Sentential Complementation: Proceedings of the International Conference held at UFSAL, Brussels, June 1983 9783110882698, 9783110131130

Citation preview

SENTENTIAL COMPLEMENTATION Proceedings of the International Conference held at UFSAL, Brussels June, 1983

Linguistic Models The publications in this series tackle crucial problems, both empirical and conceptual, within the context of progressive research programs. In particular, Linguistic Models will address the development of formal methods in the study of language with special reference to the interaction of grammatical components. Series Editors: Teun Hoekstra Harry van der Hulst Michael Moortgat

Other books in this series: 1

M. Moortgat, H. van der Hulst, T. Hoekstra The scope of lexical rules

2 Harry van der Hulst and Norval Smith The structure of phonological representation. Part I. 3 Harry van der Hulst and Norval Smith The structure of phonological representation. Part II. 4 Gerald Gazdar, Ewan Klein, Geoffrey K. Pullum Order, Concord and Constituency 6 Teun Hoekstra Transitivity. Grammatical Relations in GovernmentBinding Theory

SENTENTIAL COMPLEMENTATION

Proceedings of the International Conference held at UFSAL, Brussels June, 1983 W. de Geest Y. Putseys (eds.) UFSAL, Brussels

¥ 1984 FORIS PUBLICATIONS Dordrecht - Holland/ Cinnaminson - U.S.A.

Published by: Foris Publications Holland P.O. Box 509 3300 AM Dordrecht, The Netherlands Sole distributor for the U.S.A. and Canada: Foris Publications U.S.A. P.O Box C-50 Cinnaminson N.J. 08077 U.S.A.

CIP-data

Published with financial support from the Belgian National Fund for Scientific Research, the Belgian Ministry of Education and the Contact and Documentation Centre for Flemish Linguists (UFSAL, Brussels) ISBN 90 70176 90 4 (Bound) ISBN 90 70176 29 7 (Paper) ® 1984 Foris Publications - Dordrecht. No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission from the copyright owner. Printed in the Netherlands by ICG Printing, Dordrecht.

" . . . It seems to me that this conference has revealed that a great deal of progress has been made towards genuine explanation" Tom Wasow, Chairman of the Discussion Panel.

Contents

Preface Henk van Riemsdijk Introductory remarks

IX

1

Joseph Aoun A significant generalization in Chinese

11

Joseph Bayer Towards an explanation of certain that-/ phenomena: The COMPnode in Bavarian

23

Hans Bennis and Liliane Haegeman On the status of agreement and relative clauses in West-Flemish . .

33

Robert D. Borsley On the nonexistence of VP's

55

Andrea Calabrese Multiple questions and focus in Italian

67

Kris ti Koch Christensen Infinitival (pseudo-) complementation of noun phrases in Norwegian

75

Gerald Gazdar, Ewan Klein, Geoffrey K. Pullum, Ivan Sag Foot features and parasitic gaps

83

André Hantson Towards an analysis of retroactive gerunds

95

Teun Hoekstra Government and the distribution of sentential complementation in Dutch

105

Geoffrey Horrocks The lexical head constraint, X-theory and the "pro-drop parameter"

117

VIII

Contents

Irma Hyvarinen Infinitivsyntagmen als Aktanten in den Verbzentrierten Satzmustern des Deutschen

127

Jan Koster Infinitival complements in Dutch

141

Guiseppe Longobardi Some remarks on connectedness and c-command

1S1

James D. Mc Cawley Exploitation of the cyclic principle as a research strategy in syntax

165

Pierre Pica On the distinction between argumental and non-argumental anaphors

185

Christer Platzack The position of the finite verb in Icelandic Jerrold M. Sadock The pragmatics of subordination

205

Mark Steedman A categorial theory of intersecting dependencies in Dutch Infinitival Complements

215

Hans Tilo Tappe On infinitival clauses without COMP

227

Knut Tarald Taraldsen The internal structure and external distribution of tensed clausal complements in English, French and Norwegian

239

Hoskuldur Thrainsson Different types of infinitival complements in Icelandic

247

Johan van der Auwera Subject vs non-subject asymmetries in the relativization of embedded NP's

257

References

271

195

Preface

Complementation is an essential method of making infinite use of finite means. It offers the possibility of constructing complex propositions out of simple ones. Accordingly, it must play a central role in any theory of linguistic structure. As organizers of the June '83 conference we assumed that a common interest in such a central linguistic issue should provide the necessary stimulus to bring together a number of proponents of different linguistic theories. The success of our endeavours can be gauged from this result: one hundred and thirteen participants from fifteen different countries attended the three day convention. To be part of the social fabric at the conference and to be engaged in the invigorating contacts between such a large number of researchers sharing related interests, is one thing; to present the results of a successful conference is another. The present book aims to perform just this function, although it offers only a partially complete harvest: twenty-three out of thirty-five papers read at the conference are represented here. Furthermore, we decided against making an edited version of the contributions to the lively panel discussion at the end of the conference, because we preferred to omit what otherwise would have been too diluted a version of the pointed ad-libs of the participants. In addition we leave it to the readers' imagination to detect the points of convergence to be found in the different approaches. It may suffice to direct their attention to the question of the shared arguments in matrix and embedded clauses; to the topic of the local dependencies, information about which may be expected to be conveyed in some way through the heads of phrases; to the chains conveying information about the filler gap dependencies; to the licencing of absent arguments; to the importance of the projection principle with respect to syntactic cooccurrence restrictions. We acknowledge our debt of gratitude for the success of the final panel discussion to the contributing participants: Barbara TomaszcykLewandowska (Lodz University), Flip G. Droste (Leuven University), Louis Goossens (Antwerp University), Knut Tarald Taraldsen (Tromso University), Anna Szabolcsi (Hungarian Academy of Sciences), and the chairman, Tom Wasow (Stanford University). We include in our gratitude Dany Jaspers of the Department of General Linguistics at UFSAL and

Leen Maebe, secretary to the Contact and Documentation Centre for Flemish Linguists (CODOC), for their sincerely appreciated assistance, and the series editors Teun Hoekstra, Harry van der Hulst and Michael Moortgat for their willingness to include the proceedings of the June '83 International Conference on Sentential Complementation in their Linguistic Models Series. Wim de Geest Yvan Putseys

Introductory remarks H E N K VAN R I E M S D I J K Tilburg University

There are, roughly speaking, two types of conferences: broad ones and narrow ones. What is at stake is the relationship between conferences like the present one which are perceived as quite open and diverse and conferences like the GLOW-conference, for example, which are perceived as narrowly focused. In the former case, what is most important is to point out that the people present, speakers and audience alike, really have much more in common than they usually believe and enact. In the latter case, it is necessary to stress the same point, namely that the people present should try to refrain from presenting their work as if they were talking to their next door research companion. What I would like to do, then, in these introductory remarks is two things. First, I will list a number of basic background assumptions which I believe virtually every linguist present at the meeting, and hence every linguist represented in the present volume, shares. Second, I will attempt to present a crude overview of the main clusters of issues which arise when we discuss sentential complementation in an attempt to provide a rough descriptive background for the discussions to follow. 1. Starting out with the former, I will present a list of ten background assumptions, partly overlapping or at least interrelated, which I am confident are shared. (i) Central problem: language acquisition While stories about the speed with which the infant acquires its mother tongue are reiterated ad nauseam, there is no denying that almost everybody will have to admit that it is one of the major goals of the linguistic theory or theories that we work towards to account for this astonishing feat. As a result, we regard the object which we study as a real object, a mental organ as Chomsky has put in. From this it also follows that what we study is the grammar, not the language. (ii) Universals Nobody can help being overwhelmed by the enormous diversity of natural languages. This we share with the structuralists, the romantics of the 19th century and many other periods in linguistics. Much stronger, however, is the feeling that universals are there to be found. We may disagree as to what they are, how they are likely to look, how to best find them, but we

1

2

Introductory

remarks

all think we must find them. The linguistic theory we work on must therefore be a theory about universal grammar. (iii) Recursivity If the grammar is a mental object it has to be finite. At the same time it has to account for potentially infinite sentences, infinitely long, infinitely many. It follows that recursion has to be built into the grammar. The evidence for recursion is omnipresent and inescapable. Rather than taking this to be a discovery of vast consequence, I submit that most linguists tend to regard it as a perfectly trivial and boring fact. But we should not forget that there are millions of people out there who do not know or realize this and who it is even difficult to explain it to. (iv) Sound - meaning As far as I can tell all models around hold that the grammar has to, among many other things, provide an explicit mapping between sound and meaning. Again, we may differ on how this is best accomplished, on how much phonetics has to go into the phonology, how much interpretation into the semantics etc., but the task remains the same. (t>) Sentence grammar We conceive of the grammar as a sentence grammar. Everybody knows that texts and discourse exist, that sentences may entertain relationships with one another in some way. Nevertheless I perceive a consensus to the effect that the sentence is to a very great extent an autonomous object within any larger whole, its structure being determined by rules and principles sui generis. (pi) Structure dependency Studying these rules and principles further, I believe no one can escape the conclusion that they are mainly structure dependent. The very fact that they are is probably one of the major incentives to pursue the study of grammar. (vii) Generative grammar Though the term "generative grammar" has, in addition to its purely formal scientific meaning, taken on a social interpretation, I feel it is worthwhile to point out that anyone is a generative grammarian (with no capital letters in the term) who holds that the grammar must be a finite rule system which explicitly characterizes an infinite set of sentences. I cannot believe that there are many linguists around who aren't generative grammarians in that sense. (viii) Levels of adequacy The days in which pure description was the sacrosanct goal of any serious linguist are past. True, the idea that we must go beyond description to

Henk van Riemsdijk

3

attain the nobler goal of explanatory adequacy is, as Geoff Pullum recently put it, a "vogue-driven" one which may sooner or later disappear again since "science goes through phases like everything else does". Undoubtedly there is also a great deal of hanky-panky with the notion of explanation and the number of times it appears in the linguistic literature is probably inversely proportional to the number of times we really see the light. Nevertheless the desire to eventually see the light ought to provide a strong unifying force in the present-day linguistics scene. (ix) Competence - performance Everyone knows what this distinction is about. It is one of the most criticized distinctions in the post-Aspects model-wars. Yet everyone lives by it, so I feel it is only fair to say that we have a consensus on this point as well. I am not denying, of course, that people may differ as to whether certain performance facts can be considered relevant to the theory of grammar, and if so which ones, but the fact remains that abstraction and idealization is a pervasive characteristic of current linguistic research. (x) Liberal methodology The term may not be that well chosen, but then it is always hard to say serious things about methodology. Modern linguistic research is characterized by a comparative freedom of strict methodological imperatives. In this sense it is clearly different from corpus linguistics, the level bound nature of discovery procedures, and the statistical strait-jacket of the social sciences. This is not to say that facts obtained in these cumbersome ways are rejected out of hand, it's just that everything goes. Beyond that, of course, there is the indisputable fact that most grammarians will tend to have most faith in serious introspection. Again a considerable and farfrom-trivial consensus. It ought to be obvious that the ten points of common philosophy that I have listed are not to be taken as ten commandments. But even though nobody gets banned or exorcized as a heretic if he rejects some of them, it would not hurt if we as the linguistic community regarded them as a little bit of a credo. I am saying this because it would help to build a stronger sense of coherence in the linguistic community and to overcome all the little petty disagreements which simply do not seem to me to be worth sacrificing the essential unity of the field for. I repeat that it is my interpretation that the ten points are by and large so uncontroversial that people forget that one might actually disagree on them. But it suffices to go back only a few decades in the history of linguistics to find that it is indeed possible to reject a majority of these points. And if we embrace them, then we should remain conscious of how closely we are linked. 2. With this in mind, let us turn to the topic of sentential complementation. How does one react to such a topic? I would not be surprised if quite a few theoreticians frown upon such an essentially descriptive topic which focuses on the properties of a group of con-

4

Introductory

remarks

structions which may well turn out to be artefacts or spurious generalizations. The average well-behaved GLOW-topic, for example, usually addressed a theoretical issue: bounded vs. unbounded rules, markedness, rule ordering etc. I do not think this is a matter of having to choose; there is a necessity for both types of topics. I do believe, however, that there is certain danger immanent in descriptive topics. The danger stems from the fact that the modern linguist is brought up with the notion that facts are theory-laden. As a result it is only too easy to jump to the conclusion that infinitival complements in LFG are quite different creatures from those in GB and so on and so forth. Well, that is nonsense. It is easy to point to any number of problems in the domain of sentential complementation which anyone who holds the above credo just cannot afford to ignore. The relevance of the findings may be slightly different from so-called theory to so-called theory but first and foremost we have to struggle to understand what is going on. I found it rather difficult to find a single organizing principle on which to base a coherent overview of the issues concerning sentential complementation which to my mind stand out as the central ones, but I will try anyway. Also, I find it extremely difficult to divest myself of the vernacular of my own incrowd, the main reason being that alternative terminologies are generally characteristic of some other incrowd. So I will simply stick to the terminology of EST. What I will try to do is to review the main syntactic categories or positions involved in sentential complementation and discuss the ways in which they figure in the major issues that concern us. (i) S When a complement clause has a lexical complementizer, it is quite uncontroversial that it is of a specific category, call it S. Beyond that there is already much confusion and problems abound. Is the distribution of such complement clauses directly determined by this category, for example? In most languages true noun phrases and Ss do not have a completely identical distribution. For some languages at least a radical solution to the effect that S never appears in an N P position has appeared defensible (e.g. Emonds' earlier work). Such a solution raises many questions, of course, such as how it is possible for an S to have argument status with respect to some matrix predicate. Do the rules affecting predicates and arguments generalize freely over N P and S and their distinct positions in a clause, or must Ss be linked to a predicate-argument frame via some NP-position? To the extent that we adopt somewhat less radical alternatives, how do we account for the partial or, in some languages, total overlap between N P and S positions? Is the old approach of having S directly and exhaustively dominated by N P viable? Can we express the similarity in positioning by making use of categorial features in an expanded or modified version of category theory such as, say, the X-theory? Does such

Henk van Riemsdijk

5

an external identification of N P and S have a counterpart in the internal structure of N P and S? The answer to the latter question appears to be: yes to a certain extent, but many interesting issues remain. To mention just one, Szabolcsi has recently argued that N P s have a C O M P position in Hungarian. Can her argument be maintained and extended to other languages? Still pursuing the question of similarities between N P and S, to what extent are these categories subject to the same rules and principles? To mention just a few cases, are sentential complements subject to case matching just as N P s are? Since both S and N P serve to define domains which impose locality conditions on grammatical processes, to what extent can their domain properties be identified or correlated? Also related to the question of N P and S is the cluster of problems related to a host of constructions which are somewhat in between. Most prominent among these are the various types of nominalizations and gerund-like constructions. A final question which I would like to bring up concerns the phrasal nature of S. In other words, in what way can S be said to have a head, what is it a projection of? Here, as in so many domains of grammar, one has a feeling that anything goes. S was originally introduced as a projection of S (not your typical head, mind you), but has since been claimed to be a projection of C O M P , the A u x i l i a r y node, or of V. (ii) S (reduced) About S there will not be that much to say, meaning not that its properties are clear and uncontroversial, but rather the opposite. The only direct evidence for its existence, to my knowledge, is still Bresnan's argument from Right Node Raising based on sentences like John asked whether, but I am quite certain that, Bill left John asked when, and Bill wanted to know where, the party would be The argument is based on Postal's observation that what undergoes Right Node Raising must always be a constituent. Unfortunately this argument cannot be maintained in many languages. Dutch, for example, refutes it straightforwardly. Beyond Right Node Raising there has been mainly circumstantial or indirect evidence. Particularly frequent are statements of the sort "if X hangs from S and Y from S, then we account for the difference between X and Y " . Much more impressive, though not uncontroversial, is Rizzi's argument that languages differ as to whether they have S or S as a bounding node. Occasionally the issue of "bare-S complements" crops up. The option suggests itself in the case of complement clauses without an overt complementizer but with a lexical subject. Two main tracks can be distinguished here, I believe. One may either assume base-generated S-

6

Introductory

remarks

complements, with immediate consequences for subcategorization, Xtheory etc. Alternatively one can assume that in the absence of a lexical complementizer the S may sometimes and somehow become invisible or transparent, a process occasionally referred to as S-deletion. (ii'i) VP If I had to single out the most problematic category in this domain, I would definitely pick the VP. Its existence, at least in some languages, is firmly established. Nevertheless, its status has remained completely unknown. The majority opinion, which happens to coincide with Chomsky's view, appears to be that VP is the maximal projection of V but that, at least in complement structures, it never occurs independently as a complement without being dominated by S and S. I d o not deny that this view has much to recommend itself, but the ensuing problems are not negligeable. For one thing, if we compare VP to the other maximal projections (NP, P P , AP) the VP is clearly the odd guy out. Its structure is substantially different, especially as far as the specifier system is concerned, and if we adopt Stowell's proposal that all maximal projections have a subject N P , then the difference becomes even greater. The problem of VP complements is a much discussed one, and one that is heavily charged with theoretical issues concerning empty subjects. But if VP complements really do not exist, I feel that Brame's classical "gap in the paradigm" argument has never been successfully countered. What Brame argued was that all other maximal projection categories can be independently subcategorized for as complements, so the VP would again be the sole exception to this. Of course the argument can be countered in all sorts of ways, for example by making use of the way a VP assigns a thematic role to its subject. But this seems to me to be just shifting the problem: why does the VP behave differently with respect to its subject than the other maximal projection categories? The problem is compounded by the persistent arguments from all sorts of languages that they have no VP. This issue is further related to the presence or absence of further internal hierarchical structuring inside the VP, i.e. whether there is a V etc., but I will pass over this question since it is only marginally related to the theme of complementation. (iv) V The verb as we have come to know it from the grammar of English is a fairly dull animal. The A u x i l i a r y system is where all the action is, but the verb just sits there and doesn't move. But slowly the verb is acquiring a new image since there are at least two well-studied and wide-spread ways in which it can move. First there is the process of Verb Second (V2) which we find in some form or other in all Germanic languages. But evidence is emerging that subject clitic inversion in French might also be reduced to V2. And more spectacularly, it turns out that V2 is a central characteristic of a whole bunch of West African languages. The problem, then, appears

Henk van Riemsdijk

7

in a new light. The question is not only why the verb moves in some cases but equally why it doesn't in some others. The V2 issue is not limited to the topic of complementation, of course, but it interacts with it rather closely in the sense that it is usually properties of the C O M P or the AUX system which trigger V2, and these properties (such as being [ +tense], being realized as a lexical complementizer etc.) are in turn often characteristic for specific types of complement structures and typically imposed by the requirements of the matrix verb. There is a second way in which verbs are mobile, one which is much more intimately connected with complementation. I am referring to the various ways in which the verb can lose some of its independence and become closely connected with the matrix clause, usually the matrix verb. This is the set of phenomena often called "clause union". The facts are mainly concerned with a number of ways in which positions in the complement clause, generally an infinitival clause, become more easily accessible to positions outside. To mention just a few, we find bound anaphors (reflexives and reciprocals) which take a distant antecedent, we find clitics which show up in the matrix clause, and we find operator-like elements which take wider scope than usual. The analyses for such clause union phenomena range from actual movement of the embedded verb out of the complement into the matrix clause (Verb Raising) via various forms of reanalysis or restructuring to base-generation of the monoclausal structure. (v) COMP Starting at the top of the tree again, but descending along left branches instead of right ones, we arrive at the position of the complementizer complex. In the C O M P reside some of the essential features determining complement types. The choice of lexical complementizer, the choice between embedded question clause and declarative complement, and, directly or indirectly, the choice between tensed and infinitival clauses is made in this position. Clearly, then the C O M P must be accessible to outside action. Its features can be subcategorized for - that much is certain. Beyond that questions arise, however. Can case be assigned into C O M P , for example, as some have suggested? One of the most vexing problems concerning the accessibility of C O M P has to do with long, that is apparently unbounded, extraction of whphrases and the like out of complement clauses. I believe virtually everybody now agrees that the C O M P - n o d e plays a crucial role in long movement. Either, under a movement analysis, C O M P serves as an intermediary position into which and out of which a phrase is moved. Alternatively, the C O M P may just be a coindexed member of some chain of local links relating the wfc-phrase and its lowest gap. But what determines whether long extraction is possible? Here we enter into the

8

Introductory

remarks

domain of bridge phenomena and from there straight into pragmatics, a field that many tend to shy away from. On the other hand not all of it can be pragmatics. If in some languages long extraction is simply not possible, then some other factor must presumably be responsible, but which one? I am not trying to suggest that the problems surrounding C O M P are all of such a complex and far-reaching nature. There are some quite simple puzzles, but they are no less vexing. To mention just one, there is the question of whether COMP, as a node, dominates more than just the position of the lexical complementizer. In other words, does C O M P dominate clause-initial w/i-phrases, does C O M P branch? Such a simple question, but I do not feel it has been answered to anybody's satisfaction. (vi) AUX Closely related to C O M P are some properties of the auxiliary system, which I will refer to as AUX (I find the terminological shift to such unpronounceable creatures as INFL and AGR regrettable). AUX is where, presumably, the difference between finite and infinitival clauses resides, in agreement with the corresponding features of the COMP-node. In addition, of course, a host of properties have been attributed to this position: the presence of agreement features, for example, with extremely farreaching consequences, internal to the government-binding framework at least. The status of such nodes as C O M P and AUX remains unclear for the same reason as that of S and S. It has been suggested, for example, that AUX is the head of the S-projection, in which case S is really AUX. But then is S = AUX or is C O M P the head of that one? On the other hand if S and S are projections of V, what sort of specifiers are AUX and C O M P then? The question of accessibility from the matrix clause is, of course, closely connected with the status of AUX and C O M P as a head, since, by and large, heads are supposed to be accessible to outside government while their dependents are not. Such, at least, is the belief that I have always held. And indeed the question of direct accessibility does arise. Among others, there is one issue here which is much neglected and which, is again not addressed in the present volume. What I have in mind is the problem of consecutio temporum. (vii) The subject NP The picture which I have presented above is a gloomy one; not so much because no progress has been made but because I felt it was good to show how many questions remain, how shaky all sorts of seemingly uncontroversial assumptions are, and hence how much work remains to be done. If there is one category in my list of seven for which I think I can make an exception, it is the subject NP. That is why, to end in an optimistic mood, I have put it last. An enormous amount of progress has been made, in my opinion, which improves our knowledge of the properties of the subject position. We have

Henk van Riemsdijk

9

a quite satisfactory story to tell about when the subject may or must be lexically expressed. The theory of abstract case plays an important role in that story. Conversely, we also have an account of what the properties of a phonologically null subject NP are, and under what circumstances nullsubjects can occur. We have learned a great deal about the so-called complementizer-trace phenomena or, more generally, subject vs. nonsubject asymmetries. And we have an analysis of them in terms of the Empty Category Principle. We have an account, largely based on the Binding Theory, about when a phonologically empty subject is a PROelement which is subject to control. I hasten to add that at this point the eulogy must be interrupted to mention the fact that we must rely on a control theory to tell us which NP can, must, or may not serve as a controller for a given PRO. Only the barest outlines of such a theory are available at present. I hope that my top speed overview of some of the major problems in the area of complementation may serve as a useful background for at least some of the material contained in the present volume. Even more, however, I hope that I have shown what an impressive amount of research remains to be done and how much insight to be obtained. As long as we do not lose sight of this, it will seem obvious that here is a task which we must confront not against each other or despite each other but jointly in a concerted effort to gain a better understanding of what linguistic theory has to say about sentential complementation.

A significant generalization in Chinese JOSEPH AOUN University

of Southern

California

This paper deals with the distribution of w/i-expressions in Chinese and its theoretical implication. In the first section, a comparison between the government-binding theory and the generalized binding theory presented in Aoun (to appear) is outlined. In Government-Binding, the distribution of variables is essentially accounted for by the Empty Category Principle. In Generalized Binding, it is accounted for by the binding theory. For instance, in the latter theory, it follows that variables - like anaphoric reciprocals and reflexives -obey the Nominative Island Condition which is subsumed under the binding theory. In the former, theory, it does not obey this Condition. Thus, in Generalized Binding, both (la) and (lb) violate the Nominative Island Condition: (1)

(a) *they ; said (that (each other,- left)) (b) *who did they say (tj that (tj left))

In Government-Binding, sentence (la) violates the Nominative Island Condition and sentence (lb) which illustrates a *that-f-efFect, violates the Empty Category Principle. With respect to a language where the Nominative Island Condition does not apply, the two theories make different predictions. Generalized-Binding predicts that both (la) and (lb) will be grammatical in this language. Government-Binding predicts that only (la) will be grammatical. In case (lb) turns out to be grammatical, it will be so for independent reasons. The language in question is Chinese. The Nominative Island Condition is irrelevant in this language: both (la) and (lb) are grammatical. We claim that this state of facts represents a significant generalization which can only be captured in a generalized binding framework (section 2).

1. Government Binding and Generalized Binding In a given domain, anaphors - such as reciprocals, reflexives, NP-traces must be bound, i.e. must be c-commanded by an antecedent in an 11

12

A significant generalization in Chinese

argument position (A-position). In the same domain, pronouns must be free, i.e. must not be c-commanded by an antecedent in an A-position. Roughly, an A-position is a position which receives a grammatical function - subject, object...-. An Ä-position (non-argument-position), on the other hand, does not receive a grammatical function. According to this characterization, COMP-position is an Ä- position. The opaque domain referred to above in which an anaphoric element must be bound and a pronominal free is defined by the binding theory in terms of governing category. A governing category may, informally, be characterized as the minimal clause or noun phrase containing the anaphoric expression or the pronominal and a subject - this is the so called Specified Subject Condition - or an agreement marker which happens in English to occur in tensed clauses only - this is the so called Nominative Island Condition. To illustrate, consider the following sentences: (1) (2)

John wants [PRO, to see himself,] *John wants [PRO, to see him,]

In (1) and (2), the governing category containing the anaphoric element in (1) or the pronominal element in (2) and the subject -PRO- is the embedded clause in which these elements appear. Thus, himself must be bound in this category, which it is. But him must be free, which it is not. Sentence (2) will violate the binding requirements. Consider, now, the following contrast: (3) (4)

*they,- said that each other, AGR left they, said that they, AGR left

In (3-4), the grammaticality judgments are reversed as expected. The minimal clause containing the reciprocal in (3) or the pronoun in (4) and the agreement marker is the embedded clause. In this governing category, the anaphor each other must be bound, which it is not and the pronoun must be free, which it is. Thus, (3) - but not (4) - will violate the binding requirements formulated in (5): (5)

(A) An anaphor must be bound in its governing category (B) A pronominal must be free in its governing category.

To the above principles, we should add a principle which applies to names and to empty elements coindexed with w/¡-phrases (wh-traces or variables). It requires these elements to be free: (5)

(C) An R-expression must be free

It is to be noticed that this principle is distinct in its formulation from the first two principles (5 A-B) in that it is not formulated in terms of governing category. Actually, there are proposals which dissociate this

Joseph

Aoun

13

principle from the other binding principles. They question its existence as an independent principle in the grammar and try to derive its effects from various rules and considerations; cf. Chomsky (1982). As it will become clear in the subsequent discussions, it is the effects of this principle rather than its status which will concern us. T o use the current terminology, it is irrelevant for our purpose to characterize this principle as an axiom or a theorem. As a consequence of this principle, neither the name in (6a) nor the variable in (6b) may be construed as coreferential with the pronoun he: (6)

(a) he said John likes beer (b) which man, did he say x, likes beer

The binding principles, thus, constrain the distribution of the various nominal expressions discussed so far. Furthermore, a subset of these nominal expressions -NP-traces and w/i-traces is constrained by the Empty Category Principle (ECP) which requires these empty elements to be properly governed. Roughly, an empty element is properly governed if it is a complement of a lexical category such as V or if it is c-commanded by a local antecedent 1 . In (7), for instance, the variable is properly governed by the embedded predicate: (7)

who, do you think that Bill saw x,

Similarly, in (8a), the variable is c-commanded by a local antecedent - the empty element in C O M P - ; hence it is properly governed: (8)

(a) who, do you think [gi, [x, AGR left]]

However, in (7b), the presence of that prevents the empty element in C O M P from properly governing the variable. The representation will be ruled out by the ECP: (8)

(b) *who, do you think [g t, that [x, AGR left]]

For variables, the antecedent relevant to the binding theory is different from the one relevant to the ECP. The binding theory requires variables to be A-free — i.e. not to have a c-commanding antecedent in an Aposition. The E C P requires these variables to be a complement of a lexical category or to be c-commanded by an antecedent in an A-position. More generally, the binding theory is a theory of A-binding. It is solely concerned with antecedents in A-positions. In Aoun (to appear), various empirical and conceptual inadequacies of the theory sketched above are discussed and it is pointed out that these inadequacies may be traced back to the ECP. This principle, thus, is eliminated and its effects are derived from various grammatical principles. In particular, it is shown that there are two kinds of anaphoric relations

14

A significant generalization in Chinese

and not only one as assumed in Government Binding; cf. Chomsky (1981): the anaphoric relation which holds between an anaphor and an antecedent in an A-position (A-anaphor) and the anaphoric relation which holds between an anaphor and an antecedent in an À-position (Aanaphor). Two anaphoric systems, thus, are distinguished: the Aanaphoric system whose members are A-anaphors and the À-anaphoric system whose members are A-anaphors. It, also, is indicated that the distribution of A-anaphors and that of Aanaphors is constrained by the binding theory which thus is generalized from a theory of A-binding to a theory of A- and A-binding: (9)

Generalized Binding Principles: (where X = A or A) (A) An anaphor must be X-bound in its governing category (B) A pronominal must be X-free in its governing category (C) an R-expression must be A-free

There are A-anaphors such as the reciprocals in Italian and the French ne... personne constructions which are subject to principle A of the generalized binding theory. As such, they must be A-bound in their governing category; cf. Aoun (to appear). Variables are considered to be A-anaphors subject to principles A and C of the binding theory. More precisely, as in Government Binding, variables are treated as name - like expressions subject to principle C as illustrated in (6a-b). Generalizing the definition of anaphors to include all empty elements, they will also be treated as anaphors subject to principle A of the binding theory. Obviously, both principles can be satisfied because variables are A-anaphors and not A-anaphors: they need a ccommanding A-antecedent in their governing category. To illustrate, consider once again sentences (8a-b). The governing category for the variable in subject position is thé embedded clause: it is the minimal clause containing the variable and an agreement marker -AGR-. In (8a), the variable is A-bound by the trace in COMP. In (8b), however, the presence of that prevents the trace from binding the variable which will be free in its governing category; thus, violating the binding principle A. In the previous paragraph, we illustrated how the generalized binding theory may account for the so-called *that-i effect (cf. Chomsky and Lasnik 1977) explained in terms of the ECP in Government Binding. Let us turn now to variables in object positions. As illustrated in (lOa-b), the presence of that is irrelevant for these variables: (10)

(a) [who; [do you AGR think [§t,that [Bill AGR likes Xj]]] (b) [who, [do you AGR think [gt,- [Bill AGR likes x ; ] ] ]

The grammaticality of (lOa-b) is straightforwardly accounted for by the generalized binding theory. To show how, we need to explain with

Joseph

Aoun

15

more details how the binding theory works. In Government-Binding, it is assumed that AGR and the subject to which it is related are coindexed. This may be thought of as a means of encoding the agreement rule at work between these two elements. Moreover, a notion of accessibility is introduced to characterize governing categories. Consider (11), for instance: (11)

*(for each other to win) AGR would be unfortunate

As we said earlier, in a given clause, the subject and AGR are coindexed. If, now, AGR were to function as an accessible element defining an opaque domain - Chomsky (1981) speaks of accessible SUBJECT - , we would have a situation were the wellformedness condition ¿-within-i is violated: AGR would be coindexed with the subject and each other2. In brief, AGR cannot function as an accessible SUBJECT for the variable. But the governing category for an element a is defined as the minimal clause or N P containing an accessible SUBJECT and a 3 . However, in (11) we have no accessible SUBJECT for a - the reciprocal - . Following a suggestion of N. Hornstein, it is assumed in Chomsky (1981), that the root clause counts as the governing category for an anaphoric element with no accessible SUBJECT; cf. Chomsky (1981) for a more detailed discussion. Now in (11), the governing category for the reciprocal which has an accessible SUBJECT is the root clause. In this clause each other does not have an antecedent, therefore the sentence will be ruled out by the binding theory. Beyond all the details of the machinery, the intuitive content of the proposal is clear. Elements such as anaphors must be bound; they are bound in a local domain defined in terms of accessibility. When there is no domain definable in such a way, consider the whole root clause as the local domain. With this in mind, let us turn to (lOa-b). In both sentences, the variable does not have an accessible SUBJECT. AGR of the embedded clause cannot count as an accessible SUBJECT. Recall that AGR is coindexed with the subject it is related to. Suppose it were to count as an accessible SUBJECT for the variable; since it is coindexed with the subject which is in A-position Bill, the variable would end up A-bound by Bill. Being subject to principle C, variables have to be A-free. Thus, if AGR were to count as an accessible SUBJECT for the variable, we would have a violation of principle C. A similar reasoning will show that AGR of the matrix clause cannot count as an accessible SUBJECT for this variable either. We, thus, are in the situation mentioned in the previous paragraph where an anaphoric element — the variable - has no accessible SUBJECT. Therefore, the root clause will count as the governing category for the variable in (lOa-b). In this governing category, the variable is A-bound by who; thus, satisfying principle A of the binding theory. To sum up, variables are subject to principles A and C of the binding theory. When a variable is in subject position, the governing category for this variable is the clause in which it is subject (cf. 8a-b). When it is in non-subject

16

A significant generalization in Chinese

position, the governing category is generally the root clause (cf. lOa-b). This amounts to saying that variables are subject to the Nominative Island Condition (NIC) but not to the Specified Subject Condition (SSC) 4 in the generalized binding framework suggested in Aoun (to appear). In the Government-Binding framework argued for in Chomsky (1981), variables are neither subject to the N I C not to the SSC; the so-called NIC effects illustrated in (7a-b) are accounted for in terms of the ECP: (12)

(a) In Generalized Binding: Variables: ( + NIC), ( - SSC) (b) In Government-Binding: Variables: ( - N I C ) , ( — SSC)

The predictions of each theory are quite clear. Let us consider the *that-i effect discussed above; cf. (8a-b). In the Generalized Binding framework, it is accounted for by the N I C and in Government Binding by the ECP. Suppose, now, that for some reasons, the N I C is irrelevant in some language. According to (12a), variables will not display a *that-r effect in this language. According to (12b), the fact the N I C is irrelevant in the putative language will have no impact on the *that-r effect. That is, the *that-t effect is expected to hold in this language despite the irrelevance of the N I C and if it doesn't hold, its absence will be accidental; it will not be possible to trace it back to the irrelevance of the NIC. It is to these considerations that we will turn in the next section. Chinese, as it will appear, is a language where the N I C is irrelevant. As predicted by (12a), the *that-i effect is irrelevant for variables in Chinese. In so far that it is not possible to draw any corelation between the absence of the N I C and the absence of the *that-r effect in (12b), we will claim that a significant generalization is missed in Government Binding. Chinese, thus, will provide evidence for (12a) over (12b).

2. A Significant Generalization in Chinese: The Absence of the NIC Consider the following sentence: (1)

*John, said that himself; AGR will come

The governing category for the reflexive himself in (1) is the embedded clause: it is the minimal clause containing the accessible SUBJECT -AGRand the reflexive. In this governing category, the reflexive is free; thus, violating the binding requirement A. Surprisingly enough, the Chinese counterpart of this sentence is grammatical: (2)

zhangsan, shuo (ziji, hui lai) say self will come "zhangsan said that himself will come"

Joseph

Aoun

17

As Huang (1982) indicates, whether a clause is finite in Chinese or not, its I N F L does not contain (AGR). It therefore follows that an element in subject position has no accessible SUBJECT in its own clause. Thus, an anaphor may occur in the subject position of a clause and have an antecedent outside the clause in which it occurs as in (2) or (3). In brief, the N I C is irrelevant in Chinese: (3)

zhangsan, shuo [ s [ s z i j i , you mei you qian] mei guanxi] say self have not have money not matter "*zhangsan said that whether himself has money or not didn't matter"

Recall, now, the discussion of the previous section where a comparison between Generalized Binding and Government Binding was outlined; cf. 1. (12a-b). In the Generalized Binding framework, we expect no *that-i effect since the N I C is irrelevant. In the Government Binding framework, no such corelation is expected: the *that-i effect - accounted for by the E C P in this framework - should not be suspended and if it is, this will be merely accidental. The fact is that the *that-t effect is suspended in Chinese. To show this, we need to briefly discuss the status of the E C P which subsumes the *that-t effect in Government Binding. The E C P does not only constrain gaps generated by syntactic movement but also gaps generated by L F movement - i.e. movement which takes place in the L F component —. This can be illustrated in (4): (4)

(a) *I don't remember which man said that which woman left (b) I don't remember which man said that John loves which woman

The contrast between (4a) and (4b) may be accounted for by assuming that every w/i-element which has not been raised in syntax to C O M P henceforth w/i-in situ - is subject to an L F movement rule which adjoins this w/i-element to a C O M P already filled by a w/¡-element; cf. Chomsky (1973), Kayne (1981), Aoun, Hornstein and Sportiche (1981). Thus, the LF-representation of (4a-b) will be (5a-b) respectively (irrelevant details omitted): (5)

(a) *I don't remember [§j which woman,-, which man f [x, said [§ 0 that x ; l e f t ] ] ] (b) I don't remember [§, which woman,, which man, [x,-said [§ 0 that John loves x ; ] ] ]

It is clear, now, that the ungrammaticality of (4a/5a) may be traced back to a *that-r effect which can be accounted for by the E C P in GovernmentBinding and by the binding theory in Generalized Binding. From the above examples, it appeared that there are two components

18

A significant generalization in Chinese

where w/i-elements can be raised: in Syntax and in LF. In English an element must be raised in Syntax to C O M P unless this given C O M P is already filled by another w/i-element. In this latter case, the w/i-in situ will be raised in LF. In Chinese, however, as indicated in Huang (1982), w/i-movement is restricted to LF: Syntactic w/i-movement does not exist. This is illustrated in (6-8). All the following Chinese examples as well as the previous ones are from Huang (1982): (6)

(7)

(8)

[Zhangsan wen wo [shei mai le shu]] ask I who buy-ASP book "Zhangsan asked who bought books" [Zhangsan xiangxin [shei mai-le shu]] believe who buy-ASP book "who does Zhangsan believe bought books?" [Zhangsan zhidao [shei mai-le shu]] know who buy-ASP book (a) "who does Zhangsan know bought books?" (b) "Zhangsan knows who bought books"

The only surface difference among these sentences is in the choice of the matrix verb. In (6), wen (ask) is a verb which requires an interrogative complement. In (7), xiangxin (believe) does not permit an interrogative complement. In (8), zhidao (know) may optionally take an interrogative complement. As indicated in the glosses, these selectional properties of the matrix verb account for the fact that (6) must be interpreted as a statement taking an indirect question, (7) as a direct question and (8) as either. As argued for in Huang (1982), a unified description of these scope facts is available if the existence of an abstract LF movement rule is postulated for Chinese. The LF-representations of (6-8) will be as in (9-11) respectively: (9) (10) (11)

[Zhangsan wen wo [[shei], [t, mai-le s h u ] ] ] [[shei],- [Zhangsan xiangxin [^mai-le s h u ] ] ] (a) [[shei], [Zhangsan zhidao [t,-mai-le s h u ] ] ] (b) [Zhangsan zhidao [[shei], [t,mai-le s h u ] ] ]

The selectional properties of the matrix verb will ensure that shei (who) will not be moved out of the embedded clause in (9), must be so moved in (10) and may or may not be so moved in (11); cf. Huang (1982) for further arguments in favor of the existence of the abstract LF-movement rule in Chinese. A consequence of the postulation of an abstract LF-movement for a wh-word, is that the empty element left after such an extraction should be sensitive to the *that-i effect which applies to LF extractions as illustrated in (5a-b). More precisely, we expect sentences parallel to (5a) to be ruled out in Chinese by the ECP. The fact is that the sentences parallel to (5a) are well-formed in Chinese:

Joseph

Aoun

(12)

(a) zhejian shi [ggen [sshei lai]] zui you with who come most have "who is the person x such that this matter has most guanxi relation to do with x's coming?" (b) zhejian shi [ggen [§ni xihuan shei]] zui you like who most "who is the person x such that this matter has most you guanxi have relation to do with your liking x?"

19

Obviously, the empty category left by the LF extraction of who in (12a), would be ruled out if the E C P were to apply in Chinese. In (12a-b), no subject/object asymmetry- cf. (4a-b) - is observed. The well-formedness of (12a) shows that the E C P is violated. In Government-Binding, one must resort to an ad hoc analysis to account for the grammaticality of (12a). Thus, it is possible to claim that I N F L in Chinese has much more lexical content to it than I N F L in English: aspect markers in Chinese are derived from lexical categories and may be used as independent lexical items; cf. Huang (1982). According to this characterization, I N F L in Chinese, but not in English, would be a proper-governor. Thus, the empty category left by the L F extraction of the w/i-element in sentences such as (12a) would be properly-governed. In short, in Chinese, both subjects and objects would be properly governed. While the characterization of I N F L as a proper-governor may descriptively solve the problem raised by (12a) for the ECP, it surely is not able to relate the grammaticality of (12a) to that of (2). In other words, in a government-binding framework where the E C P is maintained, one is forced to claim for Chinese that the N I C is irrelevant because there is no AGR - hence the grammaticality of (2) - and that there is no *that-t effect because I N F L is a proper-governor — hence the grammaticality of (12a). In the Generalized Binding framework, the phenomena illustrated in (2) and (12a) fall under a unique generalization: the absence of AGR - i.e. the irrelevance of the N I C - in Chinese accounts for the grammaticality of (2) and (12a). At the beginning of this section, while discussing example (2), we indicated how (2) is straightforwardly accounted for by the binding requirements under the assumption that I N F L does not contain AGR. We will indicate now how the generalized binding theory also accounts for (12a) whose L F representation is given in (13): (13)

[shei,[zhejian s h i [ g g e n [ § x , l a i ] ] zui guanxi]]

Since the embedded C O M P is already filled with gen (with), even if we assume that there is a trace in this C O M P , the variable x would not be

20

A significant generalization in Chinese

properly bound. Now, the embedded clause cannot be the governing category for the variable since it lacks AGR which could serve as an accessible SUBJECT. The subject of the matrix clause cannot serve as an accessible SUBJECT; otherwise principle C would be violated: the variable would end up A-bound by this subject; cf. the discussion of examples 1.1 (lOa-b) of the previous section. Thus, we are in a situation where an anaphor has no accessible SUBJECT. By the extension of the characterization of governing category discussed in the previous section, the root clause will count as the governing category for the variable in (13). In this category, the variable is A-bound by the w/i-element shei. No violation thus, occurs. To recapitulate the content of this paper, in Chinese, I N F L does not contain AGR. That is, the NIC is irrelevant in this language. The Generalized Binding approach predicts that as a consequence of the irrelevance of the NIC, there is no *that-i effect in this language. This prediction appeared to be fulfilled. In the Government-Binding framework, no such prediction is made and the absence of the *that-r effect is a mere accident that has to be accounted for by an ad hoc analysis. A significant generalization is, thus, missed.

Appendix: NP-movement and the ECP Before closing this paper, some remarks are necessary. It is to be mentioned that in Chinese NP-movement - and not only w/¡-movement violates the ECP. To illustrate this remark, we need first to consider the following contrast in English: (1)

(a) *John ( is probable [§t, to win] (b) John; is likely [t,to win]

In Government-Binding, representation (la) is ruled out by the ECP since the trace is left in non-properly governed position: this trace is not properly governed in the embedded clause in which it appears or in the matrix clause since the presence of S prevents it from being properly governed by the matrix predicate. In (lb), on the other hand, a process of S-transparency or S-deletion is assumed. This process, which is triggered by raising predicates, allows the matrix predicate to properly govern the embedded trace. Consider now the following examples in Chinese: (2)

(a) ta ba shoupa,ku [de [t,- dou shi-tou le]] he BA handkerchief cry C O M P even wet-through ASP "he cried until he got the handkerchief entirely wet" (b) shoupa, bei ta ku [de [t, dou shi-tou le]] handkerchief by he cry C O M P even wet-through ASP "the handkerchief was made all wet by his crying"

Joseph Aoun

21

The grammatically of (2a-b) seems to indicate that the ECP is violated under the assumption that the presence of de in the embedded C O M P prevents S-deletion from applying. If this is assumed, then one must account for the grammaticality of (2a-b) by assuming that the trace is properly governed in the embedded clause; cf. Huang (1982). It however is not entirely obvious that the presence of a complementizer prevents Sdeletion from applying, specially if in place of S-deletion, an Stransparency process or a rule changing S to S is assumed as indicated in Chomsky (1981), p. 173. Assuming that the process of S-transparency has applied in (2a-b), the embedded trace will be properly governed in Government-Binding. In Generalized Binding, no problem arises either if it is assumed that 5-transparency applied in (2a-b). To show this will take us beyond the scope of this discussion; the reader is referred to Aoun (to appear).

Notes 1. a properly governs ji iff a governs P and a is lexical, cf. Chomsky (1981). Roughly the notion of government is that of head-government: a governs fi iff the first maximal projection dominating /? dominates a; cf. Aoun and Sportiche (1981). 2. a is accessible to /? iff /? is in the c-command domain of a and coindexing of (a,/?) would not violate the well-formedness condition i-within-i; cf. Chomsky (1981):

3. P is a governing category for a, iff fl is the minimal category containing a, a governor of a and a SUBJECT accessible to a. For ease of exposition, we, will omit any reference to the governor of o in our chracterization of governing category. 4. In Aoun (to appear) some SSC effects for variables inside noun phrases are discussed. This, however, is irrelevant for our discussion.

Towards an explanation of certain that-i phenomena: The COMP-node in Bavarian* J O S E F BAYER

Introduction In this paper an attempt is made to trace back to some minimal parametrical variations some major syntactic differences between Standard German (SG) and Bavarian (B), the variety of German spoken in the south-east of West-Germany. My main concern is to arrive at an explanation for the violations of the *that-r filter which occur frequently in B. The Government and Binding framework, as developed in Chomsky (1981) offers a theory which allows us to derive the Bavarian daß-t phenomena and certain cases of pro-drop without stipulations.

1. Extractability Linguists dealing with German syntax are sometimes puzzled by the fact that there is a great deal of variation among speakers as far as extraction from finite clauses is concerned. Both the examples in (1) are impossible in the usage of rigid German "non-extractors": (1)

(a) *Wen,- glaubst who believe " W h o d o you (b) *Wer ; glaubst " W h o d o you

du [ d a ß E m m a t, liebt] you that E. loves believe that E m m a loves?" du [ d a ß t f E m m a liebt] believe loves Emma?"

The corresponding Bavarian sentences in (2) are perfectly grammatical: (2)

(a) Weam, moanstn [ d a ß da Franz t, troffa hot] who think-you that the F. met has " W h o d o you think that Franz has met?" (b) Wea,- moanstn [ d a ß t; an Franz troffa hot] "Who d o you think has met Franz?" 23

24

Towards an explanation

of certain that-t

phenomena

The problem does not seem to have to d o with the *that-i filter directly, because SG reacts to both subject- and object-extraction negatively, whereas B seems to be quite happy with violations of the Empty Category Principle (ECP), which says that an empty category should be properly governed. 1 According to this principle, (la) should be fine, because the verb liebt properly governs the trace, whereas (2b) should be bad, because the trace is not governed by an X°-element and not coindexed in an immediately perspicuous way. Why then are the data as they are?

2. COMP in Bavarian B provides impressive examples of doubly-filled C O M P . As a rule, any finite embedded clause may be introduced with two COMP-positions. This holds, of course, also for relative clauses. My examples indicate that the complementizers daß (for V-complementation) and wo (for Ncomplementation) may be missing, but in fact there are many speakers who almost never leave them away: (3)

(a) I woaß ned [wann, (daß) [da Xavea t; k u m m t ] ] I know not when that the X. comes "I don't know when Xaver will arrive" (b) Es is no ned g'wieß [wea,- (daß) [t; k u m m t ] ] it is yet not sure who that comes "It is not yet sure who will come" (c) dea Hund [dea; (wo) [t, gestern d'Katz bissn h o t ] ] the dog which that yesterday the cat bitten has "the dog which has bitten the cat yesterday" (d) de Frau [dea,(wo) [da Xavea t; a Bussl g'gem h o t ] ] the woman to-who that the X. a kiss given has "the woman who Xaver has kissed"

Before I turn to a problem with Bavarian relative clauses let me indicate how the ungrammaticality of (1) can be derived. SG, as spoken by nonextractors, has one and only one C O M P . Therefore, (la) is not rejected, because the trace left in object-position would be ungoverned (it is governed), but rather because C O M P is already occupied by daß. The wword simply cannot escape, because the usual escape-hatch, a 0 - C O M P , is blocked by a complementizer-word. A first question about relatives is why many speakers would express (3c) with the relative pronoun missing rather than leaving out wo: (3c) dea Hund [wo [e gestern d'Katz bissn h o t ] ] This is strange, because now there is no antecedent for e. Assuming that e = P R O is unmotivated, since in (3c) there was much reason to argue that it was a w-trace. 2 On the other hand, the relative pronoun in (3d) cannot be left out. The result would be ungrammatical. Since deletion of the

Josef Bayer

25

relative pronoun in C O M P leads always to ungrammaticality if the pronoun is not in the Nominative, we can formulate the following principle: (4)

Unmarked-Case Transmission (UCT) 3 [ C O M P x,-] [ C O M P w o ]

[ C O M P X ; ] [ C O M P

w o , ]

where i = the unmarked Case (i.e. Nominative) There is a slight complication which, however, does not affect (4). Note that we can render a relative without a non-Nominative pronoun grammatical, if the head of the relative has the same Case as the deleted pronoun: (5)

(a) I sog's dem M 5 [wo [des e g'heat]] ned I tell-it to the man that this belongs-to not "I won't tell it to the man to whom this belongs") (b) *I kenn den Mo [wo [des e g'heat]] ned "I don't know the man to whom this belongs"

In (5a) the head-NP dem Mb is a Dative; and the Dative is required for a pronoun in A-position to fill the gap e in the clause, because gehort governs a Dative-object; in (5b) the head is in the Accusative; therefore there is a Case conflict between Dative and Accusative. We can account for this situation with the following principle: (6)

General Case Transmission (GCT) [NP;•] [ C O M P X ]

[ C O M P ^ ^ C N P J ]

[ C O M P X ]

[ C O M P ^ J ]

where j = any Case; X = 0 . If an optional deletion rule deletes the pronoun which precedes wo, the head-NP has a chance to transmit its Case onto wo, rendering some sentences grammatical and others not. 4 The ungrammatical examples would be ruled out by Case-theory. (5ab) have to be analyzed as: (7)

(a) I sog's [dem Mo,] ned where i=j (b) *I kenn [den Mo,] ned where i / j

[ S E C O M P W

icoMPW0jl

[sdes t.g'heat]]]

[ § [ C O M P 0 ]

[ C O M P

[s des t i g 'heat]]]

To sum up, the N-complementizer wo can inherit the feature [ +nominative] from a moved and later on deleted Nominative pronoun, such that no ECP-violation occurs if we face a wo-t structure. In any other situation the grammaticality depends on the coincidence of head-Case and the Case of the deleted pronoun.

26

Towards an explanation

of certain that-t

phenomena

3. Full Pronouns and Clitics Now I want to draw attention to the fact that a syntax dilemma arises when we follow certain historically oriented descriptions of the Bavarian clitic (Nominative) pronouns. The lists of full and clitic pronouns (8) and (9) are taken from Altmann. (8)

Pronouns/Nominative5

Full

Bavarian

(9)

singular Stand.German

Bavarian

1 pers.

i:

ich

mia

2 pers. 3 pers.

du ea/dea si/de es/des

du er sie es

ia/e:s si/de

Clitic

wir ihr sie

Pronouns/Nominative Bavarian

1 pers 2 pers 3 pers

plural Stand.German

-a/-e -s(t) -a -s -s

Singular Stand.German 0 -de -a -se -s

Bavarian

plural Stand.German

-ma -(t)s -s

-wa -a -se

Note now the following examples of the movement of the subject pronoun into the second C O M P : (10)

(a) [§ [ C O M P

(b) (c) (d) (e) (f)

d a ß ] [ s t k u m m ] ] is d'Suppn I until that come is the soup schö koid already cold "Until I arrive the soup will already be cold" [ § [ C O M P D U ] [ C O M P bis d a ß ] [ s -st kummst]] . . . [ § [ C O M P d e a ] [ C O M P bis d a ß ] [ s t k u m m t ] ] . . . [s[coMpmia] [ C O M P b i s daß] [ s t kumma]] ... [s [cOMpia/e:s] [COMPbis d a ß ] [ s - t s kummts]] . . . [ § [ C O M P d e ] [ C O M P bis d a ß ] [ s t k u m m a ] ] . . . IO [ C O M P BIS

There is no clear picture of a syntactic process. Why should a clitic pronoun remain in the S-clause, if it is 2nd person? Note further that the same thing would lead to ungrammaticality in all the other persons: (11)

(a) *i: bis d a ß [g-e k u m m ] is d'Suppn schö koid (b) *dea bis d a ß [ s - a kummt] is d'Suppn schö koid

Another reason for not being content with the present analysis has to do with phonology. I assume that clitic pronouns should derive in an

Josef Bayer

27

immediately perspicuous way from the corresponding full pronouns. This seems to be the case for all the forms but for the 2nd person forms. I have to emphasize again that I do not make any claims which touch the historical development. 6 Among the forms in (8) and (9) it should not be too difficult to derive, say, /s/ from /si/, /ma/ from /mia/ etc., but it would be difficult to derive in synchronic phonology /st/ from /du/ and /ts/ from /e:s/. 7 To summarize, we have - as (10) indicates - an inconsistent syntax, and we have a doubtful phonology, but we have nice tables of the pronominal paradigm where every full form has a corresponding clitic. My claim is that it is neither the syntax nor the phonology of B which is messy, but table (9). Looking at the verb forms in (lOb/e) it appears immediately that -st/-ts are the personal endings on the verb. Therefore I hypothesize that they are not clitics in subject-position, but rather inflections on C O M P . This is the reason why they cannot disappear under conditions of movement. 8 4. Pro-Drop Note now that in (lOb/e), which are now respectively analyzed as (12a/b) (12)

(a) [du [bis dal3-st] [t kummst]] is d'Suppn scho koid (b) [ia/e:s [bis daB-ts] [t kummts]] is d'Suppn scho koid

the subject pronouns du and ia/e:s can be dropped. In that case it is unlikely that movement takes place at all. In all the other examples where we had a movement analysis already in (10), pro-drop leads to ungrammaticality. Leaving away the inflectional morphology in C O M P in (12) again rules out the sentences as unacceptable. Without even approaching a more sophisticated level of argumentation we can conclude from the way the data pattern that there must be an intimate connection between the make-up of C O M P and the adjacent empty element to the right of it. B "becomes" a pro-drop language just in case C O M P is inflectionally enriched to such an extent that the referential properties of the governed empty element can be recovered. Although -st/-ts are not clitic pronouns, they have the quality to specify C O M P in the relevant way. What kind of empty element is e in this case? Since there is no antecedent, I assume it not to be a wh-trace. 9 As there is no sentence-level controller, it cannot be PRO. A more relevant reason which rules out P R O and P R O a r b is that P R O would appear in governed position. This will become clear in the next section. 5. Are there ECP-Violations in Bavarian? I will show now that empty subjects are properly governed in the sense of Chomsky (1981). The E C P reads as follows: 10

28

Towards an explanation

(13)

ECP [ a e ] must be properly governed a properly governs ß iff (i) a governs ß and [(ii) a = [ ± N , ± V ] ° or (iii) a is coindexed with ß ]

(14)

of certain that-1

phenomena

If we generalize our principle U C T in (4) a little bit, we can account for the transmission of Unmarked Case in all clauses introduced by a complementizer: (4')

Generalized

UCT

com [ c O M P i ] [COMP plementizer] [coMpX;] E c o M P C o m p l e mentizer,] where i = the U n m a r k e d Case; X may be trace x

(4') guarantees that the index of trace is transmitted onto the complementizer, if the moved element is in the Nominative. F o r example in

1

'

\

[ § [ C O M P

1

TF]

i

[ S [ C O M P D A ß ]

[st/VP]]] i

daß inherits index i and can therefore properly govern f;. As far as I can see, (4') and (6) can handle some important problems of abstract Case in B. We have accounted for the situation where a subject-trace is properly governed. In the light of data where subject pro-drop can only occur if a defined morphology appears in C O M P , it is evident that this morphology is the trigger for (optional) pro-drop. Since the morphology in question consists of the 2nd person verb-affixes, we might argue that there is a correspondence-feature which links I N F L and C O M P : (16) (17)

[ C O M P , « 2 person] [ s . . . [ I N F L a 2 person] . . . ] [ + 2 person/sg] —> / — st/ / [ + 2 person/pl]->/ —ts/ / [ C O M P , ] [ — 2 person] -»/ —P/ /

...

A complementizer which attracts verbal morphology in such a way can readily be assumed to be a lexical governor. 1 1 If this holds, the E C P is fulfilled without stipulations on proper government: In B empty categories are properly governed either (a) because they are traces governed by a verb which is [ —N, + V ] , fulfilling (ii) of (14), or (b) because they are subject-traces governed due to G U C T (=(4')), fulfilling (iii) of (14), or (c) because they are base-generated empty elements which are governed by an enriched complementizer, which might be [ + I N F L ] , fulfilling again (ii) of (14). Under this analysis it is evident why P R O and P R O a r b d o not qualify. They would be governed by C O M P . Rather, the empty element following a governing C O M P seems to have all the properties Chomsky ascribes to pro.12

29

Josef Bayer 6. Where is INFL?

There are competing theories as to where I N F L is in German. The most frequent proposals are: (I) I N F L is directly dominated by S; 1 3 (II) I N F L is directly dominated by VP. 1 4 The following Bavarian data are graphically arranged to make my point clear: 15 (18)

(a) (b) (c) (d) (e) (f)

0 wenn-ts 0 mech-ts wen,mech-ts mi ; wenn-ts warum; dal3-ts (3 warum,-ts

pro pro pro pro pro pro

-me mech-ts -me t; t; mech-ts -me t,- mech-ts -me t,- mech-ts

(If you like m e , . . . ) (Do you like me?) (Who do you like?) (If you like M E , . . . ) ( . . . why you like me) ( . . . why you like me)

One can see that the finite verb mech-ts can also govern pro. The problem with proposal (I) is that it predicts (19)

*pro (ia/e:s) mi mech-ts you me like "You like me"

to be a well-formed declarative sentence: If I N F L is immediately dominated by S, it should under the required conditions allow for prodrop. This prediction is wrong. The advantage of (I) is, however, that there is a solution for Nominative-assignment. Proposal (II) predicts correctly that (19) is ungrammatical, because I N F L being part of the maximal projection VP cannot govern pro. Beyond this, (II) gives at least the idea of an explanation for V/2 in German: (19) may be freely generated by X-syntax. The nature of pro and the presence of an appropriate landing site for V / I N F L in C O M P ! could then be seen to be the trigger for V / I N F L movement. 1 6 The problem of assigning Case to the subject-NP that (I) could solve seems to be a "Scheinproblem". If we assume - as Chomsky has once proposed - that [ + nominative] is an abstract property of finite sentences (if there is a phonological matrix where the feature can be spelt out) we can capture the fact that in V/end-sentences the subject-NP has Case without being governed by AGR:

[a 2 pers]

pro

[a 2 pers]

30

Towards an explanation

of certain that-t

phenomena

Here, I N F L does not govern N P j . N P j gets Case by virtue of being the subject of a finite clause. It might be governed by some percolation mechanism in the sense of Safir (1982). If I N F L is marked [ + 2 person] it acquires the feature [ + pronominal]; 1 7 after movement to C O M P j it can properly govern pro. In V/end-sentences with an overt complementizer or a w/¡-moved element in C O M P j agreement-rule (16) guarantees that C O M P j becomes [ +pronominal]. Ungrammatical examples like (21)

*pro mech-ts -me ned "You don't like me"

are now ruled out: If pro were in C O M P 2 of (20), it could not be governed by INFL. pro could, however, be moved to C O M P 2 , leaving a trace which would be properly governed by mech-ts.ls Is then (21) ruled in again? I think it is not, because pro might be argued to move into an ^-position, consequentially binding the trace non-locally. Semantically it is unclear how an undetermined element like pro could serve as a binder of a whtrace. Without doubt, there is a superior solution which might be achieved with a deeper understanding of German syntax.

7. Conclusion The results of this study of C O M P in B are: It was shown that the E C P holds in B. It followed quite naturally from the fact that B has doublyfilled C O M P (or rather two COMP-positions), a general rule of Casetransmission which percolates the Nominative in form of an abstract index onto an otherwise unmarked complementizer-word, and the fact that B has a somewhat more vivid morphology than SG. An important fact is that in 2nd person C O M P acquires inflectional morphology. This enables C O M P to properly govern an empty pronominal in an adjacent subjectposition. At the same time we could avoid the strange effect of B's allowing for pronoun-doubling and w/¡-movement irregularly. With the present account we could also avoid difficulties which arise in synchronic phonology of B, when the pseudo-clitics should be derived from the full pronouns. Although B seems to be a dialect which is quite different from SG, we could show that an important subset of the syntactic differences which any naive observer can notice, is explained on the basis of a few parametrical variations. The differences between B and SG I focused on in this study arise mainly, because the dialect of strict "non-extractors" has only one C O M P . If this C O M P is blocked by e.g. daß, the following sentence is closed for movement. Since German observes Subjacency, it is obvious why under such conditions COMP-to-COMP-movement cannot take place. In B, on the other hand, one can observe all effects of long whmovement, because there is room for trace and a complementizer. C O M P -

Josef

31

Bayer

inflection makes it possible to drop 2nd person pronouns, - a process which is unknown in SG. B gives an example of what Chomsky in Chomsky (1981) suspects to be a language with a "mixed system", i.e. a language where subject-drop does not occur regularly, but only in some constructions.19

Notes * For their help and patience I want to thank the uncorrupted speakers of Bavarian who were always ready to listen to my "sentences", and the people who heard my previous talks about this subject. Especially I have to thank Tilman Hohle, Gunter Grewendorf, Peter Staudacher, Craig Thiersch, and Therese Torris whose critical questions sharpened my understanding of the subject matter. I have to emphasize that the dialect described here is my own. There is at least one other Bavarian subdialect which is revealing as far as that-t phenomena are concerned, namely Lower Bavarian. I say something about Lower Bavarian in a forthcoming study. 1. Carl Bremer (personal communication) and Van der Auwera (this volume) in his U6 suggest that (universally) subjects are harder to move out of their clauses than non-subjects. According to Van der Auwera, this property should be linked to the fact that the semantic role of subject NP's is less predictable than that of non-subject-NP's. Therefore, subjects should be more clause-bound than non-subjects (cf. his U5). I cannot see the force of the argument. Note also German constructions like Mir ist schlecht, Mir geht's gut, Mich friert, Dem Patienten wird ein Tumor herausoperiert where the initial Dative/Accusative-NP's are far from roles like "patient" or "recipient". In the light of the advanced research on the E C P I doubt that much can be gained from semantic and functional considerations in this area. Bavarian indicates that there is no prima facie asymmetry in German between subject- and object-NP's. I doubt that a quantitative study of overall German would contradict that. For "properly governed" see Fn. 10. 2. A more serious reason is that in other cases P R O would be governed. This is not an available option in the framework I am following. 3. This rule is inspired by the que-*qui-rule in Pesetsky (1982). There are good reasons to consider the Nominative to be the Unmarked Case. It is, for example, the Case of quotation, it is free of special morphology, it appears in infinitival constructions, if there is no element which could assign another Case. Agrammatism gives psycholinguistic evidence in support of my view. 4. Some native speakers get very confused with pure wo-relatives in metalinguistic decision. They either tend to reject all the examples or to accept all of them, including the ungrammatical ones. I follow my own intuitions and Merkle (1975), p. 148f. 5. Note that 2nd person plural has e:s which is an old dualis form. In 3rd person there is no clear distinction made between ordinary and demonstrative pronouns. 6. For historical descriptions see Weinhold (1867) and Schatz (1907). 7. Hans den Besten (personal communication) doubts that in cases like wenn-st kumm-st (if you come) C O M P lacks a pronoun completely. He thinks that the proper analysis would be something like wenn-s-t kumm-st, where -t is a residual of du, the full pronoun. For other German dialects this might be appropriate, e.g. Berlinerisch wenn-s-te. Note, however, the following distribution of grammaticality in the two dialects: (i) (a) (b) (ii) (a) (b)

y

wenn-st kumm-st 1 wenn-st du kumm-i n-st wenn-s-te komm-st 1 y *wenn-s-te du kommst

Bavarian . . . . Berlinerisch

These data indicate that wennste is a COMP-?-clitic structure, whereas wennst is an inflected

32

Towards

an explanation

of certain that-t

phenomena

C O M P . O n l y the latter tolerates a following full p r o n o u n . Of course, this claim does not extend to a historical theory. 8. W i t h this view I a m not alone. Pfalz (1918) proposed an analysis along these lines. See also K u f n e r (1961) as well as Richter (1979), where other dialects and languages are mentioned which show agreement p h e n o m e n a in C O M P . Bennis/Haegeman (this volume) in dealing mainly with West-Flemish, develop an account according to which clitics in general are not m o r e t h a n feature carriers, i.e. agreement p h e n o m e n a without referential force. But also in their d a t a one finds indications that certain elements derive f r o m the verb, whereas others derive f r o m the p r o n o u n , e.g. dapk (ik) kommen (that I come), däze (zunder) kommen (that they come), where —k 3 ^

+ fera]

- it (DO) - her(IO) she(SUBJ) given has

This order is optional as is indicated in (ii) (ii) ... dase zie t-ze gegeven heet Note that a kind of reduplication operation does indeed occur in West-Flemish in case of subject pronouns. This may lead to a sentence in which the subject pronoun is in fact three times visible, once as spelling out of the features of AGR, and two times in the reduplicated subject pronoun: (iii) . . . dank ze kik gezien heen ...that(_ p | u ! pg^j her (DO) I (Redupl.SUBJ) seen have 9. See Den Besten (1982) for a discussion of these cases, which turn out to provide an argument in favour of his V2-analysis. 10. It is important to observe that we expect this complementizer agreement phenomenon only to occur in languages in which INFL is part of COMP. This expectation is borne out, since the only other language we know that shows these agreement phenomena is German, a regular Verb Second language like Dutch and West-Flemish. These German data are discussed in Bayer (this volume). 11. A third type of relation which is generally assumed to be indicated by superscripting is the relation between a pleonastic element and a phrase with which it is associated. Depending on the particular analysis of this type of construction, it is possible to extend the scope of definition (33) in such a way that these constructions are also covered by it. 12. Note that several other types of agreement phenomena fall under the definition (33).

Hans

Bennis

and Liliane

Haegeman

53

NP-internal agreement, like the agreement between the noun and its determiner and/or adjective, may be conditioned by (33) under the definition of c-command as proposed by Sportiche and Aoun (1981). Secondly, subject-verb agreement can be seen as an instance of (33), since AGR does not only govern the subject but also the head of the V-projection, at least if we do not consider maximal projections like VP to be barriers for government. Another extension of this definition might be to define agreement relations on categories of the type [ + N ] in stead of [ + N, —V], In that case we are able to extend (33) to agreement relations in predicative constructions with adjectives. Notice that such an approach enables us to make an interesting observation. We may divide the class of governors in two groups. Governors of type [ —N], which are able to assign Case and governors of type [ + N ] , which trigger agreement. So where the structural domain is defined in a uniform way by the definition of government, the type of operation that expresses relationships within this domain is dependent from the feature [ + N]. Another consequence of this rather speculative idea is that INFL has an exceptional status. INFL is verbal, [ —N], at least +Tense is verbal, and therefore INFL is able to assign nominative case to the subject. On the other hand INFL is nominal, [ + N ] , at least AGR in INFL is nominal, and triggers agreement relations, as discussed in this article. It might be possible to derive subject/object asymmetries from this exceptional status of INFL. 13. With respect to (41) there may be another, alternative explanation. On the one hand the empty element is determined by a clitic-like element (se) in AGR, and is hence pro, i.e. pronominal. On the other hand it is also A-bound by the moved element in XP, hence a variable. This should thus mean that the empty position is an empty resumptive pronoun. But West-Flemish lacks the resumptive pronoun strategy, hence the sentence is ruled out. 14. One of the leading ideas in the discussion of the pro-drop parameter is the observation that the possibility of dropping the subject pronoun correlates with the possibility of inversion of the subject. In Bennis and Haegeman (forthcoming) we shall discuss the possibility of free inversion in West-Flemish. We shall try to show that WestFlemish does indeed allow free inversion of the subject, but that it is hard to make this inversion visible. The reason is that West-Flemish is an SOV-language. The effect of "Inversion" will be to adjoin the subject N P to the left side of the VP, the canonical position of objects. Since in non-inverted structures the subject is in a position directly to the left of VP too, it is difficult to show the difference. Evidence in favour of inversion is derived from the phenomenon of object clitics (cf. note 8) and the occurrence of an optional pleonastic element in front of the lexical subject as in (i) and (ii). (i) . . . dan tet Pol en Jan kommen . . . that [4 . plu] it Pol and Jan arrive (ii) Morgen komt tet Jan Tomorrow comes it Jan 15. For a discussion about chains and their properties, we refer to Chomsky (1981). The exact definition of the notion A-chain is not very crucial for our purposes. What is important, is to extend the notion "bind" to "BIND"; a notation that is used in the same way by Chomsky (1981, p. 333) (cf. also Safir (1982b) though for a uniform indexing hypothesis).

On the nonexistence of VP's* ROBERT D. BORSLEY University

College

London

In this paper, I will argue that there is no VP category as the term is normally understood. Following McCawley (1970), I will suggest that the constituents that are normally assigned to such a category should be assigned to the S category. For convenience, however, I will refer to these constituents as VP's and use the term S for ordinary S's. I will also use the term S for complementizer + S structures. My arguments will depend on the assumption that subjectless infinitives are VP's. I will show that a conventional VP category is quite problematic of one makes this assumption. I will then show that the problems disappear at least within the generalized phrase structure grammar (GPSG) framework if VP's are assigned to the S category. The data that I am concerned with here have been seen by proponents of the extended standard theory (EST) as evidence for the view that subjectless infinitives are S's. I will comment briefly on this view in the final section. 1. I will begin by outlining more fully the position I am arguing against and the position I am arguing for. The position I am arguing against can be summarized very briefly. By a VP category, I mean a category with three crucial characteristics: it labels a constituent containing a verb and its non-subject arguments and certain adverbial expressions, it is a projection of V, and it differs from S either in bar level or in projection status. The disjunction in the last point means that there are two kinds of VP category to be considered. I will show that both face problems. The exact nature of the position I am arguing for depends on the nature of the S category. I will assume with Gazdar (1982) that it is V. The assumption that it is a projection of V is important. It would be rather implausible to assign VP's to the S category if it were not a projection of V. The assumption that it is a double bar category is not so important. It may be in fact that it should be analyzed as a single bar category. I will consider this possibility shortly. I am proposing, then, that both VP's and S's are instances of V. Clearly, however, they must be distinguished in some way since they differ both in internal structure and in distribution. In propose to do this with a feature ±Subj(ect), marking VP's as [ - S u b j ] and S's as [ + Subj], Thus, English declaratives will have the basic structure in (1). 55

56

(1)

On the nonexistence

of VP's

V r+subii N

V [ — Subj]

This proposal is an extension of an idea in Gazdar (1982). Gazdar suggests that S's and S's should be distinguished not by bar level, as they are in, for example, Koster (1978), but by a feature ±C(omplementizer), S's being [ —C] and S's being [ + C]. If we assume this feature too, VP's will be V[ —Subj, - C ] and S's will be V [ + S u b j , - C ] , and instead of (1), we will have (2). V

(2)

+ Subj -C N

V -Subj -C

In what follows, however, I will normally just use the labels V[ —Subj] and V[ + Subj] As I indicated in the introduction, the idea that VP's should be assigned to the S category was first advanced in McCawley (1970). The only differences between the proposal I am making and McCawley's are that McCawley does not assume the X-bar framework and that he assumes that VP's are transformationally derived. He proposes that English has VSO underlying structures and that the superficial SVO order is a consequence of a verb-fronting rule. This rule gives rise to VP's. 1 Since VP's are transformationally derived, there are no phrase structure rules introducing or expanding VP's. Hence, McCawley has no need for the + Subj feature that I am assuming. One question that arises within the present proposal which could not arise within McClawley's is: is there a V category? I have no definite answer to this question. If there is a constituent consisting of a verb and its nonsubject arguments but no adverbial expressions, it is possible that it should be assigned to a V category. It may be, however, that there is no such constituent. Moreover, if there is such a constituent, it may be that it too should be labelled V[ — Subj], If this proposal does mean that there is no V category, one might propose that the S category should be V so that VP's are V[ —Subj] and S's V[+Subj], It is not obvious, however, that this would be preferable. In any event, I will continue to assume the original proposal. 2

Robert D. Borsley

57

2. We can now consider the problems that face a conventional VP category if subjectless infinitives are VP's. As I noted in the introduction, there are two kinds of conventional VP category. On the one hand, there are those where VP and S differ in bar level. O n the other, there are those where they differ in projection status. The first kind is found in, for example, Jackendoff (1977), Gazdar (1982), and Sag and Klein (1982). The second kind appears in Chomsky (1981) and Bresnan (1982) among others. I will look in this section at the positions of Gazdar, Sag and Klein, and Bresnan. I will ignore those of Jackendoff and Chomsky since they do not assume that subjectless infinitives are VP's. For Gazdar, VP's are V and S's V. All lexical categories have double bar maximal projections. Within this analysis, (3) and (4) will have the structures in (5) and (6), respectively. (3) (4)

John seems to like Mary. It seems that John likes Mary.

V

V

John seems to like Mary

An immediate objection to this analysis is that it precludes a restrictive characterization of the notion possible complement. Jackendoff (1977) suggests that complements are either maximal projections or grammatical formatives. He assumes that all lexical categories have phrasal projections. There is no evidence, however, that minor categories like determiners have phrasal projections. It seems, then, that we must add minor categories to the list of possible complements. Clearly, it is undesirable to expand the list any further. Given Gazdar's analysis, however, we must add a further possibility, namely V. The situation is similar with Sag and Klein's analysis. Sag and Klein assume that VP's are V and S's V. They assume that all other lexical

58

On the nonexistence

of VP's

categories have double bar maximal projections. Within this analysis, (3) and (4) will have the structures in (7) and (8), respectively.

On this analysis, complements can be double bar categories, grammatical formatives, minor categories and V. Like Gazdar's analysis, then, it precludes a restrictive characterization of the notion possible complement. Bresnan's analysis avoids this objection. For Bresnan, VP's are the maximal projection of V. S's and S's are projections of no category. S's, however, count as a maximal projection. Within this analysis, complements are either maximal projections, grammatical formatives or minor categories. In one respect, then, this analysis is preferable to Gazdar's and Sag and Klein's. In another respect, however, it is less satisfactory. An important feature of sentence structure is that an S and its main verb must have the same specification with respect to certain features, for example the feature + finite. It is natural in a non-transformational framework to attribute this to some version of the Head Feature Convention of Gazdar (1982). For this to be possible, VP's must be heads of S's. If this is the case, however, Bresnan's analysis is incompatible with the following constraint: (9)

A head and its mother have the same projection status.

Some form of this constraint has been assumed by many linguists, includ-

59

Robert D. Borsley

ing Bresnan herself (cf. Bresnan and Grimshaw (1978: 338)). It may be that it is too strong. At present, however, it seems quite plausible. 3 Therefore, the fact that Bresnan's analysis is incompatible with the constraint casts some doubt on it. Further objections to these analyses come from distributional similarities between subjectless infinitives and S's or S's, which have been highlighted by proponents of the view that subjectless infinitives are S's. There are a number of similarities. I will note just three. Firstly, there are certain verbs, for example expect, which take either a subjectless infinitive or an S. We have examples like the following: (10) (11)

John expects to be here. John expects that Mary will be here.

Secondly, both subjectless infinitives and S's appear in embedded questions. The following illustrate: (12) (13)

John wondered whether to go. John wondered whether he should go.

Finally, as Koster and May (1982) point out, subjectless infinitives and S's can be conjoined. Consider, for example, the following: (14)

John expects to be here and that Mary will be here.

These similarities are quite unexpected within the analyses under consideration. Thus, they cast further doubt on these analyses. It seems, then, that both kinds of conventional VP category are problematic if subjectless infinitives are VP's. 3. I will now show that the problems I have just outlined disappear given the view of VP's that I am advocating. • If VP's are analyzed as I propose, (3) will have the structure in (15), while (4) will have that in (16).

(15)

John

to like Mary

60

On the nonexistence of VP's

(16)

seems that John likes Mary

Like Bresnan's analysis, this analysis allows us to stipulate that complements are either maximal projections, grammatical formatives or minor categories. Unlike Bresnan's analysis, however, it is compatible with the constraint in (9). Given certain G P S G assumptions, it also permits a natural account of the distributional similarities between subjectless infinitives and S's or S's that we noted in the last section. For G P S G , as developed in Gazdar and Pullum (1982), the phrase structure (PS) rules that form the syntactic part of a grammar are the product of a metagrammar, which can be represented as follows: (17)

The starting point of the metagrammar is a set of partially specified immediate dominance (ID) rules. This set is expanded through the operation of metarules. The rules are then filled out through a process of feature instantiation. Finally, the fully specified rules are linearized in accordance with a set of linear precedence rules. The result is a set of PS rules. In the present context, the most important element of the metagrammar is feature instantiation. Feature instantiation is constrained by general principles such as the Head Feature Convention. 4 It is also constrained by feature coefficient defaults (FCD's), statements specifying the values features have when no basic I D rule or general principle says anything different. The analyses that I will be sketching will make crucial use of FCD's. One other feature of G P S G that we need to note is its treatment of subcategorization facts. These are handled by associating lexical items with basic I D rules. Where a number of PS rules are derived from one of these rules, a class of lexical items will be allowed in a number of environments. I can now provide an analysis of verbs like expect, which take either a subjectless infinitive or an S. I propose that such verbs are associated with

61

Robert D. Borsley the following basic ID rule: (18)

V •V, V [ —Subj]

(I am assuming here that there is no V and that V is immediately dominated by V = [ — Subj]. Nothing depends on this assumption, however.) The crucial feature of this rule is that the daughter V is unspecified with respect to the feature ± S u b j . I also propose that the metagrammar of English includes the following FCD's: (19)

(a) [ — Subj] -* [—fin] (b) [—fin] -> [ + inf] (c) [ + Subj] - » [ + fin]

I assume that these FCD's only affect the daughters in an I D rule. Thus, they will not require a mother node that is [ - S u b j ] to be [ - f i n ] as well. From (18), we can derive the following PS rules: (20)

->v

V - Subj' -fin _+inf

V ->V [-Subj]

V + Subj' + fin .+ C .

V

[-Subj]

(21)

In (20), the daughter V is [—Subj] and hence by (19)a. and b., it is [ — fin] and [ + inf]. In (21), the daughter V is [ + Subj] and hence by (19)c., it is [ + fin]. It is also [ + C]. Nothing requires this, but nothing rules it out. These rules will provide for examples like (10) and (11), respectively. Notice now that we cannot derive the following PS rules from (18): (22)

[-Subj]

-Subj ,+ fin

[-Subj]

+ Subj -fin + inf .+ C .

(23)

(22) violates (19)a. and (23) violates (19)c. Since these rules cannot be derived, sentences like the following will not be generated:

62

On the nonexistence

(24) (25)

*John expects will be here. *John expects for Mary to be here. 5

of VP's

Examples like (24) are ruled out by binding conditions in recent versions of EST. Thus, (19a) does some of the work of these conditions. We can turn now to embedded questions. Here, I propose the basic ID rule in (26). (26)

V

-> whether,

[ + Q]

V [-Q]

The following PS rules can be derived from this: (27)

V

-»• whether

[ + Q]

(28)

V

V -Q -Subj -fin _+inf

whether

[ + Q]

V -Q + Subj + fin

These rules will provide for examples like (12) and (13), respectively. The following PS rules cannot be derived from (26): (29)

V :+Q]

whether

V r-Q ' - Subj _+ fin

(30)

V

whether

V Q + Subj -fin + inf

(29) violates (19a) and (30) violates (19c). It follows that sentences like (31) and (32) will not be generated. (31) (32)

*John wondered whether should go. *John wondered whether him to go.

Examples like (31) are ruled out by binding conditions in EST. Again, then, (19a) does some of the work of these conditions. Finally, we can consider coordination. For G P S G , the crucial principle here is the conjunct realization principle (GRP) discussed in Gazdar,

Robert D. Borsley

63

Klein, Pullum and Sag (1982). This can be formulated as follows: (33)

Every conjunct is an extension of its mother.

This requires a conjunct to have all the features of its mother but allows it to have other features as well. The following PS rule conforms to the CRP: (34)

V — Subj -fin + inf

+ + + _+

V Subj fin C and

This rule provides for examples like (14). The following PS rules also conform to the CRP: (35)

V -Subj] + fin J

V " + Subj" + fin

+ C _+ and _ (36)

V - Subj -fin + inf

V Subj" -fin + inf

+ C _+ and ^ However, (35) violates (19a) and (36) violates (19c). It follows that the following will not be generated: (37)

*John expects will be here and that Mary will be here.

(38)

*John expects to be here and for Mary to be here.

A minor point that should be noted is that the following is ungrammatical: (39)

*John expects to be here and Mary will be here.

It follows that the second conjunct in (34) must have the feature [ + C], Nothing I have proposed requires this. It seems, then, that we need some additional principle. I will not attempt, however, to decide what form this principle should take. It seems, then, that the distributional similarities between subjectless infinitives and S's or S's are not at all surprising on the view of VP's that I am advocating. Thus, there are good reasons for preferring this view of VP's to a conventional V P category if subjectless infinitives are VP's.

64

On the nonexistence

of VP's

4. As I noted in the introduction, the data that I have been concerned with here have been seen by proponents of EST as evidence for the view that subjectless infinitives are S's. It is clearly impossible to provide a detailed discussion of this view here. It seems appropriate, however, to include a few remarks about it. One point that is worth noting is that the assumption that similarities between subjectless infinitives and S's or S's indicate that subjectless infinitives are S's is reminiscent of the abstract syntax ideas of the late 60's. It recalls, for example, Ross's (1969) assumption that similarities between AP's and NP's indicate that AP's are NP's. In contrast, the position I am assuming - that the similarities between subjectless infinitives and S's or S's stem from the fact that they have similar but not identical feature analyses — seems very much in the spirit of early EST. Another point I would make is that the idea that recent versions of the S analysis of subjectless infinitives provide an explanation for certain facts is quite dubious. I will consider just one aspect of this matter. An important feature of English infinitives is that they are normally subjectless. Within the analysis I have sketched, this is a consequence of (19c). Within the S analysis developed in Chomsky (1981), it follows from the requirement that lexical NP's have case and the fact that no case is assigned to subject position in a non-finite clause under normal circumstances. It might seem as if this explains a fact that I have only described. It should be noted, however, that the account depends on a quite complex body of case marking conventions. As Gazdar and Pullum (1982) point, it is natural in English to analyze objective case as an unmarked case assigned to any N P which is not marked nominative or genitive. Chomsky, however, assumes two conventions for objective case, one applying to NP's governed by a transitive verb or a preposition and the other applying to the second object in a double object construction (1981:170-1). It is only because of this complexity that subject position in a non-finite clause is normally caseless. In this situation, the idea that the analysis provides an explanation for the fact under consideration is very dubious. I think, then, that the idea that similarities between subjectless infinitives and S's or S's support an S analysis of subjectless infinitives has been accepted much too readily. These similarities are perfectly compatible with the view that subjectless infinitives are VP's provided one abandons a conventional VP category and assigns VP's to the S category in the way I have done here.

Notes "This paper was written while I was supported by grant CHR 8498 from the S.S.R.C. The central idea of the paper is developed in a preliminary way in Borsley (1983). I am grateful to Ewa Jaworska and Neil Smith for helpful comments.

Robert

D.

65

Borsley

1. A related position is developed in Keyser and Postal (1976) and Pullum and Wilson (1977). These writers suggest that VP's arise through the application of subject raising to "flat" SVO structures. 2. V will presumably be unlike the other lexical categories if it has a double bar projection and no single bar projection or if it has a single bar projection and no double bar projection. I do not think, however, that this is a problem. 3. It seems, however, that it is untenable within the lexical functional grammar framework that Bresnan assumes because of its rather broad notion of head. Thus, the present point is not applicable within this framework. 4. It is assumed in GPSG that not all features are subject to the Head Feature Convention. Thus, the fact that the ± C and + Subj features do not obey the Convention is no problem. 5. The following is, of course, possible: (i)

John expects Mary to be here.

I assume that such examples involve an N P and a VP which do not form a constituent. They can be provided for by associating expect with the following ID rule:

(ii)

v [ —Subj]

-»v, nL

v [ - Subj]

Multiple questions and focus in Italian* ANDREA CALABRESE Scuola Normale Superiore,

PISA

A peculiarity of Italian 1 that differentiates it from other languages, is that it does not permit questioning of more than one constituent per sentence: namely questions such as (1)—(3) are not possible in Italian: (1) (2) (3)

*Chi ha scritto che cosa? *Chi e partito quando? * Quale ragazza ha dato un bacio a quale ragazzo?

Obviously we leave aside here echo and quiz interpretations for (l)-(3). This fact is rather strange in comparison to the perfect acceptability of English counterparts of (1)—(3), namely (4)-(6): (4) (5) (6)

W h o has written what? Who left when? Which girl gave a kiss to which boy?

It is interesting to note that, corresponding to the impossibility of (1)(3), sentences (7)—(9) are not possible in Italian with sentence-stress on the constituents in capitals: (7) (8) (9)

* M A R I O ha scritto una LETTER A (MARIO has written a LETTER) * F R A N C O e partito alle C I N Q U E ( F R A N C O has left at FIVE O ' C L O C K ) * SANDRA ha dato un bacio a C A R L O (SANDRA gave a kiss to CARLO)

(Notice that (7)-(9) should be the appropriate answers to the questions (l)-(3).) N o r does it make any difference if they are the beginning of a sequence of pairings — namely multiple answers such as in (10): (10)

* M A R I O ha scritto una LETTER A, F R A N C O ARTICOLO, C A R L O un LIBRO, 2 e t c . . . .

Also in this case they are ungrammatical. 3 67

un

68

Multiple questions and focus in Italian

I find it interesting to correlate the fact that it is not possible to have more than one questioned element in Italian with the fact that, in the same way, it is not possible to have more than one focus of new information in the same sentence. I will now try to account for this correlation by deriving it from the properties of the informational structure of the Italian sentence. In so doing, I will deal only with simple sentences. Another peculiarity of Italian is that it permits the subject to be freely posposed after the verb, namely it freely permits sentences such as (11)(12):

(11) (12)

È arrivato Mario (Has arrived Mario) Glielo ha dato Carlo (to him-it has given Carlo)

It has often been noticed that a postverbal subject is always informationally new (cf. Gruppo di Padova (1974), Antinucci-Cinque (1977)), while a preverbal one is not. In this way in (11) and (12) the phrases Mario and Carlo are understood as informational^ new. Another peculiarity of Italian noticed by L. Rizzi (1982, eh. IV) is that the movement of a wh-phrase with subject grammatical function is always from a postverbal position. Such movement can be observed in the behavior of the clitic particle ne. The clitic particle ne pronominalizes the head of a quantified N P only if it is in a postverbal position. See, for example, the case of the N P with object grammatical function as in (13): (13)

*(ne) ho visti molti (of them have seen many

)

Thus in (13) it is not possible to leave out the particle ne. In this way, taking a quantified N P with subject grammatical function, its head can be pronominalized only if it is in a postverbal position. So we can have sentences such as those shown in (14)—(15): (14) (15)

Molti sono andati via (Many have left) *(ne) sono andati via molti (of them left many )

In (14) the clitic particle ne is not necessary, in (15) it is. If we take now the quantifier w/¡-phrase quanti, we can note that it is impossible not to have the clitic particle ne as in (16): (16)

Quanti (How many

*(ne) sono andati via of them have left)

Andrea

69

Calabrese

Whatever is the analysis of the necessary presence of ne in (15) and (16) (see Belletti-Rizzi (1981), Rizzi (1982)), it follows from their similarity that we have to suppose that the initial position of the quantified N P of (16) must be as in (15). We assume now that there is a rule of assignment of a new information feature, from now on N-feature, that indicates which constituents of the sentence are new. We assume also that there is a principle of the grammar that requires that a w/i-expression, which is interrogative, must receive an N-feature, namely principle (17): (17)

If there is a w/i-expression, then it must have an Nfeature, otherwise it is a relative w/¡-expression

We can then suppose that there is a rule of assignment of the N-feature as in (18): (18)

Assign the N-feature either to a sentence S or to the last constituent a in the linear order of S (where a is either a predicate or an argument of S).4

Assume an intuitive definition of predicate and argument and consider the circumstantial complements as kinds of predicates. Rule (18) states that in the case of an abstract sentence as (19), we can have the N-feature assignment as in (20): (19)

[SNP[

(20)

(a) [ S N P [ V P V N P ] ]

V P

VNP]]

[N

]

(b)

[SNP[VPVNP]]

(c)

[SNP[VPVNP]]

[N

]

[IT—]

In (20a) the whole sentence is new, thus the N-feature is assigned by (18) to the whole sentence. In (20b) only the verbal phrase is new, thus the Nfeature is assigned to the predicate of (19); and in (20c) the N-feature is assigned to the object N P , that is, the last argumental constituent of (19). So, given a sentence such as (21): (21)

Carlo ama Maria (Carlo loves Maria)

(i) the whole sentence can be new, or (ii) the predicate ama Maria, or (iii) the object N P Maria and nothing else. In the case of a sentence such as (22), where we indicate with CC a

70

Multiple questions and focus in Italian

circumstantial complement, CC can receive the N-feature, as we can see in (23): (22)

[ s N P [ v p V N P ] CC]

(23)

(a) [ S N P [ V P V N P ] C C ]

or

(b)

CN

]

[N

]

So, given a sentence such as (24): (24)

Carlo é partito alle cinque (Carlo has left at five o'clock)

the whole sentence can be new, or the circumstantial complement alle cinque that is, the last constituent of the sentence. From rule (18), it follows that the subject in Italian can never be marked as the new constituent of the sentence in preverbal position, and it will be so marked only in postverbal position. In this way the informational difference between (25) and (26) follows directly (25) (26)

Carlo é partito (Carlo has left) É partito Carlo (Has left Carlo)

In (25) it is impossible to interpret Carlo as new if the verbal predicate is not new, namely Carlo can be interpreted as new only if the whole sentence is new; in (26), on the contrary, only the postposed subject is new, if the whole sentence is not new. So (25) can have either the informational structure (27a) or the informational structure of (27b), and (26) the informational structure either of (28a) or (28b). (27)

(28)

Carlo é partito (a) [ N (b) [N

] ]

É partito Carlo (a) [ N (b) [N

] ]

Now we turn to consider w/¡-expressions: by principle (17) a whexpression must have an N-feature for it to be interpreted as interrogative. Let us assume then that (17) applies at S-structure and that the wh in C O M P inherits the N-feature from its trace. In this way it follows that the trace of the w/i-phrase in C O M P will be necessarily in sentence final position so that the w/i-phrase in C O M P may receive an N-feature. From

Andrea

Calabrese

71

these assumption it follows that a w/i-phrase with subject grammatical function comes from the postverbal position; otherwise it could not receive the N-feature, therefore the sentence (29) must necessarily have the S-structure (30): (29) (30)

Chi é partito? (Who has left) [g [COMPChi,] [ s e' é partito t j ] ]

In this way we explain the similarity between (15) and (16).5 The impossibility of having multiple interrogation in Italian follows straight-forwardly from (17) and (18). Let us take, for example, the case in (1) repeated here as (31): (31)

*Chi ha scritto che cosa?

The two principles above predict that only one of the two expressions in (31) will receive an N-feature; if the S-structure of (31) is, for example, (32): (32)

[ s C c o M p C h i ; ] [ s e ' ha scritto che cosa t j ] ]

the w/i-expression che cosa does not receive an N-feature and cannot be interpreted as interrogative; so, given the structural context (31), it turns out to be uninterpretable; the same holds if (31) has the S-structure (33): (33)

Es [COMPChi/] [gt, ha scritto che cosa]]

where it is the w/i-expression chi that does not receive an N-feature; therefore the sentence becomes uninterpretable as with che cosa in (32). From (18) the impossibility of (7)- (9) also follows directly. Let us consider for example, (34): (34)

* M A R I O ha scritto una LETTERA

In (34) sentence-stress was assigned to the subject N P in preverbal position and to the object NP. Sentence-stress is always associated with an N-feature. But from (18) we see that constituents not placed in final position cannot receive an N-feature; therefore (34) is ungrammatical because the N P in preverbal subject position cannot receive an N-feature nor the related sentence-stress. A rather interesting fact that follows from the theory so far outlined, is the principled difference that must be made between traces of clitics and N P s in argument position on the one hand, and traces of left-periferal operators on the other. Let us consider two cases of the first type of empty categories:

72

Multiple

(35)

Carlo; e stato ucciso t,(Carlo,- has been killed t,) Lo, hai visto t, (it, have-you seen t,)

(36)

questions

and focus in

Italian

Given rule (18) and its application in cases like (29), it should be supposed that the N-feature is assigned to the trace that is actually the last constituent in the sentences (35) - (36); and we could also suppose that this feature is inherited by the binding category. But in (35) - (36) we could interpret as new either the whole sentence or the single VP; and the elements binding the traces in final position cannot be new. Consider now the sentence (37) which is a case of passive, as in (35), and its counterpart with the subject in postverbal position: (37) (38)

Molti i sono stati uccisi t,(Many , have been killed t;) *(ne) sone stati uccisi molti (of them have been killed many

)

In the usual analysis we distinguish (37) from (38) by the fact that, while in (37) the N P cannot receive any case in its initial position and therefore it must be fronted to the subject position by the rule of NP-movement for it to receive case, in (38) the N P can receive case in its initial position by some device which we are not interested in now (cf. Burzio (1981), Rizzi (1982)). Since in (38) the N P is in postverbal position, we expect then the obligatory presence of clitic particle ne. If we consider now the sentence (39): (39)

Quanti (How many

*(ne) sone stati uccisi of them have been killed)

given the obligatory presence of ne, we have to say that the quantified N P comes from the postverbal position and receives case in that position, as in (38), namely (39) has a S-structure such as (40): (40)

[§ [COMP Quanti

;M e fc*(ne) sono stati uccisi t f c tj]]

Now what is the difference between (39) and (37) and (35)? Why does not the N P molti in (37) have the N-feature as Quanti in (39)? Our proposal is as follows. We know that traces of NPs and clitics are distinguished from traces of operators in C O M P for the fact that the latter must be Case-marked, while the former must not. Therefore to solve the previous problem, we propose that the rule (18) applies at S-structure to Case-marked con-

Andrea

73

Calabrese

stituents only. It follows that non case-marked traces are invisible to the rule, while Case-marked traces are visible to it. In this way only the whtrace of (39) is checked by the rule (18), and so the wh in Comp can receive the N-feature; on the other hand the traces of the N P or of the clitic of (35), (36) and (37) cannot be checked by the rule (18) and so the binding categories d o not receive any N-feature. Such differences in the grammar have been proposed independently (cf. Aoun (1979), Chomsky (1981)) and it is not strange to find it also in this case. 6 It could be objected that Italian has another device to assign the Nfeature namely topicalization. But this is not true. In fact it can be argued that whatever analysis is made of Italian topicalization, the constituent in TOP-position must receive an N-feature from the sentence final position. That can be clearly seen from the topicalization of a quantified N P with subject grammatical function, whose head is pronominalized. Consider now (41): (41)

MOLTI (MANY

, *(ne) sone venuti , of them have come)

if the subject N P is pronounced with the emphasis and the intonational break between it and the rest of the sentence that is typical of the Italian topicalization, the clitic ne cannot be missing. So (41) is exactly parallel to (42): (42)

*(ne) sone venuti molti (of them have come many

)

this means then that the constituent in TOP-position does not receive its N-feature in T O P , but that it must receive it through the usual M-feature assignment in the sentence.

Notes *I wish to thank Guglielmo Cinque and Luigi Rizzi for their helpful comment and suggestions. I am also very grateful to Giuliana Giusti and Elizabeth Pearce for their help with the English version of this paper. 1. In some sense of the term. See, for example Chomsky (1970), Jackendoff (1972), Williams (1981). 2. Of course, for (7) - (9), we leave aside the contrastive reading of the stress. 3. We must note also that in Italian, it is normally possible to have gapping phenomena: for example (i) (i)

Mario ha scritto una lettera, Sandro Carlo un libro, e t c . . .

un articolo,

In this manner, it follows that the gapping is not a phenomenon which is dependent upon the informational properties of the sentence such as those which explain the ungrammaticality of

(7)-(9).

74

Multiple

questions

and focus in Ita ian

4. N o t e t h a t from (18) it follows that the assignment of focus in Italian is strictly determined by the syntactic structure. F r o m this point of view, Italian seems t o differ f r o m English in which, following C h o m s k y (1970) and Jackendoff (1972), the focus a p p e a r s t o be a property derived f r o m the surface structure and f r o m the intonational c o n t o u r of the sentence. 5. N o t e t h a t there is some overlap between the E C P principle and t h e theory here developed: b o t h actually say that in Italian a w/i-phrase with subject grammatical function must come f r o m the postverbal position. But these two approaches are, anyway, independently motivated. The motivation of our a p p r o a c h is t o explain the u n g r a m m a t i c a l l y of (1)—(3) and (7)—(9). The E C P a p p r o a c h for Italian, following Rizzi (1982), also accounts for the difference in grammaticality between sentences such as (i) and (ii): (i) (ii)

*non pretendo che nessuno venga non pretendo che venga nessuno

(i) is ungrammatical since the variable left in preverbal position by the quantifier interpretation of nessuno in L F is not properly governed by anything, while the variable in (ii) is properly governed in its postverbal position. O u r a p p r o a c h c a n n o t say anything a b o u t (i) and (ii). In the same way, t h e E C P principle cannot say anything a b o u t the ungrammaticality of (l)-(3) and (7)-(9). However we note that E C P and the theory of the informational structure of the sentence here suggested belong to t w o distinct modules of the G r a m m a r : the first to the syntax of LF; the second to a pragmatic c o m p o n e n t that we must introduce now in the structure of the G r a m m a r . In this sense the overlapping just mentioned can be interpreted as an indication that distinct modules of the G r a m m a r have similarities in their properties - an interesting issue that needs to be studied carefully. F r o m a methodological point of view, as L. Rizzi pointed out t o us, we have arrived at a theoretical situation in which it is possible t o say that the linguistic theory, given its m o d u l a r development, can a p p r o a c h the presence of redundances between different modules. It is not strange if a class of p h e n o m e n a explained by a hypothesis in a module has a certain intersection with a class of p h e n o m e n a explained by a hypothesis in another module. This is a direct consequence of the m o d u l a r a p p r o a c h that implies that a given object of research is a by-product of complex classes of p h e n o m e n a independent f r o m each other. 6. Alternatively, to explain it in another way, the difference between traces of clitics and N P s in argument position on the one hand and -traces of left-peripheral expressions o n the other can be dealt with by some notion of reconstruction which states that expressions in left-peripheral positions must always be reconstructed in the initial positions, while expressions moved within S must not. In this way, the observed difference can be an argument for a theory such as that of Riemsdijk-Williams (1981) which includes a level of N P - s t r u c t u r e in which the w/i-phrases are not in C O M P , but still in their basic positions and all the movements within S have already taken place. In such a theory we can say t h a t rule (18) applies at NP-structure; t h u s the difference noted above derives in a very straightforward way.

Infinitival (pseudo-) complementation of noun phrases in Norwegian KIRSTI K O C H CHRISTENSEN University of Oslo

In their discussion of infinitival relatives, Chomsky and Lasnik (1977) note that phrases like e.g. a man to do the job and a book to read differ from endocentric NPs in that their distribution does not always match that of their head NP. Still, these phrases are taken to be relative constructions, their defective distribution being attributed to "highly restrictive conditions, poorly understood, that govern the internal structures of these constructions and their relation to the verb that c-commands them" (Chomsky and Lasnik, op. cit.:460). In Norwegian, there is a class of constructions, exemplified by the italic parts of (1)—(3), that closely resemble the phrases discussed by Chomsky and Lasnik. 1 (1) (2) (3)

Vi we Vi we Vi we

har have har have har have

funnet found funnet found funnet found

en an en a en a

ekspert til h reparere bilen. expert for to repair the car bok h lese for barna. book to read to the children kniv til & skjcere bred med. knife for to slice bread with

Like their English counterparts, these phrases have been analysed as containing infinitival relatives. 2 There is, however, an important difference between the phrases in (l)-(2) on the one hand and that in (3) on the other: while the former do not always share the distribution of the NPs en ekspert and en bok, respectively, the underlined part of (3) is a fully endocentric NP. In this paper, it will be argued that if the infinitival clauses (ICs) in (1)(3) are analysed as infinitival relatives, then we have no (non-ad hoc) explanation why the distribution of the phrase type in (3) should differ from that of the phrases in (1) and (2). First, it will be shown that - despite some apparent evidence to the c o n t r a r y - the ICs in (l)-(3) are not analysable as infinitival relatives. Next, it will be argued that the difference between the underlined parts of these sentences with respect to en75

76

Complementation of noun phrases in Norwegian

docentricity follows from an independently motivated condition preventing a projection of ( + N) from heading a phrase whose complement is also a projection of ( + N). 1. We will begin by considering the (apparent) parallelisms between the ICs in (l)-(3) and finite relatives in (l)'-(3)'. (1) (1)' (2) (2)' (3) (3)'

Vi we Vi we Vi we Vi we Vi we Vi we

har have har have har have har have har have har have

funnet found funnet found funnet found funnet found funnet found funnet found

en an en an en a en a en a en a

ekspert til a reparere bilen. expert for to repair the car ekspert som kan reparere bilen. expert that can repair the car bok a lese for barna. book to read to the children bok som vi kan lese for barna. book that we can read to the children kniv til a skjcere bred med. knife for to slice bread with kniv som vi kan skjcere bred med. knife that we can slice bread with

First, the ICs in (l)-(3) are like finite relatives in that they contain an empty position ("gap") that must be related to a N P outside the clause: en ekspert, en bok, and en kniv, respectively. (The ICs in (2) and (3) also have an empty subject position, PRO). Second, the distribution of the element til that introduces the ICs in (1) and (3) appears to exhibit the same subject:non-subject asymmetry as does the distribution of the relative complementizer som.3 The antecedent-"gap" relations and the distribution of til/som in (l)-(3) and (1)'—(3)', respectively, are shown in (4). (4)

f*(til) , (a) en ekspert, < „, (*(som)

•a ) ' , > reparere bilen j kanj

, , f(til) P R O k } , (b) en bok,-T , , > lese ((som) N P k a n j

,• for barna '

. . , . [(til) PRO a ) , . . J J (c) en kniv, { V skjaere brad med ((som) N P k a n j Third, the ICs in (1)—(3> are like finite relatives in that they behave like adnominal complements with respect to e.g. Topicalization and Tough Movement. Cf. (5). (5)

(a) En an (b) En a (c) En a

ekspert (til k reparere bilen) expert for to repair the car bok (a lese for barna) book to read to the children kniv (til a skjaere brad med) knife for to slice bread with

har vi funnet have we found er vanskelig k is difficult to finne find

Kirsti Koch

77

Christensen

Thus, there may appear to be good reasons for analysing the ICs in (1)—(3) as infinitival relatives. However, when more data are taken into account, we find that the structural and distributional parallelisms illustrated in (4) and (5) do not hold up. 2. A first problem is to account for the distributional asymmetries illustrated in (6) and (7). (6)

(7)

(a) Vi we (b) Vi we (c) Vi we (a) En an (b) En a (c) En a

har lest om en ekspert (*til ä reparere bilen). have read about an expert for to repair the car har h0rt om en bok (??ä lese for barna). have heard about a book to read to the children har lest om en kniv (til ä skjaere brod med), have read about a knife for to slice bread with ekspert (*til ä reparere bilen) har nettopp kommet. expert for to repair the car has just come bok (??ä lese for barna) stär i bokhyllen. book to read to the children is in the bookcase kniv (til ä skjaere brod med) ligger der borte, knife for to slice bread with is over there

If the parenthesized ICs in (6) and (7) are analysed as infinitival relatives, then we have no explanation why there are verbs that admit the N P s en ekspert and en bok, but not the phrases en ekspert til a reparere bilen and en bok a lese for barna as objects, nor why these phrases are not freely permitted in subject position. Note that if the defective distribution of these phrases is attributed to certain inherent semantic properties of infinitival relatives, then we are left with the problem of accounting for the endocentricity of phrases like en kniv til a skjcere bred med. A second problem with analysing the ICs in (l)-(3) as infinitival relatives concerns the status of til. Consider first (8a, b). (8)

(a) Til & reparere bilen har for to repair the car have (b) A reparere bilen har vi to repair the car have we

vi funnet en ekspert. we found an expert funnet en ekspert til. found an expert for

If til were a complementizer introducing infinitival relatives, then (8a, b) should be just as impossible as are (8a, b)'. (8)

*(a)' Som that *(b)' Kan can

kan reparere bilen har vi funnet en ekspert. can repair the car have we found an expert reparere bilen har vi funnet en ekspert som. repair the car have we found an expert that

Thus, the til that appears in (1) and (8a, b) cannot be the non-finite counterpart of som. Rather, it must be a pre-position equivalent to the one in (9a, b).

78

Complementation

(9)

(a) Vi har funnet en ekspert til bilreparasjonen. we have found an expert for the car repair (b) Til bilreparasjonen har vi funnet en ekspert. for the car repair have we found an expert (c) Bilreparasjonen har vi funnet en ekspert til. the car repair have we found an expert for

of noun phrases in Norwegian

An additional argument that til cannot be a complementizer comes from sentences like (10a, b). (10)

(a) Vi har funnet kniven *(til) we have found the knife for (b) Han har alltid en bok (?*til) he has always a book for

& skjaere bred med. to slice bread with k lese for barna. to read to the children

If til were like som, then its distribution in sentences like these should be just as independent of the definiteness properties of the head noun and the semantics of the matrix verb as is the distribution of som in (10a, b)'. (10)

(a)' Vi har funnet kniven we have found the knife (b)' Han har alltid en bok he has always a book

(som) vi kan skjaere bred med. that we can slice bread with (som) han kan lese for barna. that he can read to the children

The reason why til must/cannot appear in sentences like (10a) and (10b), respectively, is semantic rather than syntactic. While som is a complementizer without lexical content, til is a preposition that, in these contexts, means "that is made for". Thus, corresponding to (10a, b), we have (11a, b). (11)

(a) Vi har funnet kniven som er laget til & we have found the knife that is made for to skjaere bred med. slice bread with ?*(b) Han har alltid et eventyr som er laget til k he has always a fairy, tale that is made for to lese for barna. read to the children

It is clear, then, that til cannot be the non-finite counterpart of som. Thus, there is structural as well as distributional evidence against analysing the ICs with which we are concerned as infinitival relatives.4 In the following sections, it will be argued that if we take ICs introduced by til, as in (1)(3), to be PPs and ICs without til, as in (2), to be clauses of the category N, then the distributional asymmetries exemplified in (6) and (7) above can be accounted for.

Kirsti Koch

Christensen

79

We will begin by showing that Norwegian finite and infinitival clauses must be [ + N ] . 3. In Kayne (1982), the obligatory presence of prepositions in phrases like the destruction *(of) the city is attributed to the principle (12). (12)

(A non-maximal projection) of N must not govern a maximal projection of N.

Here, I shall argue that Norwegian is subject to a similar principle which prevents a non-maximal projection of + N from governing a maximal projection of + N . The principle can be formulated as in (13), where "H" and "C" stand for "head" and "complement", respectively. 5 (13)

The Head-Complement Condition (HCC) * [ « H + N [ C + N ] ] , where a is an endocentric phrase.

The H C C correctly predicts that the [ + N ] categories N P and AP may not take bare nominal complements. (14)

(a) Resultatet *(av) mye spising er overvektighet. the result of much eating is overweight (b) Jeg er glad *(for) ditt naervaer. I am happy about your presence

Next, note that finite and infinitival clauses behave exactly like "ordinary" N P s with respect to government by [ + N ] categories. Thus, corresponding to (14a, b), we have (15) and (16), respectively. (15)

(16)

(a) Resultatet *(av) at man spiser mye er overvektighet. the result of that one eats much is overweight (b) Resultatet *(av) k spise mye er overvektighet. the result of to eat much is overweight (a) Jeg er glad *(for) at du er her. I am happy about that you are here (b) Jeg er glad * (for) £ vsere her. I am happy for to be here

If we take the obligatory presence of prepositions in sentences like (15) and (16) as evidence that Norwegian finite and infinitival clauses are [ + N], 6 and if, as we have argued above, til in sentences like (1) and (3) is a preposition, then the ICs with which we are concerned may be represented as in (17)—(19). (17) (18) (19)

[ _ n til [ + n & reparere bilen]] [ + N & lese for barna] [ _ N til [ + N & skjaere bred med]]

80

Complementation of noun phrases in Norwegian

4. The HCC predicts that phrases like (17) and (19), but not (18), may be governed by a projection of [ + N ] . Thus, it appears to make the wrong prediction for phrases like en ekspert til a reparere bilen. Further, it pr vides no explanation why the IC in (18) behaves like an adnominal complement in sentences like (20a, b), cf. above. (20)

(a) En a (b) En a

bok book bok book

k to k to

lese read lese read

for to for to

barna the children barna the children

har vi funnet. have we found er vanskelig k finne. is difficult to find

Note, first, that the H H C is a condition on endocentricity, not a condition on permissible sequences of phrases. It does not, of course, predict that any N P followed by a P P is a single constituent, nor does it block sequences of [ + N ] categories. Second, note that neither Topicalization nor Tough Movement is a diagnostic of constituenthood. Consider (21), where none of the italic sequences form single constituents. Still, these sequences may be topicalized. (21)

(22)

(a) Vi spiser ofte egg til frokost. we eat often eggs for breakfast (b) Det har alltid vaert en formyelse a vcere her. it has always been a pleasure to be here (c) Det er ikke umulig a reparere denne bilen. it is not impossible to repair this car (a) Egg til frokost spiser vi ofte. eggs for breakfast eat we often (b) En forneyelse k vaere her har det alltid vaert. a pleasure to be here has it always been (c) Umulig I reparere denne bilen er det ikke. impossible to repair this car is it not

If it can be shown that phrases like en ekspert til h reparere bilen and eventyr afortelle barna do not form single constituents, then the problems noted at the beginning of this section are of no consequence for the HCC. 5. Note, first, that the distributional properties of the italic part of (1) are not changed when the IC following til is replaced by a NP, as in (23). Thus, we have (24a-c) corresponding to (5a), (6a), and (7a) above. (1) (23) (24)

Vi we Vi we (a)

har funnet en ekspert til à reparere bilen. have found an expert for to repair the car har funnet en ekspert til bilreparasjonen. have found an expert for the car repair En ekspert (til bilreparasjonen) har vi funnet/ an expert for the car repair have we found er vanskelig â finne. is difficult to find

Kirsti Koch Christensen

81

(b) Vi har lest om en ekspert (*til bilreparasjonen). we have read about an expert for the car repair (c) En ekspert (*til bilreparasjonen) har nettopp kommet. an expert for the car repair has just come Examples like these show that the fi/-phrase in (1) is a constituent of VP, subcategorized by the matrix verb.7 A second argument that the italphrase of (1) does not form a single constituent, is that the NP en ekspert may be replaced by a clitic pronoun, as in (25). As shown by (26a, b), clitics may not head larger NPs. (25) (26)

Vi we (a) Vi we (b) Vi we

trenger need trenger need trenger need

en an en an en an

ekspert/'n expert /him-clitic ekspert/*'n expert /him-clitic ekspert/*'n expert /him-clitic

til â reparere bilen. for to repair the car som kan reparere bilen. that can repair the car fra Belgia. from Belgium

Thus, it is clear that the italic part of (1) does not form a single constituent. Sentences like (2) are more problematic, however. (2)

Vi har funnet en bok a lese for barna. we have found a book to read to the children

Note, first, that this phrase differs from the one discussed above in that the initial NP may not be replaced by a clitic pronoun. Thus, we have no overt evidence that it does not form a single constituent. Still, it has a defective distribution. In order to account for the "borderline" properties of phrases like this, I shall tentatively propose the following explanation: It is generally assumed that this construction has developed from Old Norse sentences like (27b) below, i.e. sentences where the direct object precedes a non-finite verb. (27)

(a) J)à then (b) J^à then

gàtu got gàtu got

menn men menn men

at sjà land. to see land land at sjà. land to see

As the VO-order got fixed, these preposed objects were interpreted as being governed by the matrix verb. As a consequence the IC became "relative-like", but, due to the HCC, these phrases have not turned into "genuine" endocentric phrases.

82

Complementation

of noun phrases in

Norwegian

Notes 1. See section 2 for a discussion of the element til that appears in (1) and (3). 2. Cf. Serland (1982) and Taraldsen (1983). 3. Like English that, som is obligatory when the relativized N P comes from the subject position of the highest embedded clause. Otherwise it is optional. 4. An additional problem concerns the nature of the "gaps" in ICs. If these ICs are infinitival relatives, then the "gaps" must be variables, i.e. traces A-bound by an operator in C O M P . There is, however, little evidence that Norwegian ICs have a C O M P position. First, there is no complementizer corresponding to English for. Second, there is no overt Wh Movement in ICs. (i) (ii)

*Vi har funnet en kniv med hvilken & skjjere bred, we have found a knife with which to slice bread *Det er uklart hva 4 gjare. it is unclear what to do

The non-existence of sentences like these would follow if Norwegian ICs have (a) no C O M P position or (b) an (obligatorily) empty C O M P position. Alternative (a) implies that empty positions in ICs cannot be variables. Alternative (b), on the other hand, is compatible with P R O Movement, i.e. the analysis that Chomsky (1981, 1982) proposes for sentences that involve Movement to C O M P with no phonetic reflex, e.g. purposives. I will not go into these problems here, however. 5. Cf. Hoekstra's and Taraldsen's papers in this volume for related, but more general proposal. 6. See Taraldsen (1983, in this volume) for arguments that that Norwegian finite clauses are [ + V, + N], 7. Note that the italic parts of sentences like (i) and (ii) are endocentric phrases The reason is, presumably, that the prepositions ph and til are selected by the N P en ekspert and the adjective rette, respectively, rather than by the matrix verb. (i) (ii)

Vi we Vi we

har have har have

lest read lest read

om about om about

en ekspert ph h an expert on to den rette mannen the right man

reparere biler. repair cars til h reparere bilen. for to repair the car

Foot features and parasitic gaps* GERALD GAZDAR (Sussex), EWAN KLEIN (Newcastle), G E O F F R E Y K. P U L L U M (Santa Cruz), IVAN SAG (Stanford).

1. Introduction An "unbounded dependency" construction (UDC, hereafter) is one in which (i) a syntactic relation of some kind holds between two substructures in the construction, and (ii) the structural distance between these two substructures is not restricted in any way (e.g. by a requirement that both be substructures of the same simple clause). Suppose that we construe the information that needs to be communicated between the two substructures of a U D C in featural terms. Then, in canonical UDC's, we will have such information percolating up (or dripping down) through a nest of sentential and VP complements. Gazdar and Pullum (1982) ["GP82", henceforth] treat the features involved as instances of a more general class of features which they call "foot features". They go on to propose a percolation mechanism, the Foot Feature Principle (FFP), for such features. The present paper shows that the F F P can be independently motivated by the behaviour of the features responsible for the appearance of WH morphology in WH phrases. Treating the latter as foot features provides a solution to the longstanding problem associated with such NP's as the lettering on the covers of which and provides an explanation for why simple N P [ W H ] VP sentences can appear as relative clauses and questions. The paper then turns to the role of the F F P in the analysis of UDC's. In earlier GPSG work the rules responsible for passing U D C information through a tree were arrived at by a metarule known variously as the "derived rule schema", or the "slash introduction metarule". This metarule has no place in the present analysis. Everything that it did, and some other things, now follow as a special case of the application of the F F P in feature instantiation. Furthermore, given a restriction on metarule application to rules which introduce a lexical category it ceases to be possible to formulate a "slash introduction metarule" with the necessary properties. The paper concludes by considering a class of sentences not immediately provided for under the earlier analysis, namely those sentences, originally noticed by Ross (1967), which contain what have come to be known as parasitic gaps following Engdahl (1983) and Taraldsen (1980). It is shown that the FFP, without additional machinery, interacts with the 83

84

Foot features and parasitic gaps

rest of the grammar of English in a way which makes the correct grammaticality predictions with respect to such sentences. 2. Unbounded dependencies It is analytically useful to think of UDC's as having three parts, the top, the middle, and the bottom. The top is the substructure which introduces the dependency, the middle is the domain of structure that the dependency spans, and the bottom is the substructure in which the dependency ends, or is eliminated. The theory of foot features elaborated in GP82 is crucial to our analysis of UDC's. Indeed, our analysis of the middle of UDC's consists of no more than the feature SLASH taken together with the Foot Feature Principle (FFP). Since this is at once the simplest and the most important part of our treatment of UDC's, we will begin with it. That is to say, we will begin in the middle. GP82 introduced categories that can be conveniently notated as ©/$, where this is an abbreviation for [CAT' © [ F O O T [SLASH $]]]. Conceptually, an ©/$ is to be thought of as an © that has the category $ as the coefficient of its SLASH feature. And the intuitive interpretation of ©/$ is that of "an © with a $ missing". Thus, canonically, a constituent of category ©/$ will be a constituent of category © which has (at least) a hole in it where one would expect to find a constituent of category $. So, for example, an S/NP is a sentence which is missing a noun phrase. (1)

Foot Feature Principle: The increment of the mother's F O O T feature is the unification of the increments of the daughters' F O O T features.

The notions "unification" and "increment" are given formal graphtheoretic definitions in GP82. However, the best way to grasp the effect of the F F P in respect of slash categories is to inspect an example of its application. Consider (2): (2)

S - N P VP

The F F P Legitimates the following instantiations of (2) with respect to the feature [SLASH NP], (3)

(a) S / N P - > N P / N P VP (b) S / N P - > N P V P / N P (c) S / N P - N P / N P VP/NP

Some representative examples of improper instantiations of (2) are shown in (4).

Gerald Gazdar, Ewan Klein, Geoffrey K. Pullum and Ivan Sag (4)

(a) (b) (c) (d) (e) (f)

85

*S -NP/NPVP *S ->NP VP/NP *S/NP - N P VP *S/NP - N P / P P VP *S/NP->NP VP/PP *S/NP-* NP/PP VP/NP

In the case of the simple rule we have been considering, the generalizations that the F F P imposes are these: (i) at least one, but possibly more than one, daughter has its slash feature instantiated in a manner identical to the instantiation on the mother, and (ii) no daughter gets its slash feature instantiated in a manner distinct from the instantiation on the mother. We will encounter the application of F F P in more complex instances, in particular its application to rules whose uninstantiated counterparts already include a slashed category, later on. We will turn our attention now to the bottom of UDC's, the part of the structure in which the chain of slashed categories comes to an end and we reach an incomplete constituent or gap. (5)

Slash termination metarule 1 (STM1)

e -w $ I V

©/$->• w This says that every rule in the domain of metarule application - i.e. every rule which introduces a lexical category - has counterparts in which a daughter is missing, but where the mother has the category of the daughter assigned to be the value of its SLASH feature. The most basic and general rule responsible for introducing UDC's in English is that shown in (6). (6)

S-©S/©

This rule says that a sentence can consist of an @ followed by a sentence with an ©-type hole in it, and it is responsible for, among other things, the topicalization construction. 3. WH constructions Apart from SLASH, there is another foot feature associated with UDC's, and that is the feature WH. This feature, according to GP82, has the internal structure shown in (7). (7)

[ W H AGR W H M O R ]

86

Foot features and parasitic gaps

We will ignore the agreement feature AGR here. The feature W H M O R encodes the morphological type of the w/i-expression involved. It has at least two distinct coefficients: (8)

[ W H M O R {Q, R}]

The two that we shall be concerned with here are Q, for interrogatives, and R, for relatives. The distinction between Q and R will allow the lexicon to exclude the possibility of what being used as a relative pronoun in standard English. We will use [ + Q ] and [ + R ] as abbreviations for [WH AGR [ W H M O R Q ] ] and [WH AGR [ W H M O R R]], respectively. And we will assume the existence of feature cooccurrence restrictions [FCR's] which forbid the appearance of the WH feature on VP or A'. The impossibility of either questioning or relativizing on VP constituents in English provides support for the VP FCR. Note the contrast between the examples in (9). (9)

(a) and going to that party, you certainly are! (b) *Going to what party are you?

And the A' FCR is motivated by the following contrast, pointed out to us by Stan Peters. (10)

(a) How fond of him, are you? (b) *Fond of whom, are you?

The proposed FCR will not forbid the appearance of WH on AP, thus it will permit (10a), but it forbids its instantiation on the daughter A', thus preventing generation of (10b). The effect of the FCR is thus to restrict the appearance of wh words in AP's to the determiner position. We will also assume an FCR making WH and SLASH mutually exclusive in English, an FCR which can be motivated by the need to prevent the generation of examples like (11). (11)

*Which books did you wonder whose reviews of had annoyed me?

The following rules will now serve to introduce relative clauses and embedded questions, respectively. (12) (13)

N ' - N ' S [ + R] VP->V[18] S[ + Q ]

These rules naturally lead one to ask what the rules are which allow S[ + R] and S[ + Q ] to be expanded. And the answer to this question, or at least part of the answer, is that these rules cannot help but include the

Gerald Gazdar, Ewan Klein, Geoffrey K. Pullum and Ivan Sag

87

rules that we already have which expand S. We have assumed two such rules in this paper already, and they are repeated in (14) below, for convenience. (14)

(a)S-NPVP (b) S - e s/e

The features [ + R ] and [ + Q ] are W H features, and thus foot features. So they will be instantiated by the F F P in exactly the same way as SLASH features, provided, of course, that the resulting instantiations are consistent with the prevailing FCR's. This means that the grammar will automatically contain the instantiations of (14) shown in (15) and (16). (15)

(a) S [ + R ] - N P [ + R ] VP (b) (i) S[ + R ] - > N P [ + R ] S / N P (ii) S[ + R ] - P P [ + R ] S/PP (iii) S[ + R] A P [ + R ] S/AP

(16)

(a) S [ + Q ] - > N P [ + Q ] VP (b) (i) S[ + Q ] - N P [ + Q ] S / N P (ii) S[ + Q ] - > P P [ + Q ] S / P P (iii) S[ + Q ] A P [ + Q ] S/AP

These rule instantiations, none of which have had to be specially listed in the grammar, interact with rules (12) and (13) so as to generate all the examples in (17) and (18). (17)

(a) The doctor who worked for Kim died. (b) (i) The doctor who Kim worked for died, (ii) The doctor for whom Kim worked died.

Note that there is no exemplar for (15c (iii)) in English, although some other Languages permit adjectival relatives of this kind. Their absence in English is explained partly by the A' F C R discussed above, and partly by the fact that the English Lexicon contains no [ + R ] counterpart to the interrogative adjectival determiner how. (18)

(a) Sandy wondered who worked for Kim. (b) (i) Sandy wondered which doctor Kim worked for. (ii) Sandy wondered for which doctor Kim worked. (iii) Sandy wondered how expensive the wine was.

It is a surprising consequence of the present proposals that matrix subject relatives and constituent questions like those in (17a) and (18a) must have, and can only have, the simple N P V P structures such as that shown in (19).

Foot features

and parasitic

(19)

gaps S.

NPDet

J5[ + R] doctor

N P [ + R] who

^ ^ V P PP

V[6]

I

worked

P[3]

I

for

NP

I

Kim

They must have this structure because of the existence of the "S -> N P VP" rule and the principle that governs the way foot features work. And they can only have this structure, and not, say, one in which the wh N P is followed by a sentence with a missing subject, because no other rules are available. In particular, STM1, being a metarule, can only apply to rules which introduce a lexical category and cannot, therefore, apply to the "S N P VP" rule so as to eliminate the subject in favour of a slash on the mother. The present framework thus offers no counterpart to the vacuous movement analysis of such sentences standardly offered in transformational accounts. The F F P will instantiate W H features on N P and P P rules, as well as the S rules shown in (14). (20)

N' ->N[2] P P N P - Det N ' PP P[3] N P

Thus among the instantiated extensions of the rules given in (20) will be the following. (21)

N ' [ + R] - > N [ 2 ] P P [ + R] N'[ + Q]->N[2] PP[ + Q] N P [ + R] Det N ' [ + R] NP[ + Q] Det[ + Q] N'[ + Q] P P [ + R ] -* P [ 3 ] N P [ + R] P P [ + Q ] - P [ 3 ] NP[ + Q]

And these rules will induce structures such as those shown in (22) and (23).

Gerald Gazdar, Ewari Klein, Geoffrey K. Pullum and Ivan Sag (22)

89

+

Det[ + Q]

N'[ + Q]

whose

N[2]

"^PP[ + Q]

picture

P[3]

^NP[ + Q]

of

whom

NP[ + R]

(23) Det I the

N'[ + R] ,PP[ + R]

N N[2] f: Lettering

NP[ + R]

P[3] on

Dét the

N'[ + R] N[2] covers

^ P P [ + R]^ P[3]

NP[ + R]

of

which

4. Missing subjects The reader will recall that we have, as yet, only provided half of our account of the bottom of UDC's, namely STM1. Although a considerable amount follows from STM1, it does not, and cannot, provide an exhaustive account of slash termination. For example, given only what we have presented so far, the grammar will not contain a rule that will permit the following sentence to be generated. (24)

The doctor who we believe worked for Kim is dead.

The reason for this is simple: STM1 is a metarule, metarules apply only to rules that introduce a lexical category, the subject of an English declarative is not introduced by such a rule, therefore it cannot be eliminated by STM1. (25)

Slash termination metarule 2

(STM2)

•

-»W S I V e / I - > W J where $ -»I J, for phrasal I, J. This says that any rule which introduces a category $ for which the grammar provides an expansion I J, has a counterpart in which $ is replaced by J, and the mother has I assigned to be the value of its SLASH

90

Foot features

and parasitic

gaps

feature. Crucially, STM2, like STM1, is a metarule and thus can only apply to rules that specifically introduce a lexical category. To show the effect it has, we exhibit in (27) a couple of the rules that result from the application of STM2 to the rules in (26) (26) (27)

VP->V[9] S[FIN] V P -> V[14] N P S [ F I N ] VP/NP->V[9] VP[FIN] V P / N P -» V[14] N P V P [ F I N ]

O u r grammar now has the rule it needs to generate (24), and will assign it the structure shown in (28). (28)

Det

I the

N' doctor

+ ^^S/NP

N P [ + R] who

NP

I

we

V[9]

I

believe

^

,VP/NP

;VP V[6]

I

worked

PP

P[3? I for

NP

I

Kim

5. Parasitic gaps Consider the following set of rules, all of them legitimate instantiations under the F F P . (29)

(i) S / N P NP/NP VP/NP VP/NP VP/NP VP/NP VP/NP VP/NP

-> N P / N P -> D e t / N P VP/NP -»V[12] -> V[13] -*• V[14] V[15] V[16]

VP/NP N'/NP PP/NP NP/NP NP/NP NP/NP NP/NP PP/NP

PP/NP VP/NP S/NP S'/NP S'/NP

These rules have two things in common: (i) we have made n o reference to any of them, nor have we invoked any of them in providing example sentences or trees, and (ii) they all introduce two daughters with a slash feature identical to that of the mother. But they exist, and they have

91

Gerald Gazdar, Ewan Klein, Geoffrey K. Pullum and Ivan Sag

consequences for what the grammar will generate. For example, they predict that the strings in (30) should all be grammatical. (30)

(a) (b) (c) (d)

Kim wondered which author reviewers of always detested. Kim wondered which models Sandy had sent pictures of to. Kim wondered which dolls house there was a replica of in. Kim wondered which authors the editor wanted reviewers of to please. (e) Kim wondered which authors the editor had told reviewers of that they should pan.

And, somewhat surprisingly, they are all grammatical, although (30c) is pretty indigestible, for reasons, we assume, that have more to do with its Escherlike semantics than with its syntax. Examples of the type illustrated in (30) were almost entirely neglected by linguists until the circulation, in 1981, of Engdahl's important paper (Engdahl 1983) which thoroughly explored both the facts and their theoretical implications. Most of the examples in this section are adapted from ones to be found in her paper. An earlier G P S G analysis of her data is to be found in Sag (1983). We assume here the correctness of Sag's arguments against Engdahl's claim that there are grammatical differences with respect to island constraints between the two single-gap-containing subconstructions in a (noncoordinate) double gap structure. We also assume that the hierarchy of acceptability for such constructions, as outlined by Engdahl, is not to be explained in the syntax itself. Accordingly, we do not concern ourselves with "degrees of acceptability" in what follows. Example (30a), for example, will be assigned the structure shown in (31). (31) ;S[+Q] Det[ + Q ] V which

NP/NP

^ >

authors ^ N ' / N P N

.1

reviewers

AdvP

PP/NP alJays

I

P[3]

I

P / N

P VP/NP vfo

I

detested

of

However, as Engdahl (1983) shows, the appearance of these "parasitic gaps" is highly constrained. For example, they only show up in UDC's, and not in virtue of passive, "Equi", or "Raising" constructions, a fact which, as Engdahl points out, supports a distinction between local and nonlocal dependencies, such as the [ + / —SLASH] distinction which is

92

Foot features

and parasitic

gaps

fundamental to the present framework. The grammaticality of the examples shown above follows from the operation of the F F P on SLASH features required by the U D C ' s involved. Passive, "Equi", and "Raising" constructions are not UDC's, nor can they be properly analysed as such, and thus they cannot, in themselves, legitimate the appearance of parasitic gaps. Engdahl also provides data bearing on an altogether more subtle constraint illustrated in the examples below, which are taken from her paper. (32)

(a) *Who did you say was bothered by John's talking to ? (b) *Which slave did Cleopatra give to ?

These examples clearly involve U D C ' s , and yet parasitic gaps are not possible. The theory of U D C ' s outlined in G P 8 2 and this paper is in the enviable position of not needing to say anything about such examples or, indeed, the examples in (30). The latter will be generated, those in (32) will not. Since this fact is not immediately obvious, we will pursue the topic a little. Consider the subject gap in (32a). This can only be induced by the rule shown in (33a). (33)

(a) V P / N P - > V [ 9 ] VP (b) V P / N P - » V [ 9 ] V P / N P (c) V P - » V [ 9 ] VP

Inspection of (25) will confirm that (33a) is indeed in the grammar, thanks to STM2. However, (33a) will not allow us to generate (32a) because the VP it introduces is not slashed, and the VP in (32a) must be slashed since it has a gap in it which is not legitimated by any U D C internal to the VP. T o get (32a), we would need the rule shown in (33b), but this rule is not in the grammar, and cannot be. It cannot arise through instantiation of (33a), given the definition of F F P . It could arise through instantiation of (33c), but the latter is not, and must not be in the grammar. If it was, then we would be claiming that strings like (34) were grammatical standing on their own. (34)

*You said was bothered by Sandy.

Consider the direct object gap in (32b). This can only be induced by the rule shown in (35a) or that in (35b). (35)

(a) VP -V[ll]pp (b) V P / N P - * V [ 1 1 ] P P / N P (c) VP - V [ l l ] PP

Gerald Gazdar, Ewan Klein, Geoffrey K. Pullum and Ivan Sag

93

Inspection of (30) will confirm that (35a) is indeed in the grammar, thanks to STM1. But (35a) will not allow us to generate (32b) because the P P it introduces is not slashed, and the P P in (32b) must be slashed since it has a gap in it which is not legitimated by any U D C internal to that PP. To get (32b), we would need the rule shown in (35b), however this rule is not in the grammar, and cannot be. It cannot arise through instantiation of (35a), given the definition of F F P . It could arise through instantiation of (35c), but the latter is not in the grammar. The structurally similar, and lexically intersecting, rule shown in (36) will be in the grammar, of course. (36)

VP/NP->V[6] PP/NP

And this rule will be responsible for examples such as (37). (37)

Which charities did you give to

?

But this only involves a single gap and is thus not pertinent to the matter in hand. (38)

Which Caesar did Brutus imply while ostensibly praising ?

was no good

Engdahl notes that the example in (38), due originally to Alan Prince, is grammatical, even though, at first sight, it appears to be similar to the ungrammatical (32a). But there is an important structural difference between them, and this difference means that the present analysis correctly predicts the grammaticality of the example. Crucially, the adverbial phrase modifies the VP imply was no good, not the VP was no good. The former is slashed, but the latter is not. (39)

(a) VP - VP AdvP (b) V P / N P ->• V P / N P A d v P / N P

The parasitic gap thus gets into (38) in virtue of the rule in (39b), which is itself and instantiated extension under F F P of the rule in (39a) whose presence in the grammar we take to be uncontroversial. There is one class of examples where the F F P would lead us to expect grammaticality, but where Engdahl claims that instances of the type in question are pretty clearly unacceptable, even to those with a high tolerance for parasitic gaps. Reaction to oral presentations of the present paper to audiences of native speakers of English suggests, however, that a number of speakers do find them acceptable. The construction is that in which a verb subcategories for two PP's. In this construction the F F P gives rise to rules which will permit both prepositions to be stranded.

94

Foot features

(40)

(a) ?Which of your neighbours did you complain about (b) ?Who do you seldom talk to about ?

and parasitic

gaps

to

?

Interestingly, Engdahl notes that the second example is acceptable in Swedish. We have no syntactic explanation to offer for the widespread unacceptability of the English examples, and we suspect that Engdahl's hypothesis that the correct explanation is tied into language-particular conditions on bound anaphora is correct.

* This paper is a condensation of part of Gazdar, Klein, Pullum, and Sag (1982). For a rather different attempt to explain the parallels between multiple occurrences of whexpressions and parasitic gap phenomena, see Kayne (1983). Also, cf. Longobardi (this volume).

Towards an analysis of retroactive gerunds* ANDRÉ HANTSON Notre Dame University,

Namur

1. Introductory A retroactive gerund occurs in sentences like: (1)

Your hair needs/wants/requires cutting.

What is typical of this gerund is that it is active in form but passive in meaning, 1 as is clear from the fact that it is often replaceable by a passive infinitive, as in Your hair needs to be cut. Mostly its object is co-referential with the subject of the main clause. As is well-known, the present-day verbal gerund has taken several centuries to develop. More precisely, its passive form is a relatively recent development. Whereas we now say He escaped being whipped, Shakespeare still wrote He escaped whipping, using an active gerund with a passive meaning. In nearly all cases this retroactive gerund has been replaced by a passive gerund. 2 But the original form has been kept in a very limited number of contexts. According to traditional grammars it occurs as the object of the verbs need, want, require, bear and deserve, the prepositions past and beyond, the adjective worth. However, it can be shown that in present-day English the construction is only productive with the verbs need, want and require. As for the verbs bear and deserve and the prepositions past and beyond, it is now restricted to a few stereotyped phrases. Compare: His words don't bear repeating/*translating, He deserves hanging/*obeying, That child is past/beyond saving and *That theory is past/beyond falsifying. As regards the adjective worth, the gerund is very current as its object, but, as I shall show in 3.3, it is structurally different from a retroactive gerund. Gerunds after verbs of the need class are typically used alone, but they may be accompanied by an agent phrase and/or either an adnominal or manner adverbial, as in This matter needs (CAREFUL) handling (BY AN EXPERT) and This matter needs handling (CAREFULLY) (BY AN EXPERT). From these instances it can be deduced that the gerund functioning as the object of need can be either nominal or verbal. 3 I will 95

96

Towards an analysis of retroactive

gerunds

concentrate on the verbal variety, whose retroactive nature will be attributed to a moved object, which is mostly PRO.

2. Retroactive Gerund Clauses 2.1 Retroactive

Gerund Clauses with a PRO

Subject

Let us consider the following instance: (2)

Your toe needs seeing to. 4

Using the government-binding framework, I suggest we derive (2) from the D-Structure under (3). (3)

_

S,

As this diagram indicates, I assume that need takes an S object. That the sentential object of need contains a C O M P node is clear from instances like (4). (4)

Howj does this matter, need [g t j ] [ s P R O , handling t; t ; ]

Further evidence for the presence of a C O M P node will be given later. Instances like (4) also show that the S is not dominated by an N P , which would cause a subjacency violation. Though the gerund clause has a C O M P node, this node may be taken to be empty at the level of D-Structure. We may therefore assume that need does not subcategorize for S, but the S, or more precisely, the head of S, which is I N G . Putting it differently, the lexical entry of need must contain the following subcategorization frame: need [ + V, + ] i N F L - I N G d ] I N G d may be looked upon as a bundle of features, one of which is [ + d ] (=dethematization). This feature is a typical, though not necessary, component of the past participle morpheme. It is present in those cases where the latter has passive meaning, as in He was beat-ENd and He saw her beat-ENd, but not in He has arrive-EN. The feature [ + d] dethematizes the subject of the verb and causes that verb to lose its caseassigning property. As a result, the direct object must move into subject position, so as to get case. Thus np was beat-ENd John is changed into Johni was beat-ENd i,. 5 Similarly in Latin the suffix -or, as in amor ' 7 am

André Hant son

97

loved", normally has the feature [ + d] but it does not when it occurs with a deponent verb like sequor "J follow", which has active meaning. Now, returning to ING, we can assume exactly the opposite of what happens with Latin deponent verbs: normally this morpheme has active meaning and does not contain the feature [ + d ] , but in a retroactive gerund it does contain it and gets passive meaning. Also in this construction affixhopping applied to a morpheme containing the feature [ + d ] results in the object N P moving into subject position. Thus (3) is changed into the following SStructure: (5)

Your toe needs [§ PRO, see-INGd to t ; ]

Having become the subject of the gerund, P R O is then co-indexed with the subject of the main clause, which is its controller. Following Chomsky (1982:256ff) and Reuland (1983:117), I assume that affix-hopping can apply to ING either before or after S-Structure. In (5) it may be taken to have applied before S-Structure. As a result, P R O is not governed by I N F L at S-Structure. On the other hand, the S-barrier prevents government by need. Consequently, P R O is ungoverned and (5) is not ruled out by the PRO-theorem. So we see that at LF need resembles any other verb taking a gerund object in that the subject of the gerund, if it is PRO, is controlled by the subject of the main clause. In that respect (5) is exactly like John doesn't mind \_gPRO opening that door]. The difference between the very restricted class of need type verbs and the regular class of mind type verbs is that at D-Structure P R O functions as the subject of the gerund with the latter while it has object function with the former. Of course, the N P movement that brings about the pseudopassivization of the gerund object of need is only possible if the subject node of the gerund is empty. But this is not a problem, given the fact that sentences such as *Your hairt needs a good hairdresser cutting PROi], where the subject is lexical, are again ruled out by the PRO-theorem. 2.2 Retroactive Gerund Clauses with a Lexical Subject So far we have been considering gerund clauses with a P R O subject at SStructure. But of course with verbs of the mind type the gerund can also have a lexical subject, as in: (6) (7)

I don't mind [ N P his opening that door] I don't mind [s him opening that door]

I shall disregard the construction instanced in (6), where the gerund has a genitive subject. This construction, probably an NP, does not occur -with retroactive gerunds, as is apparent from the ungrammatically of * This matter needs an expert's handling carefully, which is again attributable to a violation of the PRO-theorem. I shall here follow Reuland (1983:115) in

98

Towards an analysis of retroactive

gerunds

assuming that gerunds with P R O subjects are Ss, and correspond to (7), rather than (6). Let us now consider (7), where the subject of the gerund is in the objective case. Reuland accounts for this objective case as follows: the verb mind, being transitive, assigns objective case to its object, which is the gerund clause. More precisely, it assigns objective case to the head of this clause, viz. ING. The morpheme I N G then transmits the objective case to the subject of the gerund clause in the same way as a tensed I N F L transmits nominative case to the subject of a tensed clause. At this point I want to draw attention to a very interesting construction, which can be regarded as a retroactive gerund with a lexical subject. 6 This construction is illustrated in: (8)

You need your toe seeing to (by a chiropodist). (Radford (1977:59))

Other examples are: You need your head examining, You require your leg seeing to, My car wants something doing to the engine. It is evident that in the D-Structure of sentence (8) the N P your toe functions as the object, not the subject, of the verb see to. The D-Structure of (8) is therefore: You need \_gnp INGd see to your toe]. Again the object N P may be assumed to move into subject position, yielding the following S-Structure: (9)

You need [§ your toe,- I N G d see to t,]

I pointed out a moment ago that affix-hopping may apply either before or after S-Structure. In order to account for the fact that P R O was ungoverned in sentences like (5) ( = Your toe needs [ j P R O , see-INGd to t,]) we needed to apply it before S-Structure. But notice that in (9) it must apply after S-Structure, so that at S-Structure I N G can assign objective case to the S-Structure subject of the gerund. However, now we are faced with a dilemma. On the one hand affix-hopping cannot apply until after SStructure (otherwise the lexical subject of the gerund clause does not get case), on the other hand it should apply before S-Structure (or else see to will still be a transitive verb at S-Structure and the N P trace will get case, thus incorrectly marking the sentences as ungrammatical). So we shall have to find a way of dethematizing the gerund that is independent of affix-hopping. I shall come back to this problem in a moment. But first I need to have a look at the constituent structure of retroactive gerund clauses. 2.3 Structural Idiosyncracies of Retroactive Gerund Clauses - Feature Percolation We have already noted that a retroactive gerund is mostly used alone but can be accompanied by a manner adverbial and/or an agent. Also, we know that in D-Structure a retroactive gerund clause must contain a

André

Hantson

99

direct object. But let us now see if it can contain any other constituents. Taking indirect objects first, we see that these are generally impossible, no matter whether it is the indirect or the direct object that undergoes N P movement. This is clear from the ungrammaticality of *You need giving some advice and *Some advice needs giving to you. The presence of a P P that the verb s u b c a t e g o r i e s for is sometimes acceptable, as in He doesn't need reminding of his duty, These advances need bringing into focus. But more often than not the construction is rather doubtful or even ungrammatical, cf.? That novel needs translating into a world language,? He doesn't need congratulating on his results, *? He never needs talking into having another drink. Complements of place are usually low on the acceptability scale, cf. ?? Such an expensive car needs parking in a garage. Conversely, predicative complements are much better, as can be seen from: Brown bread needs slicing thin. Moreover, the following instance shows that predicative complements can combine with agent phrases: Your hair needs dyeing exactly the right shade by a skilled hair-dresser. So there appear to be a number of idiomatic constraints on the constituents ,that can occur in a retroactive gerund clause. Accordingly, we are faced with a rather unusual situation: the verb of the main clause not only selects the type of I N F L that its clausal object can take, but it also determines what constituents that clause can contain. It therefore looks as if the main clause verb should be allowed to somehow subcategorize down into the VP of the gerund clause. I suggest we d o this by means of some sort of feature percolation device, which works as follows: I N G d , being the head of the gerund clause, transmits the feature [ + d ] to the V P of that clause, which in its turn transmits it to its own head, viz. V. In this way we can ensure that the verb that will surface as a retroactive gerund already has the feature [ + d ] at the moment when the constituents of the VP of the gerund clause are selected. As a general rule, a verb with the feature [ + d ] is a passivisable verb that occurs in the following subcategorization frame:

(10)

But some verbs, like bring, for instance, have to be marked in the lexicon as capable of taking also a P P such as into focus when occurring in a retroactive gerund clause. In this section I have tried to demonstrate that an analysis of retroactive gerunds requires something like feature percolation. In doing so I have only used gerunds with a covert subject, but the argument obviously also holds when the subject is lexical. However, the amount of disagreement among native speakers then tends to be even greater, as is shown by a comparison of My wife doesn't need reminding of her duty with % I don't need my wife reminding of her duty.

100 2.4

Towards an analysis of retroactive

gerunds

Conclusion

Let us now return to the dilemma that we were faced with in connection with structures like (9) ( = You need [5 your ioe, INGd see to £,]). I recall that (a) case-assignment to the S-Structure subject your toe requires that affix-hopping should take place after S-Structure, (b) the verb see to should be prevented from assigning case to the trace, or else the sentence will be incorrectly rejected as ungrammatical. Thanks to feature percolation we are now able to get out of this dilemma. On the one hand the SStructure under (11) shows that affixhopping has not taken place before SStructure, so that the lexical subject your toe gets case. (11)

You need [5 your toe; I N G d ] v [ + d ] see to [t,]

On the other hand, feature percolation has dethematized the verb see to and has caused it to lose its case-assigning property, so that the object N P has moved into subject position, and, quite crucially, its trace does not get case. In this way we have been able to account for retroactive gerunds with a lexical subject. Notice that we have tacitly assumed that the DStructure subject of the gerund clause is empty. If this were not the case NP-movement would be impossible and the sentence would be rejected by the case filter because the object of the gerund could not get case. Another, very interesting type of retroactive gerund is illustrated in sentences like What do you need doing to your car? The S-Structure of this sentence is shown under (12). (12)

What; d o you need [ g t ; ] [ s t ; I N G d do t; to your car]]

Here what starts off as the D-Structure object of do, then undergoes N P movement, thereby becoming the derived subject of the sub-clause. Next it undergoes WH-movement, which moves it successively into the C O M P of the sub-clause and the C O M P of the main clause. It will be seen that sentences of this type provide further evidence for the S nature of the retroactive gerund clause. Let us now return for a moment to instances such as (2) ( — Your toe needs seeing to), where the S-Structure subject of the gerund is PRO. The analysis proposed earlier remains essentially unchanged. It should be pointed out, however, that the verb see to now loses its case-assigning property as a result of feature percolation rather than affix-hopping. So, as far as the trace of P R O is concerned, it no longer matters whether affixhopping takes place before or after S-Structure. But, as regards P R O itself, affix-hopping must still be assumed to take place before S-Structure, as we saw previously: otherwise P R O would be governed and (2) would be incorrectly ruled out by the PRO-theorem. To sum up: given feature percolation, retroactive gerunds behave in exactly the same way as normal gerunds at S-Structure: if the subject is

André H ant son

101

PRO, affix-hopping takes place before S-structure, so that P R O is ungoverned; if the subject is lexical, affix-hopping takes place after S-Structure, so that at S-structure ING still governs that subject and can transmit objective case to it. 3. Addenda 3.1 Parallelism with the believe [ S /VP to VP~] Construction Retroactive gerund clauses may look very unusual in that the main clause verb must be allowed to determine both the internal structure of the subclause and the choice of verb of that clause. Actually, this state of affairs is less exceptional than is seems. A similar phenomenon can be observed in a much discussed construction, viz. the accusative + infinitive construction dependent on verbs of the believe type. Consider the following instances: I believe John to be honest, I believe John to have worked hard, I believe John to be working hard, */ believe John to work hard, *l believe John to be beaten regularly. If the believe class verb is active, the infinitive clause must contain either the copula verb be, or the perfect auxiliary or the progressive auxiliary. If this is not the case, the sentence is mostly ungrammatical. But there are exceptions. Thus He found the greatest charm of the place to LIE in its social freedom (Scheurweghs (1959:225)) is fine, even though none of the conditions just mentioned is fulfilled. By contrast, *He found the stability of inert gases to RESULT from the arrangement of the electrons in their atoms (cf. Van Ek (1966:71)) is out. So it looks as if also verbs of the believe class must be allowed to determine certain aspects of the constituent structure of the VP of their clausal object, as well as the choice of the verb of that clause. Apparently, also here feature percolation may be needed. 3.2 Is the Retroactive Gerund a Case of WH~Movement? In the analysis proposed here the retroactive nature of the gerund in (2) (= Your toe needs seeing to) is accounted for by NP-movement, which also plays a central part in the analysis of passives. However, it might be objected that there is a better analysis available, viz. WH-movement, which Chomsky (1977:102 ff) proposes for sentences such as John is easy to please, where the covert object of please is also co-referential with the subject of the main clause. On the analogy of this sentence, which fulfils the criteria for WH-movement, it might be suggested that also in (2) the empty category after seeing to is a WH trace. At first sight this analysis certainly looks preferable: given the fact that a WH trace requires to have case, we would of course not need the feature [ + d] to dethematize the gerund, though we would still require it to account for its idiosyncratic sub-categorization restrictions. The point is, however, that there appears to be no independent evidence for this analysis. The diagnostic properties

102

Towards an analysis of retroactive

gerunds

(e.g. apparent violation of subjacency) which Chomsky proposes for W H movement presuppose that it is possible to insert clauses between, say, easy and meet in John is easy (for us) to meet, as in John is easy (for us) to convince Bill to arrange for Mary to meet. With retroactive gerunds this sort of insertion is totally impossible. Thus His leg needs amputating cannot be changed into *His leg needs persuading that surgeon to amputate. Conversely, the fact that the D-Structure subject is an argument against a WH-movement analysis. 3.3 The Analysis of the worth + Gerund

Construction

As I suggested earlier, the gerund that functions as the object of worth behaves differently from the retroactive gerund occurring as the object of need. The following differences should be noted: (a) the gerund can have an agent phrase after need, but not after worth, compare That point needs stressing by any linguist with * That point is worth stressing by any linguist', (b) some non-passivisable verbs can occur as the gerund object of worth, not of need, compare A degree is worth having with *A degree needs having (by all applicants), and She is worth marrying with *She needs marrying (by one of those doctors); (c) N P s other than the direct object can be gapped with worth, not with need, compare Your country is not worth sacrificing your life for with *Your country doesn't need sacrificing your life for, (d) parasitic gaps are possible with worth, not with need, compare This book is not worth praising without reading with *This book does not need praising without reading. These differences indicate that the gerund after worth is not passive-like and does not lose its object through NP-movement. More specifically, the possibility of having parasitic gaps suggests that the worth construction is a case of WH-movement. 7 Moreover, this time there is evidence of apparent subjacency violation. Thus That problem is not worth solving can be changed into That problem is not worth trying to solve. The WH-movement analysis also correctly predicts that the gerund will be unable to have a lexical subject at S-Structure, as is in fact shown by the ungrammaticality of *You are worth your eyes seeing to by the best oculist in town.

Notes * I want to thank Liliane Haegeman, Steve Harlow and Dany Jaspers for their instructive comments on earlier versions of this paper. 1. The term "retroactive" is borrowed from Jespersen (1940:221 ff. 114), who uses it for infinitives but not for gerunds. The reason why he considers it inappropriate for gerunds is probably that he does not distinguish between verbal and nominal gerunds. It seems likely that only verbal gerunds can be retroactive. 2. Visser (1973:1886-8) mentions quite a number of verbs which used to take a retroactive gerund object. 3. Huddleston (1971:153-4) analyses the verbal gerund complement of need type verbs as involving raising. One problem with this analysis is that it forces one to assign a different

André

Hantson

103

function to the gerund depending on whether it is verbal or nominal. Another, more serious problem is that the construction is as a rule (all but) impossible with idiom chunks. Consider: *Headway needs making, *? Tabs need keeping on his operations. The following instance is admittedly much better, particularly if the adverb really is inserted: (?) The cat really needed letting out of the bag. However, the cat seems to have a thematic role here, which would appear to point to a P R O analysis (which is needed if the gerund has object function, see 2.1) rather than a raising analysis. I am taking this instance rather than (1) because it is more easily transformable into 4. a sentence containing a gerund with an overt subject (cf. 2.2). The combination see to is restructured as a compound verb. 5. Dethematization also occurs in impersonal passives like Dutch Er werd gedanst "There was danced", where the subject is dethematized even though the verb is intransitive. 6. Strictly speaking this construction is participial rather than gerundial in nature (cf. Poutsma (1926:519 fT)). This is suggested by the fact that (a) a genitive subject is impossible, cf. *You need its seeing to. (b) the ing-form alternates with a past participle, as in You need your toe seen to. However, the originally participial construction may now be considered to fill a gap in the gerundial pattern. This is also the case in instances like They were arrested for the same offence, distributing leaflets and Her hand shot out again in the well-known Latin gesture - the first finger and the little finger sticking out. It is well-known that there has been quite some fusion between verbal gerunds and present participles. Interestingly, Poutsma (1926:523) observes that "on the strength of its logical relation to the preceding (pro)noun" the ing form under discussion "may, with some justice, be regarded as a gerund". Some linguists go even so far as to refuse to make a systematic distinction between gerund and participle, cf. Kruisinga (1932:55 ff) and Quirk et al. (1972). 7.

I thank Annie Zaenen for pointing this out to me.

Government and the distribution of sentential complements in Dutch TEUN HOEKSTRA State University Leiden

Government and the Distribution of Sentential Complements in Dutch* Within traditional transformational grammar, the ordering of phrases was in large part handled in terms of the PS-rules of the system, supplemented by a number of transformations that could rearrange the order of base structures. As for sentential complements, it was assumed that they were generated by the PS-rules in the positions of N P ' s with a corresponding function, i.e. subject clauses in the position where subject N P ' s are found at D-structure etc. Their postverbal position was assumed to be derived by means of an extraposition transformation. I shall not go into the motivation behind this approach, nor shall I pay attention to the problems that this approach meets with. 1 From the perspective of the GB-theory, this approach is unsatisfactory on all counts. The expressive power of the transformational component is reduced to such an extent that it is impossible to formulate categoryspecific conditions. The formulation M O V E a not only allows for a rightward shift of S's, but of other categories as well. Therefore, principles applying to the output of the transformational component, i.e. S-structure or representations derived from it, like LF, have to be formulated in such a way that the postverbal occurrence of clauses and PP's is allowed in Dutch, but not of N P ' s and AP's. It will be clear that in that case, it is these principles rather than the transformational component that carry the explanatory task. What is true for the transformational component is a fortiori true for the PS-rules. As Stowell puts it: " . . . the theory of phrase structure is largely redundant, offers no real depth of explanation, and provides little more than an arbitrary collection of generalizations about each phrasal category" (1981:50). I shall proceed assuming that the ordering of phrases is not explained by PS-rules and that we may therefore dispense with them altogether. The internal structure of phrases is in large part determined by the Projection Principle and the 0-criterion. Specifically, I shall assume a distinction similar to the distinction proposed by Williams (1981, 1982) between internal and external 0-roles. In Hoekstra (1983b), I argue that 105

106

Government and the distribution of sentential complements in Dutch

phrases bearing internal 0-roles must occur in the first order projection of the head, i.e. that these phrases must be c-commanded in the traditional sense of c-command. The external fl-role is assigned to the subject of the phrase. Following Stowell, I assume that all major categories can have subjects. 2 Having now outlined the basic working assumptions about the origin of phrases, I shall concentrate on the principles that account for the distribution of phrases in general and of sentential complements in particular, with the central focus on these matters in Dutch.

1. The Unlike Category Condition A central role in the account of the distribution of phrase types in my approach is played by what I call the Unlike Category Condition, henceforth UCC. Basically, the U C C is a generalization of the hypotheses put forward by Kayne (1982). Kayne's hypotheses are given in (1). (1)

(a) A non-maximal projection of N must not govern a maximal projection of N (b) N o projection of [ + V] can be an argument

The hypotheses in (1) show a certain resemblance to another, more familiar principle with distributional consequences, i.e. Case Theory. Case Theory says essentially that NP's can only occur in positions governed by a [ —N] category. For the category N this means that it may not take an N P complement, a well-known fact. If Stowell is right that AP's must receive Case, which I doubt, however, adjectives should not be able to take AP complements, due to Case Theory. Stowell furthermore argues for what he calls the Case Resistance Principle, which has the effect that projections of [ —N] categories, i.e. VP's and PP's may not occur in the complement of V's and P's. If we generalize these principles to their maximum 3 , we arrive at the principle in (2), the UCC. (2)

At S-structure, no element of the category [aV, jSN] may govern [aV, j3N]

In favour of the UCC, various pieces of evidence that were advanced in favour of the hypotheses that it generalizes can be put forward. So, nouns do not take N P complements at S-structure, although they may take appositional NP's as in (4).4 (3) (4)

*the destruction the city Bill, the chairman of the committee

As for adjectives, one would expect that in addition to the structures in (5), where a small clause AP follows the transitive verb consider, and (6) where

Teun

Hoekstra

107

a raising adjective takes a clausal complement, and (7) in which a raising verb takes a small clause AP complement, a construction like (8) would be possible as well, with a raising adjective taking a small-clause AP complement. There are n o raising adjectives that take AP complements, however. (5) (6) (7) (8)

I consider [ ^ p John foolish] John is likely [ s t to win the game] John seems [ ¿ p i dead] *John is likely [ A p i dead]

This fact too follows from the UCC. Turning next to prepositions, we note that in general P's cannot be followed by a P P complement. This possibility is so rare that preposition was defined in several structuralist and traditional grammars as a category introducing an exocentric construction. In Dutch, prepositions with P P complements are found in only two situations. After the prepositions van, tot and voor and after the preposition met "with" in the so-called absolute construction. Examples are given in (9) and (19).5 (9)

(10)

(a) van [voor de oorlog] from before the war (b) voor [bij de maaltijd] for with the meal met [ p p een das om zijn nek] with a tie around his neck

These examples are exceptional enough to conclude that normally P's do not take P P complements. The final category to be discussed is the category V. This discussion will be deferred until after we have introduced our second hypothesis.

2. Asymmetrical Government The standard definition of government, e.g. Aoun & Sportiche (1981), is symmetrical, in the sense that it does not imply any direction in which government holds. If the ordering of phrases is not determined by PSrules, however, there must be some way to account for the fact that in English, or VO-languages in general, the object follows the verb, whereas it precedes the verb in OV-languages like Dutch. Now it could be argued that this ordering of the verb and its object may be accounted for by leaving this order unspecified as far as U G is concerned and assuming that it is fixed on the basis of the abundant positive evidence that specific languages provide for the fixation of this order. This is certainly true, but I would like to argue that what is fixed in this way is not so much the parameter of ordering the verb and its object, but rather the orientation of government, i.e. I assume that government is an unidirectional notion, the

108

Government

and the distribution

of sentential

complements

in Dutch

direction being fixed for each category in each language on the basis of positive evidence. It would make sense to assume this position rather than the more straightforward position that merely the ordering of the verb and its object is fixed, if it can be shown that fixing the direction of government has interesting consequences beyond the determination of the order of the verb and its object. This is precisely what I would like to argue, i.e. I would like to claim that once the order of government of V is fixed, there is a range of consequences entailed by it, of which the ordering of N P ' s with respect to the verb is only one and probably also the trigger to settle the government in one or the other direction. I claim therefore that in Dutch the verb governs to the left, whereas in English it governs t o the right. Given that Case is assigned under government, the fact that Dutch has OV-order and English VO-order follows from this orientation of government. A more interesting consequence is that, although clausal complements occur in postverbal position in Dutch, both finite and non-finite clauses, there are no post-verbal infinitival clauses with a lexical subject, comparable to the English E C M constructions. Consider the structures in (11). (11)

(a) dat that (b) dat that (c) *dat that

Jan [ ^ p het verhaal] [ y gelooft] John the story believes Jan [ v gelooft] [§ dat het verhaal John believes that the story Jan [Y gelooft] [g NP het verhaal] John believes the story

waar true waar true

is] is te zijn] to be

A preverbal N P object and a postverbal infinite clause in the a. and b. examples are allowed, but c. is ungrammatical. The reason for this is that the lexical N P subject cannot receive Case from inside the embedded clause, because this is non-finite. Therefore, Case should be assigned to this N P by the matrix verb, but since this verb governs to the left, this is impossible as well. It should be noted that the assumption that there is no S-deletion in Dutch provides no real alternative. Dutch has subject raising constructions, examples of which will be discussed below, and therefore Sdeletion, or whatever kind of process one wishes to assume for the transparency of clauses. The suggestion that in Dutch, Case assignment cannot cross a major projection boundary, or that Case can only be assigned to a constituent that is thematically governed by the Caseassigning category does not provide an alternative either. Consider the following examples. (12)

(a) dat that (b) dat that

ik I ik I

het document geldig verklaar the document valid declare verklaar dat het document geldig is declare that the document valid is

(C) • • • CAP [NP het document] [A geldig]] [Y verklaar]

Teun

Hoekstra

109

The verb verklaren in (12b) assigns two thematic roles, one to the subject and another to the clausal complement. The minimal assumption is that this verb assigns the same thematic roles in (12a), where the role that is assigned to the clausal complement in (12b) is now assigned to the smallclause AP. The subject of this small clause complement receives Case from the matrix verb, without being a thematic argument of this verb. If we adopt the small-clause analysis, as in (12c), the process of Case assignment must cross the AP barrier, which illustrates that it must be possible to cross a maximal projection boundary. Another interesting consequence that I shall not discuss here is that Dutch does not have raising adjectives. In Hoekstra (1983b) this fact is traced to the unidirectionality of government. The fact that W H extraction out of sentential complements of adjectives is impossible in Dutch, but not in English, can be attributed to the same requirement, i.e. it explains that (13) is ungrammatical, but that its English equivalent in (14) is not. (13) (14)

*Wie is het waarschijnlijk dat Piet e zal ontmoeten? W h o is it likely that Peter will meet e?

For a full discussion of these matters, I must refer to Hoekstra (1983b).

3. The Position of Sentential Complements I trust that the preceding discussion is sufficient to establish the two central principles, the U C C and the hypothesis of unidirectional government, which I shall use to account for the distribution of sentential complements. Let me start with a description of the relevant data. We have to distinguish between two fundamentally distinct complement types: V-raising structures and normal sentential complements. By their nature the former are always tenseless, whereas the latter occur in both a finite and a non-finite variety, as usual. The most striking difference is that Vraising structures are limited in their distribution to complements of V. Normal sentential complements function as subject, object, object of a certain set of prepositions and as prepositional objects, but in the latter case, the situation is somewhat complicated. Starting with the normal sentential complements, we note first of all that Stowell's hypothesis that they may not occur in Case-marked positions seems to be wrong as far as Dutch is concerned. The following examples illustrate this. (15)

[§ dat we gingen eten]

that we started the meal

110

Government and the distribution

of sentential complements

in Dutch

With temporal prepositions, N P and S have an overlapping distribution, which makes it difficult to claim that S may not occur in Case-marked positions. In the case of prepositional objects, however, the S must occur in postverbal position, like subject and object clauses. The preposition is accompanied by the proform er: (16)

(a) dat that (b) *dat that (c) dat that

wij o p zijn komst rekenden we on his coming counted wij o p dat hij kwam rekenden we on that he came counted wij er o p rekenden dat hij kwam we there on counted that he came

Sentential objects occur in postverbal position as well, with an optional pronominal element in preverbal position, mainly with so-called factive predicates. Similarly, sentential subjects obligatorily occur in postverbal position, while the syntactic subject slot is filled either by the pronominal form het or er or the position is left vacant. We turn to the latter possibility later. The examples in (17) and (18) illustrate this. (17)

(a) *dat that (b) dat that

Jan [ § d a t John that Jan (het) John it

hij ziek was] betreurde he ill was regretted betreurde [gdat hij ziek was] regretted that he ill was

(18)

(a) *dat [ s d a t hij koorts h a d ] niets bewees that that he fever had nothing proved (b) dat het mij ergerde [gdat hij ziek was] that it me annoyed that he ill was (c) dat er uit dat feit volgt [5 dat Piet ziek was] that there from that fact follows that Peter ill was (d) dat (3 vast stond [gdat Piet ziek was] that definite was that Peter ill was

How can we derive the obligatory postverbal position of these sentential complements from the assumptions we have made so far? Let us start with sentential objects. What is to be explained, then, is why the configuration in (19) is not permitted. (19)

[ . . . S Vo]

This would follow from the U C C under the hypothesis that S is a category that has the features [ + V , — N ] in common with V. If that is the case, the S being governed by V because it precedes it would violate the U C C . It is quite plausible to assume that in Dutch, S is a projection of I N F L and that I N F L is verbal in the sense of having the relevant features. I want to

Teun Hoekstra

111

claim that the structure of full sentences in Dutch is as in (20). If that structure is correct, the obligatory postverbal position follows directly from the UCC. (20)

INFL" COMP'

~~

"""""" V " "*" NP(subj)

INFL' "

"

INFLU

y'

It should be noted that the structure in (20) violates the UCC as well, if INFL is non-distinct in its feature content from V. A corrolary is that V o has to be raised to INFL 0 , a process that can be regarded as an instantiation of V-raising. I claim that the relatively low degree of configurationality that has been claimed for Dutch and German is a consequence of this merging process of I N F L 0 and V o . At this point, a brief digression is in order. It will be evident that the structure of clauses in (20) cannot be maintained for English as INFL in English and other languages is not external. Whereas traditional S can therefore be regarded as a maximal projection of V in Dutch, this is not possible for English. It may be right therefore to consider the traditional VP as the maximal projection of V in English but not in Dutch. 6 A consequence of the V-raising process applying to V o is that the syntactic subject position is governed at S-structure by the verb, i.e. in this respect the structure derived by V-raising is not different from the constructions with small clauses where the subject of the small clause is governed from outside. This fact has two consequences. First, it explains why sentential subjects have to be extraposed, i.e. why sentential subjects must occur in postverbal position, just like sentential objects. The combination of V o and I N F L 0 governs the syntactic subject position and if S is non-distinct in feature content from V as we have been claiming above, the sentential subject may not occur in the syntactic subject position. Another consequence is that the subject position may be empty as far as the ECP is concerned, the position being properly governed by V o . An empty subject in Dutch is found in three construction types: (a) with long WH-movement from the subject position (b) with "extraposed" sentential subjects (c) with a 0-subject These constructions are exemplified in (21)—(23), respectively (21)

Wie dacht jij dat e dat gedaan heeft? Who thought you that that done has

112

Government and the distribution of sentential complements in Dutch

(22)

Overal wordt e beweerd dat Jan de moord gepleegd heeft. Everywhere is asserted that John the murder committed has Tijdens het eten werd e flink gedronken. During the meal was heartily drunk

(23)

Dutch is not a PRO-drop language like Italian: the possibility of dropping a pronoun with definite interpretation is typically absent. The subject position in Dutch can be empty if the position is coindexed with an expression in an A-position, as in (21) and (22), or if no thematic role is assigned to the position. In that case it need not be visible for the 6criterion whereas it may be empty as far as the ECP is concerned, at least if my claim that the subject position is properly governed is correct. 7 Having thus accounted for the obligatory postverbal position of subject and object clauses, we are left with the case of prepositional objects. As the contrast between (15) and (16) shows, the fact that the sentential complement has to occur in postverbal position cannot be due to the preposition per se. It is not immediately evident, however, how the fact that the S must occur in postverbal position in the case of prepositional objects could be derived from the UCC. Let us assume, that in the case of prepositional objects the preposition may be taken to assign Case to its complement, harmlessly in this situation, but that it is the verb that determines the 6role of the complement. While it makes sense that Case government is determined purely in terms of configurational properties, this is not true for thematic government. Given the fact that it is the verb that determines the (^-structure of the P P domain, rather than the preposition itself 8 , we may say that the clause is governed in the relevant sense by the verb. One might want to formalize this in terms of co-superscripting in the sense of Rouveret & Vergnaud (1980). The fact that a pronominal element must accompany the preposition with an extraposed sentential object of P is explained if prepositions are not proper governors, i.e. do not license empty categories in Dutch. Summarizing the main point, we have argued that sentential complements may not occur in subject position, in object position and in the complement of prepositions in the case of a prepositional object. The reason for this is that a clause in these positions is governed by the verb, which is forbidden by the UCC. This follows if clauses are taken to be projections of INFL and INFL has a feature content that is non-distinct from V. Let us now turn to V-raising structures. Two examples are given in (24b) and (25b). (24)

(a) dat that (b) dat that (c) dat

ik I ik I ik

hoor hear Jan John [Jan

dat Jan lacht that John laughs hoor lachen hear laugh lachen] hoor

Teun (25)

113

Hoekstra (a) dat ik vind dat Jan zeurt that I think that John nags (b) dat ik Jan vind zeuren that I John find nag (c) dat ik [Jan zeuren] vind

As with small clauses, we may assume that the 0-roles determined by the verbs horen and vinden are the same in the a. and b. cases, i.e. the role assigned to the post-verbal clause in the a-cases is assigned to the complement between brackets in the c-structures. However, this complement cannot remain as it is there. The verb must be extracted from it by the process of V-raising, resulting in S-structures like (24b) and (25b). The relevant questions that must be asked are (a) (b) (c)

what is the category of the bracketed expressions in (24c)-(25c)? why must this type of complement precede the verb? why does V-raising have to apply?

The answers to those questions will automatically account for the rather limited distribution of the V-raising structure. I shall discuss these questions in this order. There are two candidates for the brackets in (24c) and (25c), S, i.e. I N F L " , or V" (i.e. S). Let us look at the possibility that it is a projection of I N F L first. There are a number of problems with this assumption. First of all, it would not explain why these preverbal clauses are never introduced by a complementizer. This is a remarkable property. Some verbs, like proberen "try" can take both a clausal complement in postverbal position or a preverbal complement that undergoes V-raising. In the former case, the clause can be introduced by the complementizer om, but not in the latter, as is illustrated by the examples in (26). (26)

(a) dat that (b) dat that (c) *dat that

Piet Peter Piet Peter Piet Peter

probeerde (om) de deur te openen tried for the door to open de deur probeerde te openen the door tried to open om de deur probeerde te openen for the door tried to open

Secondly, the subject of the preverbal complement must receive Case from the matrix verb. If an I N F L projection intervenes between this subject and the matrix verb, one would expect Case assignment to be impossible. Thirdly, if the bracketed part is an I N F L projection, it is expected that the complement has to follow the verb, just like all sentential complements do. Let us therefore adopt V" as the category of these preverbal complements. The absence of a complementizer is then automatically explained. Similarly, there is no problem in assigning Case to the subject N P , as the configuration is essentially the same as with small-clause AP's as in (12c).

114

Government and the distribution of sentential complements in Dutch

A motivation for the assumption that no I N F L projection is present in the case of these preverbal complements is that the temporal interpretation of these complements is fully determined by the matrix verb, a fact that is discussed in De Geest's thesis on perception-verb complements (cf. De Geest 1973). This would be explained if each INFL projection creates its own temporal domain, INFL being the node that dominates TENSE. The second question is why these V"-complements precede rather than follow the verb. It should be noted that their preverbal occurrence violates the UCC. The fact that they must nevertheless precede the verb may be derived from our hypothesis of unidirectional government. Suppose that the bracketed part in (25c) is moved to postverbal position. The structure would then be ruled out by the Case Filter since the lexical subject of the V" would not receive Case, as Case is assigned under government and V governs to the left. This point was already mentioned above. Therefore, if an infinitival clause is to have a lexical subject, it has to occur preverbally. It is no coincidence, therefore, that all verbs that take what is traditionally called an Acl complement also trigger V-raising, or to put it differently, take preverbal complement clauses in Dutch. If this is correct, one would expect that subject raising verbs must also take preverbal complements. This expectation is borne out. Consider the following examples. (27)

(a) dat het schijnt dat zij ziek zijn that it seems that they ill are (b) dat zij [ y e ziek te zijn] schijnen

V-raising still has to apply to the structure in (27b). The reason why all subject raising triggers must also be V-raising triggers is that the trace in the embedded subject position, resulting from the application of MOVE a, must be properly governed by the matrix verb, which it can only be if it precedes the verb. We can distinguish between three classes of V-raising triggers: verbs that take Acl complements and verbs that induce subject raising, which both have been discussed above, and thirdly verbs that take a phonetically non-realized subject which is not a trace. What I would want to give as an answer to the second question, viz. why are these complements in preverbal position, is that the subject of these complements must be governed by the matrix verb. This is immediately evident in the first two cases, but not in the third. Consider the example in (28a) which is an instance of this third type. (28)

(a) dat zij de krant wil lezen that she the newspaper wants read "that she wants to read the newspaper" (b) dat zij [ y e de krant lezen] wil

The question is: what is e in the structure in (28b)? The generally adopted

Teun Hoekstra

115

answer to this question is that it is PRO. This would require that the subject is ungoverned, contrary to our assumptions. It should be noted that the idea that the subject position of these preverbal complements is PRO in these cases is somewhat odd. The structure must be transparent so as to allow the extraction of the verbal head by means of V-raising, whereas it should be opaque in order not to violate the PRO-theorem. I claim that the empty category in subject position is indeed governed, in accordance with the assumptions we have been making with respect to these preverbal complement clauses. This result can be obtained if it is assumed that this empty category is not a pronominal anaphor, a characterization that would lead to the requirement that it is ungoverned, but rather that it is an anaphor. It should be noted that these alleged P R O subjects of preverbal clauses are never interpreted as arbitrary in reference, but always as coreferential with either the subject or the object of the matrix clause, i.e. always controlled. We are dealing then with an empty category in an A position, A-bound by a local antecedent with an independent 0-role. Given the functional definition of empty categories and the general optionality of phonological features, the existence of such an entity is to be expected. The problem that this entity brings in is how its distribution can be limited to the subject position of this complement type. It would take us too far afield to adress this question. The requirement that the antecedent and the anaphor receive their 0-roles from different predicates appears to be descriptively adequate, however, and has interesting consequences. The final question to be answered, then, is why V-raising has to apply. This answer can be derived from the UCC. It was already mentioned above that the structures in (24b), (25b), (27b), and (28b) violate the UCC, as a V-projection is in the government domain of a verb. The answer to the second question makes clear that this violation of the UCC cannot be avoided by placing the complement in postverbal position, as this would lead to an ungoverned subject, given the directionality of government in Dutch. Therefore, the projection has to be eliminated. This is exactly the consequence of V-raising according to Evers (1975), a consequence of a general pruning convention that applies to any category of which the head is extracted. An interesting recent reinterpretation of this pruning process is presented by Huybregts (1983). It should be noted that the fact that the distribution of V-raising structures is limited to complements of V is also accounted for. I have just argued that the essential property of these complements is that their subject must be governed by an outside governor. Therefore, they can occur in the complement of those categories that have the ability to govern into a sister projection. This ability is restricted to verbs in Dutch, i.e. only verbs are structural governors in the sense of Kayne (1981).9 This concludes our discussion. I have shown that the interaction of the UCC and the unidirectionality of government accounts in a principled fashion for the distribution of sentential complements in Dutch.

116

Government

and the distribution

of sentential

complements

in Dutch

Notes *I would like to thank Frits Beukema and Henk van Riemsdijk for helpful discussion of the ideas in this article. 1. For a discussion of the way in which sentential complements receive a thematic role the reader is referred to Hoekstra (1983a). 2. Although each category may have a subject, this is not necessary. Specifically, it will be argued below that a particular type of P P does not have a subject. Whether or not a particular instance of XP has a subject is determined by the thematic grid associated with the head of XP. An exception is the subject of S, which seems to be obligatory, (cf. Chomsky's (1982) Extended Projection Principle). 3. It should be noted that both Kayne's hypotheses and Case Theory have scope beyond the UCC. The maximalization is therefore a matter of generality of the principle, rather than a maximalization of its scope. 4. The definition of government that I defend in Hoekstra (1982) differs from generally adopted versions of government in that it assumes that the roof for government is not the maximal projection, but rather the X' domain. Therefore, the appositional N P in (4) is not governed by the noun according to this definition of government. 5. For a discussion of the absolute met/with construction, see Beukema & Hoekstra (1983a, 1983b). 6. Various pieces of evidence suggest that such a difference exists. The processes of VP deletion and VP preposing found in English are absent in Dutch. This should not cause any surprise if VP is not a maximal projection in Dutch. A similar difference has been argued to exist between English and the Scandinavian languages by Taraldsen who conjectures that V is the head of S in all Germanic languages that display V-second phenomena. 7. For a somewhat more extensive discussion of this matter, see Hoekstra (1983a). 8. Therefore, this P P will not include a subject (cf. note 2). 9. Although the theory presented here makes rather straightforward predictions, it is not easy to falsify it. It is not sufficient, e.g. to point at a rigid SOV language like Japanese, which takes preverbal complement clauses, in order to falsify the theory. It should also be argued that these preverbal clauses are projections of [ + V, — N]. In the case of Japanese, this is not true: the preverbal clauses are really complements to a noun koto "fact", that protects the sentential projection from the governing features of the verb.

The lexical head constraint, X-theory and the "pro-drop parameter" GEOFFREY HORROCKS University of London

Introduction In recent years there has been quite extensive discussion in the literature of the so-called "pro-drop parameter". 1 Here I wish to suggest that the phenomena typically discussed under this heading are better explained as following from quite independent principles of UG, and at the same time to provide answers to a number of questions left unanswered in Horrocks (1983). For those not fully familiar with the framework of Generalized Phrase Structure Grammar (GPSG), I begin with a brief synopsis. The basic idea is that a phrase structure grammar should be defined not as a list of phrase structure rules but rather by means of a metagrammar which generates such a rule list and expresses generalisations about it. The components of the metagrammar are presented in (l): 2

(1)

ID rules

Metarules

The initial set of immediate domination (ID) rules express domination relations without reference to linear order. A subset of these ("lexical" ID rules) introduce lexical heads and their complements, and are equivalent to subcatagorisation frames. Each ID rule is assigned an arbitrary number, so we might have a V' rule numbered 6 allowing V' to dominate V and N". Verbs such as kiss and bite will then be listed in the lexicon as belonging to subcategory 6, the verb class that can appear in the context admitted by rule 6. Lexical ID rules may be operated on by metarules, whose purpose is to express generalisations about possibilities of subcategorisation. For example, the metarule in (2): 117

118 (2)

The lexical head • H, N", N", V" (bet) V' - » H , V" (say) [±C] (d) H < N " < P" < V' < V" (ii) (a) V" -> N", H' (b) a < H' (c) V' -> V, N", P", V" V' ->• V, V" [ + C] (d) H < a < V"

(stikhimatizo) (leo)

These are rules of the grammars of English and Modern Greek respectively. The sentence rules (6(i)(a)) and (6(ii)(a)) and the L P rules (6(i)(a)) and (6(ii)(b)) are identical. The I D rules for bet and stikhimatizo ( = "bet") are very similar. Where in English one bets someone so much that something will be the case in Greek one bets so much with someone that something will be the case. Again, the I D rules for say and leo ( = "say") are virtually identical, the only difference being the optionality of the complementiser in English. The L P rules for lexical heads and their complements, (6(i)(d)) and (6(ii)(d)), are similar in that heads must come first and complement sentences last, but they differ in that in English other complements are strictly ordered with respect to each other while in Modern Greek they are not. The result is the Greek has considerably freer phrase order than English (see Horrocks 1983). So far the grammars of the two languages d o not seem very different. But suppose now that, under pressure from the L H C to get subjects into the subcategorisation frame of a lexical head, English chooses option (5i) (complement liberation) and Modern Greek chooses option (5ii) (X' liberation into X"). It follows at once that in English subject N " will always precede V'. Either we have a normal "configurational" sentence whose constituents appear in the order specified in L P rule (6(i)(b)), or we have a "liberated" sentence within a V', with N " and V' as sisters of a verb like say, where once again N " will precede V' because of L P rule (6(i)(d)) (cf. also infinitival examples with the same order such as believe Noam to have recanted). It also follows that Modern Greek will have the so-called "pro-drop" property of "free" subject inversion. In configurational sentences subject N " will precede V' as in English (the relevant L P rules are identical), while in "flat" sentences resulting from the liberation of the constituents of V' the verb will necessarily precede the subject N", as well as its complements, because of the "head first" L P rule (6(ii)(d)). Free inversion is simply a consequence of the L H C and the choice of X' liberation rather than complement liberation as a means of placing subjects in properly governed positions. Notice, though, that free inversion

Geoffrey Horrocks

121

of subject is a property of those languages that choose (5ii) and have a head first LP rule; the two do not necessarily go together. Still further consequences follow from the choice within X-theory between (5i) and (5ii). If the contents of a complement V" were liberated [+c]

into a mother V', that and V" would be governed by V, but not the subject N" and V'. It is only when the contents of a complement V" are liberated [-C]

that subjects become properly governed and so accessible to the mechanism introducing unbounded dependency gaps. Hence that is obligatorily absent from structures involving a subject dependency in English.6 The *that-t filter and its successors now receive a very natural explanation that follows from the general principle of government theory that unbounded dependency gaps must be sisters of lexical heads, a requirement that itself follows from the role of metarules in GPSG as statements of subcategorisation generalisations. It is clear from consideration of co-ordination possibilities that this analysis is along the correct lines. Where there is a dependency into subject position in a complement sentence the result will necessarily be a structure involving a bare, tensed V':

Since dependencies into positions other than that of subject result in V"/N" it should not be possible to co-ordinate subject dependency structures (V') with object dependency structures (V"/N"). This is clearly the case, as (8) shows: 7 (8)

T h i s is the woman that I met and invited me to the party.

In Modern Greek, on the other hand, there is no problem with the complementiser. Since the subject of a flat sentence after the contents of a V' have been liberated is governed by its own verb:

122

The lexical head constraint V

(9)

V"/N'

N' N" V

pyos (who)

ipes (said-2-s)

V"/N'

pos (that)

V irthe — (came-3-s)

0 j

the complementiser can stay, and in fact is obligatory. Hence we derive a second "pro-drop" characteristic, namely apparent violation of the *that-t constraint, from the LHC and the choice of (5ii) rather than (5i). Notice that on this analysis subject dependency structures are V"/N" just like object dependency structures, and that the two types should therefore coordinate freely. This is in fact the case: (10)

afti ine i yineka pu sinandisa ke me proskalese sto parti this is the woman that met-l-s and me invited-3-s to-the party "This is the woman that I met and (who) invited me to the party"

Notice too that inversion of the subject is not a prerequisite for apparent violations of the *that-t constraint; again the crucial factor is proper government by a lexical head. 8 I turn finally to the phenomenon of 'pro-drop' itself. Consider the metarule (11) which handles cliticisation in Modern Greek: (11)

V'

H, X, N " M

V V' - H, X, (N") M M In effect, an object N " is optional in an ID rule if it is governed by V and that V agrees with it in person number and gender. Such object agreement features are spelled out as a proclitic pronoun by a rule schema of the type in (12): (12)

V

cl + V m

M

The input and output rules in (11) together with the spellout rule (12)

Geoffrey

Horrocks

123

therefore permit three options. Once mapped into PS rules by the LP rules, these will analyse sentences of the following types: (13)

(i)

idha tin Gzanthula saw-l-s the Blondie "I saw Blondie" (ii) tin idha tin Gzanthula her saw-l-s the Blondie "I saw her, Blondie" (iii) tin idha her saw-l-s "I saw her"

(V + N"-object)

(cl + V + N"-object, "clitic doubling") (cl + V)

Compare now (14): (14)

V' - H, X

«

V" -+ H, X, (N") M M This is the V' liberation metarule already introduced as (5ii), except that agreement features have been added and the subject N " has been made optional. Again, once mapped into PS rules by the LP rules, the input and output rules in (14) will analyse sentences of the types: (15)

(i)

i Ksanthula traghudhise the Blondie sang-3-s "Blondie sang" (ii) traghudhise i Ksanthula sang-3-s the Blondie "She sang, Blondie" (iii) traghudhise sang-3-s "She sang"

(subject-N" + V')

(V + agreement/clitic + subject-N") (V + agreement/clitic)

The formal parallelism between the output of the cliticisation metarule (11) and that of the V' liberation/flat sentence metarule (14) is obvious. In fact if a suggestion of Borsley's (this volume) were adopted, and V" and V' were treated as instances of the same category differing only with respect to a feature [ ± subject], the two metarules could be collapsed into a single statement. In any case the generalisation can be expressed very simply. Cliticisation is a process which involves the copying, and subsequent morphological spelling-out, of the agreement features of N " onto lexical governors which are their sisters; such N " then become optional. 9 Clearly subject N " in flat sentences may be "cliticised", but not the subject N " of

124

The lexical head constraint

configurational sentences, which are accordingly obligatory in a language such as English. They may, of course, be cliticised if complementisers are proper governors in the language in question, see footnote 9. "Pro-drop" then is simply the extension of cliticisation to subject N " when these are properly governed (i.e. by a lexical head). In Modern Greek this involves government by V. Once again the crucial determining factor with respect to the optionality of subject N" is the notion of government; and in the particular case of Modern Greek this is connected (though this is not a necessary connection) with the choice of (5ii) rather than (5i), a choice itself conditioned by the LHC, which in turn follows from the role of metarules in GPSG.

Conclusion The main conclusion of this paper is that there is no "pro-drop parameter" as such. The syntactic properties in question are in fact independent of each other (as is clear from consideration of languages such as West Flemish, which exhibits some, but not all, of the properties customarily grouped together under the "pro-drop" heading, see footnotes 8 and 9), and follow from the LHC, which is in effect a restrictive version of the ECP, and the choice which it forces between options made available by Xtheory for getting subjects into properly governed positions. The results of the analysis offered here are, I believe, quite compelling. If we insist on some principle to the effect that what is logically propositional must also be syntactically sentential (e.g. Chomsky's (1981) Projection Principle), then parts at least of the account offered would have to be abandoned. I think it is worthwhile questioning the validity of such principles, not only on the grounds that they appear to be global, but also on the grounds that, by introducing logico-semantic considerations into syntax, they may well prevent one from adopting the most general and best motivated solutions to certain problems of syntactic analysis. Notes 1. See e.g. Chomsky (1981, 4.3 and 4.5) for a summary and some discussion. 2. For a comprehensive presentation see Gazdar and Pullum (1982). 3. Clearly there will be some variation between languages with respect to permissible values of W in (4). Notice that the use of a metarule for slash termination, implying that we are dealing with a generalisation about subcategorisation, is entirely natural. Arguments for a level of deep structure representation and a rule of w/i-movement based on the need to express subcategorisation generalisations appear in virtually every textbook of transformational syntax. 4.

It seems that a further option, of

V" liberation into V" , may have to be allowed for [-C] C+c] in the light of West Flemish data presented by Bennis and Haegeman in this volume. This, of course, assumes that complementisers may be proper governors in some languages, see notes 8 and 9 below.

Geoffrey

Horrocks

125

5. See Pullum (1982) for some discussion of the role of liberation metarules in GPSG. 6. Notice that in a language such as French, which is very like English in being a "non pro-drop" language, a different strategy has to be adopted (at least in some dialects) because of the obligatoriness of complementisers. If subject N " cannot be placed in a governed position as in English, it seems that "dependencies" into subject position involve no displacement; if there is no gap there is no problem with the LHC. Hence examples such as: Voici la femme que vous croyez qui est arrivée here-is the woman that you believe who is arrived "Here is the woman who you believe arrived" with qui in argument position, just as if que vous croyez were missing. 7. See Gazdar (1981) and Gazdar, Klein, Pullum and Sag (1982) for extensive discussion of this and related problems. 8. Thus in West Flemish, for example, (Bennis and Haegeman, this volume), we get such "violations" of the *that-t constraint even though there is no real evidence for subject inversion of a grammatically unconditioned sort. We must take complementisers to be proper governors in this language, cf. footnote 4, and see also footnote 9. 9. It is interesting that in West Flemish (Bennis and Haegeman, this volume) there may be overt markers of subject agreement on the complementiser, suggesting once again that it can function as a proper governor.

Infinitivsyntagmen als aktanten in den verbzentrierten satzmustern des deutschen IRMA H Y V Ä R I N E N Universität Oulu, Finnland, Akademie

von Finnland

1. Einleitung In der verbzentrierten Valenztheorie wird das aus dem Finitum allein bzw. aus dem Finitum nebst weiteren Prädikatsteilen bestehende Prädikat als Hauptglied des Satzes angesehen, von dem das Subjekt, verschiedene Objekte, das Prädikativ und die verbspezifischen Adverbialien als Aktanten bzw. Ergänzungen abhängen. Das jeweilige regierende Verb, der Valenzträger, legt die Quantität sowie die morphosyntaktische und semantische Qualität der Aktanten fest. Die Aktanten sind entweder obligatorisch, so d a ß ihre Auslassung den kontextlosen, isolierten Satz ungrammatisch macht, vgl. (1) und (3) unten, oder aber fakultativ, d.h. ohne Gefährdung der Grammatizität auslaßbar. Auch die fakultativen Aktanten sind aber im Stellenplan des Verbs verankert, denn sie sind nicht in beliebigen Sätzen hinzufügbar oder durch eine andere F o r m ersetzbar, vgl. (2) unten. Auf die Grenzziehung zwischen den obligatorischen und fakultativen Aktanten gehe ich hier nicht ein, sondern konzentriere mich auf die Gesamtvalenz. - Eine besondere G r u p p e unter den Aktanten bilden die sog. disjunkten Aktanten: Von zwei Aktanten erscheint der eine oder andere allein, oder beide stehen nebeneinander. - Von Aktanten sind freie Angaben zu unterscheiden, die in Grenzen der allgemeinen semantischen Kompatibilität in beliebigen Sätzen hinzugefügt oder gestrichen werden können, vgl. (4).1 (1) (2)

(3) (4)

(a) (b) (a) (b) (c) (a) (b)

Karl besucht seine Tante, (obl. Aktant) »Karl besucht. Karl liest (das Manuskript), (fak. Aktant) *Karl arbeitet das Manuskript. Karl arbeitet (an dem Manuskript), (fak. Aktant) Karl wohnt in München, (obl. Aktant) *Karl wohnt. Karl arbeitet [eifrig] (an dem Manuskript) [in München]. (freie Angaben) 127

128

Infinitivsyntagmen

als aktanten in den

verbzentrierten

Die herkömmliche Valenztheorie befaßt sich hauptsächlich mit einer oberflächennahen Satzstruktur (ONS) und stellt den Satz als eine Dependenzhierarchie dar. Von der Wortfolge wird in der Regel abgesehen, bei Bedarf können aber auch Wortstellungsregeln mit der Valenztheorie kombiniert werden. 2 Man kann die Valenzsyntax entweder mit Hilfe von morphosyntaktischen Satzmodellen oder aber mit Hilfe von Satzgliedmodellen erfassen. Darüber hinaus kann man semantische Information entweder als semantische Merkmale oder Tiefenkasus bzw. Kasusrollen hinzufügen oder auch mit einem mehrstufigen Modell a r b e i t e n . 3 - Mein Ziel ist es, verbspezifischen Infinitivkonstruktionen Satzgliedfunktionen zuzuordnen. Hinter der O N S nehme ich eine syntaktische Tiefenstruktur (SYTS) an, die wierderum durch eine logischsemantische Tiefenstruktur (LSTS) bedingt ist, vgl. (5). Infinitivische Aktanten werden als eingebettete Sätze betrachtet. Für Vollverben als Valenzträger wird unabhängig von der Oberflächenform eine bestimmte syntaktische Tiefen- bzw. Elementarvalenz angenommen. In der Valenzhierarchie des Gesamtsatzes fungiert das infinitivische Verb je nach dem Grad der Einbettung als Valenzträger zweiten, dritten usw. Grades. Auf der Ebene der SYTS stelle ich die Elementarvalenz der Einbettung in einem umrahmten Kasten dar, vgl. (6). (5)

(6) SYTS:

Karl hat vor,

ein Haus zu kaufen.

hat. vor

ONS:

Karl

kauft / \ Karl ein Haus

hat^vor Karl

zu kaufen

I

ein Haus

2. Zum Wesen des Infinitivs Die Fähigkeit des Infinitivs, als Aktant aufzutreten, gilt als eine nominale Eigenschaft. Zu den verbalen Eigenschaften zählt wiederum seine Fähigkeit, eigene Objekte. Prädikative und Adverbialien zu bekommen. Anders als ein finites Prädikat kann der Infinitiv jedoch kein Nominativsubjekt bekommen, aber jedes infinitivische Verb, in dessen Elementarvalenz ein Subjektaktant enthalten ist, hat ein sog. logisches Subjekt. Je nachdem, wie dieses logische Subjekt im Gesamtsatz realisiert werden kann, können die Infinitivkonstruktionen gruppiert werden. Diese sog. Subjektrelation (bzw. »Orientierung« 5 ) ist eine subklassenspezifische Eigenschaft des regierenden Verbs. In der klassischen Transformationsgrammatik wird die Subjektrelation durch Tilgungs und Anhebungstransformationen erklärt. 6 In einigen Fällen m u ß aber das Untersatzsubjekt getilgt werden, auch wenn der Obsersatz kein identisches Glied enthalten kann (vgl. 3.2. unten). Deswegen erhebt sich die Frage, ob

Irma

129

Hyvärinen

etwa eine lexikalistische Annäherungsweise der transformationalistischen vorzuziehen wäre.7 Unabhängig davon, welche Erklärungsbasis gewählt wird, ist es wichtig, die regierenden Verben mit einem Subjektrelationsindex zu versehen. 8 um der Bildung von fehlerhaften Sätzen vorzubeugen. Andererseits wird die Subjektrelation weitgehend durch die Verb- und Satzsemantik festgelegt, so d a ß mindestens ihre Identifizierung in der Regel keine Schwierigkeiten bereitet.

3. Gruppierung der Infinitivkonstruktionen nach der Subjektrelation Die verbabhängigen deutschen Infinitive können in sieben Subjektrelationsklassen eingeteilt werden, die in der Tabelle (7) zusammengefaßt sind. Auf die jeweils in Frage kommende Transformation bei der Ableitung der Infinitivkonstruktion wird in einer eigenen Spalte hingewiesen. Bei den sechs ersten Gruppen Inf a-Inff fungiert die Infinitivkonstruktion als Aktant, wogegen es sich bei Inf g um ein komplexes Prädikat handelt. Inf a, Inf b, Inf c und Inf d können durch die Equi-NP-Tilgung abgeleitet werden. Auch wenn das koreferente Obersatzglied nicht explizit im Satz erscheint, kann die Konstruktion mit Hilfe einer Hinzufügungsprobe richtig eingeordnet werden. In einem Satz wie (8a) gehe ich davon aus, d a ß das indefinite Untersatzsubjekt mit dem ebenfalls indefiniten, fakultativen und hier im konkreten Satz fehlenden Dativobjekt identisch ist: 9 (8)

(a) Der Lehrer empfiehlt, dieses Buch zu kaufen, (b) Der Lehrer empfiehlt uns, dieses Buch zu kaufen.

Nur wenige Verben weisen Variation in der Subjektrelation auf, und auch dann ermöglicht der Kontext in der Regel die richtige Interpretation. So verlangt z.B. bitten in der Regel einen Inf b, enthält der Untersatz aber das Modalverb dürfen oder steht der Untersatz im Passiv, geht es um die Subjektrelation Inf a.10 (Dabei wird angenommen, d a ß die Passivierung des Untersatzes der Infinitivierung vorangeht.) Vgl. (9)

3.1.

(a) (b) (c) (d)

Ich Ich Ich Ich

bitte dich, das Fenster zu öffnen. (Inf b) bitte, das Fenster zu öffnen. (Inf b) bitte, das Fenster öffnen zu dürfen. (Inf a) bitte, informiert zu werden. (Inf a)

Infe

Die durch Inf e gekennzeichnete Subjektrelation enthält die sog. accusativus cum infinitivo-Konstruktionen (Acl) bei den Verben sehen, hören, fühlen, spüren sowie lassen und machen. In der einschlägigen Literatur

130

Infinitivsyntagmen (7)

als aktanten

in den

SUBJEKTRELATION DES INFINITIVS

Kennzeichen

Subjekt des Infinitivverbs

Inf a

Identisch mit dem Subjekt des regierenden Verbs

Infb

Identisch mit dem Akkusativobjekt des regierenden Verbs

Infc

Identisch mit dem Dativobjekt des regierenden Verbs

Infd

Identisch mit dem Präpositionalobjekt des regierenden Verbs

Infe

Teil des Infinitivsyntagmas

das logische Subj. fungiert als Aktant des Inf.

Inff

(a) Identisch mit einem Aktanten des Prädikativs, d.h. mit einem Obersatzaktanten zweiten Grades

das logische Subj. des Infinitivs ist weder Aktant noch Koaktant des Inf.

(b) Hat keine im Stellenplan des Gesamtsatzes vorgesehene Realisation

Inf 9 «

w

T3
' S o .c 'S S • SP . > ,5 u. U. t-r s f ? „ V Ö 3 -o S a " 3 c - a ^ fit, ÛÛ 1> o •CH c -o 'Sc *o o S í í ag - s c ¡ 2 c =6 — ».a « A BQ S n O ^ e-3-Sjs a a -s : 9 y . •fi «> « « .a 3 >g S M i .a» s60 s® e T3 Ë 2 o S 3 a„ 1* 3- s 3 u S •O 3 r. wï ^ O N tí "" u ï S S EocS d .a ¡s s 5 « •o « G c 3,-S 25 VcE .2 cd '5 U. W S u "S s> JJ J3 a "S0> "8 « ~ C U V V X) so 1_ C 2 S 3 > ao.o 3 '53 u. u j i u u J t L i t t J u j t g w t i j u J W W •53 i> > U J3 UJ UJ 3

JJ-H oo -3 ja ~ e V)

oc c 3 C J3 O

Cu X! 3 . Ci

S JS cd J3 3

N 8 3 ' 5 5 N E 3 S •g ¡ t u » « * S X BU o ô5 •S .S M 3 a so £ 8 • -C US Ö -C u .s wa ! "3>-o a' U â S i¡•g J5.S < J3 * 3« r S m —i l B - S M> - ''§ J is iö w K C £ " , U§o i ' g5 i: o c« T3 ^ e - g ïNJ ShaSflJ a $-2-S ! m teC n3 I tfl 3 o (A ^ M I W N C/3tu tu 2 l i S H

£

7 3

C JJ «3 •M S

•o E J3 « 00

• ¿ oA 1r : « 2 S1 û CL. O + + + 'S'S'S'S 0 0 0 0 J¿ J¿ J¿ M M M M C »« • ¡3• S3 - S3- 2 j-— J-— "sc SO >eS e ^ > 3 ¿ ¿ ' f b a h >1/11/: c/3c/3 ^ ^ 00 -S « -s Q -o w U UJ tu tu i: fl-SUfli* 73 UJw u S JS ,2 u £ u Swwh NQ 1 1

3

%

& > te o -S o3 3 ? + + + o 9 o o M ** w rt jì ed rt Q , ^ «trt'!? « » Ja »c bu « c -1 ts s . 5 '8 a N •, ai

•O O à

+ +

io1 jy 3 3 C/3 C/3

fi fi fi ^

o o o jaxixix) X> 3 3 3 3 3 V3 C/3 C/3 C/31/3

3 •o

:0 3

o

+ £

j. « + + Q 3 :« -O 3 3 3 3 3 C/5 C/1 C/3 C/5oo fl. C/5« 2

+ + + +

£ „

3 3 C/3 C/1

Ì !

O

i < CU

£

+

te ^ o. o '¿•li ? » 3 ' C/3 < i + + • IffS" I! 3C/3?C/3

138

Infinitivsyntagmen

als aktanten in den

verbzentrierten

ersetzbar und können kein Korrelat haben. Es gibt aber auch infinitivische Aktanten, die diese Eigenschaften aufweisen, deren Regens aber aufgrund von (i) als Vollverb gilt, z.B. sich weigern. (iv) Ein mangelhaftes Tempusparadigma und fehlendes Passiv sind für Hilfsverben typisch; es gibt aber auch Nebenverben mit vollem Paradigma (z.B. anfangen) und Vollverben ohne Passiv (z.B. alle Reflexiva). (v) Das Vorkommen von zu. topologische Kriterien. Auxiliarisier und Negierbarkeit des Infinitivs blieben in meiner Untersuchung außer Betracht. Auch sie könnten zur Festlegung des Selbständigkeitsgrades der verschiedenen Infinitivkonstruktionen verwendet werden. Das Resultat wäre ein Spektrum mit gleitenden Übergängen. Als Nebenverben betrachte ich: (a) das futurale Hilfsverb werden; (b) die sog. sprecherbezogenen Modalverben (z.B. Er muß/wird/dürfte/mag/kann/ soll/will uns gesehen haben); (c) die Varianten der nichtsprecherbezogenen Modalverben, die Subjektindiiferenz aufweisen, vgl. Kriterium (i) (z.B. In diesem Aufsatz darf es keinen Tippfehler geben, Darauf kann/muß verzichtet werden); (d) die nichtintentionalen Verben der aktionalen Einstufung (Es beginnt/fängt an/fährt fort/hört auf zu regnen); (e) Verben für Anschein (scheinen) Habitualität (pflegen), Tendenz (drohen, versprechen, das nichtvolitive wollen z.B. in Der Regen wollte kein Ende nehmen) sowie ( / ) das umgangssprachliche »Platzhalterverb« tun (z.B. Regnen tut es bestimmt den ganzen Tag).

4. Satzmodelle mit infinitivischen Aktanten Im beiliegenden Anhang sind die infinitivischen Satzmodelle nach der Subjektrelation und Satzgliedfunktion des Infinitivs gruppiert. 2 3

Anmerkungen 1 Zu den Grundprinzipien der Valenztheone siehe z.B. Tarvainen (1981) sowie die theoretischen Einleitungen in den Valenzwörterbüchern von Helbig/Schenkel (1982) und Engel/Schumacher (1978). Zu der Entwicklung und den verschiedenen Richtungen der Valenzforschung siehe Korhonen (1977). 2. Vgl. z.B. Engel (1970). 3. Mehrstufig ist z.B. des Beschreibungsmodell von Helbig/Schenkel (1982). Siehe auch Korhonen (1982). 4. Die in der vorliegenden Arbeit verwendeten Baumgraphen sind Dependenzgraphen mit dem Prädikat als oberstem Glied. An den Knoten der Bäume können Lexeme, morphosyntaktische Symbole, Satzgliedsymbole oder Siglen für Ergänzungsklassen hängen, vgl. (a). In einigen Arbeiten werden Bäume verwendet, in denen S ( = Satz) als oberstes Symbol steht und das Prädikat mit seinen Argumenten auf gleicher Ebene steht, vgl. (b). Konstituentenstrukturbäume wie (c) tauchen in der deutschen Valenzforschung selten auf. In einigen Fällen, z.B. bei den Anhebungstransformationen, kann der Baumtyp auf die Darstellungseleganz einwirken.

Irma

139

Hyvärinen (a2) einen

Vfin Sn

Apfel

Sa

vgl. Heibig/

V(jn = finites Verb S = Substantiv n = Nominativ a = Akkusativ

Schenkel (1982) (a 3 )

Prädikat Subjekt

V

(»4)

Akkusativobjekt vgl. Engel/ Schumacher (19781

V= < >= E= 0= 1=

Verb Satzbauplan Ergänzung Nommativ Akkusativ

S=Satz P = Prädikat A = Argument

S= NP = VP = N= V= Det =

Satz Nominalphrase Verbalphrase Nomen Verb Determinator

5. Vgl. Bech (1955 31-42). 6. Vgl. z.B. Funk-Kolleg Sprache 1 (1973 286f.) und Reis (1976 64ff.). 7 Siehe Höhle (1978). 8. Vgl. Engelen (1975/1 96) und (1975/2). 9 Z.B. Edmondson (1982 158) sieht hier eine Indefinit-NP-Tilgung im Untersatz vor, ohne auf das Vorhandensein einer identischen Leerstelle im Obersatzmodell hinzuweisen. 10. Vgl. Bech (1955 38f„ 112f.) und Helbig/Buscha (1972 94). 11 Vgl. Duden Grammatik (1973 524), Engelen (1975/1 69-79), Heringer (1970 173179, 183-192), Evers (1975). 12. Vgl. z.B. Helbig/Buscha (1972 106). 13. Z.B. Bech (1955 131-165), Lindgren (1964 322), Schulz/Griesbach (1982 79f., 340), Erben (1972 589f.) und Helbig/Buscha (1972 91) betrachten das logische Subjekt des Infinitivs als a n vom Obersatzverb verlangtes Akkusativobjekt. Ebenfalls Bierwisch (1971.122f.) leitet die Acl-Konstruktionen mit Hilfe der Equi-NP-Tilgung ab; ähnlich Harbert (1977) die Acl-Konstruktionen bei den Wahrnehmungsverben, wogegen lassen + Infinitiv durch Clause Union abgeleitet wird. 14. Gegen die Passivierung von Acl-Konstruktionen sind z.B. Folsom (1966 11), Höhle (1978 1 Vif.) und Clément (1971 248). Konstruierte Beispiele mit Passiv im Ober- oder/und Untersatz finden sich z.B. bei Teubert (1973). 15. Siehe Bruaas (1977 361). Vgl. auch Bech (1955 141f.) und Harbert (1977 131). 16. Helbig/Schenkel (1982 139). 17 Vgl. Ebert (1976 128). 18. Engel/Schumacher (1978 249, 201, 182 und 214). 19 Bruaas (1977 360). 20. Teubert (1973). 21 Zu diesen Verben siehe u.a. Helbig/Buscha (1972 92), Buscha (1973 116), Engelen (1975/2 15f.) und Bech (1957 28). 22. Auch wenn das Subjekt in der Valenztheorie als Aktant betrachtet wird, kommt ihm eine gewisse Sonderstellung zu; es gilt als primus ínter pares. Die meisten deutschen Verben verlangen ein obligatorisches Subjekt und stellen daran semantische Restriktionen. -

140

Infinitivsyntagmen

als aktanten in den

verbzentrierten

Referentielle Subjekte (Sn) und das formale Subjekt es schliessen sich per Definition gegenseitig aus, denn das formale Subjekt ist weder auslaßbar noch ersetzbar; vgl. Heibig/ Schenkel (1982:95f.), die jedoch das formale es als Teil des Verbs betrachten. 23. Zu der Liste der infinitivischen Satzmodelle (Anhang) sind folgende Anmerkungen nötig: (i)

Das formale es als Subjekt oder Objekt, vgl. die Modelle (9), (13), (25) und (36), betrachte ich als einen formalen Aktanten ohne Entsprechung in der LSTS. Es wird in der quantitativen Gesamtvalenz mitgezählt. (ii) Der Terminus Infinitivobjekt, vgl. (14) und (26), steht für Fälle, in denen das Infinitivsyntagma nicht mit einem nominalen Objekt kommutiert, so d a ß die Unterart des Objekts nicht als Akkusativ-, Dativ-, Genitiv- oder Präpositionalobjekt bezeichnet werden kann. (iii) In (15), (16), (17), (27) und (28) betrachte ich die Infinitivkonstruktion als einen mit dem nominalen Adverbial disjunkten Aktanten, d.h. der Satz enthält den einen oder anderen oder aber beide gleichzeitig. (iv) Das Korrelat (es oder Pronominaladverb, z.B. in (4), (7)) betrachte ich als einen diskontinuierlichen, formalen Teil des Aktanten.

Infinitival complements in Dutch JAN KÖSTER Tilburg University

1. Theoretical background The major theoretical background assumption of this article is the thesis of radical autonomy. According to this thesis, the major properties of grammar are construction-independent. In the earliest versions of generative grammar, there was an almost general correspondence between construction types and rules. Thus, for the passive construction there was a passive transformation, for the cleft construction there was a cleft transformation, and so on. It was soon realized that the rules in question had much in common, and that an adequate theory has to factor out these common elements. In the seventies, this led to highly impoverished rule components ("move alpha") that interacted with principles that were in part construction-independent. Or at least, many ingredients of the principles were construction-independent. The property of c-command, for instance, was part of many different dependencies, such as anaphor binding and the antecedent-trace relation (movement). In fact, the elaboration of trace theory was a major step in this development because it stressed the similarities between anaphor binding and movement. The path towards a theory of construction-independent properties was not followed to its utmost consequences, however. Construction-bound rules are definitely something of the past, but the standard collection of properties is still to a large extent constructiondependent. Many properties are not properties of constructions, but properties of construction classes. Each of the following classes, for example, is distinguished by its own cluster of properties: (1)

construction class (a) anaphor binding (b) movement (c) control

properties binding theory subjacency argument structure

In general, some overlap of properties is allowed, but the major property of locality is supposed to be different in all three cases. 141

142

Infinitival complements in Dutch

Principle A of the binding theory (Chomsky 1981: anaphors are bound in their governing category) is the locality principle for bound anaphors. In part at least, movement is also constrained by this principle because NP-traces are considered anaphors. The locality principle for movement is subjacency. It is particularly important for wfi-movement, because, contrary to NP-traces, wJt-traces are not considered anaphors. Subjacency is conceptually different from principle A of the binding theory in two respects. First, it is a condition on rule application rather than a condition on representations (like principle A), and second, it specifies two bounding nodes rather than the one node mentioned by principle A. As I have argued elsewhere, there is considerable evidence that a version of principle A is also the right locality principle for w/i-traces.1 Evidence to the contrary can only be found in SVO languages like English, Italian, French, and the Scandinavian Germanic languages. SOV languages like Dutch and German show a very strict behavior with respect to island conditions. In fact, all known locality conditions for whtraces follow from a version of principle A for these languages. The same is true for the aforementioned SVO languages, in the unmarked case. The exceptions in these languages are due to a marked strategy, the connectedness strategy discussed by Kayne (1983). Since this strategy, which treats wh*traces as pronominals, is not available in the same way for SOV languages, we predict that these languages do not have "variables" that are not strictly locally bound, such as parasitic gaps and traces bound in violation of the W^-island constraint (cf. Koster (forth-coming)). If this conclusion, based on the study of Dutch, is correct, the thesis of radical autonomy is vindicated because there are no construction-specific locality principles anymore: both anaphors and w/i-traces are governed by a version of principle A of the binding theory, and subjacency can be cancelled as a superfluous condition. In this paper, I would like to concentrate on certain locality phenomena in control constructions. According to standard assumptions, control is neither constrained by principle A nor by subjacency. The fact that control has locality constraints in many (but not all) cases is presumably regarded as a consequence of the relatively unknown principles of argument structure. Although I agree with the idea that some locality properties of control are due to principles of argument structure, I do not believe that it is true in general. It is not true, I will claim, where the local properties of control correlate with three other properties which can also be found for bound anaphors and traces (including w/i-traces). We are now coming to the core of the thesis of radical autonomy: there is a cluster of at least four properties that can be found in all three construction classes mentioned in (1), and in many other constructions. The following properties are relevant:

Jan

Koster

(2) (a) (b) (c) (d)

143 properties of core grammar obligatoriness uniqueness c-command locality

dependencies

These properties systematically correlate, and are so generally found in many different constructions that I believe that they constitute the core of the properties of dependencies permitted by universal grammar. The first two properties are usually not systematically considered, but introduced in a rather casual way. Take "obligatoriness", for instance. It has been known almost since the beginning of generative grammar that bound anaphors differ from pronominals in that they are obligatorily bound. In the binding theory of Chomsky (1981), it is assumed without comment that anaphors must be bound in their governing category. It appears that "obligatoriness" systematically correlates with the other properties mentioned in (2). The same holds for uniqueness. Bound anaphors always have a unique antecedent, i.e. split antecedents are not permitted. The same appears to hold for the antecedent-trace relation and many other dependencies. The first two properties of (2) probably indicate that the core dependency pattern of grammar has the form of a function: for each argument (a category of a certain type), there is a unique value (the antecedent) in a given structure. The third property, c-command, is well-known and the only case where the thesis of radical autonomy has not been challenged. It is generally assumed that this property is more or less uniform for all dependencies. 2 It is perhaps a historical accident that the fourth property, locality, has been considered non-uniform for such a long time. It is probably due to the "derivational history" of generative grammar. Until Chomsky (1973), anaphora and movement constructions got a rather different treatment and it was only after the introduction of trace theory that the two construction classes were systematically studied from the same perspective. But even after the introduction of trace theory the obvious step towards unification was not taken. The reasons appear to be rather complex. First of all, there was no obvious motivation left for movement transformations after the introduction of traces. The main argument has been since then, that movement is characterized by its deviant locality property, namely subjacency. The second reason has been the violation of the w/i-island condition in certain languages (cf. Rizzi 1978). Such violations seemed incompatible with the "anaphoric" view of the locality principle involved in movement. The third reason is somewhat more subtle. It has to d o with the "logistic" view of w/i-movement. According to this view, wh-movement creates operator-variable structures. A further consequence would be,

144

Infinitival complements in Dutch

then, that the alleged variables (wfi-traces) do not behave as anaphors but as the conceptually related names (Chomsky 1981). All these reasons have been bad reasons. As I mentioned before, the whisland violations are presumably due to an independent strategy for empty pronominals, so that they really tell us nothing about the nature of locality for w/¡-traces. The "logistic" argument is even worse. It has been shown, conclusively in my opinion, that wfc-traces have nothing to do with variables in the logical sense (cf. Koster 1983a, and especially van Riemsdijk 1983 for compelling arguments). Control is another example of a construction with (sometimes) local properties that are believed to be independent from the type of locality found in principle A of the binding theory. I will show within a moment that also in this case the locality properties are sometimes instances of the locality found in principle A of the binding theory. Before going into control constructions in more detail, I would like to mention some potential misunderstandings of my position. It has sometimes been said that I denied the difference between trace and PRO, but that I cannot be right because there are clear differences between the antecedent-trace and the antecedent-PRO relation.3 But that the relations are different, is exactly what was claimed in Koster (1978). My criticisms were directed against the then current view that trace and P R O are two different primitives. The differences were recognized but relegated to an independent factor, namely the nature of the antecedent (0 or 0). In other words, trace and P R O were distinguished contextually, like in Chomsky (1981, chapter 6). Similar comments can be given for the distinction between A-binding and A-binding. It is not claimed here that these relations are the same. What is claimed, however, is that these functionally different dependencies have something important in common, namely the configurational properties under (2). In other words, the thesis of radical autonomy does not entail the absurd claim that all constructions have the same properties. It only entails that many constructions have the pattern (2) in common, and that there are no construction-specific locality principles insofar as constructions also have the other properties of (2) in common. Another factor that should not be overlooked is that locality principles can be parametrized with respect to the "opacity factors" involved. I am claiming that the general form of locality principles is uniform: (3)

y must have an antecedent in: . . . . . . co... y...]... where co is the opacity parameter (values: SUBJECT, C O M P , X°) and ft the minimal category containing co and y

For English anaphors, the value of the opacity parameter is (accessible) SUBJECT. In Dutch, it also involves C O M P (cf. Koster 1983b). In both languages, it involves C O M P for wfc-traces.

Jan

Köster

145

Apart from the setting of these parameters, then, the pattern (2) is totally independent from the content of lexical items and therefore from semantics in any sense of the term. Let us see next how control is related to this central pattern.

2. Infinitives in English and Dutch In Koster (1978) it was claimed that, apart from the 0-status of the antecedent, the following two sentences have the same configurational properties: (4)

(a) John, seems [t,to go] (b) John,- tries [e, to go]

Chomsky (1981) is a serious challenge to this view, because it claims that the properties of raising (4a) are substantially different from the properties of control (4b). As said before, the difference has never been at issue as far as the properties of the antecedent are concerned, but I do not believe that there are compelling arguments for other differences between (4a) and (4b). Other alleged differences are (i) the general configurational differences between movement and control, and (ii) the fact that the embedded subject of (4a) is governed (trace), while the embedded subject of (4b) is not governed. Neither of these claims is strongly supported by the facts. Standard arguments involve the demonstration that control can be non-local as in: (5)

Johtii thinks that it is important [e,to show himself]

It has become clear since Williams (1980) that such examples tell us little about the status of PRO in (4b). The reason is that Williams convincingly demonstrates that control is not unitary in English. He distinguishes two classes, obligatory control with the properties of (2), and optional control, which does not necessarily involve the properties of (2). Example (4b) is in the former class, while (5) is in the latter class. Consequently, (5) tells us nothing about (4b). In an earlier paper (Koster 1981; cf. 1983c), I claimed that obligatory control (in Williams' sense) occurs in complements that lack the complementizer for (in underlying structure), just like raising complements of type (4a). This suggested that S-deletion (Chomsky 1981) is not distinguishable from absence of a complementizer ( f o r with infinitives). This means that complements of verbs like try (4b) have a transparent "§deletion" complement, because try does not select for. This has the nonstandard consequence that PRO (e,) is governed in (4b), and in for-less verb complements in general. PRO is, in other words, an anaphor in these contexts.

146

Infinitival complements in Dutch

This conclusion goes against a central thesis of the GB framework, but it has the nice consequence that it explains why Williams' obligatory control has the properties of anaphors (cf. 2). The standard view, on the other hand, does not explain the dichotomy in the class of control complements. An objection against anaphoric PRO (from the point of view of Chomsky 1981) is that we lose an ECP explanation for the following ungrammatical sentence: (6)

*Johni was tried [e, to go]

If try does not trigger S-deletion, e, remains ungoverned, contrary to the requirements of the ECP for such movement structures. If try does trigger S-deletion, this explanation is lost. This objection can easily be circumvented, however. The reason is that there is still an independent theory of control, minimally a theory that stipulates that the controller for promise or try is the underlying subject. This property of try is in conflict with the requirements of the binding theory, induced by the fact that e, is governed. According to the binding theory, e( must be bound by John in (6); according to the theory of control, Bj must be bound by the underlying subject of try. Since it is impossible to meet both requirements, (6) is ungrammatical. The same explanation holds for what Bresnan (1982) calls "Visser's generalization": "verbs whose complements are predicated of their subject don't passivize." Thus, we cannot derive *His friends are struck (by him) as pompous from He strikes his friends [as PRO pompons'], because the theory of control requires binding from the underlying subject. Since PRO is in a small clause, it is also in a transparent complement, and therefore governed. But if it is governed, it must be bound by an overt c-commanding antecedent, again a requirement incompatible with the requirement of control theory (for the derived structure). Another interesting contrast given by Williams (1980) is the following: (7)

(a) *It was. tried [ P R O to go] (b) It was arranged [ P R O to go]

In general, this construction is only possible with verbs that select for (like arrange), and not with verbs without for (like try). The explanation rests on the fact that for the theory of control it often does not matter how the designated controller is expressed. It can be left implicit (as in 7b), or it can be expressed by an obligue phrase as in It is difficult for Billj [PROj to gfo]. The binding theory requires an overt c-commanding antecedent for empty categories. This is why (7a) is ungrammatical: since P R O is governed, the designated controller (the subject of try) cannot be left implicit as in (7b). In fact, PRO must be bound by it m (7a), which is not the designated controller.

Jan

147

Koster

This analysis is confirmed in an interesting way by evidence from Dutch. In this language, the equivalent of (7a) is grammatical: (8)

Er werd geprobeerd [(om) P R O te gaan] There was tried CO M P to go

The reason for the grammatically seems clear: Dutch proberen (try) selects a complement with an optional complementizer om. Consequently, the complement is not necessarily transparent in Dutch, like the English complements that select for. Surprisingly, Dutch also has a counterpart of (7a) that is ungrammatical. Since Evers (1975) it has been known that in Dutch infinitival complements occur on both sides of the verb. Proberen (try), for instance, can have its complement "extraposed" to the right of the verb, in which case it optionally selects om, or it can have its complement to its left, in which case it never selects a complementizer and obligatorily ondergoes Verb Raising. The following facts illustrate the possibilities: (9)

(a) dat that (b) *dat (c) dat [yte

hij probeerde [(om) P R O het boek te lezen] he tried COMP the book to read hij [(om) P R O het boek te lezen] probeerde hij [(*om) P R O het boek [ye,] probeerde lezen,] V-raising

I will refer to these two complement types as extraposed complements and VR complements, respectively. The fact that Dutch has these two complement types makes it impossible to escape from the conclusion that P R O can be governed. It appears to be the case that VR complements are generally transparent, while extraposed complements are opaque. Since P R O occurs in the transparent VR complements (cf. 9c), it must be governed there. But then we make a prediction. Sentences like (8), the grammatical counter parts of the ungrammatical English sentences like (7a), always involve extraposed complements, as can be seen from the optional complementizer. VR complements, on the other hand, pattern like English (7a), i.e. passivization leads to ungrammatically: (10)

*Er werd [ P R O het boek [ye,]] geprobeerd There was the book tried

[ y t e lezen,] to read

In this case, the infinitival complement is transparent, which results in governed PRO, which in turn makes the sentence ungrammatical like the English example (7a). There is plenty of evidence that VR complements are in general transparent, in contrast with extraposed complements.

148

Infinitival complements in Dutch

It is remarkable, for instance, that complements that require transparency, like raising complements, only occur as VR complements: (11)

(a) *dat hij schijnt [r het boek te that he seems the book to (b) dat hij [i het boek — ] schijnt that he the book seems

lezen] read te lezen to read

There is an independent reason why the trace t cannot be governed in (11a). Dutch is an SOV language, which presumably means that verbs only govern arguments to their left (cf. Hoekstra 1982). If that is the case, the trace in (1 la) is on the wrong side of the verb. This also has a bearing on exceptional case marking in Dutch. This phenomenon, if it is a phenomenon, only affects VR complements (12a) and never extraposed complements (12b) (see also De Geest 1972): (12)

(a) dat hij [haar that he her (b) *dat hij zag that he saw

het boek — ] zag lezen the book saw read [haar het boek lezen] her the book read

Again, there is an independent cause for the ungrammaticality of (12b). This example involves a so-called te-less infinitive, i.e. an infinitive without the morpheme te (Engl. to). These te-less infinitives are more NP-like, and are in all likelihood a counterpart of the English gerund (cf. Koster 1983c for arguments). Like gerunds, te-less infinitives cannot be extraposed. What is important in this context is that exceptional case marking in (12a) shows that ie-less infinitives are transparent. This is important, because many te-less infinitives are control complements: (13)

dat hij, [PRO, het boek — ] wil lezen that he the book wants to read

There is no reason to assume that the subject position of (13) (PRO) is less accessible for government from the matrix verb than the case-marked subject in (12a). Other transparency phenomena are formed by the movement properties of clitic-like elements in Dutch (cf. Evers 1975). The morpheme er (there), for instance, can freely be moved from a VR complement (in control structures), but not from the corresponding extraposed complements: (14)

(a) dat er iemand [ P R O een verhaal t over — ] probeerde that there someone a story about tried tevertellen to tell

149

Jan Koster (b) *dat that over about

er iemand probeerde [(om) PRO een verhaal t there someone tried comp a story te vertellen] to tell

A last transparency phenomenon that I would like to mention here is adverbial scope. Sentence adverbials, like waarschijnlijk (probably) can freely occur in VR complements, and have wide scope (over the matrix clause) (15a). This is not possible with extraposed complements (15b): (15)

(a) dat hij [PRO het verhaal waarschijnlijk — ] probeerde that he the story probably tried vertellen tell (b) *dat hij probeerde [(om) PRO het verhaal waarschijnlijk that he tried comp the story probably vertellen] tell

te to

te to

It is fair to say, then that transparency phenomena are systematically found in VR complements, but not in extraposed complements. But since VR complements can contain PRO, this element is not protected against government from above. 3. Conclusion According to the thesis of radical autonomy, the four properties listed under (2) are not construction-specific. Especially the fourth property (locality) cannot be different for movement, bound anaphora, and control, provided that the three other properties of (2) can also be found in these constructions. Since it was already known that a subclass of the control structures shows the pattern (2) (Williams 1980), it was worthwhile to consider the possibility of governed PRO for these cases. The hypothesis that certain control complements are like raising complements in that the absence of a complementizer makes them transparent, did not only lead to an explanation for the occurrence of pattern (2) itself in the relevant cases, but also to an explanation for several other facts, such as Visser's generalization. These, and many other facts are more extensively discussed in Koster (1981, 1983c). On the basis of the Dutch complement system we concluded that governed (anaphoric) PRO cannot be avoided. In Dutch, there is a class of infinitival complements, VR complements, that shows the following transparency phenomena: Verb Raising, N P raising, exceptional case marking, clitic climbing, and wide scope of sentence adverbials. The transparency of VR complements correlates, like in English, with the fact that these complements never select a complementizer. But since PRO occurs in

150

Infinitival complements in Dutch

these generally transparent VR complements, there is no non-ad hoc way to avoid government of P R O by the matrix verb in these cases. Like for English, this is a welcome conclusion because this government from the matrix verb explains why PRO in VR complements always shows the pattern (2). But then there is a form of locality that cannot only be found in anaphora constructions and movement constructions, but also in certain control constructions. The existence of a common locality principle for anaphora, movement and control confirms the thesis of radical autonomy.

Notes The Bounding Condition in Koster (1978), especially in the form given in Roster 1 (1979) has the same form as principle A. See also Koster (1981a) and (forthcoming, chapter 5). 2. point of Lasnik's 3.

In Koster (1982), it was proposed that c-command is a locality principle from the view of the antecedent. Like standard locality, it can perhaps be parametrized as in notion "¿command" (cf. Lasnik 1976). Cf. for instance Chomsky (1982, 87).

Some remarks on connectedness and c-command GIUSEPPE LONGOBARDI Scuola Normale Superiore

1. In a recent article R. K a y n e (1983) has proposed to subsume Chomsky's (1973) original Subject C o n d i t i o n a n d several other syntactic constraints under a principle whose most refined formulation is the following: (1)

Connectedness Condition (CC) Given a set of empty categories / ? j . . . /?„, each locally b o u n d by a single antecedent a in a tree T, the union of the g-projection sets of every a n d the antecedent a must f o r m a subtree of T.

W h e r e in t u r n g-projection set and g-projection are defined as follows: (2)

(3)

T h e g-projection set of a category p, governed by y, is constituted by /?, every g-projection of y, and every category d o m i n a t i n g jS and n o t d o m i n a t i n g y. (a) Y is a g-projection of X iff (i) Y is a projection of X (in the sense of X ' theory) or of a gprojection of X, or: (ii) Y immediately d o m i n a t e s W a n d Z; Z is a g-projection (maximal, f r o m the X' theory point of view) of X; and W a n d Z are in a canonical government configuration. 1 (b) W a n d Z are in a canonical government configuration iff: (i) In a language with basic V O order, W precedes Z. (ii) In a language with basic O V order, Z precedes W.

As can be seen, the Connectedness Condition (1) (henceforth CC) at-

151

152

Some remarks on connectedness

and

c-command

tributes the ungrammatically of (4)

(a) * Which boy would for us to approach e be risky? (b) *The boy who for us to approach e would be most risky . . .

in contrast to the grammaticality of (5)

(a) Which boy would it be risky for us to approach e? (b) The boy who it would be risky for us to approach e

to the obvious linear asymmetry between preverbal and extraposed subjects: preverbal subjects are leftmost constituents of the node which immediately dominates them, i.e. S, a position not allowing g-projections of an internal gap to extend further up in an VO language and thus to reach their antecedent in the examples (4). Extraposed subjects are obviously never leftmost constituents and thus w/i-extraction from them is possible. Such an account automatically extends to parasitic empty categories (in the sense of Taraldsen (1981); Chomsky (1982) and Engdahl (1983)), like the ones in the following examples, which appear to freely violate Subjacency but not the Subject condition: (6)

(a) *A person I did not mention e supposing that for us to approach e would be risky (b) ?A person I did not mention e supposing that it would be risky for us to approach e.

The notion of "union" of the g-projection sets of coindexed empty categories built into the formulation of CC correctly allows, in addition, for the only kind of gap actually acceptable within a preverbal subject: (7)

?A person who [$[s' f ° r e]...

us

to approach