accounts of these many complex constructions. 3. THE CRUCIAL STEP: EXTENDING TO THE FEATURE LOCATION VERBS This section will show how syntactically based semantics can solve most of the problems in (1) through (6). The principles introduced here in a preliminary way will become the core of a syntactically based system for assigning predicate-argument interpretations. These interpretations make it unnecessary to associate syntactic structures with an additional deftvational level, such as Jackendoff's autonomous "conceptual structures". They also obviate their lexical counterparts, the linked "theta-grids" or otherwise complex (e.g., dual) formats for lexical entries which appear in many contemporary proposals, such as Rappaport and Levin's (1986) and Zubizarreta's (1989). At the same time, a full utilization of the syntaxbased principles will require incorporating two formal devices invented by Jackendoff, so his work serves not only as a foil but also as an inspiration. 13 An important part of the poor match between deep structures and semantic properties of the sort shown in (1) through (6) comes down to the variation between direct objects which are "locations" or "goals" and those which are "themes". For this terminology on the semantic (= thematic) roles assigned to NPs, see Jackendoff (1972, chs. 1 and 2). (8) a, b. C. d. (9) a. b. Objects as goal/location: John worries Bill. John smeared the wall with the paint. A UFO was approaching the train. The play pleased me. Objects as theme: John knows Bill. John smeared the paint on the wall. 13 Jackendoff's proposals are the best known to me for representing mental events and states (for h u m a n s and for primates as well). Moreover, some events are fairly faithfully represented by individual sentences (i.e., an otherwise motionless scene in which A child is chasing a dog), and thus such events m a y be associated with sets of sentences true of them. But I do not envision that any systematic or interesting subset of individual sentences can in general be m a p p e d onto events in any meaningful way (e.g., One crucial event has often influenced interest rates). Rather, a sentence itself defines a mapping (whose formal characteristics are at present impossibly obscure) of a belief system into its possible modifications. 384 (9)c. d. JOSEPH E. EMONDS A UFO was moving the train. I liked the play. Previously unremarked is the fact that this variation is no different from that found with the objects of prepositions: 14 (10) X=V X=P Objects are always goal/location (X = + L O C A T I O N ) : worry, approach, please to, through Objects may be goal/location (X = -+LOCATION): smear, pass, load by, with Objects are always theme (X = - L O C A T I O N ) : know, move, like of, about (=/=near) As for prepositions, it is customary to assign them some syntactic or semantic feature, say _+LOCATION, which correlates with whether their object is a goal/location or a theme. This feature is independently justified, at least on P, by the roles that it plays in the co-occurrence restrictions involving P. For example, only a P which is + L O C A T I O N of space or time can accept the intensifier right in SPEC(P), whether or not a given P has an object, i5 (ll)a. John spoke right to the director. The UFO flew right {by/through} (the forest). Put her book right with mine. He dashed right {upstairs/away}. 14 The converses of the statements in (10) do not always hold, for reasons that will be given later. For example, the theme of move is not always its object, which follows from (15) below. 15 PPs include intransitive P such as away, homeward, and uphill as argued in Emonds (1972), and a few unusual PPs which consist of a null P with an NP such as home or there. Such PPs occur in focus position in cleft sentences and in all other PP positions (e.g., the road {home~there} is uninviting). A number of disparate reasons are often adduced for claiming some language lacks P, but the most usual examples betray only a yearning for their language's "special status" among specialists. The uninflected Chinese "co-verbs" are homonymous with V (except for tone) but are unambiguously P by many generative tests (Huang, 1990). For diametrically opposite reasons, Sanskritists often associate PP structures with impoverished morphology, but cannot escape postulating the category unless they rename them as "adverbs" which must appear with an NP (object). Finally, a language such as Igbo has few prepositions, but then French has few eomplementizers. But should a P-less language be found, none of the thematic role principles proposed here would be prima facie falsified. THETA-ROLE (11) b. ASSIGNMENT 385 She remained right in Baku. She finished right before three o'clock. John spoke (*right) of the director. John smeared the wall (*right) with the paint. The agent arrived (*right) by train. John was working (*right) nonetheless. John knows (*right) about you. She worked (*right) for three hours. A second use of the feature LOCATION on P involves subcategorization of V. Chomsky (1965, ch. 2) observes that some verbs take obligatory PP complements. For example, put requires a PP which is +LOCATION. (However, to~from appear with put only in fixed expressions: put it from your mind; put the question to the jury.) Many of his other example verbs require PPs whose heads are necessarily (but not sufficiently) + LOCATION. By feature percolation, all these PPs themselves are + LOCATION. (12)a. Sue put the report {by/down/through/near} the mail slot. We should put the warnings {back/downstairs/at the crossroads/toward the entrance}. *Sue will put the report {of the director/by mail/about the sale}. b. John dashed {to/*of} the director. John was dashing {by/through} (the forest). *John is dashing by bicycle. *John should dash about his report. For other verbs of this sort, both transitive (place) and intransitive (glance), Chomsky's (1965) discussion of subcategorization can be consulted. Any grammatical theory that can minimally express co-occurrence restrictions involving P as in (11)-(12) must use a feature such as -+LOCATION, independently of any property of P's objects. A third use of LOCATION is that it is a necessary condition for the occurrence of -+DIRECTION. As implied in Chomsky's discussion, among the + L O C A T I O N P one must distinguish a binary (or perhaps ternary) feature -+DIRECTION. For example, only [P, + LOCATION, +DIRECTION] is fully compatible with dash (e.g., *he dashed near the beach~at home~on the road), while other verbs such as place have a cooccurrence restriction requiring [P, +LOCATION, - D I R E C T I O N ] . (13) Sue placed the chains {near the beach/at home/on the road/ *through the forest/*to the director/*away. 386 JOSEPH E, EMONDS The -DIRECTION P which satisfy place's subcategorization and the + D I R E C T I O N P which satisfy that of dash t o g e t h e r constitute the class + L O C A T I O N . T a k e n t o g e t h e r then, the need to specify these subclasses of locational P (which b o t h exclude non-locational of, about, [= concerning], for [in place of], despite, etc.) further confirms the existence of the L O C A T I O N subcategory of P. In some way, the feature L O C A T I O N ultimately must play a role in o t h e r constructions as welt. (i) F o r example, a focus PP in an English pseudo-cleft construction i n t r o d u c e d by a where-clause must be + L O C A T I O N . Where this ugly machine goes is {into the trunk~near the others/by the door/downstairs/outside~at the entrance/to the foreman/ *of the director/*by mail/*about the sale/*without postage/*with the repressive atmosphere}. (ii) T h e r e are further features which carve up the d o m a i n of [P, + L O C A T I O N ] . It is a c o m m o n p l a c e of Latin and Classical G r e e k g r a m m a r that, in contrast to Sanskrit, Ps of physical location (say, + P H Y S I C A L ) must be overt, while those P which are + L O C A T I O N but express non-spatial ideas (cause, p u r p o s e , experiencer, animate indirect objects or benefactors, etc.) can be u n d e r s t o o d (in m y view, e m p t y but present). 16 Given these several i n d e p e n d e n t justifications of the feature L O C A T I O N , it is a simple m a t t e r to account for the thematic roles on the objects of P: (14) (15) A n N P sister of a transitive P is a goal/location if and only if P is + L O C A T I O N . T h e t h e m e can a p p e a r in any N P a r g u m e n t position distinct f r o m the goal/location position. 17 A n N P a r g u m e n t of a h e a d X is a direct or indirect object for which X is subcategorized, or X's subject, defined as the lowest N P c - c o m m a n d i n g X1 in all the same N P and S as X1 ( E m o n d s 1985, p. 76). R e t u r n i n g n o w to the V / P parallel in (10), it is an equally simple m a t t e r ~6 In the pseudo-cleft paradigm with where, +PHYSICAL as well as +LOCATION is required. To see this, observe that verbs of saying used communicatively (Zwicky's term), such as yell, can appear with animate indirect objects introduced by P which are then -PHYSICAL; as a result, the pseudo-clefts with where are ungrammatical: *Where John yelled was {at me/to me/past me/for me}. For the same reason we find *Wherewe {depend/insist} is on foreign oil. Nonetheless, such alternating P can certainly be semantically meaningful. Thus, a reviewer's suggestion to identify +LOCATION both with "semantic force" and with well-formed pseudo-clefts with wherewould fail to account for the ill-formedness of the above pseudo-clefts with where. 17 As discussed in the next section, Talmy (1978) considers the distinctness to be a general principle of cognition, not specifically linguistic. THETA-ROLE ASSIGNMENT 387 to account for the thematic roles on the objects of verbs; we simply replace the P in (14) by the X (=N, V, A, P) of the bar notation. By the theory of theta-role assignment justified in Emonds (1985, ch. 1), A and N do not take NP sisters, so this generalization has an effect just where we want it to, on the objects of V. Thus, abbreviating LOCATION as L in lexical entries, we can postulate lexical entries for the verbs in (8) through (10) as follows: 18 (16)a. b. c. d. e. f. g. h. worry, V, +L, + (about)^NP know, V, - L , + (about)^NP smear, V, - L , + N P ( P P ) 19 approach, V, + L, + (NP) move, V , - L , + (NP)([PP, +L]) pass, V,-+L, + (NP)([PP, +L]) please, V, + L, + NP like, V, - L + NP In my view then, Vs and Ps share the property of uniquely determining, via inherent features such as -+L, the thematic roles of their objects. There are ample independent justifications for using features of the category P such as + L to also distinguish subclass(es) of V. Besides the original work of Gruber (1965) on the incorporation of P features into V, much of Talmy's work has demonstrated how extensively features of various adverbial constituents (among others, PPs of path) are "conflated" with a governing verb. For example, one striking generalization discovered by Talmy (1985a) involves a contrast between all Romance languages and all Germanic and Slavic languages. In Romance, a few motional verbs incorporate the path features that in English typically show up on prepositions such as the intransitive post-verbal particles (up, down, in, out, back, away, etc.). French examples are verbs such as monter, 'go up', descenclre 'go down', sortir 'go out', entrer 'go in', partir 'go away', rentrer 'go back', etc. 18 For typological convenience, I introduce a caret ^ in subcategorization features for linking grammatical formatives and phrases to replace the arch . . . . . of Chomsky (1965) and Emonds (1985). Throughout, about with the locative meaning "near" or "around" is excluded to simplify the exposition. That is, about is - L here. 19 These frames do not include details irrelevant to the main argument. The Ps after smear, for example, cannot be what is termed - G O A L in Emonds (1985, ch. 5): *He smeared the paint from the wall; *he smeared the wall of the paint. It is of interest, however, that this generalization is incapable of being expressed in a "theta-grid", since it affects the P in PP arguments which have different thematic relations. 388 (17) JOSEPH E. EMONDS Jean a sorti la bouteille John brought out the bottle. Nous avons descendu nos photos We brought down our pictures. Marie a mont6 les v&ements Mary took up the clothes. In these French examples, the path sense of the bolded English prepositions, expressed by subcategories of P such as + H I G H (up, down), is realized by the same features occurring on V. A Romance motional V with such inherent path features may co-occur with a directional PP which, as is typical, expresses path and goal: (18) Marie est entr6e dans le lac. Mary entered into the lake. La voiture est descendue en bas de la colline. The car went down the hill. Je montais derri6re la maison. I was going up behind the house. Talmy has discovered that, in contrast to what occurs in Germanic or Slavic languages, Romance verbs which incorporate a feature of manner rather than of path are uniformly incompatible with directional PPs that express path and goal; the Romance PPs in (19a) are simply adverbs of static location. (19a). Marie a nag6 dans le lac. Mary swam (around) in the lake, not Mary swam into the lake. La voiture a gliss6 en bas de la colline. The car skidded at the bottom of the hill, not The car skidded to the bottom of the hill. Je courais derri6re la maison. I was running (around) behind the house, not I was running (to) behind the house. The excluded interpretations in French must rather be rendered by paraphrase, as in (19b): (19)b. Marie est entr6e dans le lac en nageant. Mary entered the lake (by) swimming. La voiture est arriv6e en bas de la colline en glissant. The car got to the bottom of the hill (by) skidding. J'allais derri6re la maison en courant. I went behind the house (by) running. THETA-ROLE ASSIGNMENT 389 The Romance contrast in (18) and (19) seems to be best described as follows. (i) A single verb cannot simultaneously incorporate a path feature and a manner feature. (ii) A motion verb cannot occur with a directional PP constituent (expressing path and attained goal) if the verb incorporates a manner feature. 20 We need not decide which of (i) and (ii) is specific to Romance; crucially, (i) is formulated in terms of a V incorporating a subcategory (a path feature) of P; simply the existence of this incorporation, either in universal lexical semantics or in language-particular semantics, is what interests us. This account of the contrast in (18) and (19) confirms the intuition based on (17) that lexical classes of French verbs (such as monter, sortir, etc.) are distinguished by means of path features such as "--HIGH ordinarily associated with prepositions. In the next section, more evidence for the existence of the prepositional feature - L on V will be presented, which indicates that L is visible at the post-transformational "s-structure" level and hence thoroughly syntactic. Let us now examine all the lexical possibilities for a verb like pass. Since two different NP arguments of the same verb cannot get the same thematic role of goal/location (one effect of the Theta Criterion of Chomsky 1981), transitive pass and its complement P cannot both be + L. When pass is intransitive, we cannot tell if it is + L or - L . (20) pass ambiguously - L : John passed (into the closet). pass = +L: John passed the closet (*into the hallway). pass = - L : John passed the peanuts (into the hallway). *John passed (the closet) of the peanuts (P = - L ) . 21 We can see from the above that the supposed indeterminacy of the thematic relation of an NP in a given position in examples like (1) through (6) is already evaporating. Some first steps in the direction of predicting particular theta-roles in terms of grammatical configurations are taken in Culicover and Wilkins (1984, sections 1.3 and 3.1.3). Here, I claim that the syntactic location of these thematic roles can be fully determined by general interpretive principles akin to (14) and (15), and depends only 20 Talmy claims further that the Romance family is alone among Indo-European languages in behaving this way, although other languages can exhibit the patterns of Romance. 21 Throughout, I assume that prepositionless indirect objects as in John passed Mary the peanuts are derived from John passed the peanuts [e,+Lf)] Mary, as argued in Emonds (1972, i986). 390 J O S E P H E. E M O N D S on the distribution of the intrinsic, non-contextual feature -+L on categories like V and P. 22 If we can predict the thematic roles of objects from (15) and an extension of (14), we can impose the following restriction on the type of information that can be listed with lexical items, as exemplified in (16). (21) Semantic Atomism: A lexical entry may stipulate non-predictable (=item-particular) relations with a complement only by subcategorization for the complement and for its intrinsic features. 23 A consequence of (21) is that neither "theta-grids" nor item-particular "theta-linking" specifications are allowed in lexical entries, a drastic, massive, and most welcome purging of semantic information from lexical entries (what the child must learn). There are of course intrinsic and indisputably semantic features associated with morphemes (words have meaning in isolation in any case), but a word cannot idiosyncratically "project" a semantic interpretation such as a theta-role onto a complement, except perhaps in idioms with highly specialized meanings. Moreover, for the claim in (21) to hold, it doesn't matter whether L is semantic, since L is an intrinsic feature (which, as seen earlier, interacts with specifiers and higher heads) and not a feature selecting a complement. 22 I thus reject the distinction in Bresnan (1982) between "semantically unrestricted" objects of V and "semantically restricted" objects of P. Bresnan argues that "semantically unrestricted" should replace c-command as a necessary condition on predication, a view seconded in Zubizarreta (1985). She claims that c-command cannot distinguish between often prepositionless dative NPs, of which adjectives and floating quantiflers cannot be predicated, and direct object NPs, sometimes accompanied by apparent Ps (Spanish a and French de), of which adjectives and floating quantifiers can be predicated. However, in Emonds (1985, chs. 1 and 5, and 1987) I argue that dative and genitive prepositions, and oblique morphological case all realize empty deep structure Ps; c-command then perfectly well excludes these NPs from being subjects of non-c-commanded predicates. Just as datives are then universally 1,[{D]-NP in deep structure, so also the Spanish a and French de found with certain subjects and objects are not structural Ps at s-structure. Thus, c-command remains adequate for predication, and no new primitive, "semantically unrestricted", need be introduced. 23 The domain of subcategorization which relates V ° and phrases is the relation of sisterhood with (or government by) V °, as in Chomsky (1965). The sorts of features in (16), given the subsequent clarification of the notion of feature percolation, are exactly what we find in that work. His more recent Projection Principle is therefore, as in its first formulations, a requirement that subcategorization frames be respected (presumably by the "same" XPi) at all syntactic levels of a derivation. In my view, however, lexical subcategorization features are not trivial "snapshots" of grammatical deep X 1. For discussion of economy principles which reconcile lexical features with general syntactic requirements, see Emonds (1990). In particular, while NP subjects require no subcategorization, exactly as in Chomsky (1965), subjects of special form (clausal subjects and null or "impersonal" subjects) require features like +S and +g} . By THETA-ROLE ASSIGNMENT 391 In recent terms, (21) means that autonomous "semantic selection" ("sselection") by individual lexical items (Pesetsky 1982) doesn't exist; semantic selection is at most a secondary annotation on syntactic frames which selects intrinsic features of the complement, exactly as in Chomsky (1965).24 Semantic Atomism thus clashes with Grimshaw's (1979) proposal that there exist purely semantic (= syntax-independent) well-formedness conditions on predicates and their complements. While I agree with Grimshaw that item-particular syntactic subcategorization and general semantic principles "impose well-formedness conditions on different levels of representation", I argue in Emonds (1992) that only a single syntactic subcategorization feature is needed to select complement clauses, even though this feature can come into play at two different levels of a derivation, that of lexical insertion and that of semantic interpretation. The present framework, based on (21), is therefore squarely counterposed to any selection which operates in the autonomous "lexical conceptual structure" of recent work on verb classes. While one might object to the vague semantic terms in such studies ("MEANS", "STATE", "come to be" in Rappaport and Levin 1986; "undergo change", "material integrity", "linear separation" in Hale and Keyser 1986), what seems to the principle of indirect theta-role assignment justified in Emonds (1985, ch. 2), the resulting subjects have the following deep forms, respectively. (i) NP (ii) NP !1 N g N I In addition, economy principles suffice to license these empty N only at deep structure; at s-structure, further operations such as extraposition, topicalization, and /t-insertion are required. 24 My position is thus that the mechanisms for predicate-complement relations involve no more recourse to semantic features than exactly what is countenanced in Chomsky (1965); the syntactic feature X °, + YP may (only) include a further selection restriction (or impose an interpretive condition) of + F on YP, where F is an inherent syntactic or semantic feature of yO (and not a relation between X and Y). One might want to exclude semantic features even here, but examples such as the verb drink (cf. Section 5 below) seem to warrant purely semantic features such as + L I Q U I D ; a "liquid" interpretation must be imposed on the direct object in sentences such as She drank {the butter~the instant coffee/??the ice/??the pills}. Thus, I claim that any "s-selection" (as in Chomsky 1986) exists only to the extent it is present in Chomsky's (1965) selection restrictions. 392 JOSEPH E, EMONDS me unscientific about this work is the intermingling of these terms - which might be the best available to intuitively describe the meanings of words in isolation - with item-particular linking variables which supposedly account for syntactic well-formedness. My point is to show that such linking variables are superfluous. Once they are removed from predicate-argument structures (e.g., those of Rappaport and Levin 1986, p. 9), what remains is simply syntactic subcategorization, which is all that is needed for well-formedness, and a residue of traditional semantic terms ("liquid", "artifact", etc.) whose less than formal nature in no way impinges on the operation or explicitness of generative mechanisms. 25 Before continuing with more evidence for the feature L and for principles (14) and (15), it is appropriate to sharpen our terminology and to flesh out our set of lexical and interpretive principles for syntactically based lexical semantics. Given the preliminary stage of this research, I will attempt to predict only the thematic roles for verb classes which have hitherto been extensively discussed in terms of theme, goal, source, and agent - namely, verbs of physical or psychological motion, location, ownership, or communication. For expository purposes, I say that all such verbs are "thematic", notated +T, without any implied claim that there are non-thematic verbs. I will not discuss the roles of NPs in adjuncts, such as instruments, benefactives, passive agent phrases, etc. 4. GENERAL PRINCIPLES THE OF T H E M A T I C FIGURE/GROUND ROLE ASSIGNMENT: DICHOTOMY The first principle we need, discussed in Chomsky (1986, section 3.3.3.3.2), is a general requirement on interpretation, which I formulate a s f o l l o w s : 26 25 In Bresnan and Kanerva (1989, section 6), the linking conventions are called "lexical mapping (or encoding) principles". Even though the authors allow a certain freedom between theta-roles and grammatical positions, they still are curiously and needlessly restrictive: "Cross-linguistically, the theme or patient is canonically encoded as an unrestricted function, either subject or o b j e c t : . . . " ; "there is cross-linguistic evidence that locative arguments alternate between oblique and s u b j e c t ; . . . " Much discussed patterns such as they drained the sink of water and they filled the car with sand falsify such statements (which are later crucially used in one section, although hedged in another). In light of the present analysis, when these inadequacies are removed, what results is the present system, less the unnecessary "lexical role structures" which replace subcategorization. In addition, the authors extend their proposals to cover passive agent phrases, which I exclude from discussion here. 26 Actually, Full Interpretation is a special case of a more general requirement that all lexical X ° in a single root constituent (e.g., an S with no parentheticals) be connected to every other through a chain of theta-relatedness", yO and Z ° are "theta-related" if and only if the THETA-ROLE (22) ASSIGNMENT 393 Full Interpretation: Every NP which is an argument of a y 0 must be interpreted as one from a specified list of argument types, unless a particular grammatical position is lexically specified for a particular y 0 as taking no argument (a "theta-bar position"). Secondly, we must also express the definitional claim in previous discussions of thematic relations that, for "thematic" verbs at least, an NP bearing the role of theme is in some sense obligatory. In both autonomous and syntactically based lexical semantics, this is basic, although in both frameworks a theme of a +T verb can sometimes be either understood or syntactically expressed. 27 I will return to this below, after introducing an important shift in terminology. In a series of papers, Talmy (especially 1975, 1978, 1983, and 1985a) provides extensive justification for a uniform treatment of the semantic (thematic) roles of goal/source/location with predicates of motion and location. For any such phrase, which is typically an NP, Talmy proposes maximal projection of one (say Z max) is in a position to receive an interpretation as an argument or an adjunct of the other (y0). 27 My claim that the theme is always present has led a reviewer to ask if there is a covert phrasal theme in We butter bread or They skin peaches. Such denominal verbs suggest to me rather a slightly modified version of Walinska de Hackbeil's (1986) empty-headed zero derivation, as in (i)-(ii): (i) (ii) S S NP I VP V NP w°s-"-v butter NP ~ VP V NP i skin Suppose that these verbs are derived from the following lexical entries: butter, N, ([v skin, N, ([v ], +L, + G O A L , + ], + L , - G O A L , + NP) NP) By the bar notation and the right-hand head rule for compounds, the frames Iv ] can be realized only as in (i)-(ii). The induced empty lower V are unspecified for L O C A T I O N (as discussed in Section 6, - L is the unmarked value for V), so their N sister may be assigned theme, but not location. By Lieber's (1983) Percolation Convention III, the higher V will inherit specification for L and for _+GOAL from its other daughter (N), and so will assign Ground roles (goal and source respectively) to its direct objects. 394 JOSEPH E. EMONDS the single term "ground" (in contrast to "figure", his term for theme). He claims that the semantics of the path or site of motion, typically expressed with a syntactic P, are better understood separately from the notions expressed by the objects of these P, which are those of the "ground" for the movement or location. 28 It will be useful to have Talmy's (1985b) definitions at hand. (23) Figure ("theme" is the term of Jackendoff and Gruber): a moving or conceptually movable object whose site, path, or orientation is conceived as a variable the particular value of which is the salient issue. Pretheoretically, in an " u n m a r k e d " configuration, the "figure" is the subject of an intransitive verb and the object of an agentive transitive verb, as in (24): (24) The train moved. The UFO moved the train. (25) Ground: a reference object - itself having a stationary setting within a reference frame - with respect to which the figure's site, path, or orientation is characterized. Talmy observes that the "unmarked" realization of a path and a ground is in the form of an adposition P and its syntactic object NP. (Here throughout, I use the prepositional language English.) He further remarks that a path can be realized as a case inflection on the ground NP. In Emonds (1985, Section 5.7 such case inflections are shown to be realizations of an abstract case assigned by a deep structure P; this P can be phonologically zero if in fact the abstract case is "productively" realized morphologically on the object NP. 29 This P which induces case inflection can also be used as the source of the path reading so that the syntactic realizations of a path are uniformly deep structure Ps, for the types of predicates under discussion. Let us now determine which NPs in a clause are associated with Talmy's thematic roles of figure and ground. His work repeatedly stresses that language imposes the semantic figure/ground asymmetry on two different 28 In sentences such as John ran from behind the barn, I imagine we still want to say that the ground is a PP, itself a complement to the P from, as in Jackendoff (1973). 29 More generally, I there formulate and defend an "Invisible Category Principle", which licenses empty closed categories such as P, INFLECTION, D E T E R M I N E R , etc. when the syntactic features of these closed categories are "productively realized" (defined in the work cited) morphologically on the phrasal sister of the closed category (respectively, NP, VP, N 1, etc.); cf. also Emonds (1987). THETA-ROLE ASSIGNMENT 395 NP arguments of a verb of motion or location. That is, the figure and ground are always distinct, which no doubt reflects an important dichotomy present in the cognitive psychology of all higher animals. Thus, any adequate principle of interpretation must assign the figure and ground classifications to separate argument NPs. (26) Figure specification: For any lexical IX °, +T], exactly one Figure NP, distinct from the Ground, must be present among the deep structure arguments of X °. While many of the properties by which Talmy distinguishes figures and grounds seem intuitively right, those not having to do with PP structure completely escape description in formal grammatical terms. For example, he points out that figures are typically smaller and/or geometrically simpler than grounds, yet we can have Our computation center (figure) should be built near a water shutoff (ground). Talmy doesn't deny such variations; rather, I imagine he would say that in this example the overriding pragmatic impact of the water shutoff is that it is "more familiar" (another of his criteria for grounds). While grammatical theory cannot further formalize such properties, it must identify which NPs are the Figure and Ground inputs to pragmatic calculations. 3o While (26) insures that the Figure and the Ground will be separate and thus incorporates (15), it does not suggest, correctly I claim, that the Figure (theme) typically appears in a given syntactic position. Rather, the Figure may appear in any argument position. Does the same thing hold for the Ground? It is a commonplace in research on the linking of thematic roles to syntactic positions that notions such as Ground (i.e., goal, source, and location in works other than Talmy's) can occur in a verb's subject or object position, given an appropriate lexical choice. Thus, in (5a) and (6b), but not (5b) or (6d), the direct object is the goal (Ground) NP; in (la), the object Bill is also the goal or location of worry, while in (lb) the subject John has this role; (2a,b), in contrast, show no such shift and the subject John is the location (Ground) of know in both sentences. (1)a. b. John worries Bill. John worries about Bill. 30 Other researchers have proposed properties dependent on thematic roles which are more syntactic than those given by Talmy; Culicover and Wilkins (1984) claim that pairs of controlling NPs and corresponding predicates in various predication structures are limited as to which thematic roles they bear (the translation between their terms and Talmy's is straightforward). 396 JOSEPH E. EMONDS (2)a. b. (5)a. b. C. d. (6)a. b. C. d. John knows Bill. John knows about Bill. John smeared the wall with the paint. John smeared the paint on the wall. The paint smeared on the wall. Who smeared the paint on the wall? A A A A (*stationary) UFO was passing by. (*stationary) UFO was passing the (immobile) train. (*stationary) UFO was moving by. (stationary) UFO was moving the (*immobile) train. Similarly, the many studies of "psychological predicates" distinguish arbitrarily what Brekke (1988) calls "o~-Experiencer" predicates (e.g., hate, like, fear, etc: the subject NP is the "Experiencer" or Ground) from "/3Experiencer" predicates (disgust, please, scare, etc.: the complement NP is the Experiencer).31 These variations in syntactic positions for thematic roles (in Talmy's terms, in the position of the Ground in "marked cases") have served to justify "autonomous semantic" conceptions of thematic roles, whereby links of particular roles to specific syntactic positions, as effected by for example (14), are taken to be unpredictable; i.e., to hold only in the "unmarked cases". In the general case, the thematic roles of particular verbs are assumed to be unlinked to syntactic positions and represented rather as an autonomous "theta (thematic role) grid" for each lexical item (this term is found in Stowell 1981, but the concept has a longer, if vague, history). In my view, this is a radical and unreflective overstatement of the freedom of syntactic positioning available to phrases bearing particular thematic roles. The indeterminacy of linking can be eliminated, at least for the cases systematically discussed in the literature, if we know for each verb which can take a Ground whether its direct object can, must, or may not express this thematic relation. As seen in the previous section, the object of smear can be, of move cannot be, and of approach must be, the 31 I have not been convinced by attempts such as Belletti and Rizzi (1988) to derive all of both types of psychological predicates from deep structures in which themes are always direct objects, given the full agentivity (and ambiguity) possible in crucial examples such as Mary amused herself (when she tried to fix her car). Of course, if Chomsky's principle of assigning the agent role, (33) below, could apply at s-structure, I could accommodate Belletti and Rizzi's syntactic derivation, since agent interpretation is the only statement in my interpretive system which refers to subjecthood. THETA-ROLE ASSIGNMENT 397 Ground. Similarly, a direct object of worry must be the Ground, while the object of know cannot be. In the lexicon, then, a verb whose direct object can be a Ground (smear, load, drain, answer, teach, pay) is entered as (+LOCATION), and one whose direct object must be a Ground (fill, cover, rob, please, worry, obey) is, without parentheses, +LOCATION. Parentheses in lexical entries have their usual interpretation, indicating that what they enclose may or may not be present in a well-formed deep structure. Verbs which cannot have a Ground as a direct object are unmarked for LOCATION. Beyond these lexical specifications, the only mechanism needed to account for instances of "marked linking" (between thematic roles and syntactic positions) is generalizing (14) from pO to yO, a step which is in effect a simplification of (14): (27) Ground Specification. A direct object NP of a transitive y0 is a Ground if and only if yO is + LOCATION. For example, the verb smear, which is entered as V, + ACTIVITY, (+L), + NP(PP) will give rise to the following deep structures: (28)a. vI V NP I I s,. r ,a,°t b. iP v1 V NP liP tJ [?L] i on N, I the wall owal, L with I paint We observe that the Ground, determined by (27), in contrast to the Figure, may but need not be present among the deep structure arguments of a given [Yi, +T]. The Ground thus contrasts with the Theme, which, for the + T verbs under discussion at least, is generally necessarily present among a predicate's arguments. Thus, we find ?Bill smeared the wall, ?Sue stripped some trees, even though subcategorization is satisfied. We can construct an additional argument for lexically representing verbs with direct object Grounds as + L O C A T I O N by using the generalized specification for Ground (27) to explain an enigmatic pattern in derived nominals. Rappaport (1983) has observed that a range of direct object 398 JOSEPH E. EMONDS NPs cannot be introduced by of in derived nominals. In particular, goal NPs are unacceptable in this context. (29) *John's amusement (interest) of the children with his stories. *Mary's worry of her relatives comes as no surprise. *Our obedience of that law caused us grief. *John's entry of the room was a surprise. *Don't worry about a little smear of the wall. *They were surprised at my answer of a professor. *She said that your payment of the company was overdue. *A lecture of the students on this issue would be in order. In Emonds (1985, ch. 1), I argue that the usual of-phrases in derived nominals, like all other complements to N, are necessarily instances of PP structures. A transitive verb's feature + NP cannot therefore be satisfied trivially in N1; instead, an alternative "indirect" mode of thetarole assignment developed there assigns the theta-role ordinarily associated with the direct object of V to an NP in a structure [pp[pO]-NP]. For NP, appears example, the derived nominal of the verb hate, which is + as follows, where the NP is the Figure (theme). Of is later inserted under p , 32 (30) N1 [V,-L] I hate N P I1 red 0 NP imperialism Now, if our use of -+L on V is legitimate, the verbal bases of the derived nominals in (29) are all + L . By (27) right to left, the direct objects of such Vs are Grounds, and by indirect theta-role assignment, so must be the italicized NPs in (29). For example, the NP in (31) must be a Ground. 32 In derived nominals, a range of features including subcategorization and also -+ L percolate up to N from the left branch, because the suffixal head N is unspecified for such features. This is Feature Percolation Convention III established in Lieber (1983). THETA-ROLE 399 ASSIGNMENT N1 (3~) N [V,+L] I amuse ~ N p p P III ment D NP the children If the italicized NPs in (29) are Grounds, then the introductory Ps in (29), by (27) left to right, have to be + L. But, as we have seen, o f is always - L , so (27) correctly excludes the examples of (29). Only in cases where a preposition appears which is specified as + L, italicized in (32), do the appropriate derived nominals exist: (32) Our obedience to that law caused us grief. They are talking about your leap o v e r the fence. John's entry i n t o the room was a surprise. Don't worry about a little smear o n the wall. They were surprised at my answer to a professor. She said that your payment to the company was overdue. A lecture to the students on this issue would be in order. Thus, principle (27) and representing verbs as - L , taken together with the analysis according to which complements to N are uniformly required to be in PPs (Emonds 1985, ch. 1), are independently supported by the contrasts in (29) vs. (32), which are automatically predicted. 33 Finally, principles (22), (26), and (27) are complemented by one further principle in universal lexical semantics which specifies when NP arguments are agents. To my mind, the studies of thematic relations have never improved on Chomsky's formulation: (33) Agent Specification. "Thus one rule (probably universal) will stipulate that for verbs of action, the animate subject may be interpreted as the agent, etc." (Chomsky 1972a, p. 75) 33 It would seem that the of in "action nominals" (our obeying of that law, yowr leaping of the fence) is generated without - L, perhaps because it is inserted only in the phonological component, subsequent to the level of s-structure where (27) presumably still holds. C h o m s k y (1970, pp. 58-59) observes, "there is an artificiality to the whole construction that makes it quite resistant to systematic investigation. Furthermore, the construction is quite limited". 400 J O S E P H E. E M O N D S When an NP is not an agent, it must of course still be interpreted, either as a Figure or Ground. We will see how this comes about in detail in the subsequent sections. In addition, nothing here prevents a Figure or a Ground from receiving an Agent role too, a correct prediction which nonetheless has befuddled many previous attempts to "parcel out" thematic roles among NPs. Cf. the ambiguities of The bird flew across the room and Sue amused herself when she tried to fix her car. Cf. also note 39, regarding purely instrumental subjects, which I do not attempt to cover in this study. With the four above principles, we can predict the distribution of all thematic roles for + T verbs on the basis of deep structures, given the lexical device introduced here of being able to list either prepositions or verbs as +-LOCATION. These predictions are far from trivial, however; for example, the generalized Ground Specification (27) characterizes where, inside a deep structure X I, we will find a Ground NP - without mentioning under what conditions a Ground can occur in subject position. To see how the devices introduced so far allow us to eliminate entirely "thematic grids", specification of external argument, and all item-particular linking specifications, we must develop the formalism for lexical entries and examine in several cases, many of them complex, how these entries and the interpretive principles (22), (26), (27), and (33) - each in themselves simply stated and exceptionless - combine to yield the variety of role configurations that various verbs enter into. My approach conforms to a perspective implicit in Chomsky (1965, ch. 2); but, besides the principles for interpreting thematic roles just introduced, we will need certain minimal enrichments of lexical specifications beyond those proposed in Chomsky (1965), in order to overcome problems brought out in that work concerning the expression of certain semantic equivalences in deep structures. Interestingly, the modifications required are exact analogs to those proposed in Jackendoff (1987). Indeed, the formal essence of one of his proposals, as will be seen, permits us to entirely eliminate theta-grids and theta-linking from individual lexical items. Without it, this research tack, even with the modifications so far proposed, would founder on pairs like (3). But at the same time, since Jackendoff's devices do have exact counterparts in my syntactically based system, the demonstrable need for these devices in no way constitutes evidence for a separate derivational level of "conceptual structures". 5. RETHINKING JACKENDOFF~S REPRESENTATIONS My rejection of Jaekendoff's autonomous conceptual structures as a derivational level should not be understood either (i) as a claim that all THETA-ROLE ASSIGNMENT 401 semantic specificity can be reduced to syntactically justified representation or (ii) as a denial of the validity of certain of Jackendoff's proposals for capturing lexical properties. Regarding (i): there can be a great deal of non-syntactic specificity in adequate semantic representations, to which the syntax has no access. Such specificity is parallel to phonological specificity. However, much of what is purely semantic is associated with, and presumably represented in, the full lexical entries of the morphemes in the "head categories" of N, V, A, and P. Semantic features of arguments can then result from general principles operating on intrinsic lexical features of these heads and item-particular subcategorization frames. Like other general rules of semantic interpretation whose outputs are inaccessible to syntax (e.g., rules of disjoint reference, rules of quantifier scope, etc.), rules assigning semantic features to arguments can apply to syntactic structures. Such a conceptualization of phonology and semantics as specific but not autonomous was articulated in Chomsky's lectures of 1967. As an example of semantic specificity, consider the notion "patient." I completely agree with Culicover and Wilkins (1984) and Jackendoff (1987) that "patient" and "theme" are two independent semantic notions, the former related to "what is affected" in a proposition and the latter to the semantics of motion and location. 34 As discussed earlier, there is no correlation in principle between the semantic theme (=Figure) and any given syntactic position. However, as Jackendoff says, there is a correspondence between patient and direct object position (first discussed in generative terms in Anderson 1971). For syntactically based semantics, the research questions then naturally become: (a) Are there definable conditions under which patients are not direct objects? (b) Under what restrictive conditions are direct objects patients? (c) What might be the empirically based rule(s) of inference involving "patient" used in constructing full semantic interpretations? 35 In general, autonomous lexical 34 The notion of theme, which is more central in Jackendoff's work and in our concerns here, can overlap with that of patient. Thus, in John put the book in the box, the book is the patient of p u t and the theme of the motion or location expressed by p u t in the box. In John saw the book in the box, the situation is parallel, except that the book is not a patient. Rappaport and Levin (1986) try to collapse theme and patient, I feel unsuccessfully. They claim that (i) loaded a truck with the books' includes (ii) loaded the books onto a truck in its lexical conceptual structure. But (i) does not imply (ii); books may be left over in (i), and onto is not implied by (i) (replace truck by refrigerator). 35 Some detailed empirical and theoretical work on the notion patient is that of Zubizarreta (1989) on "affected" and "unaffected" arguments. She examines constructions which exclude unaffected (non-patient) NPs, which correspond to objects of certain verbs in active constructions, and uses a device of lexical incorporation into V to account for where affected patients may occur syntactically. I do not offer a comprehensive alternative to her proposal; nonetheless when she points out that "the morpheme zich cannot be syntactically bound by 402 JOSZeH z. EMONDS or conceptual representations of such restrictions and interpretations do not really explain them. Rather they tend to mask the fact that we don't understand them. Regarding (ii): rejecting a conceptual structure level of derivation does not entail rejecting all the notational devices, lexical properties, or generalizations that Jackendoff (1987) chooses to formulate at that level. Rather, I argue that such semantic characteristics, to the extent they can be interestingly formalized, are equally well (and hence preferably, by the earlier methodological considerations) expressible in syntactically based representations, and are often completely predictable from them. Naturally, syntactically based formulations of his more interesting proposals will not be immediately apparent, since otherwise he would not have proposed autonomous conceptual structures to begin with. To exemplify how a syntactically based approach can handle directly some of the lexical properties and notational devices discussed by Jackendoff, I re-examine two lexical entries for verbs which he develops in detail. One general point he makes repeatedly, also supported in Zubizarreta (1985), is that lexical entries (e.g., of verbs) can supply a semantic representation with understood arguments that are syntactically absent. For example, with drink, the direct object (the theme and patient) is obligatory in the semantics but optional in the syntax, while an understood complement of place (meaning for Jackendoff roughly "into the mouth of NP/', where NPi is the subject of drink) is obligatory in the semantics but absent in the syntax. In syntactically based semantics such requirements on a verb's arguments are all lexically expressed in subcategorization, and thus associated with syntactic structure once lexical insertion has taken place. This contrasts with Rizzi's (1986), Zubizarreta's (1989), and others' proposals that certain arguments are in no sense projected onto the syntax. In my treatment, the following notation expresses deep structure understood arguments in lexical subcategorization frames. (34) + XP: syntactically and semantically obligatory. + (XP): syntactically optional, and semantically present if and only if syntactically present. + (XP): syntactically and semantically present in any case, a subject which is a co-argumentof it" (ch. 2), I think immediatelyof disjoint reference, or of pronouns subject to Principle B of the binding theory (Chomsky, 1981, ch. 3). Since binding principles presumably hold at s-structure, I would conclude that the s-structure of the patient (affected) argument, in Dutch at least, may exhibit incorporation, rather than attributing this to the lexicon. THETA-ROLE ASSIONMENT 403 but X P m a y be covert. + XP" semantically obligatory, but X P must be covert. 36 In the trees resulting f r o m lexical insertion, c o m p l e m e n t s which are syntactically present but invisible (and thus semantically present) can also be r e p r e s e n t e d with an underline, as in the following example. (35) J o h n might suggest to leave town. S NP John I might V vP PP COMP s /5"-.. I T I to V I leave vP NP I town Empirical support for X P includes facts of control structures and of null a n a p h o r a ; e.g., in the preceding example, the u n d e r s t o o d indirect object is the " c o n t r o l l e r " of the e m b e d d e d infinitive. In Rizzi (1986), it is shown that languages can differ as to h o w and w h e n u n d e r s t o o d arguments (for me, semantically present and syntactically covert) can enter into relations of control. T h e underline symbol N P can also be recast as a lexical null a n a p h o r . Thus, + ( N P ) means that N P is optional (as is conventional), and that it m a y be a null a n a p h o r , whereas + N P means that N P is obligatory, but m u s t be a null anaphor. A s a lexical a n a p h o r , N P must receive its o w n role in o r d e r to satisfy Full Interpretation. M o r e o v e r , as with o t h e r anaphors, N P m a y be locally b o u n d , or it can sometimes have an arbitrary interpretation, as in (35). W e will see instances of b o t h interpretations just below. 36 There are some logical possibilities for subcategorization that do not seem to be realized and hence are not provided for in my notation: elements that are semantically optional can be present in the interpretation only if they are also syntactically overt. If necessary, the notation could be enriched to accommodate the possibilities excluded by (34). The notation (NP) is crucial in preventing verbs such as read and wash from undergoing "transitivity alternations" in deep structure. 404 JOSEPH E. EMONDS Drink. Returning now to J a c k e n d o f f ' s (1987) example drink, it should be observed that its place c o m p l e m e n t can be optionally present, as in John drank it down and Mary drank the scenery in. The further specificity of the verb drink (with an animate subject) should be determined by comparison to an animate subject "taking liquid in". That is, drink (compared to say absorb) means "in the way normally used by the agent NP for nourishment". I conjecture that in the c o m m o n vocabulary (the "primary vocabulary" in E m o n d s 1986), a universal convention for activity verbs supplies this proviso, except that the notion " n o u r i s h m e n t " is supplied as part of the semantic field of drink (but not of take or absorb). By placing drink in a semantic field of " n o u r i s h m e n t " , I do not claim that drinking is a subcase of nourishing oneself, but rather that the m o d e of drinking is the m o d e usual for nourishment. 37 Thus, I propose the following entry for drink, where semantic terms are in quotes: (36) drink, V, - L, " n o u r i s h m e n t " , "activity", + (NP "liquid') (P)^NP, In this entry, I crucially use J a c k e n d o f f ' s notation NPi to indicate coreference with the subject NP. Thus, NP means a null anaphor of arbitrary interpretation, and NPi has a (necessarily locally) bound interpretation. These enrichments of syntactic subcategorization, actually a simple extension of existing categories and concepts, are all we need to capture the advantages of J a c k e n d o f f ' s lexical formalisms. As in (34), underlining indicates semantic and "invisible" syntactic presence, probably implying unavailability for purely syntactic operations. P must be further specified so that just the prepositions in and down (and the completive up) can appear; this restriction is the same for both conceptual and syntactically based semantics. Except for the empirical discrepancies discussed here, all the information present in J a c k e n d o f f ' s (28) can be derived from (36) and the 37 Drink appears to impose selectional restrictions on its arguments, which Jackendoff characterizes as "explicit information that the verb supplies about its arguments" Jackendoff 1987, p. 27).-Such restrictions might well be treated as cognitive rather than linguistic information. If they are purely cognitive, then a stronger version of Semantic Atomism (21) is justified, whereby no inherent semantic selectional restrictions can be annotated to a subcategorization frame. Jackendoff casts his lexical entry for drink partly in terms of the concrete noun mouth, perhaps influenced by a tendency to accord concrete nouns some kind of prior status in semantic description. This is an error, since it is perfectly easy to imagine an animal or extraterrestrial with a mouth (for food) which drinks through its nose, its skin, or whatever. THETA-ROLE ASSIGNMENT 405 interpretive principles I have provided. 38 In a clause whose verb is drink, two deep complements are present but not necessarily overt. The correspondence between positions in the verb's lexical entry and those in the syntactic verb phrase, at least in cases like these, is trivial. With respect to theta-roles, drink, as claimed by Jackendoff, is a thematic verb involving motion of a liquid. Since drink is - L (*John drank the sink), the only role available for the object is the Figure. Since PP is semantically obligatory, it must then contain the Ground. Finally, by Full Interpretation and Agent Specification, the subject must be an agent. 39 The fully interpreted representation includes such non-syntactic semantic annotations, but no autonomous conceptual structure component augments the syntax. There is only the process of lexical insertion, which determines the presence of (sometimes covert or invisible) argument XPs, and principles of semantic interpretation to further specify syntactic structure, such as the theta-role assignments (22), (26), (27), (33) and the proviso for the "usual mode" discussed above. As Jackendoff notes, such principles are akin to what Katz and Fodor (1963) originally called "projection rules". Buy, sell. The device Jackendoff introduces to account for invisible NP complements coreferential with the subject NP (the appearance of NPi in the lexical frame, here rendered as NPi) plays an important role in simplifying the general problem of linking thematic roles to syntactic positions. We can structure discussion around a division between two types of obligatorily agentive subjects: (37) (3)a. b. Some subjects of activity verbs must be agents even when they carry another thematic role: e.g., goal as in (3a) or source as in (3b). John bought the book from Bill. Bill sold the book to John. 3s Here is (28) from Jackendoff (1987): drink, [ - N , + V ] , (NPj), [E.... C A U S E ([Thing ]i, [E.... GO ([Thing LIQUID]i [Path TO ([Place IN ([Thing M O U T H OF ([Thing ],i)])])])])] 39 A reviewer calls to mind "instrumental" subjects, as in The sponge drank (up) the milk. Many well formed sentences whose animate subjects are necessarily agents (by Full Interpretation) have counterparts with inanimate instrumental subjects. As indicated in Section 3, I am not treating roles such as instruments here. 406 (38) JOSEPH E. EMONDS Any subjects of activity verbs are agents when they carry no other thematic role (bring, move, pass). 4o As a statement which holds generally, (38) is an obvious consequence of Full Interpretation (22) and Agent Specification (33). But some other factor must affect verbs like buy and sell which conform to (37). This factor is Jackendoff's notation NPi, used similarly to how it appears in the lexical entry (36) for drink. I define the lexical insertion of a verb with an insertion frame + (NP0 as follows. If the NP in a tree to which NPi corresponds is empty, it must be coindexed with the subject. (That is, this NP is then a null anaphor. If it is not empty, the NP may take on any otherwise permissible form.) The entry for buy can now be formulated as (39), where + G O A L are features of P. (39) buy, V, - L , "activity', inherent (+ GOAL^NPi) ( - GOAL^NP) features, + NP If all the parenthesized elements of (39) are chosen and if NPi is empty, the tree (40) for (3a) results. 41 4o (37) and (38) cover all alternations previously discussed in terms of varying linkages between thematic roles and syntactic positions. However, there remain some intransitive activity verbs like run and dash whose animate themes must be agents (perhaps actor is a better term), due to the inherent semantics of the verbs involved. Hence, as observed by Hale and Keyser (1986), when such a verb is optionally transitive (and - L O C A T I O N ) , sentences such as John jumped the horse over the fence can be generated, in which two arguments apparently receive agent interpretations. 41 If GOAL?NP is chosen to be syntactically overt, NPi can be a lexical NP or it can be a lexical anaphor (John bought a book for himself'). The lexical representation of the prepositions in entries like (39) can be made more precise, but a full treatment would be too tedious. Clearly, the required choices for overt P with buy (for, from, of) are not "idiomatic". They should fall under the scope of a principle established with a detailed argument in Emonds (1985, ch. 4), the "Designation Convention": In any local language-particular statement of a transformational nature (which includes subeategorization features, as discusscd in Chomsky 1965, ch. 2, note 18), syntactic subcategories of the bar notation necessarily refer only to the unmarked or "designated" elements of that category. The unmarked English representatives of the + G O A L and - G O A L subcategories of P for non-physical location are to and from (archaically of; dialectically off of). Along this "non-physical" dimension of locational P, benefactive and other uses of for seem related to to much like the purely physical toward is related to physical to. In any case, all these elements are + G O A L , and the exact relation of to to for remains mysterious. THETA-ROLE (40) John 407 ASSIGNMENT S V NP PP PP bought the book [ P] L+GOAIJ NP i [-OAd 0 from I 0 Bill Since buy is a thematic verb which is - L , the object can only be a Figure (theme); consequently the objects of P can only be Grounds, and the subject can only be an Agent. The appearance of Jackendoff's symbol N P , not in conceptual structure, but in the subcategorization frame, provides a simple account for the entire class of verbs which follow (37). Parallel to (39), the entry for sell is (41), which gives rise to (42) as the representation of (3b). (41) sell, V , - L , + (42) "activity", same inherent features as for buy, NP(+ GOAL^Np) - GOAL^NPi S ~ Pi Bill VP V [ sold NP PP PP the book [ P ] L+GOALI I to NP John [_ooad I 0 NP. i, 0 The obligatory source in (42) explains why the subject NP must receive a separate agent role in order to satisfy Full Interpretation. Thus, (3a,b) are represented as semantically identical, except for the fact that the 408 JOSEPH E. EMONDS agents in the two sentences are understood as coreferential with different complements. Climb. As Jackendoff (1985, 1987) notes, the syntactic complements of climb are an optional NP or PP. When its syntactic complement is an NP or 0 , there is an implied sense of upward movement, while this is not necessarily present with a PP (John climbed down the rope). In addition, when the complement is an NP, the movement is "complete", in the sense that the top of the object climbed is reached, at least in simple or perfect tenses (John has climbed the rope). Let us suppose that each of the basic paired post-verbal particles in English (in, out, up, down, off, on, back, away) is characterized by some feature values; for up I will abbreviate them as HIGH. Leaving aside for a moment the completive sense of transitive climb, a first version of syntactically based insertion frame for climb is as in (43). (43)a. b. climb, V, +MOTION, + H I G H , + (NP), "clambering manner" (Jackendoff's term) climb, V, +MOTION, + PP, "clambering manner" By (43a), transitive climb then "incorporates" the path feature HIGH, just like the French intransitive motional verbs discussed in Section 3 (e.g., monter "go up" also incorporates HIGH). As has been seen from Talmy's work, such incorporation of features from adverbial PP positions is extremely common. The frame in (43b) can be revised perhaps to + PATH?(NP), where PATH is the syntactic subclass of prepositions which realize Jackendoff's conceptual category "path". 42 + H I G H is a directional subcategory of PATH (and of P). Now, using exactly the brace notation for the lexicon that Jackendoff introduces to capture mutually exclusive dependencies, we can collapse (43a,b). 42 If PATH is a semantic rather than a syntactic subcategory of P, the revised frame for (43b) would include a semantic selectional restriction, exactly analogous to the feature L I Q U I D in the entry for drink; cf. note 37. Jackendoff (1985) notes some other specifics of climb. For example, in John climbed this way and which way do we climb from here, the NP headed by way acts like a PP (downward motion is possible). According to Emonds (1987), way has the character of PP more generally; note also, *This way was climbed by John. Jackendoff further notes that the "clambering manner" is restricted to agentive subjects (we can assume this is definitional of clambering), and that climb must include either "clambering" or upward motion (my HIGH), which can be expressed by a further use of his brace notation. I see no problem with incorporating these refinements into the syntactically based approach being developed in the text. THETA-ROLE (44) climb, V, +MOTION, "clambering manner" ASSIGNMENT {+HIGH}, 409 + {PATH}^(NP), The two readings of (44) correspond to (43a,b); if HIGH is chosen, (43a) is obtained, while if PATH is chosen, (43b) is obtained. 43 It seems, therefore, that syntactically based lexical representations, incorporating both Jackendoff's "NP/' and brace notations as well as a device (34) for "understood arguments", fully express the complexities and generalizations brought out by Jackendoff. For example, it follows from these principles that Full Interpretation and Agent Specification together predict the earlier descriptive generalization (38). Moreover, as just discussed, (38) and the device of syntactically covert oblique arguments co-indexed with subjects together predict (37). A final criticism of the autonomous conceptual semantics approach refers to the fact that its basic categories, "thing", "path", and so on, are, as Jackendoff states, typically realized by certain syntactic categories. These correspondences are simply identity relations in syntactically based semantics. The latter approach has the obvious advantage of not introducing an extra set of categories into linguistic theory. 6. ASSIGNING THEMATIC ROLES TO S T A T I V E VERB ARGUMENTS To see in more detail how the principles introduced here interact to predict configurations of thematic roles, I first consider the simpler cases of intransitive non-activity ("stative") verbs. By (33), stative verbs cannot involve an agent. Ordinarily, the subject NP of a stative verb is the Figure; e.g., be, reside, go (the sense of "be compatible"), extend, dwell, occur, belong (either the locational or possessive sense), and appeal (the sense of "be attractive"). This interpretation is assured by obligatory subcategorization for a "Path + Ground" PP whose head is +LOCATION, for 43 What remains to be expressed is the sense that the "top" of the path is reached when climb is transitive. This is stipulated in Jackendoff's entry, but I think this is the wrong tack. While I can not go beyond Anderson's (1971) work on this subject, it seems that "reaching the top" for a verb of upward movement is nothing other than a subcase of completed action that is so often (but not always) part of the interpretation of a transitive structure. Cf. cross (through) the field), chew (on) the bread, learn (about) algebra, etc. Thus, a rule of interpretation (a "projection rule") operating on a syntactic combination V + NP seems more appropriate than an item-particular stipulation ("top" for climb, "side" for cross, "through" for pierce, "completely" for learn, etc.). I am not proposing a concrete rule (there are verbs which escape it, such as approach and spray), but rather am appealing to the reader's sense of the construction to see that the syntactically based approach promises to capture a generalization that the more "autonomous" conceptual semantic formulation obscures. 410 JOSEPH E. EMONDS then Full Interpretation for the subject can be effected only via Figure Specification (26). A variant of this situation occurs when the verb can appear without an overt syntactic PP, as in (45a); the Figure interpretation for the subject then results from lexically listing the PP complement of Path + Ground as semantically obligatory by means of the frame + (PP, +L). (45)a. Theta-roles exist (only in the minds of linguists). Every planet necessarily revolves (in an elliptical orbit). If an intransitive stative verb is lexically specified as taking a PP whose head is - L O C A T I O N , then its subject is necessarily a Ground rather than a Figure. Depending on the verb, a PP of this sort containing the Figure NP can be obligatorily syntactically overt (45b), or optionally overt and semantically understood (45c): (45)b. This room reeks of perfume. *This room often reeks too much. The arrangement smacked of misogynist paranoia. *The arrangement really smacked. c. John knows of (about) your plans. The Shadow knows. Mary was hearing about it. Mary will never hear. The third possibility for intransitive stative verbs is subcategorization for a PP unspecified for - L O C A T I O N . As a result, the Figure or the Ground are not required to appear in a particular position. (45)d. This room stinks (of perfume). Something stinks (in this room). The castle resounded with cries. Cries resounded through the castle. The stative verbs with unspecified PP complements as in (45d) are probably not lexically "unmarked"; this might lead to an intolerable level of ambiguity. Rather, it seems that the unmarked lexical value for a P in a subcategorization frame is +LOCATION, meaning that the first type of stative verb discussed above (be, reside, go, extend, dwell, occur, belong, appeal; etc.) is unmarked. The same comment holds for the P which occur with activity verbs to be discussed in subsequent sections. Consider next "locative inversion" sentences, whose properties in Chiche~a are analyzed in Bresnan and Kanerva (1989). The English THETA-ROLE ASSIGNMENT 411 counterparts, which the authors correctly suggest are related constructions, are given in (46): (46)a. b. C. d. e. f. g. h. In the village is a well. To the village came those visitors. Into the well fell a goat. On the beach stands a fox. Outside is cold. ??At the market will break out a fight. ??In the forest have remained lions. ??In the house are sleeping the chickens. The authors show conclusively that the s-structure of the Chiche4ca versions of (46) are as in (47), where the locative XP are certainly in subject position, and the themes (Figures) are with equal certainty arguments internal to V. (47) S XP (locative) VP I V1 V NP (theme) In addition, they cite six transformational analyses for inverted locatives in English, criticizing all of them for failing to at least partially generalize over the obviously similar Chicheqca construction, except Emonds (1976) and Bowers (1976), who propose (47) as the s-structure for examples as in (46). If, as Bresnan and Kanerva imply, the locative XP subjects in Chiche@a are NPs and include no PP structure, which is suggested by the Chiche~ea verb's overt agreement with locative subjects, then principles (26) and (27) correctly allow Figure (theme) and Ground (locative) to be assigned indifferently in subject or object positions for a verb which is -+L, 412 JOSEPH E. EMONDS + NP. In this way, we find the predicted transitive counterparts to the similar intransitive alternations in (45d). 44 Transitive stative verbs can thus behave similarly to intransitive statives, precisely because V as well as P can be + LOCATION. The unavailability of Agent Specification for statives forces upon the subject, by Full Interpretation, whichever role between Figure and Ground the object lacks. In contrast to what holds for P, however, the most likely unmarked value with V is - L O C A T I O N . Corresponding to the unmarked intransitive verbs with obligatory Figure subjects (go, reside, belong, appeal), transitives with Figure subjects are marked as + LOCATION; among them is a large class of "psychological predicates". (48)a. b. c. d. Does that path reach the garden wall? Few families inhabit that district. John doesn't resemble me. Your prowess amazes Bill. The fact that a verb is + L does not imply that its Ground must appear in direct object position, but only that its Figure may not. Thus, the verb reach is + L and is subcategorized as + ([P, +L])^NP, which allows its Ground to appear either as direct object or prepositional object. (There is no need here to distinguish between + L and the special case of + DIRECTION.) The apparently more puzzling pattern in (1), which might be taken as indicating the necessity of item-particular theta-grids, can easily be expressed with a similar formalism; worry (itself + L) is to be subcategorized as + ( a b o u t ) ^ ( N P ) , where about, a characteristic P which introduces a Figure (a "theme"), is - L . Thus, in (49a) the subject is necessarily the Figure, while in (49b) it is necessarily the Ground. 44 TWO other possibilities for (47) come to mind. (i) It may be that (47) is transformationally derived (Emonds 1976, Bowers 1976) in one or both languages. Although Bresnan and Kanerva, as non-transformationalists, would reject this option, they indicate three properties of the post-verbal NP in (47) that might be susceptible to a transformational account. If (47) is not a deep structure, then it is doubtless derived from unexceptional, uninverted strings containing [V, + L], like those in (45a). (ii) It may also be that (47) is a base-generated structure in which XP = [Np[NO]PP]. Such a deep structure, along with an explanation for the unexpected zero head, is proposed for English locative subjects in Emonds (1985, section 7.7.3); it can explain why a "locative NP" in Chiche@a fails to exhibit an otherwise pervasive singular-plural distinction. Such a PP subject gains further plausibility in Chiche@a because the locative features are morphologically realized on NP, which allows the Invisible Category Principle (Emonds 1987) to license an empty P. In this case, the Figure (theme) is realized on the only available NP with a lexieal N (the object), and the Ground is assigned by [P, + L]. THETA-ROLE ASSIGNMENT (49)a. b, c. 413 John (Figure) worries Bill (Ground). John (Ground) worries about Bill (Figure). John (Ground) worries that you can't pay your bills. An S-complement as in (49c), according to argumentation in Emonds (1985, ch. 7), has the structure P ( = COMP) + S. The G r o u n d interpretation of the subject in (49c) follows from our principles if the unmarked sentence-introducing P t h a t is - L . Such a classification for t h a t is entirely plausible because it reflects the fact that an S typically designates a Figure rather than a physical or psychological location (Ground). 45 7. ASSIGNING ROLES TO INTRANSITIVE ACTIVITY VERB ARGUMENTS In the interpretation of the arguments of activity verbs, the principle of Agent Specification (33) plays a central role in inducing a range of linking patterns. There are many variables to be considered; verbs can be transitive or not, and may or may not take oblique complements. In addition, each of these complements may be further specified in one of the four ways permitted in (34). Finally, either the verb or the P head of an oblique complement can have the values + L O C A T I O N , - L O C A T I O N , or +-LOCATION. A preliminary study of this scope cannot pretend to claim that all the theoretical lexical possibilities are realized, or have even been examined. At the outset, however, it deserves mention that non-syntactically based alternative approaches to lexical semantics rarely even acknowledge the issue of whether all the various possibilities they provide for linking thematic roles and syntactic positions are utilized. Such frameworks, including most treatments aligning themselves with Chomsky's government and binding theory, have an essentially open-ended notion of "possible lexical entry" and do not address any substantive questions about what sorts of lexical frames are impossible; they typically specify a general format for expressing almost any conceivable linking pattern and take for granted that determining restrictions on such linking is orthogonal to the proper elaboration of both syntax and lexical representations. In contrast, the syntactically based system devised here is precise enough to also bear 45 The present system naturally expresses the fact that that-clauses alternate with o f +NP and about +NP (e.g., after verbs like know and think), since in all three constructions the head is a P which is -L. To derive patterns as in (49c) from putative underlying sequences of about + that + S, with subsequent that-deletion, would lead to highly unnatural subcategorization frames in addition to the ad hoc deletion rule. 414 JOSEPH E. EMONDS scrutiny on whether the range of possible lexical entries it licenses are in fact realized, a task which escapes even preliminary formulation in both generative and conceptual semantics. My discussion of intransitive activity verbs can aspire to a certain completeness, since they appear with considerably fewer argument patterns than do transitives. Because intransitives have no direct object, the feature choice + LOCATION on the verb itself cannot effect variation in syntactically positioning Figure and Ground. When the activity verb is thematic (+T), the possible lexical variations are the choice of frames P P , __(PP), (PP), and PP, where each head of PP can be further specified for features such as LOCATION. We first discuss cases where P are +LOCATION. Intransitive verbs whose obligatory PP component is + L (or + L, + DIRECTION) include dash, lurk and glance (Chomsky 1965, ch. 2); as mentioned earlier, this positive specification for LOCATION is probably unmarked for a P appearing in a subcategorization frame. Stride and look appear with + ([PP, +L]), and verbs like jump and expand are + ([PP, +L]); the former but not the latter plausibly have an obligatory but possibly covert oblique complement. For all these verbs, their subjects are Figures and optionally Agents, as predicted by (26), (27), and (33). 46 (50) John lurked (was lurking) in the shadows. Danger lurked (was lurking) in the shadows. John looked (into the street). The garden looked onto the street. Oil companies have expanded (into arms contracts). Our sales area is expanding (into neighboring states). Is the frame + PP, + L also a possibility? In the present system, a typical obligatorily "understood" PP complement is what is postulated for certain transitive activity verbs whose direct objects are Figures (cf. discussion of buy, sell, and drink in Section 5). By means of a frame + P^NPi, where P must be + L and i indicates co-indexation with the subject, Ground Specification (27) can pick out a (syntactically covert) oblique NP, while Agent Specification (33) necessarily interprets its coreferential subject NP. Nothing prevents intransitive counterparts to buy, 46 W h e n an activity requires an + A N I M A T E subject (e.g., dash, glance, stride), an agentive reading for the subject is often the only one possible. I assume that this is pragmatic, and independent of the optionality provided for by (33), which is amply justified in other examples throughout this study. THETA-ROLE ASSIGNMENT 415 sell, and drink which differ from" them only in that their Figures are in subcategorized PPs rather than direct objects. As expected, there are such verbs: (51) A counterpart to buy: John learned {off about} your trip (from Sue). + _ _ t o ^ N P i , . . . A counterpart to sell: John taught {about/*of} Paris (to tourists). + _ _ f r o m ^ N P i , . . . A counterpart to drink: John partook {off*about} the refreshto^NPi,... ments. + Thus, the subcategorization possibilities for intransitive activity verbs whose PP complements are + L O C A T I O N all seem realized. As is quite typical, the thematic role of an oblique NP can be inferrcd from the subcategorized prepositions in (51), but not vice-versa. This asymmetry is almost too well known to require comment, and yet frameworks which promise to derive syntactic frames from thematic grids rather than vice-versa appear with curious regularity, and make no mention of such facts. There also exist intransitive activity verbs whose PP complements are lexically specified as -+L O C A T I O N . When the PP complement of such verbs is + L O C A T I O N , principles of Ground and Agent Specification insure that the subject NP is a Figure, as in (52). Of course, such a subject is optionally an Agent if animate. (52)a. b. Customers were swarming into the new store. New lights gleamed on the trees. When the PP complement of such verbs is - L O C A T I O N , it will contain the Figure and the subject NP is the Ground: (53)a. b. The new store was swarming with customers. The trees gleamed with new lights. Finally, there are intransitive activity verbs whose PP complements are necessarily - L O C A T I O N . Consequently, the Figure is in a PP and the Ground is the subject. As usual, the subject is an optional agent. 47 47 Some activity verbs whose PP complements are necessarily - L (i.e., containing a Figure), do not seem to occur with agentive subjects. In other words, when an animate subject is available for such an intransitive, the verb is necessarily stative: The streets were filling with demonstrators. John filled with pride at the sight. *John was filling wilh pride at the sight. 416 JOSEPH E. EMONDS (54) John should think {of, about} his past, and you should do so as well. Has he thought of the consequences (,*or will he do so soon)? For some intransitive activity verbs, both the Figure and the Ground are in PPs, so that the subject is necessarily, by Full Interpretation, an Agent. (55) John {talked, bragged, spoke} {of, about} his past to the entire class. John heard {of, about} his past from the entire class. Alternatively, such verbs may allow for an understood source or goal NP co-indexed in the lexical entry with the subject, as with teach and learn. 8. ASSIGNING ROLES TO TRANSITIVE ACTIVITY VERB ARGUMENTS Our final application of the principles of theta-role interpretation involves the system of interpreting thematic roles applies to transitive activity verbs. Here, we must necessarily be schematic, since, as mentioned at the beginning of the previous section, the number of possible variations precludes an exhaustive treatment. It can nonetheless be easily appreciated that the syntactically based semantic system receives strong support in this area. When an activity verb (one with a potential agent) is optionally transitive, Full Interpretation and Agent Specification interact to correctly predict the often observed "transitivity alternation" in the grammatical position of the theme (Figure) between the subject position in the intransitive use and the object position in the transitive use. Many activity verbs are - L O C A T I O N (the unmarked value for verbs) and optionally transitive. They then typically have the subcategorization frame + (NP) (PP, +L). Recall that the unmarked value for P in subcategorization frames is +LOCATION, so that it is written here just for clarity; this frame is the simplest available for a verb with three arguments. The present framework now automatically predicts, without any lexical stipulation about thematic roles or linking, that such verbs have the following alternations. 4s 48 It is of interest to note that Talmy's work, as detailed as it has become, consistently accents intransitive motional structures. I suspect this is because only ad hoc manipulations of semantic structure could bring about a uniform positioning of the theme NP with regard to the motional V. My view is that such a uniform representation would serve no purpose THETA-ROLE (56) ASSIGNMENT 417 The rope (Figure) dropped (onto the steps). The thief (Figure, optionally Agent) dropped (onto the steps). The Thief (Agent) dropped the rope (Figure) (onto the steps). During the quake, the benches (Figure) moved (down the slope). During the quake, the prisoners (Figure, optionally Agent) moved (down the slope). During the quake, the guards (Agent) moved the prisoners (Figure) (down the slope). In order for such alernation between objects and subjects to be blocked, either the direct object must be syntactically (take, lift) or semantically (push, donate) obligatory, or the subject NP must be lexically co-indexed with an understood Ground NP (cf. the discussion of drink and buy in Section 5). In the case of obligatorily transitive activity verbs (with the unmarked feature - L ) , the Ground will necessarily be in a PP which is either syntactically or semantically obligatory, and the subject will, by Full Interpretation, necessarily be an agent. Verbs of this sort are put, hand, shove, send, etc., the last two having semantically obligatory but not syntactically overt Ground complements (e.g., John sent a letter). The transitive activity verbs which are - L O C A T I O N that we have so far discussed have either an at least semantically obligatory direct object or an at least semantically obligatory PP. In contrast to such verbs we can consider get, which is as well a + T (thematic) verb and - L (its direct object cannot be a Ground), and which syntactically must have a complement, but one of any type. If get is transitive, its object is the Figure, and the Ground may be in either the subject or the PP position. (57) John (Ground, not Agent) got the letter. Seattle (Ground) gets a lot of rain. John (Agent, not Ground) got the letter to Victoria (Ground). If the obligatory subcategorization of get, namely + XP (YP) or perhaps + (NP) XP, is satisfied rather by a PP, the unmarked PP is + L, and again as predicted, the Figure subject is optionally an agent. (58) The storms often (*try to) get to Seattle. The tourists often try to get to Victoria. These variations, in a framework that takes the assignment of thematic at all, since it is perfectly possible to read the t h e m e NP off the English surface structure of both transitives and intransitives. 418 JOSEPH E. EMONDS roles to be primitive, would require tortuous linking stipulations. Moreover, the obligatory nature of the complement remains completely mysterious in such a "semantically based" framework. Get does not "obligatorily case-mark" (cf. (58)), nor is any role other than the theme obligatory; yet, we have clear evidence that some complement to get inside V ~ is necessary: (59) *Seattle sure has been getting during the winter. *Will the rains get during the weekend? A last type of transitive activity verb which is - L O C A T I O N has an at least semantically obligatory direct object, but allows the Ground to be realized either in a PP or in the subject position. (60) John (Agent) collects clothes on this porch. Sweaters (Ground) collect lint. Mary (Ground, optional Agent) has collected quite an assortment of roommates. Good examples of a transitive activity verb which are + L are fill and cover, since their object cannot be a Figure: *fill the sand into the bucket; *cover the blanket over the bed. The variations expected in the present system are again what is observed: (61) The sand (Figure, not Agent) filled the bucket (Ground). John (Agent) filled the bucket (Ground) with sand (Figure). The migrant workers (Figure, optional Agent) fill newer buses (Ground) first. The pesticides (Figure, not Agent) covered our neighborhood first. The campaign workers (Agent) covered our neighborhood with pesticides (Figure). The campaign workers (Figure, Agent) covered our neighborhood first. The difference between fill and cover is that the former is not syntactically obligatorily transitive: the bucket filled vs. *the bed covered. What is to this point unexpected is that intransitive fill with a Figure subject is excluded: (62) *The sand is filling fast. A similar restriction appears on other optionally transitive + LOCATION verbs: THETA-ROLE (63) ASSIGNMENT 419 John has worried a lot. *My assignment has worried a lot. Apparently a + L O C A T I O N verb which is thematic requires that its Ground and not its Figure be obligatorily present. Many of these verbs are subcategorized as + (NP), so that the Ground is an understood object rather than an explicit subject (e.g., the train is approaching, how long have you been climbing?), but if such a verb semantically has only one argument, then indeed the argument will be Ground. Finally, some of the most complicated cases are presented by activity verbs which are optionally or obligatorily transitive and which are also -+L O C A T I O N . Discussion of the various possibilities would take us into too much detail; nonetheless, it should be noted that some of these verbs have already been discussed in the first motivations for our interpretive system given earlier in Section 3: move, pass, smear, etc. I will terminate this section not by going through more subclasses of + L O C A T I O N verbs, but by bringing up some quite plausible objections to my interpretive system that might be made on the basis of such verbs. I have used alternations as in (64) to motivate not linking thematic roles to particular syntactic positions. (64) We loaded the books onto the truck. We loaded the truck with books. They sprayed the wall with the paint. They sprayed the paint onto the wall. However, since only the Figure is generally obligatory, and the PP seems optional, we may well ask why (65a) is not bad and why (65b) is bad: (65)a. We loaded the truck. b. *The books loaded the truck. The answer to this is that the subcategorized PP position is semantically obligatory; i.e., the frame is + NP (PP). Thus there is an understood PP of Figure in both of (65a,b). This Figure renders (65a) well-formed and (by the uniqueness of the Figure) (65b) is ill-formed. The appropriateness of such an analysis is suggested by the fact that the obligatory head of this PP may be an overt locational P, rendering (65b) well-formed with the books now serving as an "instrumental", a thematic role not discussed here. (65)c. The books loaded the trucks down. Another objection might be made on the basis of my analysis of verbs 420 JOSEPH E. EMONDS like spray. Again, since spray enters into pairs like (64), it is -+L. We observe that spray, unlike load, may be intransitive. (66) The paint sprayed (all) over the wall. The paint sprayed in all directions. So far, then, we cannot exclude (67a,b), in the face of (67c). (67)a. *The paint sprayed the wall. b. *The wall sprayed the paint. We sprayed the wall/the paint. C. But again, we can make appeal to a frame involving "underlining": + (NP) (PP). We now explain (67a,b) by virtue of their not being allowed to have two Figures (cf. 26) or two Grounds. 49 It should be observed that the two problems I have discussed in concluding, along with tentative solutions suggested in the framework developed, are at a level of discernment which semantically based approaches rarely even approach (Jackendoff and Talmy aside). The very fact that questions of such detail can serve as crucial evidence indicates the superiority of the syntactically based framework, which predicts what its competitors stipulate and allows precise statement of puzzles previously buried in a morass of unexamined lexical detail. 9. T H E SYNTACTIC BASIS OF LEXICAL SEMANTICS The principles of Figure, Ground, and Agent Specification presented here interact with Full Interpretation to provide every NP argument of a "thematic" (+T) verb with a thematic role or roles. 5o Broadly speaking, Full Interpretation insures that each NP argument receives at least one thetarole; the Figure may appear in any position; so may the Ground, but the Figure and Ground, following Talmy, are mutually incompatible on a single NP. The Ground can be the object of a V or of an expressed or understood P if and only if P or V is +LOCATION; an Agent must be the subject of an activity verb. Even though I have argued that the assignments of theta-roles are entirely predictable, the interpretive system 49 However, an unsolved problem is the unacceptability of *The wall sprayed with paint. This can be clumsily represented in Jackendoff's brace notation (section 5), but doing this sheds no light on any general problem. so As indicated earlier, verbs of physical motion and location, of ownership, of communication, and of emotion are " + T " or "thematic" verbs. It is unclear whether other classes typified by verbs such as avoid, break, complete, destroy, make, etc. can be subsumed under this class, but my working hypothesis is that they can be. THETA-ROLE ASSIGNMENT 421 developed here avoids associating given roles with particular syntactic or "conceptual structure" positions, either in syntactic structures or in lexical entries. Thus, there are no fixed "grammatical meanings" for subject, direct object, etc. In my view, recent studies which couch their findings in terms of thetagrids, item-particular lexical linking, and non-syntactic representations of understood arguments (e.g., Culicover and Wilkins 1986, Levin and Rappaport 1986, Rizzi 1986) in fact contain implicit arguments for syntaxbased accounts of their results. In these studies, the theta-roles which are crucial for expressing their results are any theta-roles associated with direct object or subject positions, which is in fact another way to say that the generalizations are to be stated in terms of deep syntax. For example, Levin and Rappaport establish that the subject of an adjectival passive is necessarily a direct object in the corresponding active verbal structure, counter to the claim of many preceding articles that particular theta-roles determine which NPs may be the subject of a passive adjective. (It is necessary to assume that "dative objects" are not deep direct objects; for example, deriving them transformationally, as in Emonds (1986), is compatible with Levin and Rappaport's generalization.) In Rappaport and Levin (1986), the same authors also demonstrate that -er nominals are formed from verbs by a process which eliminates their deep structure subjects, whatever the semantic role of that subject may be (agent, instrument, or theme). A similar conclusion indicating syntactic primacy can be drawn from the results of Culicover and Wilkins (1986) and Rizzi (1986), who show that Romance and English differ as to whether understood direct objects (not particular theta-roles) can control complement infinitives. In spite of the extensive use of theta-role vocabulary in these studies (particularly after the relevant generalizations have been established), the crucial structural identifications concern (empty) deep structure direct objects, recast as NPs which (would) receive the same theta-role as such objects. In yet another case, the results of Randall (1987) on the "thematic inheritance" from verbs in derived nominals also suggest that syntactic subcategorization can express generalizations obscured by theta-role terminology. A central finding of Randall's is that several verbal endings in English derivational morphology (that is, morphemes that are head N and A) are incompatible with, as she terms them, "all indirect arguments". (68) The plane is flyable (*by experts/*into the wind/*to Paris/*by computer). 422 JOSEPH E. EMONDS We all admired the arranger of this furniture (*with no assistant/*in a novel way/*by telephoning the maid/*for Bill). As Randall points out, some other verbal derivational endings are compatible with such complements: (69) The flying of the plane (by experts/into the wind/to Paris/by computer) caused little comment. John's schedule seemed arranged (with no assistant/in a novel way/by telephoning the maid/for Bill). It would seem that the appropriate generalization is that -able and -er are subcategorized as PP while -ing and -en are + (PP). 5~ Now, Randall observes that -er is compatible with an o f - N P that she terms the theme, but the fact is that -er accepts any NP complements resulting from the frames + NP and + NP NP of the corresponding verbs, no matter what their thematic role; likewise, -able is also compatible with indirect objects which arise from the frame + NP NP with active verbs: (70) Readers of such stories to children (Goal) need a lot of patience. The loaders of the next truck (Goal) should talk to the foreman. Any climbers of this mountain (Goal) know caution is needed. The bill is payable to me (Goal). In the theory of subcategorization of Emonds (1985, ch. 2), such double NP frames in the lexicon "'induce" PP structures wih empty Ps, so that the examples in (70) don't violate the feature PP in the lexical entries of -er and -able. (A PP phrase cannot satisfy subcategorization for PP if its head is empty.) Examples of derived nominals in which PPs with lexical heads realize the very same thematic roles are, as expected under the subcategorization account, excluded: (71) *The loaders (of books) into the next truck should talk to him. *Any climbers up this mountain know caution is needed. Thus, Randall's thematic role account of these interesting restrictions is better expressed in terms of subcategorizaion. In this as in the other cases discussed above, I remain unconvinced that any significant general51 A reviewer observes: "There are certain prevalent exceptions to the generalization about -able particularly in newer varieties of English; in these varieties adjectives formed with -able occasionally take by-phrases; in computerese, for instance, one encounters things like These files are readable by the super-user but not by anyone else". THETA-ROLE ASSIGNMENT 423 izations about syntactic well-formedness can be properly expressed in terms of theta-roles, although theta-roles certainly interact with pragmatic factors to disambiguate already well-formed syntactic structures (e.g., the usually contrasting interpretations of the bat/the hat flew across the room). 10. H U M A N S E M A N T I C S VS. P R I M A T E S E M A N T I C S To my mind, theta-roles are not part of syntax and are more likely properly associated with the cognitive referents of NPs, and assume significance only in a cognitive psychology for the most part shared by humans and other primates. If so, there is no advantage in "representing" thetaroles linguistically, either as annotations to NPs or in specifically linguistic "conceptual structures". We can conceive of Full Interpretation and of the principles of Figure, Ground, and Agent Specification as part of the speaker's knowledge about how to convert linguistic representations into cognitive representations, without supposing that these principles first assign linguistic features or marks to the structures they apply to. In this way, the theta-roles of NPs are "indirectly" rather than "directly" represented in syntactic structures. 52 For instance, we (and plausibly primates) have some notion of possession so that our reactions when mildly hungry and alone with a particular food may differ according to who we consider the food "belongs to". But a cognitive pairing of possessors and possessions does not have to be connected to any sentence. If an interlocutor enters during our (primate) hesitation and says Oh, John left that food for you, principles of past tense interpretation, deixis, conversational postulates (the speaker isn't lying, etc.), and Figure and Ground Specification can interact to change our reaction to the food, via a change in the psychological possessor/possession pairing that the interactive sum of these cognitive principles brings about. (Contrast the increased hesitation that might be attendant on Oh, you 52 In discussing the usefulness of Fillmore's (1968) case features, which are close counterparts to theta-roles, Chomsky (1972a, p. 75) makes exactly this point. If, however, other rules of interpretation need to know which NPs have which theta-roles, then such rules must have access to the interpreted outputs of the system provided here. Such an output might be simply a syntactic structure with annotated theta-roles, or, like the R-structure of Cuficover and Wilkins (1986), it might be treated as a predictable and non-autonomous level of semantic representation. Culicover and Wilkins argue that R-structure is necessary for the proper statement of Disjoint Reference and for several other well-formedness restrictions on interactions between predication and particular theta-roles. However, they maintain, contrary to my positions here, that semantically understood arguments are not represented in deep structure trees, and that obligatory control infinitives do not have empty NP subjects. Their arguments for R-structure partly depend on these other positions. 424 JOSEPH E. EMONDS leave that food for John.) But none of this requires that John as Ground/Goal/Possessor be represented as part of any linguistic structure; this "role" is rather a result of a modified cognitive uncertain assessment of the situation, which is arrived at by using the linguistic structure to map a previous assessment of the situation into a more fully specified one (the deictic indicating that the structure is relevant to the present). My indirect representation of theta-roles in deep structures resembles Jackendoff's treatment of them within his conceptual structures: In other words, thematic relations are to be reduced to structural configurations in conceptual structure; the names for them are just convenient mnemonics for particularly prominent configurations. (Jackendoff 1987, p. 15) Replacing "conceptual structure" by "(non-linguistic) cognitive representations", the above quotation expresses my position. In fact, one could characterize the difference between Jackendoff's and my positions by saying that we have the same position on theta-roles, but that he holds that some subset of conceptual (or better, cognitive) representations are derivationally linked to deep structures, whereas I hold that this derivational link remains undemonstrated. My position derives from the suggestions in Chomsky (1972a, pp. 67-68) on this matter, and is consistent with the parsimonious hypothesis I privileged at the onset, that the categories of connected thought are nothing else than the categories of syntax (possibly elaborated by general interpretive principles which modify cognitive representations). Under this view, formal semantics refers to the construction of mappings from purely syntactic structures to representations of cognitive states, or, more accurately, from pairings of cognitive states and syntactic structures to modified cognitive states. Since only the faintest beginnings have been made in representing cognitive states formally, in Jackendoff's work on cognition and perhaps in "situation semantics", the obstacles to the "formal semantic" enterprise are formidable indeed, s3 I believe that my position on thematic relations is strengthened by an assessment of Talmy's work on Figure and Ground. The fact is that most of his many properties that reinforce the Figure/Ground contrast (relative familiarity, relative prominence, etc.) hold not at all at a purely linguistic level; water shutoffs and computer facilities differ in complexity, familiarity, size, etc. only at a pragmatic (again better, cognitive) level. Rephrasing, these differences hold not of linguistic phrases but of the (cognitive or signified) referents of those phrases. If Figure and Ground play a s3 My ideas here have been arrived at in discussion with Alice ter Meulen, who is less pessimistic. THETA-ROLE ASSIGNMENT 425 predictive role in the representations of these contrasting properties, then they are most plausibly part of primate psychology, quite outside the proper domain of linguistic representation. Summarizing, it seems to me eminently possible that the purely cognitive representations that interact with at least the thematic predicates discussed here (of spatial motion and location, possession, emotion, communication, knowledge) do not vary in quality from those available to primates. What is human is the ability to conjure up complex propositional representations of possible worlds, both consciously and inadvertently, through the medium of syntax and in the absence of any external stimuli. (It is quite plausible that any dreamlike imaginings which primates share, though stimulus-free, are not propositional representations.) There is also, of course, a "human semantics"; general principles such as Figure, Ground, and Agent Specification do not apply in the primate world because there are no bar notation X °, YP, or subjects of verbs for them to apply to. But except for their meanings in isolation, individual lexical items (idioms possibly aside) do not contribute to the "human semantics" of connected thought; they participate in human syntax (contextual subcategorization features) and primate semantics. 54 In human language, the possible semantic relations of individual morphemes with other items cannot be learned or even represented except by applying item-independent and largely universal human principles of interpretation to syntactically well-formed sequences. Thus, I answer Vygotsky's question, "is there thought without language?" by saying: there is no propositional or even phrasal thought without syntax. In other words, external stimuli or even an internal state can evoke non-propositional groupings of concepts or "conceptual structures" (e.g., dog, chase, child), but only syntactically connected words have a meaning which can be checked for truth and appropriateness. For instance, the above triple can, in the sight of a dog chasing a child, appropriately give rise equally well to do children ever chase dogs? or that dog is chasing the s4 Following Chomsky (1965, ch. 2) and Emonds (1985, chs. 4 and 5), features such as A N I M A T E , P L U R A L , and C O U N T are syntactic in English because they independently appear in syntactic rules other than subcategorization features. Therefore, a requirement that a verb take an animate or plural subject or object can be stated in purely syntactic terms and so can be part of an individual lexical item, even under the strongest version of Semantic Atomism (21); see notes 24 and 37. The human semantics of individual items (such as game, physics, advantage, Islam, care, satisfy, cause, should, over, etc.) can be complex or at least intractable, but the problems in characterizing such items are not resolved by first elaborating how these elements behave semantically when standing in grammatical relations. The specifically human complexity of these items is fully present independent of any sentences they appear in. 426 J O S E P H E. EMONDS child or a chased child is like a dog. There are no relational semantic representations to "project" from chase or to be "selected" by chase because the semantics of lexical items is entirely atomic (21) and non-concatenative. I conclude that connected or rational or structured thought is completely dependent on (i) lexical items which have inherent cognitive associations but only syntactic properties of combination, and (ii) general principles of semantics which interpret syntactic structures but have no possible effect on syntactically unconnected sets of concepts. 55 The qualitative differentia specifica of human language and human thought, in light of the ability of primates to acquire and use limited inventories of meaningful names, appears to reside in the syntax, the phonology which makes a large lexicon possible (Hockett's (1960) "duality of patterning" or de Saussure's (1916) "arbitrariness of the sign"), and the principles of interpretation which interface the syntax with cognitive structures. In our present state of knowledge, fruitful study of the rational mind must be rooted in the study of syntax. REFERENCES Anderson, Stephen: 1971, 'On the Role of Deep Structure in Semantic Interpretation', Foundations of Language 7, 387-396. Baltin, Mark: 1987, 'Do A~tecedent-Contained Deletions Exist?', Linguistic Inquiry 18, 579-595. Banfield, Ann: 1982, Unspeakable Sentences, Routledge and Kegan Paul, Boston. Belletti, Adriana and Luigi Rizzi: 1988, 'Psych-Verbs and 0-Theory', Natural Language and Linguistic Theory 6,291-352. Berman, Arlene: 1974, 'On the VSO Hypothesis', Linguistic Inquiry 5, 1-38. Bowers, John: 1976, 'On Surface Structure Grammatical Relations and the Structure-Preserving Hypothesis', Linguistic Analysis 2, 225-242. Brekke, Magnar: 1988, 'The Experiencer Constraint'~ Linguistic Inquiry 19, 169-180. Bresnan, Joan: 1982, 'Control and Complementation', The Mental Representation of Grammatical Relations, MIT Press, Cambridge. Bresnan, Joan and Jonni Kanerva: 1989, 'Locative Inversion in Chiche~a', Linguistic Inquiry 20, 1-50. Chomsky, Noam: 1955, 'Logical Syntax and Semantics: their Linguistic Relevance', Language 31, 36-45. Chomsky, Noam: 1965, Aspects of the Theory of Syntax, MIT Press, Cambridge. Chomsky, Noam: 1968, Language and Mind, Harcourt Brace, New York. Chomsky, Noam: 1970, 'Remarks on Nominalization', Studies on Semantics in Generative Grammar, Mouton, The Hague, pp. 11-61. Chomsky, Noam: 1972a, 'Deep Structure, Surface Structure, and Semantic Interpretation', Studies on Semantics in Generative Grammar, Mouton, The Hague, pp. 62-119. 55 When I say "syntactic", I mean, "expressed in terms of the best justified category system used for transformational syntax", i.e., the bar notation. The general and claimless usage whereby "syntax" means only "subject to some well-formedness conditions" (often unspecified) is without scientific interest. THETA-ROLE ASSIGNMENT 427 Chomsky, Noam: 1972b, 'Some Empirical Issues in the Theory of Transformational Grammar', Studies on Semantics in Generative Grammar, Mouton, The Hague, pp. 120-202. Chomsky, Noam: 1981, Lectures on Government and Binding, Foris, Dordrecht. Chomsky, Noam: 1986, Knowledge of Language: Its Nature, Origin, and Use, Praeger, New York. Culicover, Peter and Wendy Wilkins: 1984, Locality and Linguistic Theory, Academic Press, New York. Culicover, Peter and Wendy Wilkins: 1986, 'Control, PRO, and the Projection Principle', Language 62, 120-153. Emonds, Joseph: 1972, 'Evidence that Indirect Object Movement is a Structure Preserving Rule', Foundations of Language 8, 546-561. Emonds, Joseph: 1976, A Transformational Approach to English Syntax, Academic Press, New York. Emonds, Joseph: 1985, A Unified Theory of Syntactic Categories, Foris, Dordrecht. Emonds, Joseph: 1986, 'Generalized NP-c~Inversion: Hallmark of English', Indiana University Linguistics Club Twentieth Anniversary Volume, Bloomington. Emonds, Joseph: 1987, 'The Invisible Category Principle', Linguistic Inquiry 18, 613-632. Emonds, Joseph: 1990, 'The Autonomy of the (Syntactic) Lexicon and Syntax: Insertion Conditions for Inflectional and Derivational Morphemes', Interdisciplinary Approaches to Language: Essays in Honor of S.-Y. Kuroda, C. Georgopoulos and R. Ishihara (eds.), Kluwer Academic Publishers, Dordrecht, Holland, 1990, pp. 119-148. Emonds, Joseph: 1992, 'Complement Selection and the Syntactic Lexicon: Rereading Syntactic Structures', A Festschrift for Nicolas Ruwet, Anne Zribi-Hertz (ed.). Fillmore, Charles: 1968, 'The Case for Case', Universals in Linguistic Theory, Holt, Rinehart and Winston, New York. Grimshaw, Jane: 1979, 'Complement Selection and the Lexicon', Linguistic Inquiry 10, 279326. Gruber, Jeffery: 1965, Studies in Lexical Relations, unpublished Ph.D. dissertation, MIT. Hacking, Ian: 1975, 'All Kinds of Possibility', Philosophical Review 84, 321-337. Hale, Kenneth and S. Jay Keyser: 1986, 'Some Transitivity Alternations in English', Lexicon Project Working Papers # 7, Center for Cognitive Science, MIT, Cambridge. Hale, Kenneth and S. Jay Keyser: 1987, 'A View from the Middle', Lexicon Project Working Papers #10, Center for Cognitive Science, MIT, Cambridge. Hockett, Charles: 1960, 'The Origin of Speech', Scientific American, 293(3), 88-96. Huang, Li-yi: 1990, The Deep Word Order and Some Prepositional Constructions of Mandarin Chinese, unpublished Ph.D. dissertation, University of Washington. Jackendoff, Ray: 1972, Semantic Interpretation in Generative Grammar, MIT Press, Cambridge. Jackendoff, Ray: 1973, 'The Base Rules for Prepositional Phrases', A Festschrift for Morris Halle, S. Anderson and P. Kiparsky (eds.), Holt, Rinehart and Winston, New York, pp. 345-356. Jackendoff, Ray: 1979, 'How to Keep Ninety from Rising', Linguistic Inquiry 10, 172-177. Jackendoff, Ray: 1985, 'Multiple Subcategorization and the 0-Criterion: The Case of Climb', Natural Language and Linguistic Theory 3,271-296. Jackendoff, Ray: 1987, 'The Status of Thematic Relations in Linguistic Theory', Linguistic Inquiry 18, 369-412. Katz, Jerrold and Jerry Fodor: 1963, 'The Structure of a Semantic Theory', Language 39, 170-210. Kuroda, S.-Y.: 1979, The (W)hole of the Doughnut: Syntax and Its Boundaries, E. StoryScientia, Brussels. Langer, Susanne: 1942, Philosophy in a New Key: A Study in the Symbolism of Reason, Rite, and Art, Harvard University Press, Cambridge. 428 JOSEPH E. EMONDS Lappin, Shalom: 1984, 'VP Anaphora, Quantifier Scope, and Logical Form', Linguistic Analysis 13, 273-315. Levin, Beth and Malka Rappaport: 1986, 'The Formation of Adjectival Passives', Linguistic Inquiry 17, 623-663. Lieber, Rochelle: 1983, 'Argument Linking and Compounds in English', Linguistic Inquiry 14, 251-285. May, Robert: 1985, Logical Form: Its Structure and Derivation, MIT Press, Cambridge. McCawley, James: 1970, 'English as a VSO Language', Language 46, 286-299. Milner, Jean-Claude: 1978, De la syntaxe ~ rinterprdtation, t~ditions du Seuil, Paris. Newmeyer, Frederick: 1980, Linguistic Theory in America, Academic Press, New York. Pesetsky, David: 1982, Paths and Categories, unpublished Ph.D. dissertation, MIT. Randall, Janet: 1987, 'Inheritance', in W. Winkins (ed.), Thematic Relations, Syntax and Semantics, 21. Academic Press, New York. Rappaport, Malka: 1983, 'On the Nature of Derived Nominals', Papers in Lexical-Functional Grammar, B. Levin, M. Rappaport and A, Zaenen (eds.), Indiana University Linguistic Club, Bloomington. Rappaport, Malka and Beth Levin: 1986, 'What to Do with Theta-Roles', Lexicon Project Working Papers #11, Center for Cognitive Science, MIT, Cambridge. Reinhart, Tanya: 1983, Anaphora and Semantic Interpretation, University of Chicago Press, Chicago. van Riemsdijk, Henk: 1978, A Case Study in Syntactic Markedness, Foris, Dordrecht. van Riemsdijk, Henk: 1982, 'Derivational vs. Representational Grammar', Linguistics in the Morning Calm, Linguistics Society of Korea, Hanshin Publishing, Seoul. Rizzi, Luigi: 1986, 'Null Objects in Italian and the Theory of pro', Linguistic Inquiry 17, 501-508. Roberts, Ian: 1990, Review of 'Levels of Representation in the Lexicon and Syntax', by Maria-Luisa Zubizarreta, Natural Language and Linguistic Theory 8, 465-489. Saussure, Ferdinand de: 1916, Cours de Linguistique Gdndrale, C. Bally and A. Sechehaye (eds.), with Albert Riedlinger, Payot, Paris. Stowell, Timothy: 1981, Origins of Phrase Structure, unpublished Ph.D. dissertation, MIT. Talmy, Leonard: 1975, 'Semantics and Syntax of Motion', in John Kimball (ed.), Syntax and Semantics, Academic Press, New York. Talmy, Leonard: 1978, 'Figure and Ground in Complex Sentences', in Joseph Greenberg (ed.), Universals of Human Language: Syntax, Vol. 4, Stanford University Press, Stanford. Talmy, Leonard: 1983, 'How Language Structures Space', in H. Pick and L. Acredolo (eds.), Spatial Orientation: Theory, Research, and Application, Plenum Press, New York. Talmy, Leonard: 1985a, 'Lexicalization Patterns: Semantic Structure in Lexical Forms', in T. Shopen (ed.), Language Typology and Syntactic Description (Grammatical Categories and the Lexicon), Vol. 3. Cambridge University Press. Talmy, Leonard: 1985b, Figure and Ground as Thematic Roles, distributed at the Symposium on Thematic Relations, Winter Meeting of the Linguistic Society of America, Seattle. Walinska de Hackbeil, Hanna: 1986, The Roots of Phrase Structure: the Syntactic Basis of English Morphology, unpublished Ph.D. dissertation, University of Washington. Whitney, Rosemarie: 1984, The Syntax and Interpretation of ~4-adjunctions, unpublished Ph.D. dissertation, University of Washington. Williams, Edwin: 1981, 'Argument Structure and Morphology', Linguistic Review 1, 81-114. Williams, Edwin: 1986, 'A Reassignment of the Functions of LF', Linguistic Inquiry 17, 265-299. Williams, Edwin: 1988, 'Is LF Distinct from S-Structure? A Reply to May', Linguistic Inquiry 19, 135-146. THETA-ROLE ASSIGNMENT 429 Zubizarreta, Maria-Luisa: 1985, 'Morphophonology and Morphosyntax: Romance Causatives', Linguistic Inquiry 16, 247-289. Zubizarreta, Maria-Luisa: 1989, Levels of Representation in the Lexicon and in the Syntax, John Benjamins, Amsterdam. Received 21 February, 1990 Revised 29 March, 1991 Department of Linguistics GN-40 University of Washington Seattle, WA 98195 U.S.A.