234 40 12MB
English Pages 269 [272] Year 1984
Parametric Syntax
Studies in Generative Grammar The goal of this series is to publish those texts that are representative of recent advances in the theory of formal grammar. Too many studies do not reach the public they deserve because of the depth and detail that make them unsuitable for publication in article form. We hope that the present series will make these studies available to a wider audience than has hitherto been possible. Editors:
Jan Köster Henk van Riemsdijk
Other books in this series: 1.
WimZonneveld A Formal Theory of Exceptions
in Generative
Phonology
2. Pieter Muysken Syntactic Developments in the Verb Phrase of Ecuadorian Quechua 3. Geert Booij Dutch Morphology 4. Henk van Riemsdijk A Case Study in Syntactic 5. Jan Köster Locality Principles
Markedness
in Syntax
6. Pieter Muysken (ed.) Generative Studies on Creole 7.
Languages
AnnekeNeijt Gapping
8. Christer Platzack The Semantic Interpretation 9. Noam Chomsky Lectures on Government
and
of Aspect and Binding
10. Robert May and Jan Koster (eds.) Levels of Syntactic Representation 11. Luigi Rizzi Issues in Italian
Syntax
12. Osvaldo Jaeggli Topics in Romance
Syntax
Aktionsarten
Hag it Borer
Parametric Syntax
Case Studies in Semitic and Romance Languages
1984 FORIS PUBLICATIONS Dordrecht - Holland/Cinnaminson - U.S.A.
Published
by:
Foris P u b l i c a t i o n s H o l l a n d P.O. Box 509 3300 A M D o r d r e c h t , T h e N e t h e r l a n d s Sole distributor
for the U.S.A.
and
Canada:
Foris P u b l i c a t i o n s U . S . A . P . O . Box C-50 C i n n a m i n s o n N . J . 08077 U.S.A.
CIP-data Borer, Hagit P a r a m e t r i c s y n t a x : case studies in Semitic a n d R o m a n c e languages / Hagit Borer. - D o r d r e c h t [etc.]: Foris P u b l i c a t i o n s . - (Studies in generative g r a m m a r ; 13) ISBN 90-6765-024-2 b o u n d I S B N 90-6765-025-0 p a p e r S I S O 805.4
U D C 804.56 + 809.256
Subject heading: syntax; romance lnaguages/syntax; Semitic languages.
I S B N 90 6765 024 2 ( B o u n d ) I S B N 90 6765 025 0 ( P a p e r ) ® 1983 Foris P u b l i c a t i o n s - D o r d r e c h t . N o part of this p u b l i c a t i o n m a y be r e p r o d u c e d or t r a n s m i t t e d in a n y f o r m or by a n y m e a n s , electronic or m e c h a n i c a l , including p h o t o c o p y , r e c o r d i n g , or a n y i n f o r m a t i o n s t o r a g e a n d retrieval system, w i t h o u t permission f r o m the c o p y r i g h t o w n e r . Printed in t h e N e t h e r l a n d s by I C G Printing, D o r d r e c h t .
To My Parents
Contents
Introduction
1
Chapter 1: PARAMETRIC SYNTAX - A MODEL 1. General Theoretical Assumptions 2. The Inflectional Component 2.1. The Projection Principle Revisited 2.2. Redefining Inflectional Relations 2.3. Levels o f Application 2.4. Further Constraining Inflectional Rules 3. Parametric Syntax Notes
7 7 15 15 18 21 24 27 29
Chapter 2: CLITIC GOVERNMENT 1. Introduction 1.1. Case Absorption 1.2. The Complement Matching Requirement 2. The Construct State and Clitic Government 2.1. The Construct State: General Properties 2.1.1. A Note on Genitive Case Assignment 2.2. Clitics and the Construct State 2.3. The Sei Phrase and the Position of the Clitic 2.4. Coindexing and Government 3. Extraction from Construct State Configurations 3.1. Introduction: Predictions 3.2. Free Relatives in Modern Hebrew 3.3. sei Insertion Revisited 4. Proper Government by Coindexed Clitic 4.1. Predictions 4.2. Two Clitic Configurations Notes
33 33 36 37 41 41 46 48 53 57 68 68 72 78 79 79 80 88
Chapter 3: CLITICS AND THE GOVERNMENT-BINDING MODEL 1. Introduction
101 101
VIII
Contents
2. On the Status of Variables 2.1. Variables and Case marking 2.1.1. Counterexamples 2.2. A-Chains and the Visibility Hypothesis 2.2.1. A- Chains and the Binding Conditions 2.2.2. The Visibility Hypothesis 3. Clitics and the Binding Conditions 3.1. Variables 3.2. Non-Variables 3.2.1. Clitic se Notes Chapter 4: INFLECTIONAL RULES 1. Introduction 1.1. Clitic Doubling and Extraction in Romanian 2. On Differences in Extraction between Hebrew and Romanian 3. On the Insertion of Case Markers 3.1. sei Insertion at S-structure 3.2. Free Application of pe Insertion 4. Free Clitic Spell-Out Notes
101 101 106 110 Ill 114 117 118 119 122 124 127 127 127 132 136 138 141 146 148
Chapter 5: CLITICS AND RE ANALYSIS IN ROMANCE 1. Introduction 2. Clitic Government and Reanalysis Constructions 2.1. Clitic Government in Causative Constructions 2.2. The Case Tier 2.3. Reanalysis 2.4. Parametric Variation in Causative Constructions 2.5. Permitir-Type Verbs in River Plate Spanish 3. Extraction from Clitic-Doubling Configurations in RP Spanish 4. A Note on Dative Clitics 4.1. Dative Case Assignment 4.2. Inalienable Possessive Constructions Notes
155 155 157 157 162 165 173 179
Chapter 6: PARAMETERS FOR INFL 1. Introduction 2. Existential sentences - presentation 3. Pro-Drop: Analysis 3.1. Pleonastic Subjects and Nominative Case A s s i g n m e n t . . . .
203 203 203 212 214
182 187 188 193 196
Contents 3.2. The Person Marker 3.3. Summary 4. Inflectional Clitics 4.1. Ergative Particles 4.1.1. A Note on Clitics and PF 4.2. The Negation Marker 'eyn as an Exceptional Preposition 5. Existential Sentences: the Accusative Derivation 5.1. Free Superscripting and r/zere-Insertion Notes
IX 219 223 224 225 231 232 237 242 244
Conclusion
251
References
255
Introduction
It has always been the assumption of generative linguistics that the purpose of linguistic theory is to understand the nature of the language faculty and to explain the acquisition of language. This while considering the impoverishment of the stimuli to which the language learner is exposed and the unavailability of direct negative evidence. The lack of evidence for "language learning" in the common sense of the term "learning", as well as the absence of any plausible learnability theory capable of explaining the nature of language acquisition on the basis of exposure to data alone, has led to the assumption that the language faculty is best characterized as a biological faculty, a mental organ of some sort, with inherent properties of its own. This mental organ has often been referred to as Universal Grammar (UG). UG narrowly restricts the class of possible grammars which a child can infer on the basis of limited, defective data. Informally speaking, then, the notion of UG allows us to suppose that the child, when exposed to linguistic data, does not construct models that would account for the data from scratch. Rather, he fits that data into already existing, innate slots. If one is to allow for the great level of generality which such an approach implies, and at the same time account in a natural way for language variation, the UG component must offer a rather abstract class of operations and principles. These can then be interpreted in different ways in different grammars. Within the theoretical framework of the Extended Standard Theory (as sketched in particular in Chomsky, 1973, 1975, 1976; Chomsky and Lasnik 1977; and subsequent literature), an example of a general operation that is, in turn, restricted by particular grammars is the rule 'Move a'. While the rule itself is part of universal grammar, different grammars may choose different values for a. Furthermore, they may choose to restrict the domain of application of the rule. For instance, it has been argued that in Chinese, 'Move WH' applies in the logical form (LF) component, but not in the syntax (see Huang, 1982, for a discussion). Thus we assume that UG contains abstract principles and operations which hold universally, such as 'Move a, X theory, the binding conditions,
2
Parametric
Syntax
the principles which govern the organization of the lexicon, etc. (see Chapter 1 for discussion of some of these notions). The particular interpretation of these principles, however, is subject to variation in a principled fashion, giving rise to an interaction between the general and the abstract on one hand, and the particular on the other. This interaction, in turn, results in a wide range of differing grammars. As an example to illustrate our point, let us look at the phrase-structure component, as given in UG by the X theory. We must allow for parameters of X theory that would arrange categorial components within the X system so as to permit SVO languages, SOV languages, VOS languages, etc. (For a recent account of this parameter involving the direction of government see Horvath, 1981). One might, however; imagine other ways in which the X system might vary from one configurational language to another. For instance, it is possible to imagine a system in which different languages would select a different number of bars either for a specific category or for any X. The question of whether such an option is actually realized or not is an empirical issue: if, indeed, a case could be made for this kind of parameter, then the parametric component would have to allow it. The availability of such a parameter will follow from a sufficiently general characterization of the X-system. It is clear however, that we do not expect to find a language in which X theory holds, alongside a language in which rules such as (1) may be written freely: (1)
NP
S
By placing the X system in UG we thus make a very strong claim: the existence or non-existence or X theory as a system regulating the phrase structure component cannot be parametrized. While it is possible that the ordering of words in some languages will not be determined by a phrase structure component (this is possibly the situation in the so-called nonconfigurational languages), when a language does have a phrase structure component, it must conform to the X schemata. A subsequent question immediately comes to mind: what is the source of parametric variation: which component of the grammar will assume the burden of delimiting the universal principles so as to generate the wide variety of existing grammars? What will determine the parametric range? It is this question which this study intends to answer. It is a desirable step forward to try and restrict the class of possible parameters. The strongest claim in this respect would be that there are no language-particular choices with respect to the realization of universal processes and principles. Rather, interlanguage variation would be restricted to the idiosyncratic properties of lexical items. These idio-
Introduction
3
syncracies, which are clearly learned, would then interact with general principles of UG in a particular way. This interaction would result in vastly different systems. The weakest claim about the nature of parameters would be that every single principle of UG may be true or untrue for a particular grammar, depending on the availability of input evidence that can determine it. While the latter position considerably weakens the notion of UG (note that it predicts that there may be two grammars which do not share any principle of UG, including, say, the availability of distinctive features as primitives for the application of phonological rules), the former position is extremely hard to maintain. In this study we will propose a model of parameters which restricts the availability of variation to the possibilities which are offered by one single component: the inflectional component. We will be using the notion inflection in a somewhat unorthodox fashion to indicate a particular kind of local relations and local features, whether specified as properties of lexical items or as properties of grammatical formatives. While these relations include Case relations and agreement relations (traditionally characterized as inflectional relations), they also include the relation of Θ-role assignment. The inventory of these relations is universal. However, the availability of a particular subset of these relations for any given grammatical formative or lexical item is an idiosyncratic, language-particular fact. Further, we will define the operation inflectional rule as an operation which redefines the assignment of inflectional features, and the application and the availability of inflectional rules will differ from language to language. The interaction of this set of properties and rules with the principles of universal grammar will, in turn, give rise to different grammars. Insofar as our system will account for a wide range of interlanguage variation, it is clearly a desirable system: it places the burden of learning on the grammatical component that is idiosyncratic and learned in every language: the vocabulary and its properties. In essence this perception of parameters entails a particular process of language acquisition. When a child is exposed to input data, she/he has available to her/him a preliminary grammar constructed of the principles of UG. These principles, however, are too general. The narrowing-down of the possibilities offered by UG is accomplished by learning the inflectional properties of different formatives and the inventory of inflectional rules operating in the input grammar. Note that although the role of input data in this case is vitally important for choosing the right option, the relationship between the determining evidence and the option chosen does not have to be direct. It suffices that the grammatical analysis of the input data cannot be reconciled with one of the choices. For instance, it will be shown in Chapter 6 that on
4
Parametric
Syntax
the basis of the absence of a pronoun in the subject position in certain languages, the child deduces the level at which an inflectional rule applies. The universal processes and principles, as delimited by the selection process sketched above, constitute the individual grammar of a particular language (or a particular speaker), often referred to as Core Grammar. Thus the mental device we are discussing here is not simply a predetermined, rigid endowment, but also a language acquisition device to which experience serves as input and core grammar is its output. While this study offers an explanation for a wide range of language variation utilizing a parametric model based on the properties of inflection, there are some clear cases of variation discussed in the literature which remain untouched. Thus, there is no obvious way in which our model can account for the position of the head in the X schemata. Neither do we touch on the question of whether the subjacency parameter (see Rizzi, 1982) can be accounted for, given our system (for an attempt along these lines, see Truscott, forthcoming). In our opinion it is premature to determine whether all parametric variation can be reduced to the properties of the inflectional system. Such strong claims must await the investigation of more grammatical systems, while using the model of grammar assumed in this work (see Chapter 1 for a detailed sketch). The data presented as evidence for our model are based primarily on pronominal clitic constructions in Romance and Modern Hebrew. In Chapters 2 and 3 we propose a particular analysis of clitic phenomena which is then subject to parametric variation in accordance with the parametric model developed in Chapter 1. However, we will not discuss clitics exclusively. We will analyze in detail the insertion of Case markers (Chapter 4), showing that it is parametrized in accordance with our system. We will further propose a process of reanalysis in causative constructions in Romance which involves the transfer of inflectional features. This process, as expected, is subject to parametrization as well (Chapter 5). Chapter 6 is devoted to parameters which effect the inflection node (INFL): we will review and revise Chomsky's(1981) pro-drop parameter, and propose that this parameter, which is compatible with our model, interacts in an interesting way with the cliticization phenomena, giving rise to a number of unusual grammatical constructions. This study is a revised version of my Ph.D. dissertation, completed at MIT in June 1981. There is no doubt that writing a thesis in linguistics at MIT is best captured as the peak point of a dazzling educational experience. I would like to conclude this introductory note by expressing gratitude to the people who contributed to this educational experience, as well as to those who were a constant source of help during the research, the preparation and the revision of this final version.
Introduction
5
First and formost, it is a pleasure to acknowledge my intellectual debt to Noam Chomsky. The influence of his thought as well as his intellectual style stretches well beyond these pages, and has left its mark on the entire discipline. Ken Hale and Morris Halle, the other members of my dissertation committee, have offered theoretical criticism as well as methodological comments. During my life at MIT I have benefited enormously from having known them and having worked with them. The linguistics department at MIT, offering me a haven in the storm for four years, is primarily responsible for my transformation into a linguist. I thank the people who made it possible: Joan Bresnan, Jay Keyser, Paul Kiparsky, Jim Harris and Haj Ross. For their support, for their friendship and for their linguistic comaradeship, I thank Anne Rochett, Robert May, Donca Steriade, Barry Schein, Neil Elliott, Adriana Belletti, Luigi Rizzi, Youssef Aoun, Dominique Sportiche, Luigi Burzio, Maria-Luisa Zubizarreta, Alec Marantz, Henk van Riemsdijk, Pino Longobardi, Riny Huybregts, Doug Pulleyblank, Mario Montalbetti, Lisa Travis and Ann Farmer. Thanks also to Joe Emonds and Jim Higginbotham for very useful discussions and to Edit Dorou for very insightful comments. Special mention goes to my peer group: Tim Stowell, David Pesetsky and Ken Safir. You were great company. Thanks go to the Sloan group at the University of California at Irvine for a very generous post-doctoral grant and the opportunity to discuss my ideas with Ken Wexler, Mary Louise Kean, Peter Culicover, and Mike Rochemont. And of course, my extended family in the West Coast: Nina Hyams, Osvaldo Jaeggli, David Pesetsky and Tim Stowell. Even Southern California would be icy cold without you. Finally, I would like to express my warm gratitude to Eric Wehrli, for his support, his encouragement and his patience. During the months of writing and rewiring, he proved himself to be a real friend, which I will not cease to appreciate. Hagit Borer Irvine, April 1983
Chapter I
Parametric Syntax - a Model
1. GENERAL THEORETICAL ASSUMPTIONS
This study is embedded in the Government-Binding model, as sketched mainly in Chomsky (the Pisa Lectures), Chomsky (1981) and subsequent work. 1 The great diversity of approaches with respect to the central notions in GB, however, necessitates a brief introduction to the theoretical foundation on which this work is based. In what follows, we will concentrate on what we regard as essential to the understanding of our parametric model and the subsequent chapters. It is not intended, however, as a comprehensive summary of GB theory. Various subsystems and notions will be treated in greater detail where they play a central role in our argumentation, e.g., government and c-command, the Case filter, the empty category principle, the notion chain, the binding conditions and the status of empty categories. For a more comprehensive description of the GB framework the reader is referred to the references cited throughout this section. The central concern of GB is to characterize the positions in which different manifestations of NP's can appear. These manifestations include fully realized referential expressions, lexical anaphors and the empty elements: PRO, [fjpe] (so-called NP-trace) and variables (WH traces and traces of quantifier raising). To this end GB assumes several subsystems, each predicting a certain distribution of nominal elements in a certain domain: the theory of the lexicon, containing complementation specifications and thematic specifications, Case theory, the binding theory and control theory. These subsystems interact in several ways, and this interaction is further constrained by certain well-formedness conditions on derivations. The GB framework shares with an earlier version of the Extended Standard Theory its perception of the structure of core grammar. This structure, following Chomsky and Lasnik (1977), is given in ( l ) : 2
8
Parametric
Syntax
The essential claim of the grammatical model in (1) is that representations at S-structure feed into two separate components. These components do not interact. Thus an operation in the LF component cannot trigger the application of a phonological rule, nor can an operation in PF affect rules of LF. The D-structure component of (1) is factored into the lexicon and the phrase structure component. The latter we will take to be constrained by X theory. We will deviate from most standard accounts of the X system by arguing that complements can be base-generated at any bar level, as long as they are governed by the head (see Chapter 2 for a discussion). Following ideas of Hale (1978), further developed in Chomsky (1980), we will take the inflection node (INFL) to be the head of S and S. INFL is itself composed of a TENSE component and an agreement component (AGR). Thus the basic phrase structure rules of English are as in (2): (2)
a. INFL b. INFL
• gOMP INFX * Ν INFL V
In Chapter 6 of this study the AGR component of INFL will be discussed extensively. We will say little about the TENSE part of INFL. We will return to the structure of the lexicon below, in our discussion of the application of inflectional rules in the lexical component. D-structure is best characterized as that component in which one-toone correlations hold between referential expressions and thematic roles (Θ-roles), between subcategorization frames and the categories which fulfil them. This assumption is natural; at that level of the derivation or prior to it, no operations that link two positions on a tree are available. Thus the satisfaction of thematic requirements and subcategorization frames has to be "local". We will return to the precise nature of this "locality" below. The linking of positions in the tree is a property of the transformational component and of S-structure. In the transformational component, the rule 'Move a ' maps D-structure representations onto S-structure representations. Thus it could simply be
Parametric Syntax - a Model
9
considered a mode of linking positions on the three. S-structure is now to be regarded as the level at which positions are linked. If so, it is natural to assume that at this level lexical specifications like subcategorization requirements and thematic assignment are met by linked elements, rather than by single, non-linked elements. In this sense these requirements are not met "locally" at S-structure. Let us try and make this description more precise. Subcategorization frames are specified in the lexical entry of each item. Similarly, every lexical category which can assign a thematic role is specified in the lexicon as assigning this particular thematic role in a particular position, a thematic position. The one-to-one correlation between the assignment of a thematic role and the referential expression which fills the thematic position in which that role is assigned is captured by the Θ-criterion, informally stated as in (3): (3)
The Θ-criterion i. Each Θ-position is assigned an argument. ii. Each argument is assigned a Θ-role. iii. Only arguments are assigned to Θ-positions.
(For a discussion of the Θ-criterion and its properties see Freidin, 1978, who argues for a similar principle; Borer, 1980a; and Chomsky, 1981). The argument specified in the definition in (3) we will take to be either a lexical NP (either a name or a lexical anaphor) which is not pleonastic (i.e., not it or there in English), or a pronominal element (including the pronominal anaphor PRO). We will return to the status of variables and anaphoric empty categories with respect to (3) and similar conditions in Chapter 3. The principle in (3) ensures that every Θ-position will be filled by only one argument and that every argument will be assigned only one Θ-role. The application of (3) is regulated by the Projection Principle, stated informally as in (4): (4)
Lexical requirements must be met at every level
Lexical requirements in the sense of (4) include subcategorization frames and θ-role assignment. Now recall that, whereas in D-structure no linking mechanisms were available, in S-structure, such links are established by "Move a " . The notion "assigned t o " in (3) is interpreted according to the intrinsic properties of each level. Thus at D-structure there must be a one-to-one correlation between lexical requirements and single, unlinked elements. In the absence of linking mechanisms, (4) can only be met if all lexical requirements are met locally, i.e., if all Θ-positions
10
Parametric
Syntax
are filled, all subcategorization frames are satisfied etc. The Projection Principle thus gives content to the "locality" of representations at Dstructure. At S-structure, on the other hand, a network oflinkshasbeen established. It is these networks which satisfy lexical requirements, providing there is an element in the network (whether a fully realized NP or its coindexed trace) which is in the position where the lexical requirements are met. The networks established at S-structure to which lexical requirements apply are called chains. In order to exemplify the interaction of the notion chain with the Projection Principle, consider the following sentences: (5)
a. John hit Mary b· [NP ] w a s hit Mary (by John) c. Maryj was hit [e]j (by John)
The verb hit in (5a) is subcategorized for an NP complement to which it assigns a Θ-role in the post-verbal position. (This Θ-role is presumably that of a patient. For some discussion of the semantic content of Θ-roles, see Jackendoff, 1972, and references cited there). In (5b), when the verb hit appears in its participial form, there is no reason to assume that its subcategorization frame and Θ-assignment properties have changed. In fact, the selectional correlation between the verb hit in (5a) and in (5b) is captured if we assume that there was no change in the subcategorization frame or the Θ-assignment properties of hit. (5b) is the assumed D-structure representation of (5c). In this representation the lexical requirements of the verb are met by the post-participial NP Mary. Thus the Projection Principle is met at D-structure. The application of "Move a " to (5b) yields the S-structure in (5c). We now have a chain consisting of the preposed NP Mary{ and its coindexed trace. This chain satisfies the Projection Principle, although the position following hit is not filled by the argument itself at S-structure, but rather by the trace of an argument. Since this trace is part of the chain which contains an argument - the subject Mary - the chain may fulfil the lexical requirements of hit. In essence, then, given the "local" nature of D-structure and the "non-local" nature of S-structure, "Move a " is an operation which maps D-structure representations to S-structure representations in accordance with the Projection Principle. Our system now has the effect of ruling out the representation in (6) both at D-structure and at S-structure: (6)
*Mary was hit (by John)
In (6), the lexical requirements of hit are not met either at D-structure or at S-structure.
Parametric
Syntax - a Model
11
The formation of chains at S-structure must conform to the binding conditions to which we will return below (and see Chapter 3 for an extensive discussion of S-structure and a precise definition of the notion chain). Chomsky (1981) distinguishes between Α-chains and Α-chains, where A is an argument and A is a non-argument. Α-chains are composed entirely of elements in Α-positions (positions in which an argument may appear in the base; essentially [NP,S], [NP,VP], [ΝΡ,ΡΡ] and various specifier positions). Α-chains contain elements in non-argument positions; they typically contain elements which occupy the COMP position. For some discussion of the properties of Α-chains see Aoun (1982). Thus far we have mentioned the predictions about the distributions of NP's which are made by Θ-theory, X theory and the subcategorization frames of lexical items. These different systems interact to predict the distribution of arguments at D-structure, but not the distribution of nonarguments at D-structure. With Chomsky (1981), we assume that nonarguments are not represented at D-structure. Such non-arguments (typically it and there-like elements in various languages and expletive empty categories in 'pro-drop' languages) are inserted at S-structure in non-0-positions. We will return to this point in Chapter 6. The distribution of NP's at S-structure is already partially predicted from the Θ-criterion combined with the Projection Principle. Since all Θ-positions have to be filled at D-structure, and since the movement of an argument so as to cover the trace of another moved argument will result in a violation of the Projection Principle, it follows that movement is only possible from a Θ-position to a non-0-position. Other principles which determine the distribution of NP's at S-structure are the binding conditions, the theory of control and Case theory. Consider first the binding theory. The binding theory utilizes the notion of government, a structural relation which holds between a phrasal head and an argument. This structural notion plays a central role in determining the properties of many subsystems in the GB framework. Notably, complementation requirements are met in the domain of government (where by complementation requirements we mean, again, subcategorization frames and Θ-role assignment). Case assignment is also sensitive to government, as is the application of the binding conditions. In this study we will presuppose the definition of government given in (7): 3 (7)
Government (definition): In the configuration [ . . . 0 . . . α . . . 0 . . . ] α can be said to govern β iff: i. α = X°. ii. where φ is a maximal projection, if φ dominates β then φ dominates a. iii. α c-commands β.
12
Parametric
Syntax
In Chapters 2 and 5 we will return to the definition of government and to the defintion of c-command which we assume in this study. We will show that the properties of the structural configuration defined by (7) predict an interesting and non-trivial range of constructions in clitic configurations in Modern Hebrew and Romance. The domain of government is the domain in which the binding conditions apply. The theory of binding seeks to characterize and further restrict the distribution of nominal elements at S-structure and the formation of chains. The binding conditions will specify the correct linking of moved constituents and their traces. It will further specify the correct linking of an antecedent and a lexical anaphor. Given the notion of chain described above and given the fact that chains satisfy lexical requirements at S-structure, it is clear that the binding theory plays a crucial role in ensuring the correct assignment of lexical features at S-structure. Thus the Projection Principle coupled with lexical requirements predicts the distribution of NP's at D-structure, while the Projection Principle coupled with the binding conditions determines the distribution of NP's at S-structure. The binding conditions are given in (8): (8)
A. an anphor is bound in its governing category (anaphors: NP traces, lexical anaphors, PRO). B. a pronominal is free in its governing category (pronominals: pronouns, PRO). C. an R(= referential)-expression is free (R-expressions: names, variables).
The definition of the notion bound is given in (9) and the definition of a governing category in (10): 4 (9)
a is Z-bound by β iff a and β are coindexed, β c-commands a and β is in an .^-position. X = A, Ä
(10)
β is a governing category for α iff β is the minimal category containing a, a governor of a and a SUBJECT accessible to a.
The dual status of PRO as a pronominal anaphor allows us to derive the peculiarities of its distribution. Since it is subject both to binding condition A and to binding condition B, and since it is clearly impossible to satisfy both of them, PRO may appear only where the binding conditions cannot apply, i.e. where it does not have a governing category. It follows that PRO cannot be governed. Assuming that the set of governors is N,P,V,A, and AGR, the only position in which PRO can appear is the sub-
Parametric
Syntax - a Model
13
ject of the infinitive position. This position is not governed as the infinitival INFL lacks AGR. (Note that we assume that nevertheless infinitival clauses have an INFL node, a fact that follows from their clausal status. They are expanded by the rules in (2) just like tensed clauses). We will now tum to yet another subsystem which predicts the distribution of NP's. This is the subsystem of Case. It has been proposed by Rouveret and Vergnaud (1980) that the following filter holds in core grammar: (11)
*NP where NP has a phonetic matrix. [-Case]
Case assignment is sensitive to government. Thus accusative Case is assigned when an NP is governed by a verb (and adjacent to it; for some discussion of accusative Case assignment see below and Chapter 5. For an extensive discussion of the adjacency condition on Case assignment see Stowell, 1981). Oblique Case is assigned when the NP in question is governed by a preposition or a preposition-like element (again, adjacency must be met) and nominative Case is assigned when the NP in question is governed by AGR (but see Chapter 6 for a more refined formulation). Since the notion of government plays a crucial role both in the binding theory and in Case theory, it is not surprising that the position which is not "covered" by the binding conditions is also "left alone" by Case theory: the subject position of an infinitival clause. Therefore that position is not Case-marked and does not enter the binding conditions for the same reason: it is not governed. The subject position of infinitivals also supplies us with a case in which the binding conditions will fail to rule out a sentence, but the Case filter will. Thus the sentence in (12) is ungrammatical, although it is wellformed from the point of view of the binding conditions: (12)
a. *John tried Bill to win b. *John decided Billj to be believed [e]j
Bill in (12a-b) is not assigned Case, and hence the sentences are ungrammatical. The construction does not violate any other conditions. Following proposals of Aoun (1979b), we will assume that the Case filter is located in the phonological component of the grammar. This assumption is consistent with the view of the Case filter as a morphological well-formedness condition, which does not interact with semantic considerations determined in the LF component. It further allows us to state naturally the interaction between the Case filter and morphological rules of Case assignment which apply in the phonological component. We will return to this point in Chapter 3, where we reject the attempt to derive the Case filter from the properties of the LF component.
14
Parametric
Syntax
Let us now turn to the LF component in the model in (1). (Throughout this study I will use the terms "LF component" and "interpretative component" interchangeably, referring to the righthandside of the split model in (1). The LF component contains rules of quantifier raising (QR in the sense of May, 1977), rules which prepose into COMP WH elements which are in situ at S-structure and rules which assign interpretation to focus configurations. It further contains the theory of control which will not play a significant role in this study (but see Chapter 3, section 2.1.1. for some brief comment). Crucially, our grammar contains the principle in (13): (13)
The Empty Category Principle (ECP) an empty category must be properly governed.
We will argue for the proper government definition in (14): (14)
Proper Government (definition): a properly governs β iff α governs β and: i. α is +V; or ii. a is coindexed with β.
The ECP has been utilized to explain various phenomena previously attributed to other factors. Thus it has been utilized to explain the "that t " filter of Chomsky and Lasnik (1977) (see in this respect Kayne, 1980a; Pesetsky, 1978; and Taraldsen, 1978, for accounts which utilize the aspects of the Nominative Island Condition which were later subsumed by ECP). It has been further invoked to explain the phenomenon of preposition stranding (Kayne 1981a), of quantifier raising in certain configurations (Kayne, 1981b; Rizzi, 1982, Jaeggli, 1982; and others) and other phenomena. We will show the explanatory power of this principle in Modern Hebrew and in clitic constructions in Romance. Kayne (1981b) has shown that the condition in (13) applies to empty categories which are left by movement rules in LF. Notably, it applies to variables which are left by the rule of quantifier raising. Thus there is reason to assume that the ECP holds in LF. Our discussion in Chapter 2 will supply strong additional evidence for this assumption (and see also Jaeggli, 1982, for an argument that ECP holds in LF in Spanish; Rizzi, 1982, for an argument that it holds in LF in Italian, and Aoun, 1981, for an argument that it holds in LF in Standard Arabic and Lebanese Arabic). To summarize: the model of core grammar given in (1) above contains different subsystems which are located in different modules of its structure. The different components of this model interact to determine
15
Parametric Syntax - a Model
the distribution of nominal elements at D-structure, at S-structure and in LF. In (15) we repeat the model in (1), indicating for each component the subsystems which are part of it: (15)
D-structure i. lexicon ii. phrase structure component I
"Move a " i
S-structure i. chain formation ii. the binding conditions the PF component
the LF component I QR
(deletions) 1
the Case filter i phonological rules
WH raising Focus raising Control theory 1
ECP
I
I
PF
LF
(ordering within components, where it matters, is indicated by a directional arrow). Conditions on well-formedness which hold at all levels: i. The Projection Principle ii. The Θ-criterion In the following section we will argue for an enriched model, supplemented by inflectional rules which apply in a particular fashion. These rules will assume the burden of language variation, leaving intact the universal model sketched in (15).
2. THE INFLECTIONAL COMPONENT
2.1. The Projection Principle Revisitied Consider again the Projection Principle as in (4). As stated, the Projection
16
Parametric
Syntax
Principle requires that lexical specifications be represented at every level. However, the precise nature of their representation is left open. As an illustration of this last point, consider the relationship between the verb and its argument depicted in (16): (16)
V NP
If we assume the verb to be transitive, say hit, the lexical specifications which are relevant for the configuration in (16) are, at a first approximation, the following list: (17)
a. the subcategorization frame of hit requires an adjacent NP. b. hit has the property of assigning a particular Θ-role. c. the argument of hit, the [NP,VP] position, is assigned a particular Θ-role. d. hit has the property of assigning an accusative Case.s
Note that the argument of hit in the [NP,VP] position is also assigned accusative Case. This however, is not a lexical requirement. It is a wellformedness condition which is stated as the Case filter in (11). We will return to the nature of this distinction below. In the standard, straightforward case the satisfaction of the Projection Principle would be very simple: hit would assign Θ-role and Case to the adjacent NP, resulting in the satisfaction of all the requirements in (17). It is conceivable, however, that in certain circumstances these direct relations will not hold, yet the list in (17) would be represented at every level. Thus we can imagine a situation in which the appropriate Θ-role or the appropriate Case will be assigned to the [NP, VP] position by an element which is not the verb hit, while the verb hit will assign the appropriate Θ-role or Case to something other than the [NP,VP] position. As an illustration of such a case, consider the following case of cliticdoubling from River Plate Spanish: (18)
a. vimos una casa saw-we a house Sve saw a house' b. lo vimos him saw-we Sve saw him/it' c. lo vimos a Juan him saw-we to Juan Sve saw Jyan'
Parametric Syntax - a Model
17
Below (Chapters 2 and 5) we will argue in detail that pronominal clitics such as lo in (18b-c) and similar constructions in Hebrew absorb the Case assigning property of the lexical element that they are cliticized to (in the case of (18), the verb). In turn the availability of the specificity marker a in (18c), which also has the property of assigning Case, results in the assignment of Case to Juan. Thus, although the Case property of the verb is not assigned to the direct object, a violation of the Case filter is avoided. 6 Assuming this to be correct, consider first the sentence in (18a). In (18a) the verb 'to see' has an accusative assigning property which it assigns to 'a house'. This is the standard situation: the Projection Principle requires that the verb assign an accusative Case and this Case is assigned to the [NP,VP] position. The assignment of accusative to that position results in the satisfaction of the Case filter. Now consider (18b). The lexical property in this particular instance is the accusative Case property of the verb, and it is represented at every level by the clitic, although it is never actually assigned. Finally, consider (18c). In this case the verb does not assign accusative Case to the object. Rather, the specificity marker a assigns Case to the object. Nevertheless, the sentence is grammatical and hence we must assume that it satisfies the Projection Principle. This is exactly the situation described above: the sentence is not a violation of the Projection Principle or the Case filter, since the verb has an accusative Case property as required and the direct object is assigned Case as required. The fact that the verb does not assign Case directly to the direct object turns out to be irrelevant for the satisfaction of the Projection Principle. Thus it is clear that the Projection Principle cannot be interpreted as a condition on the relation between feature assigning elements and the categories to which these features are assigned. Rather, it is a condition on the presence of the features themselves: all lexical features must be represented at all levels. The relational configurations, however, may be altered. Now consider another example. It is a well-known fact that the rule of passive, while dislocating the Θ-role assigned to the active [NP,S] position, does not actually eliminate it. Thus there is a clear difference between the meaning of (19a) and (19b): while the former does not imply an external intervention which brought about the sinking of the bottle, such an external intervention is implied by (19b). The agent Θ-role has not disappeared. It is simply unexpressed in (19b). It could be expressed optionally by a by-phrase, as in (19c): (19)
a. the bottle sunk b. the bottle was sunk c. the bottle was sunk by the children
18
Parametric
Syntax
It is further well-known that the θ-role assigned to the object of by is precisely that Θ-role which would have been assigned to the active subject, as is demonstrated by (20): (20)
a. the earthquake was felt (by all Californians) b. the present was badly received (by the betrayed husband) c. the ball was hit (by the children)
Thus it is entirely implausible to assume that the Θ-role which is assigned to the Z>j>-object is assigned by by. Rather, by clearly inherits the nonexpressed - but present - subject Θ-role of the verb and assigns it to its object. Now consider the status of sentences such as(20a-c) from the point of view of the Projection Principle. The passive verb is specified as assigning an (optional) particular subject Θ-role. The ftj-object is specified as receiving that particular Θ-role. How is the Projection Principle to be satisfied in this case, where the verb cannot assign its subject Θ-role directly to the by object? (We are assuming here that the assignment of lexical features is done under government. Note that the passive verb does not govern the fty-object. For some more discussion of this point see Chapter 2, section 1.2 and Chapter 5, section 2). Consider the following possibility: the verb transmits the subject Θ-role to the preposition by, which it governs. The preposition by, in turn, assigns it to its object. (For a more explicit proposal along similar lines, see Jaeggli, 1981). Now consider the satisfaction of the Projection Principle. If the Projection Principle is a condition on the presence of features, then it is clearly met by (20a-c): the verb assigns its subject Θ-role to the preposition by and the by-object is assigned the subject Θ-role by the preposition by. If, on the other hand, the Projection Principle is a condition on relational configurations, (20a-c) should be ungrammatical: the assignment of the subject Θ-role would be specified as a strict function of the verb or its maximal projection, the VP, and this relation would have to be observed at all levels. Thus, again, we conclude that the Projection Principle is a condition on the presence of features and not a condition on relations. 2.2. Defining Inflectional
Relations
The conclusion reached at the close of the previous subsection now allows us to postulate a grammatical operation which is clearly attested both in the clitic-doubling examples and in the passive examples in (19)(20): a grammatical operation which regulates the transfer of lexical features such as Θ-role and Case. Such an operation is constrained by our interpretation of the Projection Principle in a particular way: while it
Parametric Syntax - a Model
19
could change the relations between a feature assigner and a constituent to which that feature is assigned, it could not eliminate a lexical feature. Such an elimination would result in a violation of the Projection Principle. Note that no violation of the Projection Principle would result from adding a feature. If such a feature is not a "lexical" one, in a sense made precise below, the Projection Principle does not require that it should be present at all levels, and its addition in a post-lexical stage would not bring about a violation of the Projection Principle. If, on the other hand, it is a lexical feature, it cannot be added at a post-lexical stage. We will return to this below. A simple example of a feature transfer is the assignment of Case. Under the assumption that, say, accusative Case assignment can be described roughly as in (21), it is clear that a feature that previously was a feature of a verb is a feature of an NP after application of the rule: 7 (21)
[ v . . . +accusative]
NP
NP [+accusative]
(We make an auxiliary assumption here that only phonologically realized elements will function as recipients of Case features. The purpose of this assumption will become clear in Chapters 2 and 4). 8 Similarly, a rough description of the assignment of Θ-roles in a configuration such as (16) would entail that the Θ-role that was previously the property of hit is now a property of an NP. 9 Returning to the passive cases in (20), note that the operation which will transfer the Θ-feature from the verb to by has a similar nature: a Θ-feature that was previously a property of hit is now a property of by. Now consider again the cases in (18a-c). In (18a) a Case feature is transferred from the verb to the [NP,VP], In (18b) the Case feature is incorporated in a clitic. Though not directly assigned, it is present as part of the clitic complex. The same happens to the Case feature of the verb in (18c). However, in (18c) the marker a transfers its Case feature to the direct object. What is the formal nature of an operation which transfers features, or which redefines the assignment of features in such a fashion? It is an operation on lexical entries, and as such, it is natural to assume that it is a morphological operation. As such, we expect it to have particular properties: we expect it to avail itself only to material which is present in lexical entries. Such material would be categorial type, subcategorization frame, Θ-role specification, Case features, agreement features such as number, gender and person, selectional restrictions, tense and aspect specifications, phonological representations, etc. Indeed, morphological processes in general do avail themselves of such lexical material in determining the environment for the application of particular word-formation rules. 10
20
Parametric
Syntax
Our brand of morphological rules, however, is further restricted: it cannot alter lexical specifications. Thus it could not change category type, nor alter subcategorization frames, nor change the nature of the Case assigned or Θ-role assigned, nor change the meaning of lexical entries, etc. The distinction between morphological rules which are thus restricted and morphological rules which can freely access and alter lexical specifications is not new: it closely resembles the well-known characteristics of inflectional and derivational morphology respectively. Inflectional morphology is restricted similarly to our operations, while derivational morphology is free to change lexical specifications. Thus we will assume that the operations we have been describing are those of inflectional morphology. While our class of operations closely resembles those operations which are traditionally assumed to be inflectional, they are not identical in one important respect: the assignment of Θ-roles is not traditionally incorporated in inflectional morphology, but it is included in our understanding of the term inflection. It is crucial to bear in mind that incorporating the relation of Θ-role assignment in inflection does not entail that the nature or the number of Θ-roles assigned may be altered by inflectional rules. Only the manner of assignment may be altered; it is the relation of assignment which is treated here on a par with the assignment of, say, Case. Thus we make a principled distinction between (22b) and (22c): (22)
a. John dropped the ball b. the ball dropped c. the ball was dropped
The rule which relates active drop as in (22a) to intransitive drop in (22b) is not a rule of inflectional morphology; that rule eliminates (or adds) a Θ-role, the agent's, an operation which cannot be inflectional. On the other hand, the relation between active drop and passive dropped is inflectional: while the manner of Θ-role assignment is changed, the number of Θ-roles or their nature remains unaltered. At this point we can give a formal definition of the operation which we call inflectional rule :11 (23)A. a. Let f stand for an inflectionally specified grammatical feature b. Let F stand for an assigner of f. c. Let C stand for a constituent specified without a variable B. an operation which affects the assignment of f to C, such that it is not subject to any condition exterior to the properties of f, F or C is an inflectional rule.
Parametric Syntax - a Model
21
2.3. Levels of Application What is the level at which inflectional rules apply? Inflectional rules interact with the Projection Principle. In fact, their application is directly constrained by the Projection Principle. Now consider again the distinction between inflectional and derivational morphology, as stated in the previous subsection. As the null hypothesis let us assume that rules of morphology may apply on any representational string which meets their rule description, e.g., at any level of representation in the lexicon, the syntax or the phonology. Now recall the characteristics of derivational morphology. We do not expect these rules to apply at any syntactic level. This does not, however, follow from an independent stipulation. Rather, since the Projection Principle is a well-formedness condition on syntactic levels, and since rules of derivational morphology alter lexical specifications, derivational morphology cannot apply at syntactic levels, and must apply at the presyntactic level. Similarly, the application of rules of derivational morphology in the phonological component will be restricted by independent factors: the grammatical model sketched in section 1 (see (15) and related discussion) would entail that morphological changes in the phonology will not be represented at any syntactic level. Under the assumption that derivational morphology introduces changes in lexical specifications (in a sense of lexical specifications specified below), this would lead to ungrammatically. Thus the conclusion that derivational morphology may only apply in the lexicon, at a pre-syntactic level, can be derived from independent principles. The opposite is true of inflectional morphology. Since inflectional rules are directly constrained by the Projection Principle, their application at any syntactic level would be allowed. In some sense this result enables us to obliterate the formal distinction between inflectional and derivational morphology. We may simply say that morphological rules may apply at any level, subject to well-formedness conditions applicable at that level. Only rules whose output conforms to the Projection Principle can apply at the syntactic component. Subsequently these rules will have precisely that set of properties which is traditionally attributed to inflectional morphology. On the other hand, inflectional rules could apply freely in the pre-syntactic component and in the phonology as there are no well-formedness conditions that would block their application there. This fact would, in turn, account for the availability of certain inflectional morphemes in the lexicon and for lexical processes in which inflectional forms participate. (An example of the former are suppletive verb forms, as in the verbal paradigm of go in English. For an example of the latter see compounding in German as analyzed in Lieber, 1980). In recent literature (Lieber, 1980; Kiparsky,
22
Parametric
Syntax
1982) the availability of inflectional operations in the lexicon serves as evidence for locating inflectional morphology in the lexicon. While these facts clearly indicate that some inflectional morphemes must be listed in the lexicon, within our system, the acutal application of inflectional rules in the sense defined in this study may take place in the lexicon, in the syntax or in the phonology. Interestingly, recent studies in morphology (Pesetsky, 1979; Kiparsky, 1982; Mohanan, 1982) assume that morphological rules apply interspersed with phonological rules. In particular, they claim that some phonological rules apply to outputs of morphological cycles. Thus Kiparsky (1982) proposes the following model:
The similarity between this idea and our own model is apparent. Thus a model which would allow the interaction of morphological rules and syntactic levels would look as follows: 12 (25)
Parametric Syntax - a Model
23
Let us tentatively summarize: we have assumed a class of relation-altering and relation-defining rules, rules of inflection. The application of these rules is constrained by the Projection Principle, as is the application of all rules in the syntactic component. As only this class of rules considered above does not violate the Projection Principle, they may apply in the syntax. This result derives from the fact that they do not alter lexical specifications, they only alter relations. On the other hand, the application of inflectional rules is not restricted to the syntax. In the absence of well-formedness conditions that would bar their application in other, non-syntactic levels, we conclude that the following statement is correct: (26)
Given/?, R an inflectional rule,/? may apply at any level.
As a simple illustration of the way in which (26) interacts with other principles of the grammar, consider the sentences in (27): 13 (27)
a. John expects [g Bill to like Jane] b. whoj does John expect [g [e] j to like Jane]? c. John expects [g Janej to be liked [e]j (by Bill)]
In (27a). Bill is generated in the subject position. Given the ability of expect to assign Case to the subject of the subordinate infinitival, Bill is marked as accusative and the sentence is grammatical. Note that no movement takes place in (27a). Assuming the rule of accusative assignment which operates in (27a) to be roughly as in (21) above, it is clear that the environment for the application of (21) in (27a) is met at D-structure, at S-structure and at PF. The mechanism which checks the well-formedness of Case assignment is at PF, and thus, given the model in (25), we may assume that (21) can apply freely at the pre-syntactic level, at D-structure, at S-structure or at PF. Now consider (27b). Here, although the environment of the application of (21) is met at D-structure and at S-structure, the accusative Case can be assigned at D-structure only. At S-structure, following the application of 'Move a', the environment specified in (21) is only met by an element which is not phonologically realized; hence it cannot be assigned accusative Case. Thus if (21) fails to apply prior to 'Move a', the derivation is ruled out. 14 Now consider (27c). In (27c), the environment for the application of (21) is met at D-structure only by a phonologically-null element. At Dstructure, the structure of (27c) is as in (28): (28)
John expects [g [e] [ y p to be liked]]
24
Parametric
Syntax
Thus accusative Case cannot be assigned to the embedded [NP,S] position at D-structure. On the other hand, after the application of 'Move a ' Jane is in the embedded [NP,S] position and the environment for the application of (21) is satisfied. Thus in (27c), (21) may apply at S-structure and at PF, resulting in a grammatical derivation. In D-structure, however, its application is blocked. In Chapters 4 and 6 we will see that it is plausible that in some cases well-formed derivations will result only if the application of an inflectional rule takes place in the PF component. To conclude, although it is clear that inflectional rules are allowed to apply at any syntactic level, their interaction with other grammatical principles will result in narrowing down their domain of application. 2.4. Further Constraining Inflectional Rules Drawing again on recent morphological research, it has been proposed that different morphological rules are restricted to particular levels. (The levelordered morphology; cf. Siegel, 1974; Allen, 1978). Thus, returning to the diagram in (24), particular affixes are specified as attaching at level 1 (e.g., -ment; -ize; -ity; and others). Other affixes are specified as attaching at level 2 (e.g., un-; -ful; -less; -ness; and others). Various phonological properties are associated with different morphological levels, and these can be accounted for if we assume, as in (24), that level 1 morphology is associated with a particular set of phonological rules, level 2 with a different set of phonological rules, etc. On the other hand, some phonological rules may apply at more than one morphological level.1S The particular level at which morphological rules may apply, which is clearly a learned fact associated with particular morphemes, can be stated as a restriction on the application of morphological rules. 16 Thus one may state the restriction on the prefix un- as follows: (29)
un- may not be attached at level 1.
(Note, incidentally, that we need not specify that un- may not be attached at any syntactic level. Since un- changes the semantic content of a lexical entry, the Projection Principle would bar its attachement at any syntactic level). Thus it is clear that morphological rules may be restricted from applying at particular levels. The form of such restrictions would be as in (30): (30)
Given a morphological rule R, R may not apply at level n.
Finally, let us consider what it would mean for a certain inflectional rule to apply at a syntactic level other than D-structure. Chomsky (1965) draws the following distinction:
Parametric Syntax - a Model
25
A formative must be regarded as a pair o f sets o f features, one member consisting o f the "inherent" features o f the lexical entry or the sentence position, the other member consisting of the "non-inherent" features, (p. 182).
Borrowing the distinction between "inherent" and "non-inherent" features, we will assume that all inherent features are subject to the Projection Principle. "Features of the sentence position" we will interpret in accordance with our interpretation of the Projection Principle, i.e., as a set of features which are not necessarily associated with particular relational configurations. It is now clear that those features which are subject to the Projection Principle must be present in the lexicon and cannot be introduced by an inflectional rule at a later stage. Now reconsider again the list of properties in (17). We would like to divide these properties into two kinds: those that we will refer to as grammatical and those that we will refer to as substantive, drawing again on a distinction introduced in Chomsky (1965). While the subcategorization frame and the Case assignment properties will be considered grammatical properties, the fact that Θ-role is assigned will be considered a substantive property, linked inherently with the meaning of the lexical item in question. We would like to claim that the distinction between grammatical properties and substantive properties plays a role in delimiting the power of inflectional rules. We have claimed above that all inherent properties must be lexically represented. We will assume that with respect to a particular verb, say hit, its inherent features are both substantive and grammatical properties. Thus for the verb hit, being a substantive formative itself, the entire list in (17), as well as the intrinsic meaning of hit, must be represented at all syntactic levels, in accordance with the Projection Principle. Consider, however, a different situation: a grammatical formative which has only grammatical properties such as insertion environment, Case assignment features etc. Such a situation is attested by the genitive preposition sei in Hebrew, as in the examples in ( 3 1 ) : 1 7 (31)
a. beit ha-mora house the-teacher 'the teacher's house' b. ha-bayit sei ha-mora the-house of the-teacher c. beit-aj sei ha-moraj house-her of the-teacher 'the teacher's house'
26
Parametric
Syntax
Given our assumptions about the nature of clitics (see (18a-c) and related discussion), the phenomenon attested in (31a-c) is clear: in (31b-c) the clitic a absorbs the genitive Case which the noun assigns to its complement in (31a) (see Chapter 2 for detailed motivation). In (31c), a genitive Case assigner, sei (roughly ' o f ) assigns Case to ha-mora, 'the teacher', saving the phrase from the Case filter. Unlike a, the specificity marker in (18a-c), sei has no semantic function and its insertion in the environment of (31c) is entirely free. It is a genuine saving device. In the terminology used above, sei is a grammatical formative with grammatical rather than substantive properties, it is specified as inserted in a particular environment and as assigning Case. Since these two grammatical properties are not associated with a substantive formative, the set of properties which is associated with sei is not constrained by the Projection Principle. In other words, the insertion of sei by an inflectional rule in a post-D-structure level is grammatical. Assuming a stronger restriction on the insertion of grammatical formatives, we will take the statement in (32) to be correct: (32)
Only inherent features are represented in the lexicon
The statement in (32) is inspired by a strong formulation of the Projection Principle with respect to the Θ-criterion, given in (33): (33)
All and only Θ-positions are filled at D-structure
Although neither (32) nor (33) are logical necessities of the Projection Principle as formulated, they will be assumed in this work. (32) effectively disallows the insertion of sei at D-structure. Rather, the inflectional rule that would insert the grammatical formative i n ( 3 1 b - c ) must apply at a post-D-structure level. The detailed illustration that this is empirically true will be given in Chapter 4. An interesting question concerns the interaction of the principle in (32) with features such as gender, number and person. We will assume that verbs do not have these features, and that they are transferred from the noun, via INFL, if the agreement is with the [NP,S] position, and directly from the [NP,VP] position in languages which have object agreement. We will assume that gender, number and person features have the same status as Case features (and not the status of Θ-roles). Thus they are inherent features, and they must be represented lexically, only if they are the property of a lexical, substantive entry, i.e. a noun. Otherwise they may be inserted freely at a later stage, a phenomenon that is actually attested in the spell-out of clitic matrices in a post-lexical level. Yet another question concerns the interaction of gender, number and
Parametric
Syntax - a Model
27
person features and inflectional rules for the selection of particular agreement values for a particular noun. Since we assume these features to be inherent if they are associated with a substantive noun, it is clear that we cannot assume that they are changed, or even selected by an inflectional rule in a post-lexical stage of the derivation. Such inflectional rules would be barred by the Projection Principle from altering agreement values (i.e., changing a masculine noun to a feminine one, a singular to a plural). We would thus assume that all agreement-determining rules for a particular nominal paradigm must apply at a pre-syntactic level. Note that this need not be stipulated: post-lexical application would be ruled out by the Projection Principle. Such a pre-syntactic application for nominal paradigms does not imply that agreement between verbs and nouns must apply at the pre-syntactic level. While the selection of agreement values for a noun involves the choice or perhaps the change of a lexical feature, the noun-verb agreement involves the transfer of an existing feature.
3. PARAMETRIC SYNTAX
The remainder of this study is devoted to showing that a very wide range of parametric variation can be accounted for by using the inflectional system sketched above. Languages will differ in the availability of particular inflectional rules and in the application of the restriction in (30). These differences, while interacting with general principles of UG, will give rise to different grammatical systems. As a simple example of how such differences may be generated, consider the following cases: (34)
a. hkit talked-I Ί talked b. hkit talked-I c. hkit talked-I Ί talked
ma9 Karim with Karim with Karim' ma9-o with-him ma9-o la Karim with-him to Karim with Karim' (Lebanese Arabic, Aoun, 1982)
(35)
a. dibarti talked-I b. dibarti talked-I c.*dibarti talked-I
'im Anna with Anna 'im-a with-her 'im-a (1 e/sel) Anna with-her to/of Anna (Modern Hebrew)
28
Parametric
Syntax
(36)
a. for John to win would be a disaster b. it is convenient for the rich for the poor to do the hard work
First, compare (34) and (35). Again, given our assumptions about the nature of clitics, the ungrammatically of (35c) follows from a principle of UG: since the Case feature of the preposition 'im is absorbed by the clitic, the NP complement Anna cannot receive Case and the sentence violates the Case filter. In the equivalent Arabic sentences, however, a 'saving device' is available: the preposition la which is inserted preceding Karim. Although the Case features of the preposition are absorbed, the sentence is grammatical due to the insertion of la. The insertion of a Case marker preceding an object of a preposition in Hebrew is impossible. Thus we assume that the grammar of Lebanese Arabic contains the inflectional rule stated roughly in (37), but that the grammar of Hebrew does not contain a similar rule: 18 (37)
0
"la
I [γρ....
NP]
Thus the presence vs. absence of an inflectional rule results in the availability vs. non-availability of doubling in prepositional phrases in the two respective grammars. Now consider the more familiar case in (36). Here an infinitival subject is licensed directly by the availability of a Case marker: again, an inflectional rule present in the grammar of English, resulting in a construction which is missing from the grammar of other languages. As a more intricate example, consider the preposition stranding phenomenon as analyzed primarily in Kayne (1981a) (but see also Hornstein and Weinberg, 1981). Kayne argues that the crucial fact which licenses preposition stranding in English but not in French is the fact that in English verbs and prepositions assign Case similarly, and hence they govern similarly. This fact allows the verb to govern the [ΝΡ,ΡΡ] position, thus avoiding ECP violations. Now consider a slight variation of this analysis, utilizing a thematic restructuring similar to that proposed in Rouveret and Vergnaud (1980). In effect, consider the possibility that the reanalysis of the sequence V-P in English involves a transfer of the Case assignment features and the Θ-role assignment features (if such exist) of the preposition to the verb, bringing about the effective merger of these two lexical entries. This operation will not be a violation of the Projection Principle as interpreted in this study as the set of lexical features will remain the same. However, the relationship between the element which assigns the feature and the constituent to which the feature was assigned will have changed. If we assume that this merger
Parametric Syntax - a Model
29
between the verb and the preposition effectively allows the verb to govern the [NP.PP] position, the possibility of prepositions stranding in English will follow. The absence of preposition stranding in French is now reduced to the absence of an inflectional rule, redefining the Θ-role and Case assignment relations in the configuration V-P. In Chapter 5 we will propose a detailed analysis of reanalysis in causative constructions in Romance, elaborating on the feature merger proposed above. Small differences in the application of inflectional rules account for a range of reanalysis phenomena in two dialects of French and in River Plate Spanish. It is worth concluding this chapter by reiterating the conceptual advantage of a system that reduces all interlanguage variation to the properties of the inflectional system. The inventory of inflectional rules and of grammatical formatives in any given language is idiosyncratic and learned on the basis of input data. If all interlanguage variation is attributable to that system, the burden of learning is placed exactly on that component of grammar for which there is strong evidence of learning: the vocabulary and its idiosyncratic properties. We no longer have to assume that the data to which the child is exposed bear directly on universal principles, nor do we have to assume that the child actively selects between competing grammatical systems. Rather, just by learning the inflectional rules operating in her/his environment, the possibilities offered by UG are narrowed down so as to give rise to Core Grammar.
NOTES 1. The name "Pisa Lectures" refers lo a manuscript of the original lectures on government and binding given by N. Chomsky at the GLOW conference, Pisa, April 1979. This manuscript was prepared by J.-Y. Pollock and H. Obenauer. The lectures were then expanded in a book, referred to in this study as Chomsky (1981). We refer to the "Pisa Lectures" only when we discuss matters whose treatment differs in the earlier manuscript from their treatment in the more recent book. 2. As is clear from the model in (1), the terms PF and LF denote levels of representation. However, these terms are often used in the literature to refer also to the set of rules which map S-structure representations onto LF and PF, respectively. In this study the terms "LF" and "PF" as well as the "LF component" and the "PF component" are often used in this ambiguous way. 3. The definition of government in (7) is a development of an idea of A o u n and Sportiche (to appear). The intuition behind their definition is that heads govern everything in their maximal projection. In (7) this intuition is expanded to allow government of adjoined phrases as well. For more extensive discussion, see Chapter 2. 4. The definition of governing category in (10) is a tentative formulation, later replaced in Chomsky ( 1 9 8 1 ) by a definition of a "binding category", in which the government requirement is derived from other factors. For our purposes, however,
30
Parametric
Syntax
the definition in (10) suffices. Similarly, we do not discuss in this study the motivation for the notion accessible SUBJECT in (9) as this issue is by and large irrelevant to our investigation. For an extensive discussion of these topics see Aoun (1982). 5. In section 2 of Chapter 5 we will argue that in the unmarked case, verbs are not specified as assigning a particular Case. Rather, they have a certain number of Case slots which are linked with argument positions. In turn, the Case slot adjacent to the verb is interpreted as accusative and non-adjacent Case slots are interpreted as dative. The distinction between assuming that verbs assign a particular Case (here, accusative) and the system we propose below is not important for the discussion in this section. 6. For an extensive discussion of this proposal see below, as well as Aoun (1979a) and Jaeggli (1982). The proposal that the insertion of a preposition licenses cliticdoubling is from R. Kayne. 7. We will crucially assume that nominative Case assignment is not an inflectional rule of the type described here. For some discussion see Chapter 6, sections 3 and 4. 8. We know of only one case in which this assumption may be problematic: the case of parasitic gaps, as in (i): (i)
this is the paper whichj I filed [e]j without reading [e] j
If we assume, as in Chomsky (1982), that the first [e]j is the extraction site and the second [e]j is a base-generated empty category, interpreted as a variable bound by whichj, then one may question the Case status of the base-generated [e]j. (Note that no such question arises with respect to the extraction site. It could be Case-marked at D-structure while it is phonologically realized). Although this question will not be pursued in this study, it will become evident in Chapter 3 that in our system (unlike that in Chomsky 1982) there is no compelling reason to assume that the second [e]j must be Case-marked. Thus we could assume that the accusative Case feature of reading in (i) remains unassigned. 9. In section 1 of Chapter 2 we will make the mechanism by which ©-role isassigned more specific, adopting a proposal of Stowell (1981) that Θ-role assignment involves the transmission of a referential index from an argument to the verb. A Θ-role assignment to the preposition by by the verb (or more concretely, by the passive morpheme) could be achieved in several ways. For instance, the argument (the [ΝΡ,ΡΡ]) could transmit its referential index to a 'place holding' slot in the lexical entry of by, which would, in turn, transmit it to the verb. The direct formulation of this proposal will not be pursued in this work. Transmission of Θ-roles from an assigner to an assigner will be used here primarily to explain reanalysis in causative constructions in Romance. We will simply assume that the Θ-role slots of two assigners may merge. See also below, section 3, for an analysis of preposition stranding along these lines. 10. See in particular Aronoff (1976), for examples of morphological rules which utilize these different categories. As one illustrative example consider the suffix -ee, which only attaches to transitive verbs which take human objects. Hence employee, payee, but *sleepee, *eatee. This suffix makes use of subcategorization information as well as selectional restrictions. See also Williams (.1973) for restrictions on the prefixation of (level 2) re-. 11. While the rule in (23) bears some resemblances to Emonds' (1976) local rule, it should be kept in mind that (23) is a rule of morphology and not a syntactic rule. 12. It is worth pointing out some differences between the diagram in (24) and the
Parametric
Syntax
- a
Model
31
diagram in (25). In particular, note that in (24), the arrow leading from level m phonology to level m morphology is bidirectional. This bidirectionality is intended to capture the cyclic nature of phonological rules level internally. Thus some phonological rules will re-apply after every affixation. In our syntactic model, however, this aspect is missing. Rather, we assume that D-structure morphology applies prior to D-structure syntax, S-structure morphology applies prior to S-structure and PF morphology applies prior to PF. In effect, we assume that well-formedness conditions which are applicable at these different levels (in particular, the Projection Principle and the Case filter) apply to the output of morphological rules at that level. The reasons for the differences between the two diagrams will not be pursued here, but it is plausible to assume that they should be attributed to the different properties of the phonological and syntactic components respectively. 13. The constructions in (27) are so-called 'exceptional Case-marking' constructions. For a discussion of their properties, see Chomsky (1980), where it is suggested that the right way to capture the property of verbs like believe and cxpect, which allow accusative Case assignment to a subordinate subject, is to assume that they take a nonmaximal projection as their complement. This non-maximal projection permits the government of the subordinate subject by the matrix verb and hence Case assignment is possible. It will become clear below that this proposal does not fall within the restricted class of parameters argued for in this study. Kayne (1981a) argues that the effects of exceptional Case marking are achieved by the presence of a 0-complementizer which assigns accusative Case. This account of exceptional Case marking in English does fall within the restricted class of parameters suggested here. 14. This account is neutral with respect to the question of whether Case requirements are met by Case-marking the WH elements or by Case-marking the variable left behind (see Borer, 1981, for discussion). If one adopts the requirement that variables must have Case (as a general principle of grammar or as a consequence of the Visibility Hypothesis, cf. Chapter 3 ), it is still clear that if w/ioj in (27b) is not assigned Case in the base, there is no way to have a Case-marked variable in the subject position, if empty elements cannot be Case-marked directly. If, on the other hand, accusative Case is assigned to who at D-structure, one could assume that after the fronting of the WH element its trace retains a copy of the Case assigned to it by the verb prior to the application of 'Move a \ Since traces retain indices and a set of ^-features (^-features = gender, number and person features. See Chapter 4 for a discussion), it is not implausible to assume that they retain Case marking as well. 15. As an illustration, consider the nasal assimilation and the stress retraction attested when the negative prefix in- is attached to an adjective (illogical, impotent). No such phenomenon is attested with the negative prefix un-. Allan (1978) argues that the former is a level 1 affix whereas the latter is a level 2 affix. On the other hand, stress assignment rules apply at both levels, albeit somewhat differently. 16. Or conversely, as a positive statement, specifying that a rule R must apply in a particular level. While this positive formulation seems more appealing for derivational affixes, which typically attach only at one level, the statement in the text seems more appealing for inflectional rules, which typically apply at more than one level. Although it would be desirable to account for this difference in a principled fashion and to unite the statement of the restriction, this will not be pursued in this study, which concentrates on syntactic phenomena. It is worth pointing out that
32
Parametric Syntax
some derivational affixes attach at both levels, e.g. the prefix re·, as in return (level 1) and redo (level 2). Most studies in derivational morphology assume that these are two different prefixes with different properties. Another possibility, however, would be to claim that it is one affix and that the phonological and semantic distinctions between level 1 re- and level 2 re- are entirely derived from the distinct properties of level 1 morphology and level 2 morphology. If this is indeed the case, then derivational morphology follows the same pattern described here for inflectional morphology: free application which interacts with universal principles to bring about particular grammatical phenomena. 17. The transliteration of Hebrew used in this study seeks to characterize spoken Hebrew. Thus some distinctions which are preserved in orthography (and perhaps preserved in the underlying forms as well) are eliminated in our representation. The transliteration is not intended as a phonological characterization of underlying segments. 18. Interestingly, the rule in (37) is a general rule in Lebanese Arabic, which inserts the preposition la preceding [NP,VP], [ΝΡ,ΡΡ], and [ΝΡ,ΝΡ], The preposition la, however, is never inserted preceding a nominal element in the subject position. For a discussion, see Aoun (1982).
Chapter 2
Clitic Government
1. INTRODUCTION
In recent years, the study of clitic phenomena in view of the clitic-doubling construction has enjoyed a lot of attention within the Extended Standard Theory tradition (to mention only a few: Strozer, 1976; Rivas, 1977; Aoun, 1979a; Jaeggli, 1982; Steriade, 1980; Borer, 1981a; and others). In this study, we will suggest yet another analysis of clitics, inspired by doubling phenomena as they appear both in the Romance languages (River Plate Spanish, Romanian) and in Modern Hebrew. The investigation of clitic-doubling will motivate a theory of clitics that will then be extended to explain clitic phenomena which are not directly related to doubling in Modern Hebrew, Spanish and French. A sample of clitic configurations of the kind we will be discussing in this study is given in (la-f) below. In (la-c) we have structures in which the clitic alone seems to satisfy the subcategorization or complementation requirements of a lexical head, (ld-f) present constructions known as clitic-doubling. In these constructions, we find a clitic alongside an NP, both of them satisfying the complementation requirements of the head and understood to corefer. (This coreference is marked henceforth by identical indexing.) 1 (1)
a. lo vimos him saw-we \ve saw him' (River Plate Spanish; Jaeggli, 1982) b. 1 -am väzut him-have-I seen Ί have seen him' (Romanian; Steriade, 1980) c. beit-o 'omed 'al ha-giv'a house-his stands on the-hill 'his house stands on the hill' (Modern Hebrew)
34
Parametric
Syntax
d. loj vimos a Juarij him saw-we to Juan Sve saw Juan' e. L-am väzut pe Popescu· him-have-I seen OM Popescu (OM = Object Marker) Ί have seen Popescu' f. beit-Oj sei ha-more^ 'omed 'al ha-giv'a house-his of the-teacher stands on the-hill' 'the teacher's house stands on the hill' A major shift in the analyses of clitics, which resulted from the consideration of clitic-doubling constructions, has been the abandonment of the movement analysis of clitics (as suggested, in particular, in the seminal study of Kayne, 1969, 1975; see also Quicoli, 1980 and others). Advocates of the movement analysis would argue that the clitic in sentences (la-c) is a pronominal element, base-generated in the regular object position and then moved to a position adjacent to the head (the verb in (la-b) and the head noun in (lc)). However, as pointed out by Strozer (1976), Rivas (1977) and Jaeggli (1982), a movement analysis of this sort simply cannot account straightforwardly for clitic doubling (and, as pointed out by Jaeggli, 1982, this analysis was in fact constructed to account for the complementary distribution of clitics and complement NP's in French, where the sentences corresponding to (ld-f) are ungrammatical). Thus the clitic doubling construction (primarily as it appears in River Plate Spanish) motivates a base-generation analysis for clitics. (For a more complete review of the movement analysis see Jaeggli, 1982). However, the movement analysis has one elegant result which base-generation analyses cannot achieve quite as easily. Since the clitic in the movement analysis is considered to have originated in the argument position, the fact that it satisfies the subcategorization frame of the head and is assigned a Θ-role by it is captured naturally. Furthermore, the coreferentiality (coindexing) between this clitic and the argument position follows clearly from a movement analysis, but not from a base-generated one. Let us review the base-generated structure that was suggested for the clitic doubling configurations in River Plate Spanish both by Rivas (1977) and by Jaeggli (1982). It is roughly as in (2):
(2)
V NP;
V CLj
V
(See Rivas, p. 34; Jaeggli, p. 98, fn. 10.)
Clitic
35
Government
Jaeggli (1982) argues that in (2) the clitic does not c-command the coindexed NP. (We will return to the motivation for this proposal in section 3, and in Chapter 5, section 2.) The lack of a c-command relationship or any other independently motivated structural relationship between the clitic and the coindexed NP results in the need for a special rule of coindexation and Θ-role transmission which is not structuredependent (see Jaeggli, p. 66 for the latter). 2 The movement analysis does not confront this problem. Independent considerations require that an antecedent c-command its trace. Thus proponents of a movement analysis would either have to reject the structure in (2) as the output of a cliticization rule, or, alternatively, they would have to alter the definition of c-command so as to allow the clitic in (2) to c-command the coindexed NP position. Note that even if the definition of c-command is extended to cover the relationship between the clitic and the coindexed NP in (37), this structure still gives rise to some serious questions: what is the relationship between the clitic and the head V? Is the clitic in an argument position (A-position)? Does it enter into the binding conditions? Note that if we were to assume that the clitic occupies an argument position, all the clitic-doubling configurations would be problematic from the perspective of the Projection Principle. We would have a situation in which a verb selects two argument positions and assigns the same Θ-role to both of them. In this study we will advocate an analysis of clitics in which the clitic c-commands the coindexed NP. Furthermore, it will be shown that clitics are best characterized as part of the head constituent. 3 In this we will follow Kayne's suggestions that the derived structure of clitic configurations is as in (3). We will differ from Kayne in assuming that the relevant structure is base-generated. 4
(3)
X [χ di. [χχ'
cl
χ
])
NP;
il ) lexical NP S
Some things should be clarified with respect to (3). First we will take X in (3) to stand for V, P, or N. Both the relevant data and the theoretical motivation will be discussed below. Secondly, note that the clitic in (3) governs (and c-commands) the coindexed NP position. This follows from the fact that it is part of the head. Furthermore, the expansion of (3) in which NPj dominates an empty category (and which corresponds to sentences (la-c) above) is in many respects similar to the output of movement rules: we have an antecedent coindexed with an empty category which it c-commands. The following sections will be devoted to the clarification
36
Parametric
Syntax
and the justification of the structure in (3). We will investigate the nature of the [χΰ1|,Χ] combination, and the relationship between the clitic and the coindexed NP, and we will elaborate the empirical consequences of our proposal. 1.1. Case Absorption R. Kayne has observed that constructions such as (ld-f) above - cliticdoubling constructions - can only occur if the NP which is doubled is preceded by a preposition. This generalization (which Jaeggli calls "Kayne's Generalization") is accounted for by Chomsky (the Pisa Lectures), Aoun (1979a) and Jaeggli (1982) by assuming that in clitic-doubling constructions the clitic, in a sense to be made precise, absorbs the Case features of the head (the verb in (ld-c), the noun in (If)). Following the essentials of their proposals, the structure of clitic-doubling configurations is roughly as in (4): X
(4)
Case absorption (Note that (4) is neutral with respect to the status of the X + clj combination. The status of this combination, assumed earlier to be the relationship between a head and a feature, is not directly relevant here.) It is argued that in (4) the clitic absorbs the Case features of the category X (or is itself the spell-out of Case features). Note that if we assume the Case filter, it follows that no lexical material can appear in NPj unless an independent device is found which can assign Case to it, since the Case features of X are absorbed by the clitic. Just such a Case-assigning device is the dummy Case marker, which can be seen in examples (ld-f): in River Plate Spanish it is the preposition a, in Romanian it is the object marker pe, and in Modern Hebrew it is the genitive preposition sei. Indeed, the absence of these dummy markers leads to ungrammatically: (5)
a. *lojVimos Juanj 'we saw Juan' b. *lj-am väzut PopescUj 'I have seen Popescu' c. *beit-Oj ha-more^ 'omed 'al ha-giv'a t h e teacher's house stands on the hill'
(River Plate Spanish) (Romanian)
(Modern Hebrew)
Clitic
37
Government
We will adopt here the essentials of this intuition. s We will assume that, in some sense, the clitic "deprives" the coindexed NP of its Case. In effect, we will assume that the clitic is a spell-out of the Case features of the head, and, as such, truly a feature of the head. The rule of clitic spell-out is given in (6): (6)
Clitic Spell-Out [ χ Χ, α Case] —> [ χ Χ [ α Case,β gender, γ number, δ person]] Χ = [+V ] in Romance 6 X = V, Ρ, Ν in Semitic
The rule of Clitic Spell-Out is an inflectional rule (see Chapter 1 for discussion). In certain configurations the features number, gender and person are inserted and combined with the already present Case features. Then they are given a specific phonological representation. As an inflectional rule, (6) may apply at any stage of the derivation; in Chapter 4 we will discuss some cases in which (6) cannot apply in the base, but rather must apply at S-structure or in the PF component. Since we perceive of the clitic as a spell-out of features, we do not expect it to satisfy subcategorization or complementation requirements. Rather, it is the complement NP node in (1 a-f) which is generated by the base rules in the usual way and which is assigned Θ-role in the usual way. The relationship between this node and its selecting head is thus the usual relationship between a selected complement and its head (more on this in subsection 1.2 below). 1.2. The Complement Matching
Requirement
Let us now turn to the nature of the coindexing in structures such as (3). Complementation requirements are met within the government-domain of the lexical head which selects such complements. It follows from the X system that every head has to govern its complement. Although this state of affairs is derived from other principles of the grammar, we would like to state it explicitly. The methodological value of an explicit statement will become clear below, where we discuss structures in which an argument cannot satisfy complementation requirements because it is not governed by the complement-selecting head. Let us then state this structural observation as the government requirement in (7): (7)
A head must govern its complements.
In defining the notions head and complement, we will crucially rely on the
38
Parametric
Syntax
X system coupled with the assignment of Θ-roles. Thus a head is and a complement is an argument that is assigned a Θ-role by X° and which meets (7). 7 In defining complements as those arguments which bear a thematic link to the head, we seek to distinguish between elements which are selected by the head and assigned a Θ-role by it, and arguments which may be complements of the head in a broader sense, but nevertheless are not assigned a Θ-role by it. Following this distinction, we assume that the PP in (8a) is a complement of the verb dedicate, and we will further assume that Mary is assigned the Θ-role of goal by this verb. On the other hand, in (8b) Paris is related semantically to from, and the PP is not a complement of the verb in the sense implied above. 8 (8)
a. John dedicated his dissertation to Mary b. John drove from Paris
(In this study, the term complement, when used without further elaboration, refers to a complement in the strict sense defined above. In referring to complements such as the PP in (8b), which are not linked thematically to the head, we will distinguish between thematic complements and nonthematic complements. This distinction is particularly relevant in the discussion of causative constructions in River Plate Spanish). Returning now to the structure in (3), recall that we assume that the clitic is a feature on the head. As such it governs the doubled NPj. Furthermore, as part of the head it takes the doubled NPj as its complement. However, we must allow the complex [^clj,X] in (3) to contain two sets of features. The first set of features is associated with the head itself. It contains the "inherent" features of the head such as Θ-role assignment features and categorial selection. It further contains the "non-inherent" features of the head, i.e. the verbal inflection in the case where X = V and the number, gender and person features of the head when X = N. The second set are the features of the nominal element attached to the head, i.e. the clitic. The clitic is crucially a spell-out of "non-inherent" features only, as is clear from rule (6) above. We will further assume that like all nominal elements, the clitic contains a referential index. It is particularly important to separate these two sets of features in the case of a noun head and a clitic. In these cases the noun will have its own set of "non-inherent" features and its own referential index, both distinct from those of the clitic. Note, however, that the affixation of the clitic will leave the "inherent" features intact. As such it will display a behavior which is typical of inflectional affixes but which differs from that of derivational affixes. There is a legitimate question with respect to the location of the matrix of features of the clitic in the noun. As an element which contains a referential index, the clitic must be associated with a Θ-role. However,
Clitic
39
Government
since it is an inflectional affix which does not interfere with "inherent" features it cannot be linked to a Θ-role which is not otherwise assigned by the head. It is therefore natural to assume that the clitic is linked with a thematic matrix of the head. Stowell (1981) suggests that the assignment of Θ-roles to complements by a head can be captured if we assume that a complement transfers a referential index to an available thematic slot in the head. Informally speaking, this proposal implies that every head contains as many empty slots as Θ-roles which it assigns. These empty slots have to be filled by referential indices transferred from the complement. If the selected complement is not generated, or if it does not have the right Θ-role, the empty slot cannot be filled and the derivation is ruled out. Given that the clitic, as formed by the rule in (6), must be linked to one of the thematic slots available in the head, we propose that the structure of the [^clj,X] combination in (3) is as in (9):
(9)
[χ X,
Θι Δ clΉ
The symbol in (9) stands for the particular Θ-role assigned by X. The empty space indicated by Δ is the space into which the index of the complement has to fit in accordance with our assumptions about the assignment of Θ-roles. The clitic is attached to that position as an additional element rather than as an element which fills the referential empty slot. Since the clitic is not an argument, it is not a full NP, nor can it be seen as satisfying complementation requirements. Rather, the complement still has to transfer its index. Now consider a situation in which the complement of X contains an index j and j f i. Fitting the index j into the empty slot in (9) will result in conflicting indices being associated with one thematic slot. Consequently the derivation will be ruled out. We can now view the obligatory coindexing in (3) as a condition on Θ-role assignment: if the complement and the clitic do not agree in index we would have the thematic matrix in (10) which contains conflicting indices, and which is ruled out: (10)
* ΓΘ,'
40
Parametric
Syntax
Some heads select more than one complement, and can assign more than one Θ-role. In this case the complement need not agree with the clitic. Rather, it can agree with the other thematic slot. This situation is illustrated by (11): (11)
Paullaj luij presenters a Juanj Paul her him introduced to Juan 'Paul introduced her to Juan'
The thematic matrices of pwsenter in (11) are as in (12): ΘΓ
02" j
-cli-
LcljJ
In (12), the index of Juanj fills the referential slot in the thematic matrix of θ 2 (Θ 2 -goal) which is in tum associated with the clitic luiy The clitic /flj, on the other hand, is associated with the thematic matrix of 0 ! ( 0 ! = theme). Note that if (11) is the S-structure representation for that sentence, there is no source for a referential index for 0 ! in (12). We will argue, however, that the correct S-structure representation of (11) is as in (13): (13)
Paul laj luij presentera [e]j a Juanj
The referential index i is supplied by the empty category. We will return to the motivation of the empty category in constructions such as (13) below, in section 3.1. To make our claims concrete, let us define the Complement Matching Requirement as in (14): (14)
given a thematic matrix T, *T if Τ contains referential indices i, j, and i f j
It is worth mentioning that the CMR is not an independent stipulation and that it follows in a natural way from the Projection Principle. As in the case of the Government Requirement for complements in (7), it is stated explicitly for expository purposes. Note further that CMR renders a coindexation rule between the clitic and the object NP unnecessary: the obligatory character of that coindexation now follows from independent principles. In Chapter 5, section 3.1 we will return at great length to the CMR, elaborating its explanatory value for the inalienable possession constructions in Romance.
Clitic
Government
41
We will conclude this introductory note by summarizing our claims with respect to the structure of clitic configurations. We assume clitics to be affixes, the output of an inflectional rule which inserts number, gender and person features and associates them with an already existing Case feature on a lexical head. Clitics are a spell-out of Case features in the sense that once the Case feature is associated with the inserted number, gender and person features, it is given an independent phonological representation and can no longer be transferred to a complement of the head. The clitic, a nominal element, has its own referential index. Since it is part of the lexical head, this referential index (and the clitic associated with it) govern the complement NP. The clitic and the NP complement are obligatorily coindexed with each other, and they agree in gender, number and person features. Rather than assume a special coindexation rule, we claim that this coindexing follows directly from properties of Θ-role assignment. If the clitic and the NP are not coindexed the Projection Principle is violated. For concreteness we defined that aspect of the Projection Principle which is relevant for this result as the Complement Matching Requirement (CMR). The remainder of this chapter is devoted to motivating the analysis of clitic configurations proposed above, as well as to proving its central claims on the basis of empirical evidence from Modern Hebrew. We will center on the following issues: clitics govern coindexed NP complements (section 2); clitics do not occupy an Α-position (section 2); when the complement NP node (NPj in (3) above) is not lexically filled, it is expanded as [e] (and not, say, as PRO) (section 3); clitics can function as proper governors for the purposes of the Empty Category Principle (section 4).
2. THE CONSTRUCT STATE AND CLITIC GOVERNMENT
Clitic constructions in Semitic languages have not been widely researched within the framework of the Extended Standard Theory. In the following sections we will present a detailed analysis of clitic configurations in Modern Hebrew. In Chapters 4 and 5 below it will be shown that, given a few parameters, this analysis can be extended to account for clitic configurations in the Romance languages as well. Although the clitics in Hebrew behave differently from clitics in the Romance languages, there are nevertheless striking similarities. In particular, we will show that the clitics themselves are the same: in all cases they have the properties outlined in section 1 of this chapter. 2.1. The Construct State: General Properties The Construct State in Modern Hebrew indicates genitival relations between
42
Parametric
Syntax
the head noun and the complement noun. The phrase in ( 1 5 ) has roughly the structure in ( 1 6 ) : 9 (15)
beit ha-mora house the-teacher(fem) 'the teacher's house'
(16)
Nt Nj
beit
ha-mora
house
the-teacher
The structure in ( 1 6 ) yields itself to further embedding: (17)
delet beit ha-mora door house the-teacher 'the door o f the house o f the teacher'
Sentence ( 1 7 ) has the structure shown in ( 1 8 ) :
(18)
N, N,
beit
ha-mora
house
the-teacher
Ever further embedding is possible, as in (19a-b): (19)
a. yadit
delet beit
ha-mora
handle door house the-teacher 'the handle o f the door o f the house o f the teacher'
Clitic
Government
43
b. ceva yadit delet beit ha-mora color handle door house the-teacher 'the color of the handle of the door of the house of the teacher' Note that all these structures are right branching. Thus they are given a specific bracketing; for example, the head of a complex such as (19b) is ceva 'color', and its complement is 'the handle of the door of the house of the teacher'. The head of the complement is yadit 'handle', and its complement is 'the door of the house of the teacher'. This is the only way to form construct states. This requirement for right branching is captured in our diagrams by generating the complement NP under the Ν rather than under the Ν node. We will return to the motivation for this structure. An interesting property of the construct state follows from the requirement of right branching. Note that if the structures in (16) and (18) are correct the head cannot be directly modified. Any modification of the head alone (as opposed to a modification of N) would result in right branching. For adjectival modifiers this situation is exemplified by the ungrammatically of (20): (20)
*ceva color
yadit yafa ha-delet handle beautiful the-door
(20) would have the structure in (21), which contains right branching. Consequently the sentence is ungrammatical:
(21)
yadit handle In order to specify that the color of the handle is beautiful, yafe, the adjective has to appear at the end of the complex: (22)
ceva yadit delet beit ha-mora ha-yafe color handle door house the-teacher the-beautiful 'the beautiful color of the handle of the door of the house of the teacher'
44
Parametric
Syntax
In (22) the modifying adjective can be construed as belonging to any level of bracketing in a multiply embedded structure; thus (22) is actually ambiguous. The adjective yafe could refer to any noun in the complex which agrees with it in gender and number. In this case it is masculine singular, so it could refer to the color or to the house itself, both being masculine singular. (Hebrew does not have a neuter gender.) These two interpretations would have bracketings (23) and (24), respectively: (23)
[ceva [yadit [delet [beit [ha-mora]]]] color handle door house the teacher 'the beautiful color of the handle etc.'
ha-yafe] the-beautiful
(24)
[ceva [yadit [delet [beit [ha-mora] ha-yafe]]]] color handle door house the-teacher the-beautiful 'the color of the handle of the door of the beautiful house etc.'
Similarly, if we used the feminine counterpart of yafe, yafa, it could be construed with yadit 'handle'; delet 'door'; or ha-mora 'the teacher', all being feminine singular. A similar restriction holds for determiners. Only a non-head constituent in structures such as (17) can be accompanied by a determiner. This means that only the last NP in a chain of construct nouns can be definite. This situation is exemplified by the contrast between (25) and (26): (25)
(26)
ceva yadit ha-delet color handle the-door 'the color of the handle of the door' *ceva color
ha-yadit ha-delet the-handle the-door
The ungrammatical sentence (26) would have the ill-formed structure in (27):
Clitic
Government
(27)
*
45 Ν Ν
Ν
Ν Ν
ceva color
Ν
DET
Ν
hathe
yadit handle
ha-delet the-door
Since in structures such as (25) the determiner can only appear attached to the last constituent, the sentence is vague with respect to the identity of the definite element. While in these constructions the head is always construed as definite (or, in other words, the full NP is always definite), the complement itself may be understood as indefinite, although it is directly attached to the determiner. Thus the phrase in (28a) has the interpretation of both (28b) and (28c), depending on whether melex, 'king' is construed as definite or not: (28)
a. ben son b. 'the c. 'the
ha-melex the-king prince', 'the son of a king' son of the king'
If, on the other hand, no determiner appears at all, as in (29), both the head and the complement NP are construed as indefinite: (29)
a. ben melex son king 'a king's son', 'a prince'
As demonstrated by (20) and (26) above, any attempt to break the succession of bare nouns in a phrase such as (19b) with a modifier or a determiner will yield ungrammatically. Alternatively, it can bring about the "closure" of the construct state: any further genitival relationship will have to be expressed in a different way: by using the genitival proposition sei: (30)
*ceva color
ha-yadit ha-yafe delet beit the-handle the-beautiful door house
ha-mora the-teacher
46 (31)
Parametric
Syntax
ceva ha-yadit ha-yafe sei delet beit ha-mora color the-handlethe-beautifulo/ door house the-teacher 'the beautiful color of the handle of the door of the house of the teacher
.10 (31), presumably, has roughly the structure in (32):
ha-yafe the-beautiful
ceva ha-yadit color the-handle N3 beit house
ha-mora the-teacher
2.1.1. A Note on Genitive Case Assignment What is the structural relationship between the head and the complement in structures such as (16)? In particular, could we assume that the head is assigning Case (presumably, genitive) to the complement? It has been argued (first, to our knowledge, in Emonds, 1970) that the rule which assigns genitive Case is a structural rule. Thus it is claimed that in a configuration such as (33), genitive Case is assigned to NP2 : (33)
a.
NP2 [ N i
]]
b. John's writing In (33) it is desirable to claim that the head noun does not assign Case to the possessor John, since John is the specifier of writing. If we wish to restrict Case assignment by heads to their complements alone, it is plausible to assume that in (33) genitive Case is not assigned to the possessor by the head noun. Alongside (33) we have (34), in which book is the complement of the
Clitic
Government
47
head noun. However, in this instance of insertion is necessary in order to assign Case to the complement: (34)
the writing of the book
Thus (34) seems to provide additional evidence that nouns in English do not assign genitive Case, even if they can be argued to take complements. In the Semitic languages, however, the situation is not as clear. Note that the distinction between complements and specifiers in the construct state is not reflected structurally. Thus (35a-b) have an identical structure: (35)
a. ktivat Dan writing Dan 'Dan's writing' b. kativat ha-ma'amar writing the-article 'the writing of the article'
In Standard Arabic, where Case marking is morphologically overt, the complement of the construct state is marked as genitive. This is true for derived nominals as well as regular nouns. Again, the distinction between complement and specifier is obscured: (36)
a. maktabu muhammadin office Muhamad (gen) 'Muhamad's office' b. kitaabatu l'kitaabi writing the-article (gen) 'The writing of the article' c. kitaabatu muhammadin writing Muhamad (gen) 'Muhamad's writing'
As observed above there is a strict condition on construct states. Any expansion which results in right branching leads to ungrammatically, or, alternatively, to the appearance of the genitive preposition sei. This state of affairs is reminiscent of adjacency conditions on Case assignment (see Chomsky, 1980; Stowell, 1981 and references cited there). When some local restriction is obeyed, genitive Case can be assigned directly by the head noun. However, when such locality is not present the sentence is ungrammatical, unless a preposition is present to assign Case to the complement. What is the locality condition for the assignment of genitive Case?
48
Parametric
Syntax
Linear adjacency alone clearly does not suffice. Even government combined with linear adjacency is not enough, as is demonstrated by the ungrammatically of (26) above: although ha-yadit, 'the handle', is adjacent to ha-delet, 'the door' and although it governs it (for the relevant discussion of government see section 2.4 below), the phrase is ill-formed. The correct version is either as in (25), where 'handle' is indefinite, or, alternatively, the phrase can be rendered grammatical if sei is inserted preceding 'the door', as in (37): (37)
ceva ha-yadit sei ha-delet color the-handle of the-door
Let us assume that genitive Case in Semitic can be assigned by a head noun only if the first node which dominates that head (N) immediately dominates the complement which is assigned genitive Case. Thus genitive Case assignment can only occur inside N. The nature of our claim is as follows: we assume that genitive Case can be potentially assigned by all nouns in Semitic. However, this potential must be "activated" by a certain structural environment. When this environment is present genitive Case may be assigned. When it is not present, the genitive Case features cannot be assigned and a dummy Case marker, sei, is required. The genitive Case features, however, even when not assigned remain a potential property of all nouns in Semitic. Note that we are now equipped with an explanation for the right branching constraint: the structures (27) and (21) above are ungrammatical because the complement NP, ha-delet, cannot be assigned Case. In both cases, the presence of sei will render the structure well-formed. 11 2.2. Clitics and the Construct
State
As in other Semitic Languages, nouns in Modern Hebrew take clitics. Following the discussion in the previous section we would like to claim that these clitics are a spell-out of the genitive Case features of the head noun in the sense of (6) above. Consequently we propose that the combination of noun+clitic as in (38) has the structure in (39): 1 2 (38)
beit-a house-her 'her house'
Clitic
49
Government
(39)
Ν
i
Ν Ν + clj
Nj
beit-a house-her
(9
(We shall return to the symbol 0 and to what it stands for in section 3.1.) The structure of (39) appears to allow clitic doubling, in the sense discussed above. Thus, parallel to (15) and identical to it in meaning we have (40): (40)
beit-a^ sei ha-moraj house-her of the-teacher 'the teacher's house'
Following our assumptions about the structure of clitic-doubling constructions, (40) has the structure in (41): (41)
Ν Ν N + cL
I
beit-a house-her
ha-moraj the-teacher
Recall that we are assuming that the clitic in (41) absorbs the genitive Case that otherwise would be assigned to the complement NP. Hence it_is necessary for sei to be inserted in order to assign Case to the coindexed Nj. Failure to insert sei would lead to ungrammaticality, which we have predicted using the Case filter: (42)
*beit-a| ha-mora^ house-herthe-teacher
Recall that earlier we argued that sei, the genitival preposition meaning roughly ' o f , appears in another environment in Modern Hebrew. When a construct state is "broken", sei is available to express genitival relations in a way syntactically different from that expressed by the construct state.
50
Parametric
Syntax
In fact the availability of sei in Modern Hebrew results in two alternative means for expressing genitival relations: by means of the construct state, as in (15) (repeated here as (43a)), or by means of the genitival preposition sei, as in (43b): (43)
a. beit ha-mora house the-teacher 'the teacher's house' b. ha-bay it sei ha-mora the-house of the-teacher
The structure of (43b) can be roughly illustrated as in (44) (and see also (32) above); (44)
l'i sei- phrase ha-bayit the-house ha-mora the-teacher
(Note that in (44) ha-bayit, 'the house', is construed as the head of the adjoined phrase as well. We shall return to this point below; see also fn. 10.) The nature of the sei phrase in (44) and in (32) is still open. Note that now we seem to have two sources for sei: one in structures such as (44) and (32), where the sei seems to head a phrase, perhaps a PP, and another in structures such as (41), for which we would like to argue that the sei is inserted for Case purposes and does not change the NP nature of the category which it is adjoined to. The structure in (41) shares an important property with the structure in (44): both behave as "broken" construct states in the sense that they do not have to be uniquely right-branching. Thus in (41) the complex noun + clitic can be modified directly by an adjective as in (45a) and in (44) 'the house' can be directly modified as well (as in (45b)): (45)
a. beit-aj ha-yafe sei ha-moraj house-her the-beautiful of the-teacher 'the beautiful house of the teacher' b. ha-bayit ha-yafe sei ha-mora the-house the-beautiful of the-teacher 'the beautiful house of the teacher'
Clitic
Government
51
Note that the intervening adjectival material in (45a) does not prevent the coindexing of 'her' and 'the teacher'. In fact this coindexing seems completely oblivious to any stacking of intervening adjectives: (46)
beit-aj ha-yafe ve-ha-meruxak min ha-'ir sei house-her the-beautiful and-the-far from the-city of ha-moraj the-teacher t h e teacher's house which is beautiful and which is far from the city'
The availability of left-branching for phrases such as (40) seems to present a problem for our proposal t h | t (41) is the right structure for these phrases. Given the requirement that Ν complements be generated under N, as in (18) above, it is not clear why adjectival material cannot appear between Ν and Ν in (18) but can appear between the N+cl and Ν in (41), if (41) is indeed the right structure. Furthermore, the need for two distinct sources for the preposition sei indicates that perhaps some generalization is being missed. One could argue on the basis of these problems that a more plausible structure for (40), and in general for clitic-doubling cases in Modern Hebrew, would be one closely resembling the structure in (32) and (44). Proponents of such an analysis would argue that the clitic on the noun in cases such as (40) is not a spell-out of a feature but rather a base-generated pronomial which appears in the regular argument position, as in (47):
I be it house
I a her
the-teacher
(Note that this is essentially the same as (32).) Since Modern Hebrew does not have clitic climbing or any other evidence that would indicate that the clitic is not in the original argument position, it would seem that there is no evidence that any syntactic operations apply to (47). The morphological process of adjoining the clitic to the head would be a non-syntactic phenomenon which is not related to the insertion of sei or to any other syntactic process.
52
Parametric
Syntax
There are, however, some important differences between (45a) and (31) which clearly indicate that the two phrases have to be somewhat different in structure. These differences highlight the fact that construct states in which the complement is a full lexical NP (as in (43a)) should be treated differently from those in which it is a pronominal clitic (as in (38)). The remainder of this section will deal with an investigation of these differences. It will be shown that the relationship which holds between the clitic and the NP of the sei phrase in sentences such as (45a) is entirely different from that which holds, say, between N s and N2 in (32). If, indeed, the structure of (45a) was as in (47), we would not expect this difference. This difference consists of obligatory coindexing of the clitic and the NP complement in (45a) which is impossible in (32) and which is stated in clear syntactic terms: it can only hold if the clitic governs the complement NP. Before turning to the matter of coindexing, which involves rather complicated data, a clear difference between clitic complements and lexical NP's should be pointed out: whereas a chain of construct states with lexical NP's can always be expanded, providing that the structure remains right-branching (cf. examples (19a-b) and structure (18)), the introduction of a clitic brings about the immediate "closure" of the construct state. Thus (48) in a reading that would correspond to (17) (where it refers to, say, house) is ungrammatical: (48)
*dalt-o ha-mora door-it the-teacher 'the door of it of the teacher'
The ungrammaticality of (48) vs. the grammaticality of (17) would follow immediately if we assume that the clitic absorbs the Case features, since in this case 'the teacher' in (48) would not be assigned Case and thus would violate the Case filter. On the other hand, if one assumed that clitics occupy the same position that full lexical NP's do, having the structure in (18) for (48), this fact cannot be readily explained. 13 Now let us turn to the coindexing argument. Consider the following phrase, which consists of the construct state along with a sei phrase: (49)
tmunot pictures
ha-yalda sei ha-mora the-girl of the-teacher (fem)
(49) can be construed with either of the following bracketings: (50)
a. [pictures [[the girl]] of the teacher]] 'the pictures of the teacher's daughter'
Clitic
53
Government b. [[pictures [the girl]] of the teacher] 'the girl's pictures of the teacher'
Now compare (49) with a phrase in which yalda, 'girl', has been replaced by a feminine clitic: (51)
tmunote-ha pictures-her
sei ha-mora of the-teacher (fern)
(51) cannot have the meaning of either (50a) or (50b): the clitic ha in (51) can only refer to the teacher, ha-mora. In other words, (52a-b) are not possible meanings of (51): (52)
a. herj pictures of the teacher j b. the teacher's^ pictures of herj
In fact, if the clitic is replaced by a masculine one (referring, say, to 'boy', yeled), the sentence results in ungrammatically, because the masculine clitic cannot be coindexed with ha-mora, 'the teacher', which is feminine: (53)
a. tmunot ha-yeled sei ha-mora pictures the-boy of the-teacher (fem) 'the pictures of the son of the teacher' t h e boy's pictures of the teacher' b *tmunot-av sei ha-mora pictures-his of the teacher
An interpretation of (51) in which the clitic is disjoint from the complement of the sei phrase, as well as a grammatical reading of (53b), are possible only with a very sharp intonation break between the clitic and the sei phrase, and even then it is only very marginal. Thus, it seems we have an obligatory coindexing of the clitic with the complement object of sei in structures which correspond to (51) but not in structures which correspond to (49). 2.3. The sei Phrase and the Position of the Clitic We have observed above that the coindexing relationship which holds between the clitic and the object of sei in sentences such as (51) is obligatory. We have further argued that such coindexing does not hold obligatorily between the NP which is the complement of the construct state in (49) (ha-yalda 'the girl') and the object of the sei phrase (ha-mora 'the teacher'). In fact such coindexing between these two lexical NP's is gram-
54
Parametric
Syntax
matically impossible even if logically possible. Thus, for instance, if the object of sei is a pronominal element and the complement of the construct state is a full NP they cannot be understood to co-refer: (54)
*beit ha-moraj house the-teacher
sel-aj 14 of-her
(54) has only two possible interpretations. The first, more obvious one, can be translated as 'the house of her teacher' and is completely irrelevant for our purposes. The second one, which has the structure in (55) (which roughly corresponds to (32) above) means 'the teacher's house which she owns.' In this latter reading, the teacher cannot be coreferential with the pronoun.
The unavailability of a coreferential reading for (55) follows from the binding conditions, if we assume the relevant definition of c-command to be a slight revision of the definition of c-command suggested in Reinhart (1976): (56)
C-command (definitionj a c-commands β iff: i. a does not contain β ii. Suppose that 7 ι , · · · , 7 η is a sequence such that: a. 7j immediately dominates b. for every > 1, γ η is the head of then if δ immediately dominates α then either: I. δ dominates 0;or II. δ = 7j and y{ dominates β
The c-command configurations allowed by (56) are illustrated in (57) (Χ,Υ,Ζ, stand for categories in the X-system):
Clitic
Government
(57)
a.
55
γΓΠί
Ζ-β
d.
c. Χη"-7;=δ
Ymax_(
Ζ η =β X=Tn
ΥΠ=/3
(56) differs from the definition suggested by Reinhart only in the introduction of clause (iib). This clause requires that c-command be effectively contained within the domain of the head of the phrase. Note, however, that such a definition would still allow the head to c-command into adjoined phrases (for instance, it would allow the verb to c-command postposed subjects adjoined to VP in Italian): although the head is dominated by a maximal projection which does not dominate the adjoined phrase, it is the head of the maximal projection which dominates and therefore c-commands the adjoined phrase. On the other hand, the definition in (56) would prevent a head of a maximal projection from c-commanding an element in a dominating maximal projection, which has a different head but which is of the same categorial type. Note that this situation occurs in structures such as (55): we would like to block the head of N 2 (as opposed to the N 2 itself) from c-commanding the sei phrase. 15 Adopting the definition in (56), N 2 in (55) c-commands N 3 . Thus both these nodes are subject to the binding conditions and are marked disjoint in reference. If this is i n d e e d j h e case we expect it to be possible to place a reflexive anaphor in the N 3 position, bound by N 2 . This is in fact possible, as is demonstrated by (58). (The marginality of (58) is, I believe due to independent reasons. See fn. 16 and the discussion in section 4.2 below.) 16 (58)
?re'iyat ha-mora sei 'acma view the-teacher of herself 'the-teacher's view of herself
Thus the impossibility of coindexing between N 2 and N 3 can be attributed to the binding conditions. Note, however, that if this is the case we obviously cannot hold that the clitic in phrases such as (51) occupies the same position that the lexical NP ha-mora, 'the teacher,' occupies in (55): whereas the lexical NP enters the binding conditions, the clitic does not.
56
Parametric
Syntax
Hence it seems obvious that the structure of clitic-doubling configurations such as (51) cannot be represented by (55) or (47). Furthermore, j i n c e definition (56) would include any possible argument position inside Ni in (55), we have to conclude that the clitic is not in an argument position. This leads us to the conclusion that the clitic is part of the head noun as we have suggested before. What is the status of the sei phrase? In diagram (41) sei is assumed to be an inserted element, which does not change the NP-status of the category it is adjoined to. Another possibility would be to claim that it is heading a base-generated PP. Our demonstration that this is not so is based on the binding conditions. In (59) it is shown that the object of sei can serve as an antecedent for a lexical anaphor: (59)
re'iyat 'acma sei ha-mora 1 7 view herself of the-teacher 'the teacher's view of herself
Recall that we have argued that (58) is grammatical due to the fact that N 2 in the structure in (55) c-commands N 3 . I n o r d e r to account for (59), we have to assume that N 3 also c-commands N 2 . Note, however, that if the sei phrase were a PP its object could not c-command N 2 , according to the definition in (56). Indeed, objects of genuine PP's cannot c-command N 2 in a similar structural configuration, as is demonstrated by the sentences in (60) and the diagram in (61): (60)
a.
xasivat ha-mora 'al 'acma thinking the-teacher about herself 'the teacher's thinking about herself b. *xasivat 'acma 'al ha-mora thinking herself about the-teacher
xasivat thinking
*ha-mora the-teacher acma herself
Clitic Government
57
This contrast between the behavior of genuine PP's and s"el phrases can be readily explained if we assume that while true PP's adjoined to construct states have the_ structure in (61), in which no c-command relationship holds between N 3 and N 2 , sei phrases are not true PP's: they are NP's to which the dummy Case marker sei has been adjoined at a level irrelevant for the binding conditions. We thus conclude that the structure in (41) above is indeed the structure of clitic-doubling constructions. Our proof consists of two stages: first it is shown that the clitic and the complement of sei are obligatorily coindexed, a condition which does not hold for lexical complements of the construct state and for the object of sei in equivalent configurations. It is further shown that the impossibility of coindexation between the complement of the construct state and the object of sei follows directly from the binding conditions. Since the relationship between the clitic and the coindexed NP is not sensitive to the binding conditions, we conclude that the clitic cannot possibly occupy an argument position. Thus we confirm our claim that it is part of the head noun. The second stage of the proof consists of showing that sei phrases demonstrate different behavior from PP's with respect to the binding conditions. Whereas PP's demonstrate typical behavior, preventing their objects from c-commanding argument positions outside the PP, the objects of sei phrases behave as bare NP's, thus entering into a binding relationship with elements which share the same governing category - in this case, the higher N. Thus we conclude that at the level where the binding conditions apply, i.e. S-structure, sei is not present: it is inserted later, for purposes of Case assignment, and its insertion does not affect the output of rules which apply in the syntax, at S-structure or in logical form. (But see Chapter 4 for a detailed discussion of sei insertion as well as a final conclusion with respect to the level at which it applies.) 18
2.4. Coindexing and Government In section 2.2 above we have shown that a relationship of obligatory coindexing holds between the clitic and the associated NPj in cliticdoubling constructions. Is this obligatory coindexing subject to any conditions? Consider the following sentences: (62)
misgeret tmunot frame pictures
ha-yalda sei ha-mora the-girl of the-teacher
(62) (a regular construct state formation without clitic or doubling, combined with a sei phrase) permits the following bracketings:
58
Parametric
Syntax
(63)
a. [[frame [pictures [the girl]]] of the teacher] 'the teacher's frame of the pictures of the girl' b. [frame [[pictures [the girl]] of the teacher] ] 'the frame of the girl's pictures of the teacher' c. [frame [pictures [[the girl] of the teacher]]] 'the frame of the pictures of the teacher's daughter'
Now compare the corresponding sentence with a clitic (and coindexing): (64)
misgeret tmunote-ha frame pictures-her
sei ha-mora of the-teacher
Theoretically the same range of bracketing should be possible for (50) if we ensure the coindexing of the clitic and the complement of the sei phrase. Note, however, what happens in (65), which is the list of possible bracketings for (64): (65)
a. *[ [frame [pictures-her [ 0 ] ] ] of the teacher] 'the teacher's pictures frame' b. c.
[frame [[pictures-her [0]] of the teacher] ] 'the frame of the teacher's pictures' [frame [pictures-her [[0] of the teacher]]] 'the frame of the teacher's pictures'
Interpretation (a), (which is a definite logical possibility) is excluded. Interpretations (b-c) are equivalent, in spite of the different structure. Let us first consider why (a) is impossible. (65a) would have the structure in (66):
(66)
N'x
pictures-hen
0
Clitic
59
Government
We claim that the reason for the ungrammatically of (66) is because the clitic on N 2 does not govern N 3 which is coindexed with it. In this we adopt the definition of government suggested in Chomsky (1981) and given in (67): (67)
Government (definition) In the configuration [.../?... a . . . 0 . . . ] a can be said to govern β iff: i. α = X° ii. where φ is a maximal projection, if φ dominates β then φ dominates a iii. a c-commands β (We are assuming a revised definition of c-command, as given in (56) above.)
The assumption that the clitic in the complex N 2 + clj in (66) is a feature on the head N 2 _will immediately lead to the conclusion that, since N 2 does not govern N 3 , the clitic which is coindexed with N 3 does not govern it either. Now consider (68), which is the structure corresponding to (65b):
(68)
N, i Ni
pictures-herj
0
In this case the clitic does govern the coindexed argument. Thuscoindexing is possible. Now, as a last point, consider the structure of (65c):
60
Parametric
(69)
Syntax Ν! Ν Ν2
Ν frame
Ν2 Ν 2 + Cl;
N3
I
N3
pictures-herj
N3 0
(sei) the-teacherj
Again, c/j governs the coindexed argument. Let us now turn to the node in ( 6 6 ) , ( 6 8 ) and ( 6 9 ) which is marked as 0 . What is the status o f this node? Recall that we are assuming the Complement Matching Requirement (see ( 1 4 ) above). Following this requirement, an element and its complement cannot contain conflicting indices. Now recall that the domain o f complementation is that o f government. Thus the requirement that the clitic in ( 6 5 ) must govern the coindexed NP position follows immediately from the fact that the doubled NP is the complement o f that head t o which the coindexed clitic is attached. Note that this is not the case when no clitic is present. In that case the complement is free to be associated with any nominal head, providing that it is governed b y that head. I f the complementation requirements are met b y the doubled NP, the node 0 in ( 6 6 ) , ( 6 8 ) and ( 6 9 ) cannot be assigned any index. We have assumed as a first approximation that the following principle holds in the grammar (see Chapter 1):
(70)
The Empty Category Principle (ECP) [e] must be properly governed α properly governs β iff a governs β and: i. α is ± V ; ± N ; or ii. α is coindexed with β
Following Kayne ( 1 9 8 1 ) , we assume that E C P holds in L F . We further follow Kayne ( 1 9 8 0 b ) in assuming that prepositions are not proper governors. T h e
sentence in ( 7 1 ) further illustrates that at least in Hebrew
nouns are not proper governors:
Clitic
Government
(71)
*mi ra'ita 'et beit [e]? who saw-you acc house 'whose house did you see?'
61
(We shall return to this matter in great detail in sections 3 and 4. For our present purposes it will suffice to claim that nouns in Hebrew are governors but not proper governors. Hence (71) is ruled out as a violation of the ECP). Now reconsider the 0 node in (66), (68), (69). This node clearly cannot be [e]. Since nouns in Hebrew are not proper governors, it cannot be properly governed by the head noun. Furthermore, it cannot be properly governed by the clitic since it is not coindexed with it. On the other hand, it cannot be PRO either. PRO is the pronominal anaphor which is not realized phonologically and which cannot be governed. The 0 position in (66), (68), and (69) is governed by the head noun. We thus have a position which is not Case marked, which bears no referential index, which is governed but is not properly governed. _ Let us in fact assume that the node in question (N 4 in (66), (68) and N 3 in (69)) simply does not exist. In other words, let us assume that complementation requirements can be met whenever the complement is governed by the head of the construct state and that the precise position of the complement in the tree is irrelevant, as long as this position is governed. Note that the phrase-structure rules can still generate the nodes dominating 0, since base rules are optional. However, nothing can appear in this position: lexical NP will not be assigned Case, PRO will be governed and [e] will not be properly governed. So if the node is generated, every possible derivation will be ruled out. 1 9 Now let us return to the structures in (66), (68) and (69): in (66) the existence or the non-existence of N 4 is irrelevant: in no configuration is N 3 governed by N2 and hence it cannot be perceived as its complement. Thus the sentence is ruled out. 2 0 In (68), on the other hand, the derivation in which N 4 is generated is ungrammatical since no element can appear in this position. However, if the position is not generated the sentence is grammatical: N 3 is governed by N2 and hence can be interpreted as its complement. Now consider (69): if N 3 is generated, the sentence is ruled out as no element can appear in this position. If, on the other hand, N 3 is not generated the N 3 is deprived of its head: it is a genuinely 'headless' phrase. The latter situation is ruled out by independent considerations related to X theory. Hence (69) is an impossible derivation, unless N 4 is directly attached to N, resulting in a structure that is virtually identical to that of (68), as is shown by (72):
62
Parametric
Syntax
(72)
frame
pictures-herj
(sei) the-teacherj
A derivation along these lines of the identity in structure between (68) and (69) and thus of their identity in meaning supplies further evidence that the clitic should be viewed as a feature on the head, rather than as an argument filling an argument position. If one wished to argue for the latter analysis, one would have to say that the structures of (65b) and (65c) are as in (73a) and (73b), respectively: (73)
N2
(gel) the-teacher
her
(sei) the-teacher
Clitic
Government
63
Proponents of this analysis would then have to explain the unavailability of an interpretation corresponding to (73b) and to (65c) above, although (73b) is a possible structure, given the assumption that the clitic occupies an argument position. The analysis which holds that the clitic is not an argument and is not in an argument position, but rather a feature on the head, and which holds that it is the sei phrase which satisfies the complementation requirements thus accounts satisfactorily for the unavailability of a third reading. Let us summarize our conclusions so far: 1. Clitics are generated as features on the head of their phrase. They do not fill the argument position which is the complement of this head. This position is independently generated and can be independently filled if a Case-assigning device is available. Clitic-doubling constructions thus have the structure in (74): xn
(74) [χΧ,^Ι
NP; dummy Casemarker insertion
2. The coindexing of the clitic and the argument NP is obligatory and subject to the condition that the clitic governs the argument with which it is coindexed. Given our assumption that clitics are generated as features on the head of their phrase, both the coindexing requirement and the government requirement follow naturally from the Complement Matching Requirement, which prevents a head from containing a referential index which conflicts with that of its complements. 3. The genitive preposition sei is not available in clitic-doubling constructions in the base. Rather, it is inserted in the phonological component. Thus the structure which it creates is irrelevant to the binding conditions: the NP's which participate in clitic-doubling constructions behave in all respects as bare NP's and differ in this sense from NP's which are objects of base-generated prepositions. 4. The domain of complementation is the government-domain of the head. Given this minimal restriction, the complement phrase can appear in any position which is governed by the head. Assuming the definition of c-command in (56) above, this means that they may appear at any level of projection of the head, including adjoined positions. The latter hypothesis would seem to be incompatible with our assumption that in the construct state (where no doubling occurs and no sei
64
Parametric
Syntax
phrase is adjoined) the complement has to be attached at the Ν level (see examples (19a-b) above and related discussion). Recall, however, that strict adjacency in the construct state was necessary for the assignment of genetive Case. By contrast, in structures such as (68) sei insertion crucially takes place, rendering strict adjacency unnecessary. Thus the value of X n in (74) may be either Χ, X or X' (see footnote 10 for the meaning of X'). Let us now return to examples (45a) and (46) above which were cited as possible counterexamples to the structure in (74). The relevant contrast is repeated here as (75a-b): (75)
a. beit-aj ha-yafe sei ha-mora^ house-her the-beautiful of the-teacher 'the beautiful house of the teacher' b.*beit yafe ha-mora house beautiful the-teacher 'the teacher's beautiful house'
At this stage of the analysis it is clear that the grammaticality of (75a) as opposed to the ungrammatically of (75b) does not present any problem: we have derived the ungrammatically of (75b) from the fact that strict adjacency is required in order to assign Case to ha-mora 'the teacher' in (75b). Since there is no such strict adjacency in (75b) the sentence is ruled out. On the other hand, we have argued that where strict adjacency is not required, in the cases where the NP is assigned Case by sei, the complement node can be adjoined to any expansion of N. Thus we expect AP to be impossible in regular, non-doubled construct states, but we expect its occurrence to be entirely grammatical when sei is present. Thus the structure of the grammatical sentence (75a) is as in (76): (76)
the-beautiful
(§el) the-teacher
The possibility of modifying the head in (76) by an AP can assist us in constructing yet another test that will prove that clitics have to govern the coindexed NP in clitic-doubling constructions. Thus consider the following contrast:
Clitic Government (77)
65
a. tmunot pictures
yaldat-aj ha-ktana sei ha-mora^ girl-her the-little of the-teacher (fern) t h e pictures of the teacher's little girl' b *tmunot yaldat-aj ha-ktanot sei h a - m o r a j pictures girl-her the-little of the-teacher (pl) (pl) 'the little pictures of the teacher's girl'
The contrast in (77) will follow immediately if we compare the structures of (77a-b): (78) (=77a) Nt pictures pi
the little (fem)
Si
(79)(=77b) Ni AP the little
ΝΓ^ 1 pictures (pl)
(sei) the-teacher
1 N2
(PO
1 Ν,+cl; I 1 girl-her
Note that the gender and number markers on the adjective force us to argue for the structure in (78) for (77a), since in that case the AP clearly modifies yalcla 'girl'. For the same reason the AP in (79) has to be generated adjoined to N, since it modifies tmunot 'pictures'. Note that as a result of
66
Parametric
Syntax
this configuration i the clitic governs the coindexed N 3 in (78); hence the corresponding sentence (77a) is grammatical. In (79), on the other hand, such a government relation does not hold, and hence (77b) isungrammatical. 21
2.5. Three Genitive Constructions in Modern Hebrew Recall that in the derivation of (68) and (69) we have invoked the Complement Matching Requirement, along with the assumption that complementation requirements can be met by any NP which is governed b y the head, quite independently of the position of this NP in the tree. These assumptions have some interesting consequences. We can now reduce all the genitive constructions in Hebrew to the structure in (74). In essence, then, we claim that the sentences in (80a-c) all have the structure in (74) (assuming that insertion of the dummy Case marker is optional): (80)
a. beit ha-mora house the-teacher b. beit-a· sei ha-moraj house-her of the-teacher c. ha-bayit s"el ha-mora the-house of the-teacher
The structure of (80a-c) is illustrated by (81a-c), respectively: (81)a.
(=(80a))
N, N,
Ni
N2
In (81a), N 2 has to be generated under N j in order to be assigned Case. However, due to the availability of sei insertion such strict adjacency does not have to hold in (80b-c): (81) b.
Ni+clj
(=(80b))
N2 j
N+clj
N+clj
As demonstrated by (81b), (80b) is structurally ambiguous. This situation,
Clitic Government
67
however, is irrelevant as all the derivations satisfy the government requirement and the Complementation Matching Requirement. The same holds for (80c), which is structurally ambiguous as well: (81) c.
(=(80c))
N, Note that (81a-c) are all manifestations of the structure in (74) in that in all tjiree tjie complement of the head is governed by some expansion of N: Ν, Ν or N'. Only one of these constructions is limited; in the construct state proper the complement can only be generated under N, as in (81). This limitation, however, has an independent explanation. Note that, although we now argue that (80b) and (80c) are manifestations of the same structure, we avoid the pitfalls of an attempt to collapse these structures as briefly sketched above (see diagram (47) and related discussion). The earlier analysis was incapable of capturing the differences between the cliticdoubling construction and the regular genitive constructions using sei. Within our analysis, however, these differences are captured by assuming that the clitic in clitic-doubling constructions is a feature on the head rather than an argument occupying an argument position. A very simple rule will now account for the insertion of sei b o t h in (81b) and in (81c): (82)
sei Insertion
0
(3l) - applies in the phonological component 2 2
> sei
/ [Np.
NPj]
With respect to (82) we make one auxiliary assumption, i.e. that the Case features of sei have to be phonologically realized. This implies that s e / h a s to have a phonologically realized object or an attached clitic. Thus if sei is inserted preceding an empty category it will obligatorily include a clitic. (We will return to this assumption and to sei insertion in Chapter 4.) Note that the rule of Si as formulated in (82) does not preclude the structure in (83), in which N 2 is marked twice with genitive Case - once by the head N j and once by the inserted preposition sei:
68
Parametric Syntax
A phrase identical to that in ( 8 3 ) can, however, be generated by the structure in ( 8 4 ) :
more teacher Both structures are grammatical.
3. EXTRACTION FROM CONSTRUCT STATE CONFIGURATIONS
3.1. Introduction: Predictions In the last section we concluded that the structure o f genitive constructions in Hebrew: construct states, doubled construct states and regular genitives, can be captured by the diagram in ( 7 4 ) above. Let us now turn to a particular subset o f that structure, i.e. that illustrated by the diagram in ( 3 9 ) above, repeated here as ( 8 5 ) : (85)
[NKc\]
Nj
beit + o house-his
0
What is the status of 0 in ( 8 5 ) ? Recall that when discussing the 0 node which appears in structures ( 6 6 ) , ( 6 8 ) and ( 6 9 ) above, we concluded that
Clitic Government
69
the 0 node in these structures is simply not generated, and that complementation requirements are met by the governed, coindexed NP. In (85), however, there is no such governed, coindexed NP apart from the 0 n o d e itself. Recall that we are assuming as a first approximation the ECP as formulated in (70) above. Given this principle and the analysis of clitics proposed so far, we can now put to the test our assumption that the clitics have to govern the coindexed position. Three hypotheses (at least) are logically possible with respect to the 0 in (85): Hypothesis A: 0 is in an ungoverned position. The process which allows the clitic to absorb the Case of the head of its phrase also absorbs the government properties of the head. It follows that [e] cannot appear in this position, since government is a prerequisite to proper government. An [e] in that position would thus result in a violation of the ECP. Thus only PRO can appear in that position. Note that according to this analysis extraction from the NPj position in (85) is never possible: such extraction would leave behind an empty category [e] that could not be properly governed, thus violating the ECP. (This analysis is proposed by Jaeggli, 1982 for River Plate Spanish). Hypothesis A is illustrated in (86a-b): (86)
a.
NJ N+c^
N,
L_J
I
government/Case absorption b. *
PRO
NJ N+cL
N;
government/Case absorption
[e]
Li
I
In (86b), [e] is not properly governed (in fact it is not governed at all), since the government property was absorbed by the coindexed clitic. Thus the construction is ruled out as a violation of the ECP. Note that hypothesis A is incompatible with our conclusion that the clitic must govern the coindexed NP position: if the clitic did govern the coindexed position, and we were to assume hypothesis A, then, given the definition of proper government, a special mechanism would be required to block proper government by the coindexed, governing clitic in (86).
70
Parametric
Syntax
Proponents of hypothesis A would thus have to argue that the structure of the N+cl combination is branching and for a stricter definition of c-command than the one we have been assuming. Some theoretical disadvantages of a branching structure were discussed above (see example (2) and related discussion). Some others will be discussed in Chapter 5. Empirically a branching structure clearly does not enable us to state in a natural way the fact that the clitic has to govern the NP with which it is coindexed in clitic-doubling configurations. In fact for supporters of a branching structure it is crucial to claim that no government relationship holds between the clitic and the coindexed NP, in order to block proper government in that position. Yet another empirical problem of the configuration in (86) is the fact that it makes a patently wrong prediction with respect to extraction from clitic-doubling constructions: it predicts that extraction of NPj in (85)-like structures is impossible, which is incorrect. We shall return to the proof that this extraction is possible below. Hypothesis Β: The clitic in (85) absorbs only Case but not government. The 0 is thus governed and properly governed by the head N. It follows that PRO cannot appear there, since it would be governed in this position; [e], however, can appear there, and indeed it does. From this analysis it follows that extraction from this position is possible. These predictions made by hypothesis Β are illustrated by (87a-c):
(87)
a. * N+cL u Case absorption (proper) government
PRO
J
b. * N+cL LJ Case absorption
lexical NP [-Case]
Clitic
71
Government c. N:
N+cl:
Li
Case absorption
[e]
(proper) government
I
Unlike proponents of hypothesis A, proponents of Β do not have to argue for a branching structure for the N+clitic combination. Thus they also avoid the problem of accounting for the government relationship which determines the coindexing between the clitic and the doubled NP. Note further that hypothesis Β correctly predicts that extraction from the coindexed NP is possible. In (87c), however, [e] is governed twice. Even if we assume with proponents of hypothesis Β that the Ν node in (87) can properly govern the Nj position, there is no way to block proper government by the clitic as well. Thus we expect extraction to be possible even if it is shown that the Ν node is not a proper governor. Recall, that in (71) above we presented some evidence that nouns in Hebrew are not proper governors. In particular, "noun stranding" results in ungrammatically. Combined with the results of Kayne (1981a), these facts lead us to propose a modification in the definition of proper government: 2 3 (88)
a properly governs β iff a governs β and i. a is [+V], or ii. a is coindexed with β. 24
Thus it appears that hypothesis Β cannot be maintained. It crucially entails proper government by nouns in Hebrew, which is incompatible with the ungrammaticality of (71). Hypothesis C: 0 is properly governed by the governing coindexed clitic. Hence PRO cannot appear there, nor can lexical NP but [e] can. Extraction from this position is possible. Note that the configurations which hypothesis C permits are essentially identical to those allowed by hypothesis Β with one exception: in (87a) and (87c) it is not the head Ν which governs the NPj position; rather it is the coindexed, governing clitic. Thus, this hypothesis avoids hypothesis B's wrong predictions with respect to proper government by nouns in Modern Hebrew. The discussion in the following sections will consist of two major points:
72
Parametric
Syntax
In the remainder o f section 3 it will be shown that extraction o f Nj in structures such as ( 8 5 ) is indeed possible. The evidence will consist o f an analysis o f free relatives in Modern Hebrew. (See below, Chapter 4 , section 1, for direct evidence that extraction from clitic-doubling constructions is grammatical in Romanian as well.) In section 4 , proper government by clitics will be argued for directly by adducing evidence from movement in syntax and logical form in Modern Hebrew, thus showing that clitics in configurations like ( 8 5 ) have to be allowed to properly govern the coindexed position.
3.2. Free Relatives in Modern Hebrew Modern Hebrew allows for two relativization strategies, as observed in Hayon ( 1 9 7 3 ) and Chomsky ( 1 9 7 7 ) : ( A ) a movement strategy, in which all the usual constraints on movement are obeyed (see ( 8 9 ) - ( 9 1 ) ) , and ( B ) a no-movement strategy using resumptive pronouns (resumptive clitics for PP's and NP's and free-standing pronouns for direct objects), 2 5 where all the usual constraints can be violated. 2 6 This is demonstrated in ( 9 2 ) - ( 9 4 ) : (89)
a.
ha-'is
se- ( ' o t o j )
the-man that-himj
pagasti
tj
met-I
tj
'the man I met' b . *ha-'isa
se- ( ' o t a · )
the-woman that-herj
pagasti 'et met-I
ha- 'is·
s'e
tj ra'a t j
acc the-marij that tj saw t j
'the woman that I met the man who saw her' (Complex NP Constraint violation; Ross, 1 9 6 7 ) (90)
a.
ha-'i?
se-'it-Oj
rakadti
tj
the man that-with-himj danced-I
tj
'the man with whom I danced' b. *ha-'isa
se-it-aj
ra'iti 'et
ha-'is^
se-tj
rakad
tj
the-woman that-with-herj saw-I acc t h e - m a n j that-tj danced tj (Complex NP Constraint violation) (91)
a.
ha'is
se-'et
the-man that-acc
['axot-o]j sister-hisj
ra'iti
tj
saw-I
tj
'the man whose sister I saw' b. *ha-'is
se-'et
the-man that-acc
['axot-o]j
ra'iti
sister-hisj
saw-I acc the dogj
nasax tj bit
tj
(Complex NP Constraint violation)
'et
ha-keleVj se
tj
that t
Clitic Government
73
(92)
a.
ha-'is" se-ra'iti the-man that-saw-I b. ha-'is^ se-pagasii the-man that-met-I
(93)
a. ha-is se-rakadti 'it-o the-man that-danced-I with-him b. ha-'isa se-pagasti 'et ha-'is se-rakad 'it-a the-woman that-met-I acc the-man that-danced with-her
(94)
a. ha-'is the-man b. ha-'isj the-man
se-ra'iti that-saw-I se-ra'iti that-saw-I
'oto him 'et ha-'isaj se-t· ra'ata 'otoj acc the-womanj thatj saw himj
'et acc 'et acc
'axot-o sister-his ha-kelev se-'axot-oj the-dog that-sister-his
'imca adopted
Note that Modern Hebrew does not have a relative pronoun, and that the free-standing accusative pronoun 'oto is fronted and optionally deleted (as in (89a)). (See Borer, 1984, for a detailed discussion of the conditions under which 'oto is deleted and for arguments that it is deleted from the COMP position.) When the relativized element is an object of a preposition or of a noun, pied piping is obligatory. The obligatoriness of pied piping in these cases would follow if we assume that neither prepositions nor nouns are proper governors in Modern Hebrew (see fn. 23 for a discussion). Stranding prepositions or nouns would thus result in a violation of ECP. Interestingly, these environments, i.e. following nouns and prepositions, are precisely the environments which allow for cliticization in Modern Hebrew. Modern Hebrew verbs (unlike those of earlier stages of Hebrew) no longer take clitics. Instead, they take the free-standing form 'oto. We will return to this point below. Although b o t h movement and non-movement strategies are available for relative clauses, only the movement strategy is possible in questions: (95)
a. 'et mi ra'iti? acc who saw-I 'who did I see?' b.*mi ra'iti 'oto? who saw-I him
(96)
a. 'im mi rakadti? with who danced-I 'with w h o m did I dance?' b.*mi rakadti 'it-o? who danced-I with-him
74
Parametric
Syntax
(97)
a. 'axot mi 'imca kelev? sister who adopted dog 'whose sister adopted a dog?' b.*mi 'axot-o 'imca kelev? who sister-his adopted dog
As for free relatives, the situation is considerably more complicated. At first glance it seems that the same options are open for free relatives which are open for regular ones - the movement strategy and the non-movement strategy (with resumptive elements). However, there are significant differences between free relatives and regular relatives which surface under closer investigation. First of all the resumptive pronouns appear in free relatives only inside NP's and PP's, as in (98a-b). It is precisely in these environments that the resumptive element is a clitic on the head of its phrase. In direct object position, where the resumptive element is an independent pronominal form, there is an obligatory gap: 27 (98)
a. ma se-hexlatnu 'al-av what that-decided-we on-it 'whatever we decided on' b. mi se-'axot-o mazkira ba-memsala who that-sister-his secretary in-the-government 'one whose sister is a secretary for the government' c. ma Se-raciti (*'oto) what that-wanted-I (*it) 'whatever I wanted'
Furthermore, violations of the usual constraints are completely impossible in free relatives, regardless of the presence of resumptive clitics. Thus (99a-b) are ungrammatical (and compare with (93b) and (94b): (99)
a. *ma^ se-pagasti 'et ha-'is se-hexlit 'al-avj nimkar what that-met-I acc the-man that-decided on-it sold 'etmol yesterday 'whatever I met the man who decided on it was sold yesterday' b. *keday le-hityaded 'im mi· Se-'e'evod be-misrad worth to-befriend with who that-work-I in-office se-'axot-oj menahelet that-sister-his runs 'it is worth it to befriend a person whose sister runs an office in which I will work'
Clitic
75
Government
In view of (99a-b), a natural assumption would be that free relatives are formed by movement and that the clitics in (99a-b) are not "real" resumptive pronouns. One could argue that they are the result of some tracespelling rule, or of a shadow-pronoun copying rule in the sense of Perlmutter (1972). Note, however, that this explanation leaves the asymmetry between questions and free relatives unexplained. If the clitics in (99a-b) are a result of a copying rule, why isn't a similar mechanism available to questions - in other words, why are (96) and (97) ungrammatical? We have an explanation for all these facts if we assume that the structure of both construct-state formations and PP's in Modern Hebrew is as in (100):
(100) [X, clj] Recall that for construct formations we have independent evidence that this is indeed the correct structure. Although no such direct evidence is available for PP constructions, I will assume that they have exactly the same structure. This implies, in effect, that in PP's, as in the construct state, clitics are a spell-out of Case, gender, number and person features on the head itself, and that the subcategorized NP complement is coindexed with the clitic and governed by it. The structure of (101) would thus be as in (102): (101)
'it-o with-him
(102)
P-
'it-o with-him - whereas the structure of (103) would be as in (104): (103) 'im Dan with Dan
76
Parametric
Syntax PJ
(104)
Ν
Ρ
Λ Dan
'im with
Unlike construct-state NP2_s, PP's do not have a "saving device" similar to sei that would enable the Ν in (102) to surface alongside the clitic. The absorption of Case features which surface as a clitic thus excludes the surfacing of the Ν complement itself. 28 We can now assume that the Ν position in (102) is the position from which extraction in free relatives takes place. Extraction from this position will leave a clitic behind, thus accounting for the apparent "resumptive clitic" in (98a-b), in spite of the fact that extraction has taken place. In this way we can explain why, in spite of the availability of resumptive pronouns, constraints on movement cannot be violated. On the other hand, extraction from direct object position leaving a resumptive pronoun is impossible. This follows immediately from the fact that verbs in Modern Hebrew no longer take clitics. The structure of VP in Modern Hebrew is thus as in (105): (105) Ν
V
Now consider the structures in (106a-c), where the pre-extraction configuration of (98a-c) is illustrated (irrelevant details omitted): (106)
a. (=98a))
Ρ
[pp.cy
N:
'al-aVj
maj what
on-it b.(=(98b)) [NN,cli] 'axot-oj sister-his
m
H who
Clitic Government
77 yj
c. (=(98c)) V
Ν
raciti wanted-I
ma/'oto what/it
Whereas structures (106a) and (106b) have a position distinct from the resumptive clitic from which extraction can take place, (106c) does not have such a position: both the WH word and the resumptive pronoun are generated under the same node, thus accounting for their complementary distribution. Note that, although we have established the existence of an extraction site and explained the ungrammatically of (98c) and (99a-b), we still have to explain the ungrammaticality of the parallel questions as in (96) and (97). To do so let us assume that free relatives in Modern Hebrew possess a mechanism which enables WH words in COMP to receive Case from the matrix. (Such a mechanism is argued for in detail in Groos and Van Riemsdijk, 1979.) We are now equipped with a system that can explain the difference between questions and free relatives. Note that since Case is absorbed by the clitics in structures such as (106a) and (106b), the WH word generated under the Ν position will not have Case. Unless a special device is available to assign Case to it, it will be ruled out by the Case filter. Such a device is available to free relatives but not to questions. It follows that when Case absorption takes place, only free relatives are grammatical. Questions are ruled out by the Case filter. The derivation of (98a), following our assumptions so far, would be roughly as in (107): (107)
X ... [g [ C O M P m a i ] I ί
s e
" t s h e x l a t n u [pp ' a l - a v i
l
i]]]29
Case assignment, where X has Case-assignment features Let us now return to our point of departure. With respect to the identity of the 0 in structures such as (85) above, two hypotheses were contrasted: one claimed that 0 stands for PRO (A), and the other claimed that 0 stands for [e] (B,C). It was pointed out that the two hypotheses make different predictions with respect to extraction from the Nj position in (85). Whereas the PRO hypothesis predicts that extraction is impossible, the [e] hypothesis predicts that it is possible. The data presented above indicate that we can account for interesting facts in Modern Hebrew if we assume that the 0 stands for [e]. The availability of extraction from that position enables us to explain the occurrence of apparent resumptive clitics inside free relatives, which differ in their characteristics from
78
Parametric
Syntax
regular resumptive pronouns. It also enables us to explain the impossibility of violating the constraints on movement despite the occurrence of such clitics. Finally, it provides an explanation for the fact that these clitics, although they appear in free relatives,_do not appear in questions. Thus we conclude that extraction from the N j position in (85) is indeed possible, preferring the [e] hypothesis over the PRO hypothesis. 3.3. sei Insertion
Revisited
In analyzing the difference between questions and free relatives in Modern Hebrew we crucially relied on the fact that the WH element which is fronted - both in questions and in free relatives - is not Case-marked when extraction takes place from structures such as (106a) and (106b). Hypothetically, however, there could be a way around this "caselessness" at least for (106b) if sei, the genitive preposition inserted to assign Case to Nj in (106b), is present. In this case the WH element would have Case, and precisely then we would expect questions to be grammatical. In fact, in these cases we would expect only questions to be grammatical. Free relatives would be ruled out, since the fronting of sei would yield genitive Case marking on the head which we would expect to be grammatical only when the free relative as a constituent appears in a genitive position with respect to the matrix. This is due to the "matching effect" requirement (in the sense of Grimshaw, 1977, and Bresnan and Grimshaw, 1978). 30 Nevertheless, (108) is ungrammatical just as (97b) is: (108) a.
b.
*sel mij 'axot-oj [ejj 'imca kelev? of whom sister-his adopted dog 'whose sister adopted a dog? *sel mij Rina makira 'et 'axot-oj [e]j of whom Rina knows acc sister-his 'Whose sister does Rina know?'
What is the reason for the ungrammatically of (108)? Why can the sei phrase not be fronted in its entirety? Note that in this respect sei phrases behave differently from PP's: the latter can be extracted from NP's. 31 (109) a.
Itamar nitbakes le-hachir 'al 'eize Itamar was-requested to-declare on which hu kana sfarim [ppe] he bought books b. *Itamar nitbakes le-hachir s'el 'eize Itamar was-requested to-declare of which hu kana sfarim [ s W p h r a s e e ] he bought books
mesorerim poets
mesorerim poets
Clitic
Government
79
The impossibility of extracting sei along with the fronted WH element follows from the fact that sei simply does not exist at the level of the grammar at which extraction takes place, i.e. syntax. Recall that we have already argued that sei phrases dot not behave as branching structures with respect to the binding conditions (section 2.3 above); rather, they behave as bare NP's. The ungrammatically of (108) presents us with additional evidence that sei is not available at D-structure: it cannot serve as an imput to syntactic rules. Let us look again at the rule for sei insertion (§1) (see (82) above). If we assume (82) to be a rule of phonology (but see Chapter 4, section 3 for further discussion), sensitive to local context, it is clear that an extracted NP no longer satisfies the environment for sei insertion. Thus, although sei insertion is available as a 'rescueing device' for clitic-doubling constructions, it is no longer available for the fronted WH element in free relatives or in questions since the extracted NPj does not satisfy the environment specified in (82). Note that sei insertion can still apply preceding an empty category dominated by NPj in (82): it can apply in the post extraction structure. Given that the Case features of sei have to be phonetically realized this would yield a sc.'+clitic combination, and indeed such se/+clitic combinations are possible in free relatives, as in (110a). However, in questions their availability would not change the fact that the fronted WH element is caseless. Hence (110b) is ungrammatical regardless of the insertion of sei: (110) a.
mij se-beit-Oj sel-Oj [e]j nisraf who that-house-his of-his burned 'the one whose house was burned' b. *mij beit-Oj sel-Oj [e]j nisraf? who house-his of-his burned %vhose house was burned?
The properties of sei insertion as well as the insertion of Case markers in general will be picked u p again in chapter 4, where the analysis presented here will be reviewed and slightly altered to accomodate a more complex array of data.
4. PROPER GOVERNMENT BY A COINDEXED CLITIC
4.1.
Predictions
We have established the fact that, in configurations such as (85), 0 may stand for [e] since extraction from this position indicates clearly that this
80
Parametric
Syntax
position is properly governed. Since proper government entails government, it follows that whenever no phonologically realized element appears dominated by Nj, this node dominates [e] rather than PRO; PRO in this position would be governed and hence ruled out. Thus 0 not only may stand for [e], it must stand for [e], since it cannot stand for anything else. What properly governs [e] in this position? We have shown that nouns in Modern Hebrew, like prepositions, cannot function as proper governors. Thus, the proper government of [e] cannot fall under clause (i) of the definition of proper government in (88). In section 2 we have shown that in (85) and similar structures the clitic is a feature on the head noun and that, as such, it governs N - Recall that clause (ii) of the definition of proper government in (88) allowed for an element to be properly governed if it is governed by_a coindexed element. Since the clitic both governs and is coindexed with Nj, it is a plausible assumption that the clitic does indeed properly govern [e] in this position. This is compatible with hypothesis C which was illustrated in section 3.1 above?2 In fact, Modern Hebrew offers direct evidence that clitics are indeed proper governors. This evidence comes from movement both in syntax and in LF. 4.2. Two clitic
Configurations
Consider again the construct-state constructions illustrated in section 2 above. An interesting property of the Ν complement in these constructions is that it is perfectly ambiguous between two possible interpretations: if the head noun is a_derived nominal which can take both object and subject, the complement Ν can be construed either as its subject or as its object. Thus the phrases in (111 a-b) have identical structures, represented in (111c): (111)
a. ktivat Itamar writing Itamar 'Itamar's writing' b. ktivat ha-ma'amar writing the-article 'the writing of the article' c. Ν ktivat writing
Ν Itamar ha-ma'amar' the article
Clitic
Government
81
Example (112a) with the structure in (112b) is entirely ambiguous: (112)
a. ktivat-o writing-his/its 'his writing' 'its writing' b.
NJ N+cL I ktivat-o writing-his/it
Clitic doubling is equally possible with both interpretations: (113)
a. ktivat-o writing-his b. ktivat-o writing-it
sei of sei of
Itamar Itamar ha-ma'amar the-article
Note, however, that if one of the arguments is generated as the complement of the head it is assigned genitive Case features; consequently the other argument cannot be assigned Case. It can, however, be rescued either if sei is inserted ((114b-c)) or if the accusative dummy Case marker 'et is available to assign Case to the understood object, as in (114)a: (114) a.
ktivat Dan 'et ha-ma'amar writing Dan acc the-article 'Dan's writing of the article' b. ktivat ha-ma'amarim sei Dan writing the-articles of Dan 'Dan's writing of the articles' c. ?ktivat Dan sei ha-ma'amar 3 3 writing Dan of the-article 'Dan's writing of the article'
We will assume the structure of (114a) to be as in (115): 3 4
82
Parametric
Syntax
(115)
n;
'et Ν
Na
writing
Dan
n2 the article
(The structures of (114b-c) are essentially identical, with sei substituting for 'et. Note that the structure proposed for the construct-state constructions generates (115) straightforwardly.) In the structure corresponding to (115), doubling is possible, as is illustrated by (116): (116) ktivat-Oj sei Danj 'et ha-ma'amar writing-his of Dan acc the-article 'Dan's writing of the article' Note again that (116) is generated by our construct-state structure without any complications. In the structure corresponding to (115), the subject in N 3 can be cliticized. The resulting situation is given in (117): (117) ktivat-Oj 'et ha-ma'amarj hirgiza 'et Dan writing-his acc the-article annoyed acc Dan 'his writing of the article annoyed Dan' The structure of (117) is as in (118): (118)
'et
N, Ni+clj
N3
writing-his
[ej
N2i the article
A few things should be noted with respect to (118). The Complement Matching Requirement is met by N 3 ; nevertheless the relationship which holds
Clitic Government
83
between the cliticj and N 2 is that of government (although not of coindexing). 3 5 One could argue that since the accusative marker 'et is available in Modern Hebrew anyway as a base-generated marker there is no evidence that, in constructions like (118), it is inserted for Case purposes. Our main argument that sei is inserted in the phonological component was based on the fact that sei phrases behave as " f l a t " structures. They are NP's with respect to the binding conditions (see example (59) and related discussion). This argument, however, cannot be extended to 'et. It does not behave as a " f l a t " NP with respect to the binding conditions: (119)
a. *re'iyat 'acmaj
'et
ha-mora^
view herself acc the-teacher 'the teacher's view of herself Further, 'et clearly can be fronted with moved WH elements, indicating that it is an input to syntactic rules: (120)
a. ra'iti
'et
ha-xatul
saw-I acc the-cat Ί saw the cat' b. 'et ma ra'iti? acc what saw-I 'what did I see?' Thus no argument can be constructed to show that 'et is inserted in the phonological component. In fact we would like to argue that 'et is basegenerated and that it is adjoined to its phrase, as illustrated by (121): 3 6 (121)
5'
'et
Ν
The structure in (121) is the input b o t h to the binding conditions and to movement rules. The latter will move the full N' constituent, yielding (120b). The binding conditions, in turn, will treat the structure in (121) as branching and will block Ν from functioning as an antecedent for 'herself" in (119). 3 7 There is, however, strong evidence that 'et in the environment of (115) and (118) is obligatory rather than optional, as it is elsewhere. Thus consider the following sentences:
84
Parametric
Syntax
(122) a. ('et) ma Eliseva ra'ata? acc what Eliseva saw-she 'what did Eliseva see?' b. Eliseva ra'ata (*'et) kof Eliseva saw acc monkey 'Elyseva saw a monkey' c. Eliseva ra'ata *('et) ha-kof Eliseva saw acc the-monkey The generalization characterizing (122a-c) is that the accusative marker 'et appears only preceding definite NP's. When the direct object is indefinite 'et cannot appear. In structures like (118), however, the presence of 'et is obligatory. (123), corresponding to (117) but lacking 'et before ma'amar 'article', is ungrammatical: (123) *kitvat-Oj ma'amarj hirgiza 'et Itamar writing-his article annoyed acc Itamar 'his writing of an article annoyed Itamar' In effect, the obligatoriness of 'etjn structures such as (117) results in a rather strange requirement on the N2 in that structure: it has to be definite, since 'et cannot appear preceding a non-definite object._Logically there seems to be no obvious reason to exclude an indefinite N 2 , and, in fact, the English translation of (123) is perfectly grammatical. If, however, we argue that the presence of 'et in these configurations is obligatory for syntactic reasons (namely the necessity of marking N2 with Case), then the definiteness restriction on N 2 can be naturally explained in terms of the definiteness restriction on the occurrence of 'et. Since there is no evidence that 'et is inserted in the phonology in this case, and since there is evidence that elsewhere 'et is base-generated, we will assume that in (118) it is base-generated as well. However, any failure to base-generate 'et in this position - an option which is otherwise available in the grammar for indefinite objects - will in this case lead to ungrammatically since it will result in a caseless NP violating the Case filter. We thus conclude that in (115)-like structures, 'et functions as a Case-marker, independent from the level at which it is inserted. Compare now the structure in (118) to a possible expansion of (74) above, given in (124):
Clitic
85
Government
(124)
N j +clj
(sei)
the-articlej
writing-itj (118) and (124) seem identical in most relevant respects. Nonetheless there_is one crucial difference between them. Whereas in (121) the clitic and N 2 carry the same index, in (115) they carry distinct indices, and N 3 satisfies the Complement Matching Requirement. We thus have a minimal pair whose members differ only with respect to whether there is coindexing by the governing clitic or not. Can it be shown that these two configurations differ with respect to extraction? Note that, since clearly this is the only relevant difference between these two structures, a difference in extraction can be attributed only to the difference in coindexing. Consider first the sentence in (125): (125)
*'et maj ktivat-Oj [e]j hirgiza 'et Itamar? acc what writing-his annoyed acc Itamar 'his writing of what annoyed Itamar?'
Since we assume that 'et unlike sei, is base-generated, there is no reason why questions should differ from free relatives in extraction from structures such as (118). Nevertheless, (125), in which N 2 of (118) is extracted, is ungrammatical with or without 'et preceding the fronted WH. Further consider the free relative construction corresponding to (118), when contrasted with a free relative formed from the doubling configuration in (124). The relevant pair is given in (126). Note that both extractions would result in an identical surface string. However, that string can only have the meaning in (126a), corresponding to (124). The meaning in (126b) is not possible: (126)
a.
maj se-ktivat-Oj [e]j hirgiza 'et Itamar what that-writing-it annoyed acc Itamar 'that which its writing annoyed Itamar' b. *maj se-ktivat-Oj [e]j hirgiza 'et Itamar what that-writing-his annoyed acc Itamar 'the thing [his writing of which] annoyed Itamar'
86
Parametric
Syntax
The contrast between (126a) and (126b) can be readily explained if we assume that the post-extraction structure in (126b) violates the Empty Category Principle. Whereas in (126a) the [e] j is properly governed by the coindexed clitic, in (126b) the governing clitic is not coindexed with [e]j, and hence it cannot properly govern it. 3 8 A similar contrast between extraction from (118) and extraction from (124) is found in cases where extraction takes place in logical form. Thus compare the following two sentences: (127)
a.
lo barur la-nu mi not clear to-us who eize sefeq which book 'it is unclear to us who b. *lo barur la-nu mi not clear to us who 'eize seferj which book 'it is unclear to us who
biker criticized
'et ktivat-Oj sei acc writing-it of
criticized the writing of which book' biker 'et ktivat-Oj 'et criticized acc writing-his acc
criticized his writing of which book'
Assuming that WH words in situ are moved by a rule applying in logical form, and that this movement rule leaves behind a variable (see for discussion, May, 1977; Aoun, Hornstein and Sportiche, 1981; and others cited there), and further assuming that this variable is subject to the Empty Category Principle (see Kayne, 1981b, for discussion), the difference in grammaticality between (127a) and (127b) can be readily explained. (128a) and (128b) are the relevant logical form representations of (127a) and (127b), respectively: (128)
a. ( f o r w h i c h x - ) , b. (for which X j ) ,
Xjabook... Xjabook...
[ N p N + clj [NpN+clj
Xj] Xj]
(=(127a)) (=(127b>)
Whereas in (128a) the clitic is a proper governor since it is coindexed with the empty category, in (128b) it is not coindexed with it, and hence it cannot properly govern it. Thus (128b) constitutes a violation of the Empty Category Principle and the corresponding sentence, (127b) is ruled out.39'40 Concluding that clitics can function as proper governors for coindexed empty categories provides interesting predictions with respect to the three genitive constructions in Modern Hebrew - the construct state, the doubled construct state and the regular genitival structure, seen here in (129a-c):
87
Clitic Government (129)
a. beit ha-mora house the-teacher 'the teacher's house' b. beit-aj sei ha-moraj house-her of the teacher c. ha-bayit §el ha-mora the-house of the-teacher
We predict that extraction, in syntax and in logical form, would be possible only in (129b), since only in (129b) will the empty category be properly governed. This prediction is verified. Thus, of the three free relatives corresponding to (129a-c), seen in (130a-c) below, only (130b) is grammatical. This proves that syntactic movement is only possible in (129b): (130)
a. *mij who b. mij who c. *mij who
se-ra'iti that-saw-I Se-ra'iti that-saw-I se-ra'iti that-saw-I
'et acc 'et acc 'et acc
beit [e]j house beit-Oj house-his ha-bayit (Sei) [e]j the-house (of)
Similarly, in (131a-c) wide scope is only possible in (131b). Thus (131a) and (131c) are semantically deviant. The obligatory narrow scope interpretation in (131a,c) - contrasting with the possibility for a wide-scope interpretation in (131b) - proves that movement in logical form is only possible in (129b) as well: (131)
a. #τοξ sloSa ' a n a ä m nir'a miba'ad la-xalon head three men was-seen through to-the-window t h r e e men's head was seen through the window' b. ros"-anij Sei slosa 'anasirrij nir'a miba'ad la-xalon 'the head of three men was seen through the window' c. #ha-ros sei Slosa 'anasim nir'a miba'ad la-xalon (meaning as in (128a)).
Let us now conclude our discussion in sections 2.4. In section 2 it was established that the relationship which holds between the clitic and the NP with which it is coindexed is that of government. In section 3 it was shown that the NP position which is coindexed with the clitic must be governed, and, in fact, must be properly governed, in order to account for the fact that extraction can apply to it. In section 4 it was shown that this position is indeed properly governed, in accordance with clause (ii) of the definition of proper government in (88) above: the clitic which governs
88
Parametric
Syntax
this position is also coindexed with it, thus satisfying the definition o f a proper governor. T o summarize, in sections 2 - 4 we have argued for a specific analysis o f clitic constructions in Modern Hebrew. In particular, w e assumed that the clitic in some sense deprives the c o m p l e m e n t NP o f Case, so that insertion o f a d u m m y Case marker is necessary if this N P is to be p h o n o logically realized w i t h o u t violating the Case filter. We differed from Jaeggli's analysis, however, in assuming that the c o m p l e x head+clitic is non-branching. We assumed that the clitic is a feature on the head and that, as such, it governs the N P c o m p l e m e n t which it is c o i n d e x e d w i t h . We further argued in detail that the insertion o f sei, the genitive d u m m y Case marker, has to take place in the phonological c o m p o n e n t , since it d o e s not interact with the binding conditions. However, as was pointed o u t , the status o f d u m m y Case markers may vary in this respect. For instance, although it w a s argued that
'et, the accusative marker, can
function as a "saving device" for Caseless NP's in certain environments, it is most probably base-generated. Here, t h e n , w e have a morphological property o f a grammatical formative which may generate a parametric variation. In Chapter 4 w e shall show that this is e x a c t l y the case.
NOTKS 1. In examples (lc) and ( 1 0 it is a noun which takes a complement rather than a verb. One may raise a question with respect to the availability of complementation requirements and Θ-role assignment, when the head is a noun and the complement is a possessor. Clearly such a selection by the head must be allowed in the case of derived nominals, as in (i): (i)
the destruction of the city
It is not clear, however, if the same treatment can be given in the case of (ii): (ii)
the tail of the dog
In this study we will assume that both in (ii) as in (i) the complements are best characterized as selected by the head noun and as assigned a Θ-role by it. The question of whether this assignment is triggered by a structural environment, or whether these complementation requirements are properties of particular lexical items is left open. For our purposes it suffices to state that we hold all complementation requirements which are valid for verbs and prepositions to be valid in cases such as (lc) and (If), regardless of the derivational history of the head nouns. This is particularly important for the government requirement and the Complement Matching Requirement discussion below. On Case Assignment by nouns, see section 2.1.1 below. 2. One could argue that no special rule is needed in Jaeggli's analysis. Instead
Clitic
Government
89
Θ-role assignment and indexing are done freely, and any assignment which does not result in an identical index and identical Θ-role for the clitic and the object NP is ruled out by the Projection Principle (see Chapter 1 above). Recall that the Projection Principle requires that lexical specifications be adhered to at every level. If a distinct index or Θ-role is assigned to each member of the pair clitic/NP, we will have a referential expression which is not selected by any verb and an "extra" Θ-role which is not assigned by any lexical head, in clear violation of the Projection Principle. Note, however, that even if we assumed free indexing we would need a checking device at some level to ensure that the Projection Principle is obeyed and that the pair clitic/NP is assigned the same index and the same Θ-role. The lack of a structural relationship between the clitic and the NP would then be reflected as a special, nonstructure-dependent checking device, rather than as a special, non-structure-dependent coindexing and Θ-role assignment rule. 3. Our argument for the clitics as being part of the head constituent is independent from the question of whether in a structure such as (2) the clitic c-commands the coindexed NP. As we will argue below, an extention of the definition of c-command is independently required and such an extention would allow the clitic to govern the coindexed NP in (2). However, the status of the clitics as parts of lexical heads is independently motivated. 4. The structure in (3) was suggested as a base-generated structure in the early version of the Pisa Lectures. 5. Kayne's generalization is clearly not a universal. Several languages arc known to have both clitics and doubled NP's without the appearance of a dummy Case marker (cf. Macedonian, as described by Berent, 1980). Although we do not discuss these languages in detail, a possible direction for the investigation of clitic phenomena in these languages is suggested by the analysis of the particle 'eyn in Hebrew in Chapter 6. Following the analysis there it is plausible to assume that in some languages the presence of a Θ-index (in a sense specified below) or a complement index (see Chapter 6, section 3) is sufficient for the formation of a clitic which does not contain a Case feature and thus does not absorb Case. 6. Clitics on adjectives are not discussed directly in this study. Note that in the Romance languages they never surface on the adjective itself. Rather, they are attached to the auxiliary verb. This is due to the fact that in the Romance languages clitics attach obligatorily to the auxiliaries. For a discussion of this phenomenon see Chapter 6, section 3.1.1. That treatment carries over the adjectival clitics as well. 7. Aoun ans Sportiche (to appear and Chomsky (1981) argue that a lexical head must be allowed to govern everything within its projection (including, for instance, its specifiers). Accepting their argument (and see below for independent justification), it is clcar that a complement in the sense defined in the text can be generated not only as a sister to X°, but also as a sister to X. One of the premises of the X theory has been that complements can only be generated as sisters of X, as is implied by the formulae in (i) (see in particular Chomsky, 1970; Jackendoff, 1977;Stowell, 1981). (i)
a. X b. X -
SPEC X X COMPLEMENT
Such a requirement, however, will not follow from the requirement that heads govern their complements and assign Θ-role to them. It would have to be stipulated independently. Below we will present some evidence that such an additional stipulation is not required and that in fact, as predicted from our definition of a complement, complements can be generated as sisters to X.
90
Parametric
Syntax
8. The distinction between the object of to in (8a) and the object oifrom in (8b) is very clear intuitively. In fact, it is hard to see in what sense to can assign Θ-role to Mary in (8a). In what follows we will assume that the verb rather than the preposition assigns the Θ-role to the [ΝΡ,ΡΡ] in such cases and that the preposition serves as an auxiliary Case marker and perhaps assigns some secondary semantic role. This proposal, however, will not be justified in detail. For more discussion of this see Chapter 6, section 3 . 1 . F o r a different approach see Marantz (1981). 9. In many respects, the head of the construct state and its complement behave as one lexical unit. Thus, for instance, main stress falls always on the complement, and thus the head is subject to various reduction rules which operate in non-stressed environments and certain lengthening rules and deletion rules which are sensitive to stress do not apply. Hence e.g. the alternation between bayit, 'house' as an independent form, and beit, construct form; i/j>a, 'appartment' vs. dirat etc. (see Prince, 1975 and McCarthy, 1979 and references citcd there for a discussion). Furthermore, there is a strong tendency to lexicalize construct-state combinations, treating them as a single lexical entry: beit sefer, 'school' (literally 'book house'); 'orex din,' 'lawyer' (literally 'law editor'); etc. These lexicalized combinations, I believe, do not have the structure in (15). Rather, they have properties of regular nouns (for instance, a definite article can precede the full noun and not just the complement: ha-beit-sefer 'the school' but not ha-beit more, 'the tcacher's house.' (See a discussion of the definite article in construct states below). We will briefly return to these reanalyzed forms in fn. 21. 10. In complex structures such as (32) I have tried to use a consistent notation to indicate the structural relationship between different elements in the tree. All expansions of the same maximal projection arc numbered with the same subscript. The maximal projection is marked Xy (y being an integer). In cases with an adjoinedJel phrase like (32), the node dominating the adjunction is marked with a prime: X'y. There is an implicit theoretical assumption here that set phrases may be adjoined to maximal projection, an assumption that will not be defended directly in this study but which will lead to certain explanatory gains. Nor will we discuss the question of whether the structure in (32), which we assume to be S-Structure representation, is base-generated or not. Note that the adjunction of these/phrase could, presumably, result from a movement rule that would postpose the set phrase from the specifier of N , . This proposal, however, will not be pursued in detail. 11. Although the constraint on the assignment of genitive Case explains the rightbranching restriction as in eases (27) and (21), note that it cannot explain the ungrammatically of the structure in (i): (i)
*
Ν
ha-
Ν
NP
ben son
melex king
In other words, we do not explain why Ν in (i) can only be rendered definite by attaching the definite article to the complement. This study does not offer an explanation for this fact. For some attempts see Dresher (1973), Aoun (1978), Berman (1978), and references cited there.
Clitic
91
Government
12. Clitics, like other complements of the head in the construct state, change stress patterns. There are also phonological and morphological factors which determine the form of the clitic which do not interact with any syntactic phenomenon. A table of clitic forms is given in (i): 1 2 2 3 3
sg: sg sg sg sg
masc: fem: masc: fem:
-i, -xa, -ax, -o, -a,
-ay -ex a -ix, - e x -av -ha
1 2 2 3 3
pi: pi pi Pi pi
masc: fem: masc: fem:
-nu -xem -xen -am, - h e m -an, - h e n
The table in (i) roughly represents the ways in which these clitics are pronounccd in Modern Hebrew. It should not be taken to represent their underlying forms. 13. The value of this argument is weakened by the ungrammatically of the English phrase correponding to (48): (i)
a. *the teacher's its house (and compare: the teacher's dog's house) b. *the house of it of the teacher (and compare: the house of the dog of the teacher).
This state of affairs may indicate that there are some restrictions on the occurrence of pronominal elements in possessive constructions (for instance, a pronominal cannot be the head of a possessive construction). If this is indeed the case, then (48) may be ungrammatical regardless of the structure which we choose to assign to it. However, as will be shown below the argument for clitics as affixes and as non-arguments does not depend solely on the ungrammatically of (48). 14. One might argue that the ungrammaticality of (54) is a reflection of a more general restriction which does not allow the clitic on Sei to be corefcrential with the complement of the construct. This, however, is not the case. If the se/+clitic combination appears following N+clitic combination, the clitics may corefer, as is illustrated by (i): (i)
beit-ij sel-ij house-me of-me 'my house'
A question which arises with respect to (54) and (i) involves the structure of se/+clitic combinations. Do they involve the structure in (3)? In other words, is there, in these cases, an argument position which is here an instance of 0? I believe that this is indeed so. Note that it follows that if another sei is inserted, we should be able to get "clitic tripling" - and, in fact, (ii) is not bad: (ii)
?beit-am· sela-henij sei ha-talmidinij house-tnem of-them of the-students 'the students' house'
The maiginality of (ii), it seems to me, is entirely due to its extreme redundancy but it is quite grammatical. Another question which arises with respect to (54) involves the stage at which the clitic is spelled out on sei. We will return to these questions in Chapter 4, where the precise nature of sei insertion will be discussed in detail.
92
Parametric
Syntax
15. Yet another piece of evidence for clause (lib) in (56) is provided by Reuland (1981). This clause is required in order to prevent the head of the specifier in possessive constructions in English from c-commanding (and thus governing) elements which are in the domain of the head of the full NP. The relevant configuration is as in (i): (i)
[ N P [ N P [ N John]'s] [j^- [ A p beautiful] [ N brother]]]
The definition in (56) differs from the definition in Chomsky (1981) in one crucial respect: Chomsky's definition, like our definition, allows for "long distance" ccommand in cases such as (57b). However, this holds for heads alone. Thus his definition excludes (57c-d), which are allowed by Reinhart (1976). The explanatory value of our proposal will become clear as we procede. 16. It will be shown in section 4.2 that when the head of the construct state is a derived nominal two complements of the head can appear, one of which is construed as the subject and the other as the object. When this is the case there is a preferred order: the object usually follows the head, and the subject is expressed by means of a sei phrase. Violations of this order do not lead to ungrammatically but result in marginal sentences. Thus (i) is marginal, while (ii) is perfectly grammatical: (i)
?'axilat Dan sei ha-tapuax eating Dan of the-applc 'Dan's eating of the apple'
(ii)
'axilat ha-tapaux sei Dan eating the-apple of Dan 'Dan's eating of the apple'
The meaning intended by (i) can be rendered without marginality if the accusative Case marker 'et is used rather than sei: (iii)
'axilat Dan 'et ha-tapaux eating Dan acc the-apple (meaning as in (i))
Returning now to (58), the marginality of this configuration is similar in nature to that of (i), and seems to derive from the same source: since the teacher in (58) is construed as the subject, and herself as the object, the order is marked. If, as in (iii), we replace sei with 'et the sentence is perfectly grammatical: (iv)
re'iyat ha-mora 'et 'acma view the-teacher acc herself 'the teacher's view of herself
Note that, since obviously in (iv) ha-mora 'the teacher' has to c-command 'acma 'herself, it would be hard to argue that such a relationship holds in (iv), but not in (58). This peculiarity of the structure of the construct state, like other peculiarities noted before (e.g. the location of the determiner in construct states, see fn. 11) will not be pursued here (but see Berman, 1978, for discussion). 17. The reflexive form itself in Modern Hebrew appears to be a construct state of
Clitic
93
Government
the form noun+clitic ( ' e c e m , literally ' b o n e ' + clitic). Note, however, that the coindexation between the clitic a in ( 5 9 ) and ha-mora,
'the teacher' is quiet different
from the coindexation between a clitic and a doubled NP, as in, say, ( 4 7 ) . Whereas in the latter we have an instance o f doubling, which, as will be argued below, is sensitive to government b y the clitic, the coindexation in ( 5 9 ) is between an antecedent and an anaphor. In fact, the entire phrase 'acma is the antecedent, the noun 'ecem,
'bone', having no reference here other than the reference o f the clitic. In view
o f this we claim that the combination 'ecem + clitic in ( 5 9 ) is not a manifestation o f the N+cl
structure we have argued for above and is not accompanied by an empty
node. R a t h e r , the clitic on 'ecem is reanalyzed as a pronominal agreement marker, rendering the reflexive form a free standing pronominal which cannot be further analyzed. T h e structure o f ( 5 9 ) is thus as in (i): (i) se/-phrase
N, srcl
N,
N,
of the teacher.
N, N,
I
acma^ herself 18.
Note that we argue that in structures such as ( 5 5 ) , N , and N 3 c - c o m m a n d each
other, and that consequently each o f these positions can serve as antecedent for the other, as demonstrated by (i) and (ii): (i)
re'iyat
ha-mora
view
the-teacher acc
'et
'acma herself
'the teacher's view o f h e r s e l f (ii)
re'iyat
'acma
sei ha-mora
view
herself
of
the-teacher
'the teacher's view o f h e r s e l f This situation seems t o present a problem for the binding conditions: in a situation o f mutual c - c o m m a n d t h e antecedent NP is c-commanded by the anaphor, and hence it is not free. Since it is not free it violates the binding conditions. Furthermore, such a situation may give rise to two anaphors, binding each other, as in t h e ungrammatical case in (iii): (iii)
*sin'at
'acma
sei
'acma
hate
herself
of
herself
T o prevent these undesirable results it would be useful to interpret the binding conditions as a process o f index transmission from c-commanding nodes to c-commanded nodes within
a governing category
(essentially
following Chomsky
1 9 8 0 ) . This
assumption would further entail that anaphors do not have an inherent referential
94
Parametric
Syntax
index and that such an index can only be assigned to them by a c-commanding antecedent. Returning to (55), it is clear that either Ν can be an antecedent to the other in t h e sense of index transmission. This situation would never result in an index conflict between t h e anaphor and the lexical NP, and thus the combination would be grammatical. On the other hand, disjointness in reference would be granted in t h e way suggested in Chomsky (1980): in (iv) t h e anaphoric index transmitted to John by t h e coreferential he will cancel the referential index and that state of affairs will lead to u n g r a m m a t i c a l l y : (iv)
he. saw J o h n · J •) !,
n
Interestingly, similar cases exist in English. T h u s consider t h e following pair, which most speakers find grammatical: (v)
a. that picture of Rembrandt by himself b. that picture of himself by Rembrandt
19. Yet another possibility would be to adopt t h e p r o p o s a l o f S a f i r ( 1 9 8 2 ) , a c c o r d i n g to which expletive [c] does not have to be properly governed. Since Safir's proposal entails a different treatment of the pro-drop p h e n o m e n o n than the one advanced in Chapter 6 of this study, his proposal will not be adopted here. 20. Note that our explanation does not account for the reading according to which the empty category is regarded as t h e complement of t h e head n o u n (and is governed by it and by the clitic) and the sei phrase is assigned a different referential index. It seems to me that pragmatic factors are at play here. Whenever a lexically realized NP can be construed as the complement, that NP rather than the empty category is construed as that complement. A similar p h e n o m e n o n exists in River Plate Spanish. In a situation where an a phrase can be construed either as a PP or as a doubled element, t h e latter reading is greatly preferred: (i)
lo envie a Juan him send-we to Juan 'we send it/him to J u a n ' / \ve send J u a n '
Where River Plate Spanish shows a preference. Modern Hebrew shows a sharper contrast, actually ruling out all other interpretations in t h e absence of a sharp intonational break. 21. There arc, in fact, some counterexamples to our analysis. Thus t h e following is a grammatical sentence, although one could argue that its structure is identical to that of (79): (i)
signon ktivat-Oj ha-maksim style writing-his the-charming (masc) (fem) (masc) t h e charming writing style of Dan'
sei Dan^ of Dan
The class of cases which violate the government requirement is scmantically restricted. It consists solely of " m a n n e r " nouns as heads and gerunds as complements: dcrcx ha-rica 'way of running,' ofen ha-halixa 'manner of walking,' etc. Each of these elements when appearing with o t h e r nouns obeys t h e government constraint:
Clitic (ii)
*signon kis'ot-aVj style (sg)
(iii)
95
Government
chairs-his (pi)
*masluley ricat-Oj tracks (pl)
ha-xadas sei Dan^ the-new of (sg) ha-'arukim
running-his the-long (sg) (pl)
Dan
sei Darij of
Dan
Since the class of counterexamples is seniantically so restricted, I will assume that only elements which obey these semantic restrictions are analyzed as a compound of sorts, and that this compound occupies the head position. In this sense the clitic is actually a clitic on the full compound which occupies the head position; thus it does govern the coindexed NP. These configurations would then have the structure in (iv) (corresponding to (i)): (iv)
22. A virtually identical rule can be formulated for of insertion in English. For some discussion of of insertion along these lines and for a possible additional restriction on that latter rule, see Stowell (1981). The question of where in the grammar of insertion applies will not be discussed in this study. 23. Kayne argues that the availability of preposition stranding in English and its absence in French can be explained if wc assume that prepositions arc not proper governors. If follows that an empty category following a preposition is ruled out as a violation of the ECP unless some other mechanism is available to properly govern it. Such a mechanism, Kayne suggests, is the transmission of superscript from the verb to the preposition. This transmission is only possible in English, since the English prepositions assign Case in the same way verbs do. (See Kayne, 1981a for a detailed discussion.) This compatibility of Case assignment procedures allows for the transmission of superscripts. In F rench, on the other hand, Case assignment by prepositions differs from that of verbs; hence the transmission of a superscript is impossible. In languages which allow the transmission of the superscript once it has been transmitted, the verb itself can properly govern the empty position which follows the preposition and the violation of the Empty Category Principle is avoided. For a variant of Kaync's analysis which is formulated in terms of our parametric model, sec Chapter 1, section 3, where we propose that a parametrized inflectional rule is responsible for the transmission of Θ-role features as well as Case features from the preposition to the verb. Since in Hebrew both preposition stranding and noun stranding arc impossible, wc will adopt the assumption that prepositions are not proper governors, and that furthermore the inflectional rule which allows reanalysis in English does not exist in Hebrew. Interestingly, verbs in Modern Hebrew do not take clitics, unlike nouns and prepositions (see the discussion in the text below). Given our assumption that clitics are spell-out of Case features, and assuming with Kayne that the inflectional
96
Parametric Syntax
rule in q u e s t i o n is sensitive to t h e 'similarity' o f Case assignment properties, it is possible that t h e absence o f verbal clitics reflects these differences in Hebrew. 24.
T h e q u e s t i o n o f whether t h e change in t h e definition o f proper government
should be e x t e n d e d to English o r not is left open in this study. N o t e , however, that t h e main ease in which proper government b y nouns is required is in phrases such as (i): (i)
t h e c i t y ' s destruction [ e ]
With respect to (i), sec Jaeggli ( 1 9 8 2 ) , who argues that t h e e m p t y category in such
e x a m p l e s is properly governed b y its a n t e c e d e n t rather t h a n b y t h e n o u n destruction. 25.
T h e free-standing direct o b j e c t f o r m s are given in t h e chart b e l o w : 1
sg:
'oti
1
pi:
2
masc:
'otxa
2
2
sg sg
fem:
'otax
2
pi
3
sg
masc:
'oto
3
3
sg
fem:
'ota
Pi
3
Pi
pi
'otanu masc:
'otxem
fem:
'otxen
masc:
'otam
fem:
'otan
Although historically it is clear that t h e free-standing direct o b j e c t forms derive from t h e c o m b i n a t i o n o f t h e o b j e c t m a r k e r 'et with clitics, I believe that t h e y n o longer admit o f this analysis, and that they are now on a par with t h e nominative pronouns, which arc also free-standing. T h u s , in contrast t o o u r t r e a t m e n t o f prepositions with clitics, we will not regard t h e direct o b j e c t p r o n o m i n a l form as having the structure W + c l [ e j . R a t h e r , t h e y are full NP's, containing no empty category. 26.
T h e discussion o f relative clauses in this section is restricted to relativization o f
non-subject
constituents.
The
relativization
o f subject position o b e y s
somewhat
different constraints which are irrelevant for our discussion. We are assuming here the analysis o f relativization in Hebrew in B o r e r ( 1 9 8 4 ) . F o r some o t h e r analyses, see Hayon ( 1 9 7 3 ) , B e r m a n ( 1 9 7 8 ) and references cited therein. 27.
S o m e o c c u r r e n c e s o f direct o b j e c t
pronouns in free relatives are attested in
phrases such as ( i ) : (i)
mij
se-racitem
le-hakot
who that-wantcd-you t o - h i t
'otoj
be-yaldut-Oj
him
in-childhood-his
In fact t h e direct o b j e c t p r o n o u n in (i) is obligatory, and (ii) is ungrammatical: (ii)
* nil· se-racitem le-haktot
[e]j
be-yaldut-Oj
T h e regular relative clause corresponding t o (ii) is ungrammatical as well: (ii)
*ha-'is
se-racitem la-hakot
[e]
be-yaldut-o
the-man We do not o f f e r a d e t a i l e d e x p l a n a t i o n o f these cases in this s t u d y . N o t e , i n t e r e s t i n g l y , that the elimination o f
bc-yaldut-oj
from (i) will result in the same u n g r a m m a t i c a l l y
as ( 9 8 c ) in the t e x t : (iv)
*mij
se-racitcm
who that-wanted-you
lc-hakot
' o t o j ba
to-hit
him
' t h e o n e y o u wanted t o hit c a m c '
came
Clitic
97
Government
i am indebted to Edit D o r o n for pointing out these examples to me. F o r an interesting discussion of these examples see D o r o n (1980). 28. The availability of a "saving device" for various categories appears to be language specific. Thus, in Tigre, there is a saving device for PP's as well, as d e m o n s t r a t e d by (i). T h e same holds for Lebanese Arabic, as d e m o n s t r a t e d by (ii): (i)
(ii)
Lilat '+gil 9alr warakat nad'at Lilet to Ali(m) letter ( 0 sent 'Lilet sent a letter t o Ali' hkit talked-I
'+t-tUj to-him (Jake, 1980)
ma9-Oj fo-Karirrij with-him
to-Karim
Ί talked w i t h Karim'
( A o u n , 1982)
29. In diagram ( 1 0 7 ) , we left o p e n t h e q u e s t i o n of t h e structure of free relatives: are they t r u e NP n o d e s which have an e m p t y head (as assumed by Gross and V a n Riemsdijk), or are they instances of S marked w i t h t h e f e a t u r e [ + N ] , taking COMP as its head? The latter was proposed t o m e by K. Hale (personal c o m m u n i c a t i o n ) ; sec also Fassi Fehri ( 1 9 8 0 ) , where such a proposal is pursued. A l t h o u g h we lean t o w a r d s the latter hypothesis, we will not aigue f o r this analysis here. N o t e t h a t assuming that Case is assigned to t h e NP dominating t h e free relative (or, in t h e case of t h e S proposal, t o t h e S marked [ + N ] ) , b o t h proposals can c a p t u r e t h e generalization expressed by (107). T h e first proposal would claim that t h e Case features percolate to t h e first phonologically realized element, whereas the latter proposal would claim that they are manifested on t h e head, that head being t h e WH-word in COMP. T h e proposed analysis of free relatives sketched in these paragraphs has some interesting consequences f o r t h e requirement that variables have Case, suggested in C h o m s k y (Pisa lectures). These consequences will be discussed in detail in C h a p t e r 3. 30. Grimshaw ( 1 9 7 7 ) shows that t h e WH element in free relatives has to satisfy b o t h t h e categorial requirements of the gap and the categorial requirements of t h e matrix. This requirement docs not hold for e.g. e m b e d d e d questions. T h u s we find t h e following contrasts: (i)
1 asked how tall Bill is [ e j
(ii)
I asked where you p u t y o u r coat [ e |
(iii)
1 asked what you ate for lunch [ e j
(iv)
*I will hit however tall Bill is [e]
(v)
*I will hit wherever you leave y o u r coat [ e |
(vi)
I will hit whatever you t h r o w m e [e]
(vii)
I will b e c o m e however wealthy you b e c o m e [e]
(viii) (ix)
*I will b e c o m e wherever y o u p u t y o u r coat [e] I will b e c o m e whatever you w a n t me to b e c o m e [c]
98
Parametric
Syntax
The ungrammatical cases (iv), (v), and (viii) are free relatives in which the matching requirement is not met; the matrix verb does not subcategorize for an AP (iv) or a locative phrase (v, viii), while the fronted WH leaves a gap of this type. 31. Extraction of PP's from nominal structures is rather restricted in Modern Hebrew by conditions that are poorly understood. Note, however, that there are no structural considerations that would render the extraction in (109a) better than the extraction in (109b). In both cases, the extraction is clearly from the nominal phrase (as opposed to the VP), and thus the contrast between them is telling. Edit Doron (personal communication) notes the following case in which extraction of the sei phrase is possible: (i)
lo barur sei mi hu xaver not clear of how he friend 'it's not clcar whose friend lie is'
Crucially, cases in which sei extraction is possible involve copula constructions. Some exceptional behavior of sei phrases in copula construction will be pointed out below (Chapter 4, fn. 4). Apart from the sketchy account suggested there, we will not pursue this phenomenon further. 32. Note that if it could be shown directly that the clitic in (85) and not some other clement properly governs the empty category under N^, it would shed interesting light on the distribution of clitics in Modern Hebrew. It would suggest that clitics were preserved in all and only the environments where the lexical category does not function as a proper governor. Thus clitics on verbs disappeared while clitics were retained on nouns and prepositions. Further note that the clitics in post-nominal and post-prepositional positions enable extraction to take place from a position that otherwise would not allow extraction. If we assume that languages strive to avoid redundancy, we may gain some insight into the nature of this distribution. Note that, following verbs, clitics are redundant as proper governors. Furthermore, the language has developed a parallel way to express direct object pronouns. For this reason clitics on verbs began to disappear. On the other hand, clitics on nouns and prepositions are essential as proper governors. Consequently they have not disappeared and the language has not developed a parallel way of expressing pronominal objects of prepositions and pronominal objects of nouns. 33. The marginality of (114c) is due to the preference of the subject as t h e s e / o b ject in these configurations. See footnote 16 above for some discussion. 34. Note that the diagram in (115) gives only one possible derivation of (114a). Another possibility is to generate the 'et phrase under N , . The same holds for the diagram in (118), which gives only one structural representation of (117). 35. Clearly 'et phrases do not lend themselves to doubling in the way that sei phrases do. In fact, doubling with 'et is impossible. Thus, in (118), N'2 cannot meet the Complement Matching Requirement, and if N 3 is not generated the sentence is ruled out. It will be argued below that 'et, unlike sei, is available in the base. It is quite plausible to assume that since 'et is an accusative marker its object cannot meet the Complement Matching Requirement of a genitive assigning head (for example, the head noun of the construct state). 36. Adjunction of dummy Case markers was proposed by Vergnaud (1974), Jaeggli (1980), Manzini (1981), Stowell (1981) and others. 37. One could argue that the structure in (121) should enter into the binding conditions since although Ν does not c-command herself in (119), N ' does. In Chapter 4
Clitic
99
Government
some evidence will be presented that the latter hypothesis should be rejected. For a discussion of some problematic cases, see Chapter 4, footnote 7. 38. One could argue that (126b) is ungrammatical because it contains two instances of [e], only one of which can be properly governed: one instance of [e] is coindexed with the clitic (N 3 in (118)), while the other is [e]j. Note, however, that if the head rather than the clitic is the governor in (126b), there is no a priori reason why it should not properly govern two empty categories. Furthermore, since under any plausible account the clitic will properly govern the empty position coindexed with it (although it might do so redundantly), why cannot the head properly govern [e]j in (126b)? Thus, the presence of two empty categories in (126b) is, in fact, irrelevant for proper government by the clitic. 39. It has been brought to my attention that for some speakers the combination of the accusative marker 'et with the question words ma 'what' and 'eyze Vilich' is ungrammatical. Clearly for these speakers V h a t ' and 'which' in Hebrew are inherently indefinite and when preceded by the definite accusative marker 'et they lead to ungrammatically. Thus, one might claim, for these speakers the contrasts illustrated in (126) and (127) as well as the ungrammatically of (125) would follow from the ungrammatical combination 'et + what/which. For those very speakers, the relevant contrasts may be reproduced with mi 'who', which freely allows the attachment of 'et. Thus (i), (ii) and (iii) are ungrammatical, on par with (125), (126b) and (127b): (i)
*'et mij 'ahavat-Oj [e]j hirgiza 'et acc who lovc-his annoyed acc 'his love for whom annoyed Itamar?
Itamar? Itamar?
(ii)
*mij se-'ahavat-Oj [ e j j hirgiza 'et Itamar who that love-his annoyed acc Itamar 'the one his love to whom annoyed Itamar
(iii)
*lo barur la-nu mi biker 'et 'ahavat-Oj 'et not clear to-us who criticized acc lovc-his acc 'it is unclear to us who criticized his love of w h o m '
mij whom
40. The contrast between (127a) and (127b) extends to the contrast between (i) and (ii) as well: (i)
(ii)
lo barur la-nu mi biker 'ct ktivat ha-ma'ararim not clear to-us who criticized writing the articles scl 'cizc sofer of which writer 'it is unclear to us who criticized the article-writing of which writer' *lo barur la-nu mi biker 'et ktivat Dan 'et 'eize ma'amarim 'it is unclear to us who criticized Dan's writing of which articles'
Note, however, that, following the requirement of proper government of the extraction site, (i) should be ungrammatical as well, if only the WH in situ is fronted. We believe that another derivation is possible, in which the entire phrase 'writing thcarticles of which writer' is pied-piped in logical form. In (ii), however, this derivation is blocked.
100
Parametric
Syntax
The availability of pied piping in (i), but not (ii), is c o n f i r m e d by t h e grammatic a l l y of syntactic pied piping in t h e former b u t not in t h e latter: (iii)
(iv)
'et ktivat acc writing *'et acc
ha-ma'amarim sei m i the-articles of w h o
Dan biker? Dan criticized?
ktivat
Ran 'et
ma
writing
R a n acc
what Dan criticized?
Dan biker?
It has been pointed o u t to m e by N. C h o m s k y that t h e availability of pied piping in (i) and (iii) and the u n g r a m m a t i c a l l y of (ii) and (iv) might follow f r o m t h e clausal nature of (ii) and (iv) and t h e phrasal nature of (i) and (iii).
Chapter
3
Clitics and the Government-Binding Model
1.
INTRODUCTION
Assuming the general framework sketched in Chapter 1 and the references cited there, several questions can be raised with respect to the analysis of clitics presented in Chapter 2. In particular, some aspects of clitic constructions have not been dealt with: the status of the empty category which is left in doubled constructions following an extraction with respect to the typology of empty categories and the status of this empty category with respect to the binding conditions. In this chapter we will discuss these questions, elaborating on our theoretical assumptions and suggesting some modifications in the Government-Binding model. Section 2 of this chapter is devoted to the question of whether variables should be Case-marked. We will review the arguments for the stand that they must be Case-marked, either as a theoretical primitive or as a derivative of the Visibility Hypothesis of Chomsky (1981). We will conclude that the generalization that variables must have Case is false and that its correct predictions can be derived from the Θ-criterion, from the Case filter and from the typology of empty categories. We will further show that the Visibility Hypothesis can be dispensed with without cost to the grammar. In section 3 we will consider the application of the binding conditions in doubled constructions. We will show that in clitic+[e] combinations, either [e] is a variable, or the entire combination is interpreted as a discontinuous pronominal element. This result will enable us to state naturally some of the properties of the reflexive clitic in Romance. It will also indicate that the assumption that clitics are Α-binders is untenable.
2. ON THE STATUS O F VARIABLES
2.1. Variables and Case Marking Chomsky (the Pisa Lectures) suggests that the following principle holds in the grammar:
102
Parametric
Syntax
(1)
[e] is a variable iff it has Case
The strongest motivation for the principle in (1) comes from the distribution of empty elements. (1) makes a clear prediction outlined in (2): (2)
a. WH movement and Quantifier movement are only possible from a Case-marked position b. NP movement is only possible from non-Case marked positions
An illustration of the predictions made by (1) are the following sentences: (3)
a. *Johnj killed [e]j b. *whoj did you try [e]j to win?
Following (1), the sentences in (3a-b) are correctly ruled out. In (3a) [e]j is Case-marked by the verb kill and hence it is a variable. As a variable the following two principles are relevant (see Chomsky, 1981 and some discussion in section 2.2 below): (4)
α is a variable iff a = [ ^ p e ] in S bound by an operator 1
(5)
a must be A-free
Given (4) and (5), the ungrammaticality of (3a) follows straightforwardly from (1): [e]j is a variable (Case-marked trace) which violates both (4) and (5). It is not bound by an operator and it is bound by an antecedent which is in an Α-position: John-v The ungrammaticality of (3b) can be derived in a similar fashion: [e]j in (3b) is not Case-marked and hence it cannot be a variable. Rather, it is an anaphor. As an anaphor it has to be Α-bound but in (3b) it is A-free and hence the sentence is ungrammatical. Clearly. (1) is a desirable principle to the extent that it makes correct predictions about the distribution of NP movement vs. WH movement and Quantifier movement. However, there pre some counterexamples to (1). One of these is extraction from clitic-doubling constructions in Modern Hebrew, as sketched in Chapter 2 above. Recall that in Modern Hebrew extraction in free relatives is possible from the following configuration:
(6)
X [ χ X, c y
N: WH
X = N, P.
Clitics and the Government-Binding Model
103
In ( 6 ) , the Case features that would otherwise be assigned t o N- are absorbed b y the clitic and consequently N j cannot be Case-marked unless sei is inserted (which is only possible when X = N ) . However, sei insertion cannot apply at D-structure. Consequently the fronted WH element cannot be Case-marked b y sei and has t o receive Case in its landing site. T h e trace left behind, however, is not Case-marked, since sei insertion did not apply in the base. 2 T h e unavailability o f Case assignment to the fronted WH element in ( 6 ) in its initial position led us to conclude that the extraction in ( 6 ) is grammatical only when the fronted WH element can be marked for Case b y an independent device. This device is Case-marking into COMP which is available for free relatives but not for questions. This analysis accounted for the contrast between ( 7 a ) and ( 7 b ) : (7)
a.
maj se-xasavti 'al-avj
[e]j
'whatever I thought a b o u t ' b . * m a j xasavti 'al-avj
[e]j?
'what did I think a b o u t ? ' T h e trace in ( 7 a ) is not Case-marked, but nevertheless it is a variable bound b y ma,
V h a t ' , and it satisfies b o t h ( 4 ) and ( 5 ) above. Thus it is
clear that ( 1 ) cannot be maintained for free relatives in Hebrew. Y e t another problem for ( 1 ) arises if we consider the analysis o f existential sentences suggested in Stowell ( 1 9 7 8 ) . Following his suggestion, existential sentences in English are cases o f clause-internal raising (leftward movement). Existential sentences according to this analysis are generated as in ( 8 ) : (8)
[ N p ] was a man in the garden
( [ j ^ p ] = null category. See footnote 5 for discussion)
T w o operations may occur following the generation o f ( 8 ) : the post-verbal NP can be raised to subject position, leaving a trace behind and yielding ( 9 a ) ; o r . if movement has not taken place, a non-referential dummy,
there,
is inserted to yield ( 9 b ) : (9)
a. a manj was [ e ] j in the garden b. there was a man in the garden
Note that in order for a man in ( 9 b ) t o receive Case in the position following the verb to be, we have to assume that this position is a Case position. Furthermore, WH movement is possible from this position, as is demonstrated b y ( 1 0 ) :
104
Parametric
Syntax
(10)
whatj was there [e] j in the garden?
Thus a variable is possible in the post-fte position. But as we saw in (9a), NP movement is also possible from the post-öe position, precisely the situation which should be excluded by (1). Note, incidentally, that the analysis of existential sentences proposed by Stowell is entirely compatible with the Θ-criterion. Assuming that the subject position of the verb be is not a Θ-position, (as is evidenced, for instance, by the rule of passive, in which an NPis fronted to that position), the movement advocated by Stowell is indeed from a Θ-position to a nonΘ-position, in accordance with the Θ-criterion (see Chapter 1 for discussion). A rightward movement analysis, as argued for by Milsark (1974) and others, either violates the Θ-criterion or would have to claim that there are (at least) two distinct be's: one that does not have a Θ-subject (passive) and one that does have a Θ-subject (existentials). (For some more discussion of this point see Borer 1980). Thus there are strong independent reasons to prefer Stowell's analysis within the Government-Binding model. Thirdly, consider the following cases of quantifier lowering discussed in May (1977). Note in particular that (11) may have the interpretations in (12): (11)
some senatorj is likely [e]j to speak at every rally
(12)
a. It is likely that there is a senator S such that for every rally R, S speaks at R. b. It is likely that for every rally R, there is a senator S such that S speaks at R.
May suggests that the narrow scope interpretation of some senator in (12ab) is achieved by a rule of quantifier lowering which moves the quantifier some senator from its position in the matrix into a position adjoined to the S of the embedded clause. As observed by May, this is possible (crucially) only in raising structures, where a [e] coindexed with the lowered quantifier is available in the embedded [NP,S] position to serve as the variable. Note, however, that in this case (1) is violated as well, since the trace of raised constituents is not Case-marked. Nevertheless it functions as a variable. Again, there are independent reasons to favor May's analysis of the facts in (11)-(12). Thus Chomsky (1981) points out that the analysis proposed by May supplies yet another distinction between PRO and trace. Only the latter can serve as a variable. Thus "quantifier lowering" interpretations are possible only in raising structures, where such trace is available, but not in control structures, where the subject position of the
Clitics and the Government-Binding
Model
105
infinitive is occupied by PRO. Insofar as May's analysis verifies this aspect of the model, it seems desirable to retain it. Chomsky (1981), following a proposal of N. Hornstein, suggests that the "Caselessness" of the variable in quantifier lowering cases can be made to conform to the requirement in (1) (or rather to the requirement in (1), as derived from the Visibility Hypothesis; see below for a discussion), if we assume that the Case-marked antecedent of the trace in (11) transfers Case to its trace, making it Case-marked in the required sense. Note, however, that if we allow for such a mechanism, then the status of the predictions in (2) becomes unclear. Consider the sentence in (13): (13)
JohrijWashit [e]j
The transferring of Case in (11) would also transfer Case from John to its trace in (13). Thus the trace in (13) would be Case marked and the sentence should be ruled out: it contains a variable which is Α-bound. Allowing for a variable t o inherit Case from an Α-antecedent greatly weakens the predictive power of (1). All these problems indicate that (1) is too strong and that the predictions illustrated by (2) should be derived by other principles which would allow for the grammaticality of extraction from doubling configurations and from existential sentences, and for the range of interpretations in quantifier lowering cases. We would like to propose that these other principles are already independently motivated in the Government-Binding model: the Case filter and the Θ-criterion. Following these two principles, (3a) is ruled out as movement from a Θ-position to a Θ-position, yielding two distinct Θ-roles assigned to an antecedent and its trace (or, alternatively, to one A-chain; see Chapter 1 and below for a discussion). (3b), on the other hand, is ruled out on the grounds that the WH antecedent lacks Case, since it originated in a non-Case position and there is no device available to assign Case to it in COMP following the extraction. Assuming that the Case filter applies to WH elements as well as to other NP's, we expect the ungrammatically of (3b) (see also Borer, 1981c. for detailed discussion). Consider the execution of our proposal. We claim that the definition of variables is as in (4) above. Thus any [e] which is bound by an operator is a variable. (5), on the other hand, is a well-formedness condition: a variable, as defined by (4), must be A-free. 3 Now consider again the cases sketched above which do not conform to the generalizations in (2). Note that the empty categories left following extraction from clitic-doubling configurations will be variables, since they are operator-bound. As such they must conform to (5): they must be Α-free, which they are. The ungrammaticality of (7b) does not follow
106
Parametric
Syntax
from (1) now: it follows from the fact that the antecedent of the variable lacks Case. Existential sentences will be accounted for straightforwardly as well. In (9a) [e] j does not meet the requirement in definition (4), and hence it is not a variable. If it is not a variable it must be an anaphor (see section 2.2. below for some more discussion of this point). As an anaphor it must be Α-bound. Since [e] j in (9a) is Α-bound by the antecedent a man-v the sentence is grammatical. Note that since a manj moves into the subject of be position, which is a-0-position, the derivation of (9a) is not a violation of the Θ-criterion. Now consider (10). In (10), [e] j is a variable in accordance with (4), and thus it must be Α-free, which it is. Again, the grammaticality of (10) follows. Lastly, consider quantifier lowering. The trace of the raised NP, some senator, will satisfy the definition in (4) following the lowering of some senator to the position adjoined to the embedded S in LF. The grammaticality of the interpretations in (12) follows. 4 2.1.1.
Counterexamples
Let us now turn to some potential counterexamples to our analysis. First consider cases in which the Case of the antecedent cannot be checked by the Case filter, and nevertheless extraction from non-Case position is ruled out (these cases are first discussed in Freidin and Lasnik, 1981): (14)
*the manj that John tried [e] j to win
Recall that we assume that the Case filter applies in the phonological component. Thus the ungrammatically of (14) cannot be explained by the Case filter: although the moved WH in (14) is not Case marked, it would be deleted prior to the application of the filter by the rule of free deletion in COMP suggested in Chomsky and Lasnik (1977), thus blocking the application of the filter. Chomsky (1981) suggests that, in fact, in sentences such as (14) the moved WH element is an abstract operator. This assumption makes it possible to eliminate free deletion in COMP since it entails that whenever that complementizers appear the WH element is abstract, and whenever an overt WH element appears in COMP, that has not been generated. Following this proposal (and assuming that abstract operators are transparent in the phonological component), it is again impossible to determine the ungrammatically of (14) by any Case-checking mechanism, regardless of its location in the grammatical model: if abstract WH elements do not have to be Case-marked, it follows that their failure to be Case-marked cannot rule out (14).
Clitics and the Government-Binding
Model
107
Although the ungrammatically of (14) cannot be derived from the Case filter, it can be derived from the fact that the structure in (14) is a control structure, coupled with the typology of empty categories. In Borer (to appear) we argue that the properties of control are derived from the anaphoric nature of the INFL node in infinitival clauses. It is the INFL node which inherits an index from a matrix controller. This index is then transmitted to the [NP,S] position of the infinitive (the PRO) by a general rule which coindexes the [NP,S] position and INFL (see Chapter 6 for extensive discussion). A specific ARB-assigning convention will assign arbitrary index to INFL when no controller is available. If INFL is not assigned an index (either by a controller or by the ARB convention) the sentence is ruled out. Now consider again the sentence in (14). The only possible controller for INFL is John. Thus, if INFL inherits the index of John, the embedded [NP,S] position will be coindexed with John by the transitivity of indexing. However, if the embedded [NP,S] position is coindexed with John it is Α-bound. If it contains a variable (i.e., an element which satisfies the definition in (4)), then the sentence is ruled out as a violation of the well-formedness condition in (5). If, on the other hand, the position is occupied by PRO, INFL will be controlled as required, but the sentence will be ruled out as a case of vacuous quantification: an (abstract) WH element which does not bind a variable (see Chomsky 1982 for a discussion). Finally, an anaphoric non-pronominal [e] cannot appear in that position since it will not be Α-bound in its minimal governing category. Thus (14), while definitely ungrammatical, does not bear on the choice between (1) and the system proposed here. Rather, the ungrammatically of (14) derives from the fact that the embedded subject position must be controlled (via INFL) and the fact that variables must not be controlled. Similar considerations will rule out the sentence in (15)
*the man| that it is impossible [e]j to solve the problem
In (15) INFL in the embedded infinitive is interpreted by the ARBconvention. The interpretation of fronted WH elements, however, is incompatible with arbitrary reference, as is illustrated by the contrasts in (16X17): (16) (17)
one usually considers oneself very virtuous a. *who usually considers oneself very virtuous? b. v/ho usually considers himself very virtuous?
(The example in (16) illustrates that the contraint on the occurrence of
108
Parametric
Syntax
arbitrary referents is not structural). The assignment of arbitrary reference to the variable in (15) would thus result in immediate ungrammatically. We do not take a stand here on whether the impossibility of arbitrary reference is a condition on the fronted WH element or on the variable itself, but note that the ungrammaticality of (15) will follow in either case. Now consider a second class of potential counterexamples. These are cases in which the antecedent is clearly Case-marked, the variable is not Case-marked and the sentence is ungrammatical. The cases we have in mind are free relatives in which the extraction is from the subject-of-theinfinitive position. Thus compare the (a) sentences in (18)-(19) to the (b) sentences: (18)
a. a man to fix the sink came yesterday b. *whoever to fix the sink came yesterday
(19)
a.
pakid le-tayek 'et ha-mixtavim yagi'a maxar clerk to file acc. the-letters will-arrive tomorrow 'a clerk to file the letters will arive tomorrow' b. *mi (se-) le-tayek 'et ha-mixtavim yagi'a maxar who that to-file acc. the-letters will-arrive tomorrow
In what follows we will show that infinitival relatives are largely irrelevant to claims in this section. Note, firstly, that some independent restriction blocks the sequence WH-to-VP, both in English and in Hebrew. Thus infinitival free relatives corresponding to extraction from object (Case-marked) positions are ungrammatical as well (and compare with the grammatical non-free relatives): (20)
a. a toy to give to the baby was on the table b. *whatever to give to the baby was on the table
(21)
a.
ca'acu'a la-tet la-tinok haya munax 'al ha-äulxan toy to-give to-the-baby was lying on the-table 'a toy to give to the baby was lying on the table' b. *ma (se-) la-tet la-tinok haya munax 'al ha-Sulxan what that to-give to-the-baby was lying on the-table
The contrast between extraction from subject position and non-subject position surfaces again when the "offensive" sequence disappears under further embedding: (22)
a. *whoeverj it is possible fg [e]j to fix the sink] b. whateverj it is possible fg PRO to give [e] j to the baby]
Clitics and the Government-Binding (23)
Model
109
a. *mij se-'efsar fg [e] i le-tayek 'et ha-mixtavim] who that-possible to-file acc. the-letters b. ma- se-efsar PRO la-tet [e]j la-tinok] what that-possible to-give to-the-baby 'whatever it is possible to give to the baby'
However, further embedding will also render ungrammatical the infinitival relatives in (18a) and (19a): (24)
a. *a man (who) it is possible to fix the sink b. *pakid se-ef$ar le-tayek 'et ha-mixtavim clerk that-possible to-file acc the-letters
The ungrammatically of (22a), (23a) and (24) follows directly from the fact that they are all structures of arbitrary control (see example (15) and related discussion). Similarly, if a control predicate was selected in (22a), (23a) and (24) (e.g., I told Mary replacing it is possible) (22a), (23a) and (24) would be ruled out on a par with (14). Note that this problem does not arise when extraction is from the object precisely because the object position is not coindexed with INFL and thus is not subject to control (and subsequent Α-binding or arbitrary reference). Thus the subjectobject assymetry in infinitival free relatives is reduced to another wellattested assymetry between subjects and objects in infinitival constructions: only the former are controlled. Note that the nature of the WH-to-VP constraint was not explained, and it is quite possible that once it is explained, the assymetry between subjects and objects will resurface in non-control structures such as (18) and (19). It is clear, however, that at this point it is impossible to construct examples which would bear on the choice between (1) and the analysis presented here. Furthermore, even if such cases could be constructed, it is not clear that they can be instrumental as an argument in favor of (1). Infinitival relatives such as (18a) and (19a) are prima facie counterexamples to (1): if their derivation involves the movement of an abstract operator to the COMP position from the [NP,S] position of the infinitive, that abstract operator does not leave a Case-marked trace which could serve as a variable according to (1). Given our analysis, however, the derivation of (18a) and (19a) by movement into COMP is non-problematic: although the trace is not Case-marked, it qualifies as a variable in accordance with (4). An analysis for infinitival relatives is outside the scope of this study. The purpose of the above discussion is to point out that although many of the properties of these constructions remain open for investigation, they do not supply evidence against our analysis or for the claim i n ( l ) . For
110
Parametric
Syntax
some analyses of infinitival relatives see Faraci (1974), Chomsky (1980, 1981), and more recently, Kirkpatrick (1982). 2.2. Α-Chains and the Visibility
Hypothesis
Chomsky (1981) argues that the requirement that variables have Case, which is a subpart of the biconditional in (1), follows from another principle of the grammar: the Visibility Hypothesis. Loosely stated, the Visibility Hypothesis is the assumption that elements of the form [^a] are 'invisible' to Θ-role assignment in the LF component unless they have a feature. Such a feature can be gender, number or person on the one hand, or Case on the other hand. Thus, for instance, PRO is visible since as a pronominal anaphor it contains features such as number, gender and person. Similarly, Case-marked traces are visible since they contain the feature Case. On the other hand, non-Case-marked traces do not have any features in the relevant sense and hence they cannot be seen. The latter conclusion is rather problematic with respect to the assignment of Θ-roles. Although one may plausibly argue that non-Case marked traces do not bear a Θ-role themselves (assuming (1) above such traces are always NP traces), nevertheless they are the elements that are in the particular position in which a Θ-role is assigned. Thus in a sentence such as (25) the Θ-role is assigned by the participle killed in the Θ-position immediately following it, although it is the full NP, Johny which fulfills the requirement that a Θ-role be assigned to every referential expression: (25)
Johri| was killed [e]j
Further, recall that the Projection Principle requires that all lexical properties be present at every level of the derivation. Since the position occupied by NP-traces is usually a subcategorized position, it must be visible in LF in order to ensure the correct application of the Projection Principle. Thus non-Case-marked traces are visible in some sense to the rules which assign Θ-roles at LF. Yet another problem for the version of the Visibility Hypothesis presented above is the assumption advanced in Chomsky (1981) that all empty NP elements (PRO, NP-trace, variable) are different tokens of the same type of category and their different properties are determined by the various contexts in which they occur. The following are the definitions of the environments which distinguish different empty elements: (26)
a. α is an empty category if a = J ^ p F], where F C φ, F non-null, b. i. a is a variable iff it is locally A-bound and in an A-position. ii. if α is not a variable then it is an anaphor.
Clitics and the Government-Binding
Model
111
c. if α is free or bound by a local Α-binder in a Θ-position then it is a pronominal. d. if a is locally Α-bound by an antecedent in a non-Θ-position then α is a non-pronominal anaphor. (As the reader will no doubt notice, (d) in (26) is stated redundantly and may be derived from (a-c). It is stated here explicitly for expository purposes). The set of features ψ referred to in (26) are features such as number, gender and person.5 Recall that by an Α-position we refer to a position in which an argument may appear at D-structure (essentially, [NP,S], [NP,VP], [ΝΡ,ΡΡ] and various specifier positions). Note that the definitions in (26) also capture the character of PRO as a pronominal anaphor and hence its properties with respect to the binding conditions discussed in Chapter 1. If all the empty elements have φ features it is no longer clear in what sense NP trace is less visible than a variable. Thus the crucial property for visibility is Case for variables and independent Θ-role for PRO. This latter assumption amounts to predicating the assignment of Θ-role to PRO in LF on the fact that it has an independent Θ-role. 2.2.1. A-Chains and the Binding Conditions In order to capture the observation above, Chomsky proposes that the Visibility Hypothesis applies to A-chains rather than to isolated elements. The definition of an Α-chain is given in (27): (27)
C = ( 1 a. Oj is a non-pronominal empty category; or b. Q!j is A-free. iv. C is maximal (i.e., is not a proper subset of a chain meeting i-iii).
The definition of Α-chains as it appears in (27) intends to cover two kinds of chains which have somewhat different properties. The first kind is a chain headed by a lexical NP and composed of the lexical NP itself and its trace(s), if it has such traces. For this kind of Α-chain the definition in (27), intuitively speaking, states that an antecedent constitutes a functional unit with the traces it binds. The second kind of chain defined by (27) is a
112
Parametric
Syntax
chain which is headed by a pleonastic element in subject position (either PRO or a phonologically realized pleonastic element such as it or there in English) which is coindexed with a postverbal position (an NP or a clause). In the latter case the coindexing relation which holds between the pleonastic element and the coindexed element does not enter into the binding conditions (but see Safir, 1982, for a different approach). It is a chainforming relationship which is henceforth referred to as co-superscripting. If one assumes the notion BIND which generalizes over binding relations and co-superscripting relations, then the definition in (27) applies to both types of chains. (The co-superscripting chain is not relevant to our discussion in these sections. We will return, however, to co-superscripting relations and what they stand for in Chapter 6). The Α-chain as a whole is now the unit which satisfies various lexical requirements in accordance with the Projection Principle. This is captured by the following principle: (28)
The chain C = ( a j , . . . ,a n ) has the Case Κ iff for some i, a j occupies a position assigned Κ by β.
(29)
Suppose that the position Ρ is marked with the Θ-role R and C = (öj,... ,α η ) is a chain. Then C is assigned R by Ρ iff for some i, a j is in position P.
Combining the definition of an Α-chain in (27) with the Projection Principle ensures the right application of'Move a ' . Since at D-structure lexical specifications have to be met, and since the binding conditions and the notion BIND are only relevant at S-structure, it follows that chains cannot satisfy lexical requirements at D-structure. Rather, at D-structure lexical specifications have to be met by the NP's themselves, generated in the position required by the lexical specifications. It follows that at D-structure every Θ-position must be filled by a referential expression and every referential expression must be in a Θ-position. On the other hand, at S-structure chains can be formed. Thus at that level A-chains can satisfy the lexical requirement although the referential expressions themselves may no longer be in Θ-positions. Following the principle in (29), the chain as a whole can satisfy the requirement that Θ-role assignment be met at every level. In this sense S-structure can be factored into D-structure and 'Move a'. The existence of 'Move a ' , on the other hand, can be derived from the different properties of D-structure and S-structure: whereas in the former the relationship of antecedenttrace is missing entirely, it is represented in the latter. We are assuming that the binding conditions hold at S-structure. This is evidenced by the following contrast (these arguments are from Chomsky
Clitics and the Government-Binding
Model
113
1981, who, in turn, credits them to M. Brody and D. Sportiche): (30)
a. which book that John read did he like? b. he liked every book that John read c. I don't remember who thinks that he read which book that John likes d. John said that Bill had seen HIM. (HIM with focal stress)
In (30a) WH movement
applies in the syntactic component and the
representation of (30a) at S-structure is essentially as given in (30a). In this representation he can be understood as coreferential to John, a fact that follows in a straightforward fashion from the binding conditions: following WH movement, John no longer c-commands he; thus coreferential interpretation is allowed. In (30b-c), on the other hand, such a coreferential reading is blocked. (30b) is a case of quantifier raising, whereas (30c) is a case o f WH movement in LF. The LF representation of these two sentences is given in (31 a-b): (31)
a. for every book χ that John read, he liked χ b. I don't remember for which person y and which book Λ: that John liked, y thinks that he read χ
Note that with respect to c-command, the configurations in (31a-b) are identical to the representation in (30a). In neither case does John ccommand he. So if the binding conditions hold in LF, we expect coreferential reading between John and he to be possible in these cases. Nevertheless, such a coreferential reading is impossible. If, on the other hand, we assume that the binding conditions hold at S-structure, the impossibility of coreferential reading in (30b-c) will follow immediately: at S-structure, he c-commands John both in (30b) and in (30c) and thus the coreferential reading is impossible. A similar argument can be constructed for (30d). In (30d) it is possible to have a coreferential reading between John and HIM. This follows from the fact that at S-structure, HIM in (30d) is a pronoun and thus it can be coreferential with an NP outside its governing category. On the other hand, assuming a rule of focus raising in LF, the LF representation of (30d) is given in (32). HIM is replaced by a variable and variables have to be free. So if the binding conditions held at LF we would expect the coreferential reading between John and HIM to be impossible: (32)
for χ = he, John said that Bill had seen χ
While the contrast between (30a) and (30b-c) presents evidence that the
114
Parametric
Syntax
binding conditions must hold at S-structure, the grammatically of (30d) under the coreferential interpretation of John and HIM provides evidence against locating the binding conditions in LF. From a conceptual point of view it is desirable to assume that the binding conditions hold at S-structure. Since the antecedent-trace relations are an inherent property of S-structure, we expect the binding conditions, which are an extension of these relations, to hold at that level. Recall now that the notion of Α-chain as defined in (27) crucially utilizes the relation BIND, which is composed of binding and co-superscripting. Since the binding conditions hold at S-structure, we will assume the BIND relation to hold at S-structure as well. It follows that superscripting relations, regardless of the level at which they are established, are checked at S-structure. Thus Α-chains are formed at S-structure and not at LF. This conclusion is quite natural: given the Projection Principle, Achains have to exist at S-structure in order to satisfy lexical requirements. 2.2.2. The Visibility Hypothesis Chomsky's notion of Visibility would be formulated now as an additional condition on (29): the principle of Θ-role assignment to chains: (29')
and C has Case or is headed by PRO
Note that an important consequence of the definition of variables in (4) and the definition of Θ-role assignment in (29) is that variables have to head their Α-chains themselves. This follows from the fact that variables cannot be Α-bound (if they are Α-bound they violate (5) above). If one assumes the additional condition in (29'), bearing in mind that the definition of A-chains requires that the A-chain be headed by an element in an Aposition, it follows that Case-marking on the antecedent WH would not suffice to make a variable visible. Therefore it follows from (29') that the variable itself has to be in a Case position in order to be visible. In this way the effects of the principle in (1) can be derived from left to right (i.e., if [e] is a variable then it must have Case). Let us now consider the properties of the Visibility restriction as expressed in condition (29') on the assignment of Θ-roles in (29). Note that this condition includes a rather unnatural disjunction between PRO and Case, which is clearly undesirable. We have further pointed out that the condition in (29') coupled with the contextual definition of PRO given in (26c) makes the assignment of Θ-role to PRO in LF dependent on the fact that PRO has an independent Θ-role, a rather unwelcome consequence. A greater problem for the Visibility Hypothesis in (29'), however, is its mode of interaction with the Case filter. Given the Case filter in the
Clitics and the Government-Binding
Model
115
phonological component, as we have assumed thus far, it is clear that the Visibility Hypothesis is designed to block precisely one sort of configuration, namely a configuration in which a variable is in a non-Case marked chain. In order to ascertain this consider all the types of NP's covered by the Visibility Hypothesis: lexical NP's, PRO, Case-marked empty categories and non-Case-marked empty categories. Lexical elements, if they do not have Case, will be ruled out by the Case filter, quite independent of whether they are visible in LF or not. Thus the Visibility Hypothesis is not required in order to rule out nonCase-marked occurrences of lexical NP's. PRO is visible by stipulation, as is stated in (29'). Thus the Visibility Hypothesis cannot be utilized to rule out ungrammatical occurrences of PRO. Now consider non-Case-marked traces. First, consider a non-Case marked trace which is not a variable by the definition in (26b). Such an element, if not a variable, is a non-pronominal anaphor and hence it has to be bound, in accordance with the binding conditions. In order to be bound and nevertheless violate the Visibility condition it has to be bound by an NP in a non-Case position, as in (33): (33)
*it seems [g Maiy^ to be left alone [e]j]
In (33), the non-Case marked trace is bound and nevertheless it is a chain which does not have a Case. Hence it is not visible. However, (33) is clearly a violation of the Case filter, since the antecedent in the A-position, Mary, is not Case-marked. Alternatively, a non-variable trace may be bound by PRO. But since PRO is visible by stipulation, the grammaticality of such constructions does not come as a surprise. Now consider non-Case-marked traces which are variables by the definition in (26b). These cases are ruled out by the Visibility Hypothesis. These are, in fact, the only cases which are ruled out by the Visibility Hypothesis. Given this state of affairs the Visibility condition is equivalent to a condition which restricts the distribution of variables in LF and which is roughly as in (34): (34)
Variables have to be in Case-marked chains
Since variables must head their Α-chains (see discussion above), they must themselves be in the position in which Case is assigned. Thus (34) is exactly equivalent to the left-to-right subpart of (1). In order to extend the explanatory power of the Visibility Hypothesis, Chomsky (1981) suggests that the Visibility Hypothesis subsumes the Case
116
Parametric
Syntax
filter. Thus instead of having a filter which is best characterized as a morphological filter, the requirement that NP's have Case should be regarded as a well-formedness condition on the assignment of Θ-roles in the LF component. This approach crucially entails that the Case filter holds for Α-chains only, since non-A-chains do not have to be assigned a Θ-role. It further entails that Α-chains which are composed of pleonastic elements are not subject to the Case filter. Within such a system all Case assignment is prior to the LF component and there are no Case assignment rules which apply in the phonological component. In fact Chomsky assumes that there are no Case assignment rules as such. Rather, lexical NP's are base-generated with Case features which are then checked at S-structure. Note that this assumption is compatible with the assumption that WH elements, which are not part of Α-chains, do not have to have Case: at S-structure the WH words are in COMP and the Case assignment is checked on the variable which is left behind. The constructions discussed in this study indicate that this approach is inadequate. We have shown that the differences between free relatives and questions in Modern Hebrew can only be explained if we assume that WH elements have to be Case-marked. Thus the Case filter has to hold for WH elements in spite of the fact that they are not part of an A-chain. Note that this result is quite desirable. In many languages WH-elements are overtly Case-marked, as are topicalized elements and dislocated elements. None of these elements are part of A-chains but nevertheless they have to be Case-marked. The prediction that pleonastic Α-chains do not have to be Case-marked is easily falsified as well; thus consider the following sentences in Dutch and German: (35)
a.
er there 'there b. *er there
(36)
a.
wordt gedanst was danced was dancing' gedanst te worden danced to be
es wird getanzt it was danced 'there was dancing' b. *es getanzt zu werden it danced to be
The pleonastic element in the subject position in (35) and (36) must be Case marked, as is proven by the ungrammatically of (35b) and (36b), where it occurs in the Caseless subject position of an infinitive. Nevertheless it is clear that no Θ-role is assigned to it.
Clitics and the Government-Binding
Model
117
Turning now to the prediction that all Case-marking rules apply prior to the LF level, recall that we supplied ample evidence that the Case marker sei is inserted in the phonological component and is not present at S-structure. In (37), the presence of sei at S-structure would lead to a violation of the binding conditions: (37)
a. re'iyat 'acmaj sei ha-moraj view herself of the-teacher 'the teacher's view of herself b. ha-re'iya sei ha-moraj 'et 'acma^ the-view of the-teacher acc herself 'the teacher's view of herself
We will return to the insertion of sei in the phonological component in Chapter 4. At this point it suffices to say that we find the prediction that Case assignment rules must apply prior to the phonological component incompatible with the evidence presented in this study. To conclude, it is our firm belief that the Case filter should be viewed as a morphological operation located in the phonological component. Thus we find the assumption that various Case-marking rules operate in the phonological component quite natural. Given the location of the Case filter in the phonological component it is clear that it cannot be derived from the Visibility Hypothesis. Thus the Visibility Hypothesis performs one and only one task: it derives a subpart of the generalization in (1). However, we have shown previously that (1) is a false generalization and that the definition of variables as in (4) is sufficient and more descriptively adequate than (1). We thus conclude that the Visibility Hypothesis can be dispensed with, and that variables need not be Case-marked.
3. CLITICS AND THE BINDING CONDITIONS
What is the status of the [e] in clitic constructions with respect to the binding conditions? Interestingly, given the typology of empty categories in (26), the [e] in the argument position of clitic constructions may be either a variable or an anaphor. Thus consider the minimal pair in (38): (38)
a. mij se-'axot-oj [e]j hayta xola who that-sister-his was sick 'one whose sister was sick' b. 'axot-oj [e]j hayta xola sister-his was sick 'his sister was sick'
118
Parametric
Syntax
According to clause (bi) of (26), (and see also (4)), [e]j in (38a) is a variable. According to clause (bii) of (26), [e]j in (38b) is an anaphor. In section 2 above we discussed extensively some of the properties of the variable [e]j in (38a)-type configurations. In this section we will present evidence that it functions as a variable with respect to the binding conditions and the weak crossover phenomenon. We will then turn to [e]j in (38b) and show that it behaves as a pronominal and not as an anaphor. This fact will lead us to the conclusion that the configuration [clj [e]j] is a discontinuous pronominal element, defined on thematic grids. This treatment will enable us to offer a natural explanation for some properties of the reflexive clitic in Romance. The behaviour of [e] in clitic constructions with respect to the binding conditions will indicate the untenability of recent proposals that clitics are A-binders. 3.1. Variables Binding condition C requires variables to be Α-free, due to their name-like properties (this particular requirement of binding condition C is reproduced in this chapter as the well-formedness condition in (5)). The sentence in (39) demonstrates that the [e] in (38a)-like constructions obeys the strong crossover constraint, a derivative of binding condition C: (39)
a. *kol mij se-keday se-huj yityaded 'im miipaxt-Oj [e]j all who that-worth that-he befriend with family-his muzman la-mesiba invited to-the-party 'any person that it is worth it that he would befriend his family is invited to the party' b. kol mij se-keday se-huj/Davidj yityaded 'im all who that-worth that-ne/David befriend with mispaxt-Oj [e]j muzman la-mesiba family-his invited to-the-party 'any person that it is worth it that he/David would befriend his family is invited to the party'
While the extraction in (39a) is over the pronominal hu which is coindexed with the variable, in (39b) neither hu nor David are coindexed with the variable. The binding conditions predict the ungrammatically of (39a) if the [e]j in (39a) stands for a variable, since the variable is not free: it is Abound by huj. [e] j in (39b), on the other hand, is not Α-bound;hence we do not expect the sentence to be ungrammatical. Thus the contrast in (39) strongly supports the conclusion that [e] in clitic constructions is a variable if it is bound by an operator, in accordance with (4).
Clitics and the Government-Binding
Model
119
(40) demonstrates the weak crossover effect in clitics (for a discussion of the weak crossover phenomenon, see Higginbotham, 1980, Koopman and Sportiche, to appear, Chomsky, 1982 and references cited there). 6 (40)
a. *'av-ivj [e]j 'ohev 'et 'im-Oj sei kol yeledj father-his loves acc mother-his of every boy 'his father loves every boy's mother' b. *lo barur 'eyzo matana 'av-ihaj [e]j natan le-im-aj not clear which gift father-her gave to-mother-her sei 'eyzo 'isa? of which woman 'it is unclear which present her father gave to the mother of which woman'
We do not take a stand here on the nature and the source of the weak crossover phenomena. It is clear, however, that it can be used as a diagnosis for variables. Thus insofar as [e] in(38a)-like constructions is sensitive to weak crossover effects, it is clearly a variable. 3.2. Non-Variables The empty category in (38b) is clearly not a variable, since it does not confirm to (4) (or to (26bi)). Thus, it seems, it must be an anaphor. On the other hand, it is governed, and hence it cannot be a pronominal anaphor (PRO). Thus it appears that it is a non-pronominal anaphor. As such it is subject to condition A of the binding conditions and it must be A-bound in its minimal governing category. However, as evidenced by (38b), the empty category does not have an A-binder. It is further clear that the combination [clj [e]j] in (38b)-like constructions has a pronominal meaning and exhibits typical properties to pronominal elements. Thus consider the contrast in (41): (41)
a. *Rinaj xaiva 'ale-haj [e]j 'Rina thought about her' b. Rina^ pagsa 'et 'axot-aj [e]j Rina met acc sister-her 'Rina met her sister'
In (41a) the governing category for [e]j is the entire S. Subsequently the sentence is ruled out since a pronominal must be free in its minimal governing category. In (41b), on the other hand, we assume that the governing category for [e] j is the NP 'axot-a, 'her sister'. Subsequently [e]j may be coindexed with the subject Rina, which is outside its minimal
120
Parametric
Syntax
governing category (and compare (41a-b) to their English counterparts). 7 Further, an embedded clitic+[e] combination can refer freely to an NP in the matrix: (42)
Rinaj 'amra se-Dan 'Rina said that Dan
xas'av 'ale-haj [e]j thought about her'
In view of these facts it is plausible to assume that the combination [clj [e]j] is pronominal. In a sense, then, this combination is a discontinuous element in which the clitic supplies the gender, number, person and Case features, and the empty element supplies the relevant argument position. Let us try and formulate this proposal. Recall that in Chapter 2 we argued for a particular process of Θ-role assignment. Following that process, every complement-selecting head has a thematic grid with an empty slot for the referential index of the selected complement. The assignment of Θ-roles is achieved by transfering this index into the Θ-slot. Recall further that we assumed that when a clitic is attached to the head, it has to be associated with a thematic matrix of this sort in order to be wellformed. We then defined the Complement Matching Requirement as a condition on the well-formedness of thematic matrices: (43)
Given a thematic matrix T, *T if Τ contains referential indices i, j and ift
The Complement Matching Requirement ensures that the clitic and the complement will not carry conflicting indices. Let us now review the structure of thematic matrices. Recall that they have the structure in (44):
It is natural to assume that along with the referential index in (44), the NP transfers some vital semantic information. Note that this assumption is independently necessary if we expect the thematic process to account for selectional restrictions as well. Thus in a sentence such as (45a) the theme, 'the boy', will transfer along with its index some semantic information, e.g. +animate, in order to prevent the ungrammatical (45b), in which such features cannot be transmitted by the direct object: (45)
a. the tiger frightened the boy b. the tiger frightened sincerity
Clitics and the Government-Binding
121
Model
Given the nature of the combination [clj [e]j], no additional information is transferred with the index of the empty category. The only information which exists in a thematic matrix of the sort produced by this combination is the gender, number, person and Case which are part of the clitic. It is this thematic matrix which is in turn given a pronominal interpretation. It contains all the information and only the information which would be contained in the thematic matrix of a free (non-cliticized) pronominal form: gender, person and number markers, Case features and a referential index of an argument. The structure of the thematic matrix both for the [clj [e]j] combination and for a free standing pronoun is as in (46a) (and compare with (46b) which is the (partial) thematic matrix of the boy in (45a) above): (46)
a. [ χ
X,
[gender, number, person, Case]j
NP;
ι 02 b. [ γ frighten
"masc, sing, 3rd person, Case
NP:
+animate +human +young etc. i θ2 As a discontinuous pronominal element the combination in (46a) is no longer treated as a non-pronominal anaphor. Rather, it is a pronoun and as such. It is subject to binding condition B: it must be free in its governing category. The pattern illustrated by (41)-(42) follows. Crucially, in cases in which the [e] in the clitic constructions is a variable, the formation of a discontinuous element is impossible, since at D-structure the Θ-index transfered by the argument NP contains nonpronominal information. Thus e.g. in (47a) the quantifier contains the information +human while in (47b) the WH-word contains the information -human: (47)
a. 'ahavot-oj (sei) kol exadj love-his of every one 'everybody's love' b . harisat-Oj
(sei) m a j
destruction-it of what 'the destruction of what'
122
Parametric
Syntax
3.2.1. Clitic se The process suggested above for the formation of a discontinuous element can be extended to clitic configurations in other languages as well. Note in particular that it will also allow us to state idyosyncratic lexical properties which are associated with [clj [e] j] combinations. Thus consider the reflexive clitic se in French (and similar reflexive clitics in other Romance languages) as given in (48): (48)
a. Jean se lave 'Jean washes himself b. Marie s'habille 'Marie dresses herself
Both in the case of laver, 'to wash' and in the case of habiller, 'to dress' the verb takes a thematic object. Following our assumption that thematic requirements cannot be met by the clitic but have to be met by an element in an argument position, we would like to claim that the structure in (48a-b) is as in (49): (49)
a. Jean se lave [e] b. Marie s'habille [e]
It is the empty category which satisfies the requirement for a thematic object in (49a-b). The structure which we suggest for thematic matrices now enables us to state in a natural way the fact that the combination se + [e] is assigned an anaphoric reading. The process by which this combination is interpreted as a discontinuous element will specify that grids such as (50) are anaphoric: (50)
NP=
The anaphoric interpretation given in (50) will then correctly rule out any occurrence of a lexical NP in (50) in the NPj position. Such an NP will be both an R-expression, which has to be free, and an anaphoric expression, by virtue of the particular interpretation assigned to (50). Thus it will have to be free and bound at the same time. On the other hand, [e] can freely appear in the NPj position in (50), since here it would not be a variable and thus not subject to binding condition C. One could argue that a lexical NP cannot appear in the NPj position in (50) due to Case absorption by se. This solution, however, is not tenable
Clitics and the Government-Binding
Model
123
since it does not account for the complete absence of clitic-doubling with reflexive clitics, even in languages which allow for doubling, such as Romanian and River Plate Spanish. Thus the only explanation for the obligatory absence of lexical NP's in that position is the nature of the combination se + [e]. Before we conclude this section, it is worth pointing out that the facts presented here strongly suggest that clitics cannot be viewed as A-binders (see, in particular, Aoun, 1982, who develops ideas of M.A.C. Huybregts). In this section we have presented evidence that the [e] in clitic constructions acts differently with respect to the binding conditions, depending on whether it is a variable or not. When it is a variable, it does not form a discontinuous element with the clitic and it conforms to condition C of the binding conditions. When it is not a variable, it behaves as a pronominal element. In either case the behavior of [e] is independent of the properties of the governing clitic and is determined solely by the contextual definition of empty categories in (26). An analysis which claims that clitics are Α-binders would have difficulty accounting for these facts. First and foremost, it would not predict the difference between variables and non-variables. Rather, both types of [e]'s should behave alike if clitics are Α-binders. Secondly, the behavior of [clj [e] j] combinations as pronominals would be completely unaccounted for. It is certainly not the property of Α-chains to act as pronominal elements. Rejecting the participation of clitics in binding relations is, on the other hand, extremely natural in our system. Since clitics are morphologically part of another word it would be surprising if they could enter binding relations which the word they are attached to does not enter. In fact, within our system clitics can only participate in thematic relations due to their attachment to thematic grids. To summarize this chapter, we have shown that some apparent conflicts between the Government-Binding model and our analysis of clitics can be solved. The reconciliation of these conflicts, however, motivated the preference of a simple definition of the notion variable over the one suggested in the Pisa Lectures and implied in Chomsky (1981). We reiterated the theoretical advantages of placing the Case filter in the phonological component and rejected the proposal of Chomsky (1981) to derive the Case filter from the Visibility Hypothesis. A clarification of the status of [e] in clitic constructions with respect to the binding conditions lead to the proposal that at D-structure, [clj [e]j] combinations are assigned pronominal interpretation. This conclusion, coupled with the variable properties of that [e] following extraction, rendered the assumption that clitics are Α-binders untenable.
124
Parametric
Syntax
These theoretical conclusions are assumed throughout the remainder of this study.
NOTES 1. The requirement that a be in S is designed to exclude empty elements in COMP from functioning as variables. We adopt this definition, including this condition, although nothing we say in this study bears on it. 2. Recall that sei can be inserted preceding the empty category. In this case, however, its Case features will be spelled-out as a clitic, resulting in the construction in (i): (i)
In (i) the clitic on sei absorbs the Case features of sei, and thus they cannot be assigned to Nj anyway. See Chapter 4 for further discussion. 3. Other well-formedness conditions constrain the occurrences of variables, e.g. some version of subjacency will determine the well-formedness of the operatorvariable relationship. We assume that these constraints are properties of levels of application and particular rules (i.e. 'Move a ' and QR) and should not be encoded into the definition of variables. 4. A problem for the lowering analysis which we will not discuss here is the status of the empty category in the subject position of the matrix following the lowering. Note that this [e] is governed, hence not a PRO, but it is not properly governed since AGR is not a proper governor (see Chapter 6 for discussion of this point). Therefore this [e] violates the ECP. For a suggestion that this position is an expletive [e] and that expletive [e]'s are not subject to the ECP, see Safir (1982). We are skirting here an important issue that concerns the level at which the binding conditions apply. If the binding conditions apply both at S-structure and at LF, as is argued by Aoun (1982) and others, then it is clear that the variable in quantifier lowering cases is Α-bound by the trace of the lowered quantifier in the matrix [NP,S] position. In order to overcome this problem Aoun (1982) and Aoun and Sportiche (1981) argue that the binding conditions apply for a variable only up to the binding operator. Note, incidentally, that if one assumes that the binding conditions apply at L F but that the formation of a chain between the matrix [NP,S] and the embedded [NP,S] in quantifier lowering cases is blocked by independent considerations, the transmissions of Case from the antecedent to the trace argued for in Chomsky (1981) becomes even more problematic, since at LF the two are no longer part of the same chain. We will return to chains in subsection 2.2.1. below. 5. By having φ features, empty elements such as PRO, NP-trace and variables differ from a null category. The latter is simply a non-expanded node which has no features at all. In Chapter 6 we will briefly return to this distinction. The null category is marked in this work as ( χ ], where X is the non-expanded category. 6. Weak crossover can be checked in Modern Hebrew only with quantifiers and WH-
Clitics and the Government-Binding
Model
125
in-situ. In order to see that this is so, consider the structure of syntactic WH-movement given in (i): (i)
mij sve-av-ivj [ej^ 'ahav 'et 'im-Oj [e]· who that-father-his loved acc mother-his 'one whose father loved his mother'
In (i), there is no way to determine whether movement occurred form the [e]j in the subject position or the [e]j in the object position. Note that only the latter will be ruled out by weak crossover. Aoun and Sportiche (1981), while arguing for clitics as Α-binders, cite sentences from Lebanese Arabic which are parallel to the sentences in (40). Interestingly, however, in Lebanese Arabic these sentences (with the intended coreference) are grammatical. At present we have no account for the difference between Lebanese Arabic and Hebrew. Note, however, that it does not weaken our point, i.e., that the [e] in (39) and (40) is a variable. 7. We refer the reader to Fiengo and Higginbotham (1981) and to Chomsky (1981) for extensive discussion on the selection of a governing category for sentences such as (41b). For our purposes it suffices that the clitic+[e] combination clearly patterns here with pronominal expression and not with anaphoric expressions.
Chapter 4
Inflectional Rules
1.
INTRODUCTION
In this chapter we will give a systematic account of variations found in clitic constructions. In particular we will concentrate on clitics in Romanian and Modern Hebrew, attempting to show that the differences between some aspects of cliticization in these languages are attributed to the operation of inflectional rules in the sense discussed in Chapter 1. In particular we will show that although Romanian, like Hebrew, allows for extraction from clitic-doubling constructions, a crucial difference in the distribution of such extraction follows from the different properties of the Case marker pe in Romanian vs. the Case marker sei in Modern Hebrew. The restriction of the application of the sei insertion rule, in accordance with the characteristics of inflectional rules discussed in chapter 1, will account for these differences. Furthermore, some peculiar binding properties of Hebrew and Romanian will be given a natural explanation as a result of the distinct properties of sei and pe. 1.1. Clitic Doubling and Extraction in Romanian The basic paradigm of clitics and clitic-doubling in Romanian (based on Steriade 1980) is given in (1): (1)
a. lj-am väzut pe Popescuj him-have-I seen OM Popescu (OM = Object Marker) Ί have seen Popescu' b. 1-am väzut him-have-I seen Ί have seen him' (We return below in sections 2 and 3 to the nature of pe in (1).)
Clitic doubling is attested in Romanian in certain environments which are semantically specified (see Chapter 5 below for similar semantic restrictions
128
Parametric
Syntax
operating in River Plate Spanish). The clitic doubling environment is a subset of the environments in which the preposition pe is available (e.g., doubling entails pe, but pe does not entail doubling). In these environments doubling is obligatory. The semantic environments are specified in (2)-(3), and exemplified by (4)-(6):
(2)
a. Γ+specific Ί am l+pronominalj I-have
vazut-Oj seen-her
b.r+specific nlj-am väzut +human him-have-I seen L-pronominalJ c. Γ+specific -definite L+human J
(3)
pe
ea i her
Popescu^ Popescu
Oj caut pe ο fatä her I-am-looking-for a girl de la noi din satfrom our village
a.r-specific η caut -definite I-am-looking-for L+pronominal J b. -specific -definite +human -pronominal_
pe
caut I-am-looking-for
c. Γ+specific "1 am I-human I I-have l-pronominalj
pe
altcineva somebody else
un bucätar a cook
väzut cmele lui Popescu seen the-dog of Popescu
(4)
lj-am väzut pe Popescuj him-have-I seen pe Popescu b *lj-am väzut Popescuj c.*am väzut pe Popescu
(5)
a. am väzut un bucätar have-I seen a cook Ί have seen a cook' b *am väzut pe un bucätar have-I seen pe a cook c.*lj-am väzut pe un bucätar? it-have-I seen pe a cook Ί have seen a cook'
Inflectional
129
Rules
d *lj-am väzut un bucätaij it-have-I seen a cook (6)
a. am väzut pe altcineva somebody else b *am väzut altcineva c.*lj-am väzut pe altcinevaj
As can be seen from (2), clitic doubling occurs only in the cases in which the NP direct object satisfies both requirements, i.e. when it is [+human/ pronominal] and [+specific/definite]. This is the case in (4a) but not in (6c). (Clitic doubling in Romanian occurs in indirect objects as well under certain conditions. These constructions, however, will not be discussed in detail here. For some discussion of dative clitics see Chapter 5, section 4.) The Romanian data clearly lend themselves to an analysis in terms of clitic spell-out of Case features and the availability of an independent Case assigner: it is plausible that the clitic in sentences such as (4a) absorbs Case. Consequently the NP position which the verb 'to see' subcategorizes for remains caseless. The availability of an independent Case-marking device, the marker pe, renders clitic doubling possible. Clitic-doubling seems to occur in post-extraction configurations in Romanian as well. As a generalization, it can be demonstrated that postextraction clitics appear when the extracted object NP satisfies the [+specific/definite] requirements. This is shown in (7)-(9): (7)
a. b.
casa pe care credeai the-house which thought-you pe care credeai cä which-one thought-you that
(8)
a.
pe cine who b. *pe cine who
(9)
a.
ce what b. *ce what
credeai thought-you credeai thought-you
cä that cä that
cä am väzut-o . that have-I seen-her am väzut-o? I-have seen-her
am I-have am 1-have
väzut? seen väzut-o? seen-her
credeai cä am väzut? thought-you that I-have seen credeai cä am väzut-o? thought-you that I-have seen-it
With respect to the differences between (7) on the one hand and (8) and (9) on the other, Steriade states the following:
130
Parametric
Syntax
The difference between (6b) (=7) and (6c) (=8 and 9) is that in [(7)] the question word quantifies over a set of known membership ... it is appropriate only in a context where the common background of the conversation includes the information that the referent of 'you' has seen at least one of a previously mentioned set of objects. One overt indication of this is that the ... clitic of a definite question like [(7)] agrees in gender with the NP that constitutes the previous mentioned set in question: thus from [(7)] we can gather that the set has been referred to by a noun whose grammatical gender is feminine.
This obviously does not hold for (8) and (9), where the set to which the questioned element belongs is not known and thus cannot be conceived as specific. It follows that only in (7) does clitic doubling take place, not in (8) and (9). Let us assume for a moment that the requirement for [+human/pronominal] in these configurations is met by the WH word itself which is fronted to COMP and which is considered a pronominal element (but see the discussion in section 4 below). We will now turn to the analysis of the post-extraction configurations. Sentences (7a-b) seem at first glance to utilize a resumptive pronoun strategy. However, if one wished to advocate such an approach to these configurations, two serious problems would immediately present themselves: first, why is the resumptive pronoun strategy available in precisely the same environment in which clitic doubling is allowed? The second question concerns the unavailability of subjacency violations in sentences such as (10): (10)
*omulj pe carej oj-cunosc pe femeiaj carejtj lj-a the-man OM whom her-know-I OM the-woman who him-has intflnit tj a venit met has come 'the man that I know the woman who met him came'
Ross (1967) observes that constructions which utilize the resumptive pronoun strategy may violate constraints on movement such as the Complex NP Constraint and the Island Constraints (subsumed by the subjacency principle of Chomsky, 1973). If the clitics in (7) were a real manifestation of a resumptive-pronoun strategy in WH constructions, we would expect (10) to be grammatical although it is a violation of the CNPC. Nevertheless (10) is ungrammatical, as are, systematically, all other sentences which contain an antecedent and a clitic inside a Complex NP or an island. This indicates that (7a-b) are generated by movement and not by a resumptive pronoun strategy. On the other hand, the analysis proposed for clitic doubling in Chapter 2 accounts straightforwardly for the grammaticality of (7) on the one hand and the ungrammatically of (8b), (9b) and (10) on the other.
Inflectional
Rules
131
Recall that the proposed structure of clitic doubling constructions is as follows: (11)
' [ χ cli.X]) Further recall that the clitic in (11) is a spell-out of the Case features of X, depriving the complement NP of its Case. An independent device is needed in order to assign Case to the NP if it actually contains lexical material. We assume that for Romanian X in (11) is a verb and pe, the direct object marker, is precisely the independent Case-assigning device we are looking for: it assigns Case to the complement NP if the C^_se features are spelled out as a clitic. Recall that extraction from the Nj position in (11) is possible and that indeed such extraction occurs in Modern Hebrew free relatives. Now consider again the sentences in (7), (8) and (9). The fact that clitics appear in post-movement configurations only when the extracted elements satisfy the semantic requirements of direct objects in cliticdoubling configurations is now explained entirely naturally; it follows from the fact that structures such as (11) are available only in cliticdoubling configurations: in these structures the clitic exists alongside the extracted NP. The empty category left after the extraction of Nj in configurations such as (11) is properly governed by the coindexed, governing clitic. This situation is illustrated in (12): (12)
VJ [y c f V]
Ν Wi
— proper government •
J
Extraction in Romanian is possible in non-clitic-doubling configurations, as is demonstrated by (8a) and (9a). Thus in Romanian verbs are proper governors, and the empty category in (12) is properly governed by V as well as by the clitic. Given this, Romanian cannot supply direct evidence for proper government by clitics. However, since the clitic clearly governs the coindexed NP position and is coindexed with it, it is unclear how such proper government could be blocked. We conclude that in the post-
132
Parametric
Syntax
extraction configurations in Romanian the [e] is properly governed twice, i.e., redundantly.
2. ON DIFFERENCES IN EXTRACTION BETWEEN HEBREW A N D ROMANIAN:
sei vs. pe. We have shown above that extraction from clitic-doubling constructions in Romanian is possible. In Chapter 2 section 3, we have shown that such extraction is possible in Modern Hebrew as well. Recall, however, that in Modern Hebrew such extraction is only possible in free relatives. Thus in Modern Hebrew we have the following contrast:
(13)
a.
kaniti maj se-xasavt 'al-avj [e]j bought-I what that-thought-you about-it Ί bought whatever you thought about' b. *sa'alti m a j xasavt 'al-avj [e]j asked-I what thought-you about-it Ί asked what you thought about'
The contrast between (20a) and (20b) was explained b y utilizing the Case filter (see Chapter 3 for discussion): Since the clitic in b o t h (13a-b) is a spell-out of the Case features of the preposition 'al 'about', the fronted WH element cannot receive Case from 'al. Since the Case filter applies to WH elements, every fronting from a non-Case position should result in ungrammatically, unless an independent device is available to assign Case to the fronted WH element. Such a device is Case assignment into COMP available (in Modern Hebrew) for free relatives but not for questions. Consequently the extraction from non-Case positions is grammatical in free relatives but not in questions, a situation illustrated by (13). There was, however, an important difference in Modern Hebrew between the examples in (13), in which the doubling construction is in a PP, and cases in which doubling takes place in NP's. For the latter but not for the former, there is a rescuing device: sei insertion. The availability of the Case marker sei and the fact that it can be inserted preceding the NP complement in clitic-doubling configurations inside NP's enables actual doubling to surface in NP's but not in PP's. One could raise the question whether, due to the availability of a Case assigner to the doubled NP in the construct state, we would expect the difference between questions and free relatives to disappear when extraction takes place from these constructions. However, the extraction from construct state NP's shows exactly the same distribution as ex-
Inflectional
133
Rules
traction from PP's: questions are ungrammatical and free relatives are grammatical. In fact sei cannot be extracted with the fronted WH, nor can it be left behind: (14)
a.
'anaxnu 'ozrim le-kol mij we help to-every who 'we help everyone whose house b. *sa'alnu mij beit-Oj [e]j asked-we who house-his 'we asked whose house burned' c. *sa'alnu sei mij beit-Oj
of d. *sa'alnu mij beit-Oj Sei of
[e]j
Se-beit-Oj [e]j that-house-his burned' nisraf burned
nisraf burned
nisraf
[e]j
nisraf
The properties of rule (15) account for the facts illustrated by (14): the rule which inserts sei operates in the phonological component and the environment for the insertion is dependent upon string adjacency. (15)
sei Insertion 0
(SI)
**el / [ n p . ι
NP- 1 J
Since in (14a-b) the environment for sei Insertion is not met, sei is never inserted and the status of extraction from NP's is made equivalent to the status of extraction from PP's. In Romanian, however, there is no such difference between free relatives and questions. When the fronted WH element satisfies the semantic requirements for doubled objects, both are possible. Thus alongside (7a-b) we have (16) as a specific free relative in which doubling is possible, and alongside (8a-b) we have (17a-b), demonstrating that when the free relative is non-specific doubling is blocked: (16)
am väzut-Oj have-I seen-her vazut-o [e]j
pe carej credeai cä am pe which-one thought-I that have-I
seen-her
Ί have seen whichever person you thought I have seen' (17)
a.
am väzut pe cinej credeai cä am väzut [e]j have-I seen pe who thought-you that have-I seen Ί have seen whoever you thought that I have seen'
134
Parametric
Syntax
b. *am väzut-Oj have-I seen-her väzut-Oj
pe cinej credeai cä am pe who thought-you that have-I
[e]j
seen-her (18)
a.
am väzut cej credeai cä am väzut have-I seen what thought-you that have-I seen Ί have seen whatever you thought that I have seen'
[e]|
väzut-Oj c e j credeai cä am väzut-Oj [ e ] j 2 b. *am have-I seen- it what thought-you that have-I seen- it her her
Note that the ungrammatically of (18b) cannot be related to the unavailability of the marker pe in these configurations; in (17b) pe is available, and yet the sentence is ungrammatical. Again, this situation is completely parallel to that of (8b) above: there doubling was impossible as well, regardless of the existence of the marker pe. The contrast between (8) and (17), in which pe appears, on the one hand, and (9) and (18), in which pe is absent, on the other illustrates the environments in which pe is present but clitic doubling is impossible: pe is present preceding direct object which satisfy all the requirements in (2) above. In addition to those environments, it appears preceding a direct object which is [+pronominal, -specific] and which is morphologically marked as [+human]. This latter requirement is demonstrated by the contrast between (19a) and (19b): (19)
a. caut I-am-looking-for b. caut
pe altcineva pe somebody-else (*pe) altceva something-else
We will assume that the environment in which pe occurs is a homogeneous semantic class, characterized as [+P], and that pe contains the semantic features [+P]. We will further assume that as a part of the interpretive component the [+P] features of the marker pe are checked against the availability of these features in the NP object of pe. A [+P] marker adjoined to a [-P] object results in ungrammaticality. On the other hand, a [+P] direct object which is not marked by a [+P] marker is ruled out as well. 3 The grammaticality of (8a), (17a), and (19a) indicates that pe is available in environments which do not allow for clitic doubling and in which there is no need for an independent Case marker. Thus clearly the occurrences of pe in the grammar of Romanian cannot be accounted for
Inflectional
Rules
135
solely by assuming a rule of pe insertion which operates in the phonological component and which assigns accusative Case to Caseless direct objects. Rather, it seems that clitic doubling in Romanian direct object configurations is a "byproduct" of the availability of an independently existing object marker. Recalling further that pe has [+P] semantic features and thus must feed the interpretive component, it is plausible that pe is available at D-structure in the [+P] environments illustrated above, and that it has - as one of its lexical specifications - the property of assigning accusative Case to its complement. We further assume that accusative Case assignment by pe is obligatory. Digressing briefly, recall that the same holds for sei (see Chapter 2, section 2.5 for a discussion). It is plausible that this is a general property of inserted Case assigners. Since the availability of grammatical formatives is conditioned by their function (unlike the insertion of regular lexical items, which is conditioned by their semantic function and their categorial selection), the obligatoriness of that function is natural. Below we will show that both the insertion of sei and the insertion of pe are inflectional rules (in the sense discussed in Chapter 1). Thus it would be natural to assume that the obligatoriness of the inflectional function is a property of all elements inserted by inflectional rules. In essence, the obligatoriness of Case marking in the case of pe would entail that whenever an element X has Case features a, a must be realized phonologically, either as a clitic or on a phonetically realized NP. Further assuming that Case conflict leads to ungrammatically, the accusative Case assignment features of pe will predict the ungrammatically of pe occurrences in all environments in which an NP is otherwise marked for non-accusative Case: (20)
*am dat cartea lui pe Popescu have-I given book to pe Popescu Ί gave a book to Popescu'
In (20), Popescu is marked as dative by the preposition lui 'to', and an additional accusative marking by pe rules out the sentence. (Note that we are tacitly assuming that double assignment of the same Case does not lead to ungrammatically. Thus in (19b) altcineva 'somebody else' is marked accusative twice: once by the verb and once by pe. There is, however, no reason to assume that such redundant marking is ungrammatical.) The availability of pe in the base makes a clear prediction: we expect pe tc be fronted alongside the NP which it precedes, and this prediction is confirmed. As we have seen in (7)-(8) above, pe is indeed fronted with WH elements. Since pe is available when WH fronting occurs from
136
Parametric
Syntax
clitic-doubling configurations such ad (7a-b), we expect such extraction to be entirely grammatical. Although the clitic absorbs the Case features of the verb, the WH element is marked for Case by pe. Hence there is no need for an independent device marking Case into COMP. In this fashion we can account for the difference between Modern Hebrew and Romanian: in the former, sei is not available in the base and hence cannot be fronted with WH elements; in the latter, pe is available and hence we expect both questions and free relatives to be grammatical.
3. ON THE INSERTION O F CASE MARKERS
In the previous section we have shown that sei differs from pe in that it is not available in the base and hence cannot function as a Case marker for fronted WH elements. We have further shown that pe in Romanian is available in environments in which it clearly does not function as a "rescuing device" for the purposes of the Case filter. In Hebrew, on the other hand, there are no such cases: all occurrences of sei, whether in cliticdoubling configurations or in other possessive constructions, fall under the generalization expressed by the environment in rule (15). 4 All these factors favor the hypothesis that, whereas sei is not available prior to the application of "Move a " , pe is available at that stage. In Chapter 2, section 2.5, we have shown that sei phrases act as NP's with respect to the binding conditions (see also (21a) below). They behave differently from objects of prepositions (as in (21b)) or from objects of an adjoined specificity marker (as in (21c)). Since in sei configurations sei insertion takes place in the phonological component, the structure of selphrases at S-structure is not branching, and the NP object of sei c-commands the reflexive anaphor. We have crucially claimed that in structures such as (22c) N2 (and not N j ) is an A-position which enters into the binding conditions. The structure in (22c), then, counts as a branching structure, and N 2 cannot c-command elements outside its projection: (21)
a.
re'iyat 'acma^ sei ha-moraj view herself of the-teacher 'the teacher's view of herself b. *xasivat 'acmaj 'al ha-moraj thinking herself about the-teacher c. *re'iyat 'acmaj 'et ha-moraj view herself acc the-teacher 'the teacher's view of herself
Inflectional (22)
137
Rules
a. (=(2la)) N2
Ni Nq
N,
the-teacherj
herself: (sei is inserted in the phonology) b. (=(2lb))
Νχ
*
N,
N3 herselfj
the-teacher ;
(OM = object marker) In this study we argue for a theory of parameters in which languagespecific variations are determined by the nature of inflectional rules and by their mode of application. The definition of inflectional rules is repeated in (23): (23) A.a. Let f stand for an inflectionally specified grammatical feature b. Let F stand for an assigner of f c. Let C stand for a constituent specified without a variable B.
An operation which affects the assignment of f to C, such that it is not subject to any condition exterior to the properties of f, F or C is an inflectional rule
The application of (23) is according to the principle in (24): (24)
An inflectional rule may apply at any level (lexicon, S-structure, PF)
138
Parametric
Syntax
(See Chapter 1 for a more precise formulation of the notion level as it applies to inflectional rules.) The principle in (24) is subject to parametrization in a particular way: the application of an inflectional rule in a given grammar could be restricted at certain levels. The pattern of such a parameter would be as in (25): (25)
For R, R an inflectional rule, R may not apply at level L
A certain statement on markedness is incorporated into the system proposed by (23)-(25): a more marked system is the one which restricts the application of R. A less marked system is the one which allows R to apply freely at any level. This statement of markedness, however, was modified to some extent (see Chapter 1 for a discussion): we assumed that when C has no semantic content, C is not available at the lexical level. An example of a parameter consistent with our system is the pro-drop parameter discussed in Chomsky (1981) (and see Chapter 6 for a detailed discussion). The insertion of dummy Case markers as well as the insertion of specificity markers is another instance of this system. Given our analysis of sei and pe so far, it would seem that sei is restricted to apply only in the phonology, and that pe can apply only in the base. There is, however, some evidence indicating that the correct formulation of the insertion of these two markers is considerably less restricted. In the following two sections we will argue that although sei insertion is barred from applying in the lexicon due to the lack of semantic content for sei, it is otherwise unrestricted and may apply either in the syntax or in the PF component. The marker pe, which has the semantic feature [+P], may be inserted in the lexicon. Further, it may be inserted at any level. The formulation of pe insertion is given in (26) (and see (15) for the formulation of sei insertion): (26)
pe Insertion (PI) 0
> pe / [
w
NP]
3.1. Sei Insertion at S-structure The evidence presented so far to the effect that SI applies in the phonology ((21) and (22) above) is entirely compatible with the assumption that it may apply optionally either at S-structure or in the phonology. Only in the latter case would we have the structure in (22a) at S-structure, and thus (21a) and (22a) would be grammatical only if Si applied in the phonology. There is, however, an additional derivation which yields a structure similar to that in (22c). Under this_derivation (21a) is ungrammatical since N 2 cannot be the antecedent of N 3 .
Inflectional
Rules
139
Interestingly, there is some direct evidence that, in some cases, Si must take place at S-structure. These are cases in which application of §1 in the phonological component would lead to ungrammaticality due to independent factors, but where insertion at S-structure would result in grammaticality. In Chaper 2, footnote 7, we briefly mentioned that sei may itself take a clitic. This is illustrated in (27): (27)
ha-xataltul fel-o / Sel-a / sel-i etc. the kitten of-him/ of-her/ of-me 'his I her / my kitten'
In view of the analysis of clitics argued for in this study, the proposal for the structure of the sei + clitic combination should be identical in the relevant sense to the structure proposed for combinations such as preposition+clitic. This structure is given in (28): (28)
a.
P-i P+clj
b. Nj
(Recall that we are assuming that sei is adjoined to N, as are specificity markers.) For (28a), we argued that the clitic is available at the LF component, since it can function as a proper governor for Nj if the latter dominates an empty category. This analysis was used to account for the grammatically of (29a) as opposed to the ungrammaticality of (29b). (See Chapter 2, section 3, for a detailed discussion.) (29)
a.
mij se-xasavti 'al-avj [e]j who that-thought-I about-him 'whoever I thought about' b. *mij se-xasavti 'al [e]j who that-thought-I about 'whoever I thought about'
Since in Hebrew prepositions are not proper governors, only the availability of the clitic in (29a) makes the extraction from PP's possible. Otherwise the output structure as in (29b) is ruled out by the ECP. If, indeed, (281?) is the right structure for se/+clitic combinations, we would expect the N t node in (28b) to be expanded as an empty category in cases such as (27). And in such cases we would expect the coindexed clitic to properly govern this empty category. If it could be shown that
140
Parametric
Syntax
the clitic adjoined to sei does in fact properly govern an empty category, this will indicate that the clitic has to be present at LF, and hence that it is also present at S-structure. Otherwise it could not properly govern an empty category. Given that the clitic is a spell-out of the Case features of sei, its existence at S-structure would indicate that sei itself is present at S-structure. Testing whether (28b) is the right structure for se/+clitic configurations can be achieved by extraction from Nj when the clitic is present. If such extraction is possible, it would indicate that the sei in se/+clitic configurations j s present at S-structure. And, as indicated by (30a), extraction from the Ν position in the se/+clitic configuration is indeed possible. (That proper government in this case is not by sei itself or by the head Ν is demonstrated by the ungrammatically of (30b-c.) (30)
a.
'anaxnu we 'we help b. *'anaxnu
'ozrim le-kol mij se-ha-bayit s'el-Oj [e]j nisraf help to-all who that-the-house of-his burned everybody whose house was burned' 'ozrim le-kol mij §e-ha-bayit sei [e] · nisraf that-the-house of burned c. *'anaxnu 'ozrim le-kol mij §e-ha-bayit [e]j nisraf that-the-house burned
Thus, when a se/+clitic configuration appears, S i must apply at S-structure. If we were to assume in this case that S i applied in the phonology, we would have the following configuration at S-structure: (31)
(= (30c)) N, the-house
N2 [e]
In (31) Si might still apply in the phonology. Since the Case features of sei have to be phonologically realized we would still derive at PF the combination sel+clitic. However, in this case the empty category in (31) would not be properly governed, since the spelling_out of the clitic at PF would i p t affect the application of the ECP, and N! could not properly govern N 2 , as is demonstrated by (30c). (This situation contrasts sharply with the situation in which N t itself takes a clitic. In this situation, resulting in the grammatical 'house-hisj [e]j' combination described in Chapter 2, the clitic itself properly governs [e] j ) Given the optional application of S i , either at S-structure or in the phonological component, we expect the combination se/+NP to be either
Inflectional
141
Rules
branching (as in (22c)), or non-branching (as in (22a)). Only in the latter cases, however, can the object of sei (N 2 ) serve as antecedent for a reflexive anaphor (N 3 ) outside its own maximal projection. On the other hand, the combination se/+clitic can only be a branching one: a non-branching combination would yield a structure such as (31), in which an empty category is not properly governed. Since se/+clitic combinations are always branching, we do not expect the empty category in structures such as (28b) to serve as an antecedent for a reflexive anaphor: we expect that the se/+clitic counterpart of (2la) would be ungrammatical, as, indeed, it is: (32)
*re'iyat 'acmaj se 1-aview herself of-her 'her view of herself
[e]j
Note that the rather puzzling contrast between (32) and (21a) can be explained straightforwardly if we assume that only in (32), but not in (21a), the sei phrase is branching. 5 This branching structure, which is independently needed to supply a proper governor for an empty category, has the effect of blocking the [e]j node in (32) from serving as an antecedent for a reflexive anaphor. 6 To summarize our discussion of se /-insertion: we have argued that Si can apply either at S-structure or in the phonological component. In the former case, the se/-object cannot serve as an antecedent for a reflexive anaphor outside its own maximal projection, but in the latter case the object of sei behaves as a regular NP and can serve as an antecedent for a reflexive anaphor. This difference stems from the fact that when Si applies in the phonology the structure is non-branching at S-structure. When Si applies at S-structure, it creates a branching structure, thus preventing its object from c-commanding elements outside its maximal projection. When a clitic is adjoined to the sei, S i must apply at S-structure. Its failure to do so would result in the structure (31), in which an empty category is not properly governed. It follows that when a clitic is adjoined to sei and the structure is branching, the empty category-object of sei can never serve as an antecedent for a reflexive anaphor. We thus conclude that the application of S i is not restricted to the PF component. Rather, S i may apply freely at S-structure and at PF. Its application at lexical level, however, is blocked as sei carries no semantic information. 7 3.2. Free Application of pe Insertion In this section we will show that the best characterization of the insertion
142
Parametric
Syntax
of pe is as a rule free to apply at any level where the insertion of grammatical formatives is possible: in the base, at S-structure, and in the phonological component. In section 2 we showed that the fronting of pe along with fronted WH elements indicates that it must be available in the base. Note, however, that our treatment of the fronting of pe in Romanian is entirely compatible with the assumption that pe is inserted in the base optionally rather than obligatorily. Consider again the cases of fronting of pe along with a WH element, as in (7b) above, repeated here as (33): (33)
pe carej credeai pe which-onej believe-you V h o do you believe I saw?'
cä that
am have-I
väzut-Oj? seen-her-
Now let us assume that in (33) pe insertion in the base is optional. It is clear, however, that if "Move WH" applied before pe insertion, the fronted WH element would no longer be able to receive Case, since it would no longer be in the environment of [ y p NP] in which pe is inserted. Thus, PI is effectively forced to apply in the base for (33) to be grammatical, but we do not have to assume that it obligatorily applies in the base. The derivation in which it does not apply in the base is independently ruled out by the Case filter. After the extraction, pe may still be inserted preceding the empty category. However, since the Case features of pe, like the Case features of sei, have to be phonologically realized, and since pe does not take a clitic (as in fact no non-verbal elements in Romance ever do), it can be inserted only in front of phonologically realized NP's: Only in these cases will its Case features be realized. Now consider the situation in free relatives in Romanian. The relevant sentence, (16), is reproduced here as (34): (34)
am have-I väzut-Oj
väzut-Oj seen-her
pe carej credeai cä am pe which-one thought-you that have-I
[e]j
seen her Ί have seen whoever you thought I have seen' If we argue that pe is consistently inserted in the base, we expect two pe's to appear in (34): the first pe resulting from the specificity and [+human] value of the free relative itself and the second one resulting from the fronting of a [+specific, +human] direct object. Rather than stipulating that a sequence pe pe is reduced to One pe, we prefer to claim that one of these pe's is simply not inserted. Since the matching effect requires that the free relative, as an NP, will satisfy the same
Inflectional
Rules
143
categorial requirements and the same semantic requirements as the gap, it follows immediately that one pe, inserted preceding the WH element in (34), suffices. This pe can be inserted in the base in the matrix only, inserted preceding the WH element (prior to fronting of pe+WH) only, or, alternatively, inserted before the free relative constituent at a later point of the derivation. Since the post-extraction configuration in free relatives (but not in questions) satisfies the environment for PI, insertion of pe at the base, at S-structure or in PF will result in grammatical outputs with respect to the Case filter. Recall, however, that pe has certain semantic features, previously marked as [+P]. We have to represent these semantic features in the interpretive component. Thus these features have to be present at S-structure. In other words, we expect pe to be present at S-structure unless [+P] is represented in LF in some other way. Further, recall that insertion prior to the phonological component would yield a branching structure at Sstructure, blocking the object of the inserted formative from serving as the antecedent to lexical anaphors outside its maximal projection. This makes it possible for us to test whether pe is inserted at S-structure or in the phonological component: if it is inserted at S-structure, we would expect its object to be restricted and not to be able to serve as an antecedent for a reflexive anaphor outside its maximal projection. If, on the other hand, pe insertion may take place in the phonological component, we expect the object of pe to be able to function as an antecedent for a lexical anaphor. Now recall that Romanian has three kinds of direct objects with respect to pe: the kind not marked at all by pe (exemplified in (35a)), the kind marked by pe but in which there is no doubling (exemplified by (35b)), and the kind in which there is pe and there is doubling, (exemplified by (35c)) (35)
a. Ion a arätat fetifa publicului John has shown the-girl to-the-public b. am väzut pe altcineva have-I seen pe somebody else Ί have seen somebody else' c. Ion a aratat-Oj pe fetijaj publicului 8 John has shown-her pe the-girl to-the-public 'John showed the girl to the public'
Doubling is a subclass of the cases in which pe occurs. This subclass satisfies [+P] and is further [+specific/definite], [+human/pronominal]. Assuming that PI can freely apply either at the base, at S-structure or in the phonological component, and further assuming that if it applies prior to
144
Parametric
Syntax
the phonological component its structure interacts with the binding conditions as a branching structure, our proposal makes a clear prediction, i.e. that in cases such as (35a) the direct object can serve as an antecedent for a reflexive anaphor. Since there is no pe insertion, the structure never branches. In (35b), however, the application of PI in the phonology will result in not representing [+P] in LF. We assume that this situation will result in ungrammatically due to independent interpretive considerations, thus effectively forcing PI in these cases to apply either in the base or at S-structure. We thus expect the object of pe in (35b) never to function as an antecedent for a reflexive anaphor, since it will always be part of a branching structure at S-structure. The situation in (35c) is somewhat different. Here [+P] is represented both by pe and by the doubled clitic, since all cases of clitic doubling are a subset of [+P] cases. Thus, if PI applies in the phonology, the required [+P] information is represented in LF by the doubled clitic. In these cases we expect the application of PI in the phonology to be grammatical as well and we expect the object of pe in (35c) to function as an antecedent of a lexical anaphor. These predictions are verified. The object in (35a) may serve as an antecedent, the object in (35b) may not and the object in (35c) may. These respective configurations are represented in (36a-c). 9 (36)
a.
Ion a arätat fetita- eij in^i^iin oglindä John has shown the-girl her/dat her/emphatic in mirror dat 'John showed the girl to herself in the mirror' b. *Ion a arätat rpe altcineva ;ι lui ;ι insusi > >
c.
John has shown pe somebody else him/dat him/emphatic dat in oglindä in mirror 'John showed somebody else to himself in the mirror' r Ion a arätat-o-1 pe fetita ei ;1 insisi; ' 1; ' '1 John has shown-her pe the-girl her her/emphatic ϊη oglindä in mirror 'John has shown the girl to herself in the mirror'
It could be suggested that (36b) owes its ungrammaticality to the fact that altcineva 'somebody else' cannot serve as an antecedent for a reflexive anaphor. Presumably its lack of specificity would contrast with the specification of gender on the reflexive pronoun. However, when the objective altcineva controls a PRO (which is in a non-branching con-
Inflectional
Rules
145
figuration), this PRO can serve as an antecedent for a reflexive anaphor, although it is equally non-specific. This situation is demonstrated in (37): (37)
Ion a väzut pe altcineva^ PROj vorbind cu John has seen pe somebody else talking to el insusij him/emphatic 'John saw somebody else talking to himself
When a pe phrase is fronted from a non-doubling position, it leaves behind an empty category which no longer contains a branching structure. Thus, in these cases as well we expect the occurrence of a lexical anaphor which is understood to corefer with the object of pe. Although pe is inserted in the base, the antecedent for the reflexive anaphor is not the object of pe itself, b u t the variable which is coindexed with it and which is not branching. And indeed in these situations, coreference between [e]j and the reflexive pronoun in (38) is grammatical: (38)
pe cine^ credeai cä Ion a aratat [e]j luij pe who thought-you that John has showrij [e] j him insusij in oglindi? him/emphatic in the-mirror 'who did you think that John showed to himself in the mirror?'
We conclude that pe can be inserted at any level: at thebase, at S-structure, and in the phonological component. The failure of PI to apply at any of these given levels will bring about the exclusion of certain configurations for which the application of PI at that level is crucial. So if PI failed to apply in the base, a fronted WH element in questions could not receive Case, resulting in ungrammatically. Effectively then, PI has to apply in the base for structures in which 'Move WH' applies. On the other hand, it has to apply at S-structure in configurations which do not contain any other way to render in LF the [+P] features associated with the direct object. These are precisely the cases in which there is no doubling, yet pe precedes the direct object. Only in these cases does pe contain crucial semantic information which cannot be deduced otherwise. However, when pe co-occurs with doubling configurations, the [+P] information is obtained by the doubled clitic, since clitic-doubling cases are a proper subset of pe-insertion cases. In doubling constructions, then, pe insertion is free to apply in the phonological component. Since only in cases which contain pe but no doubling PI has to apply at S-structure, we expect that here the object of pe could not serve as an antecedent for a reflexive anaphor: PI at S-structure results in a branching structure. And in fact this is correct:
146
Parametric
Syntax
in these cases altcineva 'somebody else' as in (36b) cannot serve as an antecedent. 10
4. FREE CLITIC SPELL-OUT
In section 1 above, we assumed that in extraction from clitic-doubling configurations in Romanian the WH element itself, the fronted element, served as an environment for clitic doubling. Note that such an analysis implies a mechanism that will check the appropriateness of clitic doubling in the base: under such an analysis, the environment for clitic doubling is only met in the base and not, say, in LF. However, the rule which spells out various features as clitics is an inflectional rule. As such it is free to apply at any level. Neither do we expect it to be sensitive to factors such as semantic environment. In addition to that the reader may recall that while discussing cliticization to sei, we assumed that, for sei, clitic spellout applies at S-structure or at PF. In the base, sei is never present since §1 cannot apply at the base. Thus it would be desirable for clitic spell-out to occur at any stage, freely and optionally. Mechanisms independently motivated in the grammar would then check the spelling out of the clitic for appropriateness. Such mechanisms are semantic requirements checked in LF (as is the case in Romanian), proper government and violations of ECP (as holds for clitic spell-out in Hebrew) or the Complement Matching Requirement discussed in Chapter 2. In this section we will present evidence that this characterization of clitic spell-out is the right one. It will be shown that the semantic requirements for clitic doubling are checked at a late stage of the derivation, such as LF, and that these semantic requirements should not be viewed as a triggering environment at any particular stage. Rather, they should be viewed as well-formedness conditions on interpretation. Steriade (1980) aigues quite convincingly that although at first glance the WH element seems to satisfy the semantic requirements for doubling prior to its extraction, this is in fact not correct. The cases which she cites as counterexamples to this statement are cases of nominal piedpiping. In these cases the constituent as a whole does not satisfy the semantic requirements for doubling, and hence, if it is not extracted, it does not trigger doubling. However, the WH element inside the nominal expression satisfies these semantic requirements. The relevant case is given in (39): (39)
a. Popescu mi a comunicat rezultatele studiului säu Popescu to-me has communicated the-results of-study his 'Popescu communicated to me the results of his study'
Inflectional
Rules
147
b . un studiuj [ale cäruij rezultate]j mi le^-a comunicat [e]j a study whose results to-me them-has communicated Popescu 'a study whose results Popescu has communicated to me' In (39a) we do not have clitic doubling. The reason is that the NP rezultatele studiului säu 'the result of his study' does not satisfy the semantic requirements for clitic doubling: although it is definite, it is neither human nor pronominal, and doubling is blocked. However, in the extraction construction corresponding to (39a), (39b), such doubling is attested and is in fact obligatory. A failure to have a clitic in these cases would result in ungrammaticality. Steriade suggests that the requirement for clitic doubling in these postextraction cases can be expressed if we assume that, rather than the fronted WH element it is the trace left behind which has to satisfy the semantic requirements for doubling. In order for the trace left behind to satisfy these requirements, she proposes a rule of shadow pronoun copying, essentially following proposals of Perlmutter (1972). Note that if the trace can indeed be perceived as pronominal, and if we assume that this trace retains the specificity feature of its antecedent, ale cärui rezultate 'whose results', it will satisfy the semantic requirements for clitic doubling. In essence we will adopt this analysis here: with Steriade, we will assume that semantic requirements on clitic-doubling are checked at LF. We will assume that there is a device at LF which fails to assign interpretation to clitic-doubling configurations unless the semantic requirements are met. We will deviate from Steriade's analysis only on one point: rather than assuming that there is a rule of shadow pronoun copying which assigns pronominal features to the trace of WH movement, we will assume that the relevant semantic features for clitic-doubling are inherently present. We assume with Chomsky (1981) that traces are marked for features such as person, gender and number. It is this specification which enables them to be perceived as satisfying the requirements for clitic doubling rather than the pronominal feature. In fact we assume that for pronominal elements as well the relevant semantic requirement is the presence of all and only the features number, gender and person, and that the disjunction [+human/+pronominal] should be replaced by the disjunction [+human/ α number, β gender, y person]. 1 1 Let us sum up our discussion of clitics in Romanian. It has been shown that the Romanian phenomena fit naturally into the analysis proposed for clitic doubling in Chapter 2. Furthermore, insofar as our analysis can explain extraction from clitic-doubling constructions in Romanian, Romanian supplies evidence for this analysis. While discussing the differences in extraction configurations between
148
Parametric
Syntax
Romanian and Modern Hebrew, we pointed out that these differences can be explained by utilizing the properties of the grammatical formatives sei and pe. It was suggested that the insertion of these formatives is the output of inflectional rules and that the application of inflectional rules may be restricted to apply at certain levels but not at others. Thus S i was restricted to apply either at S-structure or in PF but not in the base, whereas PI was not restricted and could apply at the base, at S-structure or in PF. Independent components of the grammar, however, forced PI and §1 to apply at certain levels rather than at others in order to yield a wellformed derivation. In these cases we expected different structural properties resulting from the different levels of the application of PI and SI. These structural properties were attested. Our last section dealt with the status of the semantic requirements on clitic doubling: it was shown that these requirements are best characterized as a device operating after the application of syntactic movement rules. Naturally we assume that it is located in the interpretive component: LF. In effect, locating this device in LF allows us a free spell-out of Case features (where such exist); independently motivated considerations then rule out ill-formed outputs at a later stage of the derivation. Such considerations are the ECP (as in the cases of [e] which is not properly governed) or semantic constraints. Other devices will be discussed in Chapters 5 and 6 below.
NOTES 1. (5c) is grammatical if un bucätar 'a cook' is specific. 2. The sentences in (16)-(18) also demonstrate that the 'matching effect' in the sense of Grimshaw (1977) holds in Romanian free relatives with respect to the pe marker (see Chapter 2, footnote 30 for some discussion of the 'matching effect'). So only when the matrix takes a pe-type object can there be a free relative where extraction is from a pe-type object. It follows that in free relatives doubling is attested both in the matrix and in the subordinate clause. If we assume that in (16) fronting has taken place, and that pe is fronted along with the WH element, we should end up with a sequence of two pe"s: one resulting from fronting and one base-generated in the matrix. We will argue below that since pe insertion is free to apply at any level (base, S-structure or PF) although the derivation which contains pe pe sequences may be ruled out, there is an alternative derivaiton in which only one pe has been inserted and which is grammatical. 3. The question of how to characterize the semantic environment which we have labelled [+P] is not a trivial one. A functional explanation of pe occurrences may shed some light on this matter: it seems that object markers are usually available in relatively free word-order languages, as a device to disambiguate sentences in which the direct object could plausibly be misinterpreted as the subject. Thus in (i) there is no ambiguity, even if the word order is free, since only John can satisfy the agentive reading which eat requires. On the other hand, in (ii) the marking of Bill as the di-
Inflectional
Rules
149
rect object disambiguates the sentence. In a language in which the position of NP's in a string is irrelevant to their grammatical role, the presence of an object marker preceding Bill is desirable: (i)
John ate the apple
(ii)
John killed Bill
This issue will not be pursued further in this study. 4. There are some cases in Hebrew in which sei does not seem to conform to the environment specified by SI in (15). However, upon a closer inspection all these cases are of an equative nature, as is demonstrated by (ia-b): (i) a.
b.
ha-bayit haya sei Xana the-house was of Xana 'The house was Xana's ha-bayit nir'a sei Xana the-house seems of Xana 'The house seems to be Xana's
(ia-b) are entirely synonymous to (iia-b): (ii) a. b.
ha-bayit the-house ha-bayit the-house
haya ha-bayit sei Xana was the-house of Xana nir'a ha-bayit sei Xana seemed the-house of Xana
The only occurrences of sei in an environment that does not appear to fall within the environment specified in (15) are in this class: they all have an entirely synonymous corresponding sentence in which the occurrence of sei does satisfy the environment in (15). In view of this correlation we conclude that the sentences in (ia-b) are cases of elipsis, and that ha-bayit 'the house' was deleted under identity with the subject in these cases. The structure of (ia) is thus as in (iii): (iii)
ha-bayil· haya [j^p 0 sei NPj] the-house was J of NP
In (iii) the environment for sei insertion is met in the usual way, independently from the deletion of the head. (Yet another possible analysis for the sentences in (ia-b) involves raising of the head of the NP to the [NP,S] position. This analysis too will yield a structure that conforms to these/ insertion environment. The precise analysis of these constructions will not be pursued here.) 5. Native speakers of Hebrew vary with respect to the acceptability of 'backward anaphora'. For some speakers the sentence in (21a) is ungrammatical. For these speakers the contrast between (21a) and (32), while present, is less crucial, b o t h being ungrammatical (although all agree that (32) is worse). A significant number of speakers, however, systematically accept (21a) and reject (32). 6. One could plausibly argue that when the sei phrase_is branching, the clitic attached to N, in (i) below cannot govern the coindexed N 2 position. This would be due to the fact that only the N , position could be governed since government cannot 'enter' maximal projections:
150
Parametric
Syntax
(i)
N, + clj Note that if the clitic attached to N, does not govern N 2 , the latter cannot be perceived as satisfying the complementation requirements of N, and we do not expect the Complement Matching Requirement to be relevant. If it were not relevant, one could inquire why the coindexation between the clitic and N 2 is nevertheless obligatory. We claim that the clitic og N, governs N 2 and that N 2 satisfies the complementation requirements. Thus N 2 has to be coindexed with the clitic. The fact that N 2 and N 2 share the same index will then follow from the percolation of indices from the maximal projection to the head through all intervening nodes. As for proper government, note that if the clitic on N, does not govern N 2 , it cannot properly govern it either. Nevertheless the position is properly governed by the clitic on set. It follows that the clitic on s'el must have the same index as the clitic on Ν , , as N 2 and as N 2 . Interestingly, the obligatory presence of a clitic attached to set in this case does not follow from the assumption that Case features on grammatical formatives have to be realized. Rather, it follows from the ECP. The assumption that Case features have to be realized, however, cannot be dispensed with as it holds for grammatical formatives which are inserted in the phonological component as well. In these cases the obligatoriness of Case realization cannot be derived from the ECP. 7. Our account for the insertion of Case markers in Hebrew is clearly not complete. A notable problem is presented by the definite object marker 'et in Modern Hebrew. Since it supplies crucial semantic information (and hence is available at the base) and since it does not 'trigger' doubling (and hence it must be available prior t o LF), we expect 'et+NP configurations to branch always; we do not expect the NP complement of 'et ever to serve as an antecedent for a lexical anaphor. This may be true in cases such as (i), but it does not hold for (ii): (i)
(ii)
*re'iyat 'acma- 'et ha-mora^ view herself OM the-teacher 'the teacher's view of herself Dan her'a 'et Dan showed OM
ha-tinoketj le- 'acma^ the-baby to herself
Note, however, that the reverse situation, in which the 'et phrase contains the anaphor and the P-object the antecedent, is grammatical as well: (iii)
Dan her'a la-tinoket 'et 'acma. Dan showed to-the-baby OM herself
For (ii) and (iii) a strong precede condition holds. Thus compare (ii) and (iii) with the ungrammatical (iv) and the very marginal (v): (iv)
*Dan her'a Dan showed
le-'acma· 'et to-herself OM
ha-tinoket the-baby
Inflectional (ν)
151
Rules
•Dan her'a 'et 'acmaDan showed OM herself
la-tinoketj to-the-baby
It seems that where no clear c-command relationship can be established, a certain precede principle may still render anaphoric coindexing grammatical. This precede condition (in the absence of clear c-command) is nevertheless constrained in a very particular way. Thus consider the contrast illustrated by (via-d): (vi)
a.
Dan Dan b. Dan Dan c. Dan Dan d. * Dan Dan
siper told siper told siper told siper told
le-Rina 'al Dalya to-Rina about Dalya 'al Dalya le-Rina about Dalya to-Rina la-tinoket 'al 'acma to-the-baby about herself 'al ha-tinoket le-'acma about the-baby to-herself
From the ungrammatically of (vid) it is clear that "real" objects of prepositions cannot participate in anaphoric relations, regardless of the precede relationship. The only NP's which do participate in these anaphoric relationship which are subject to weakened conditions are arguably anaphors dominated by adjoined maximal projections, i.e._the Ν in structures such as (vii), where the c-command requirement is met by the N' node, and where the preposition (such as le, 'to') or the dummy Case marker (such as 'et or sei) is inserted and does not change the categorial nature of the phrase. (vii) inserted marker The "precede anaphors" are further constrained in ways which are poorly understood and which seem to obey certain aspects of the thematic hierarchy (see Jackendoff, 1972, for discussion). There are some cases in which the combination sel+ clj [e]j, as in (28b), can serve as an antecedent for a lexical anaphor. These are cases of precede and of thematic superiority of the object of sei. Thus compare the sentence in (viiia-b), (ixa-b): (viii)
a.
b.
(ix)
a.
ha-xasiva sei Rina. 'al 'acma· the-thinking of Rina about herself 'Rina's thinking about herself ha-xasiva sel-i 'al 'acmi the-thinking of-me about myself 'my thinking about myself'
ha-xasiva the-thinking meaning as in b. *ha-xasiva the-thinking
'al about (via) 'al about
'acmaherseli
sei Rina^ of Rina
'acmi sel-i myself of-me
152
Parametric
Syntax
In cases (viiia) and (ixa) Rina c-commands 'acma 'herself. Thus we expect considerations of precede to be irrelevant; changing the order of the constituents, as in (ixa), is indeed irrelevant to the grammaticality of the sentence. In (viib), on the other hand, the empty category corresponding to Nj in (28b) does not c-command the reflexive anaphor since it is part of a branching structure. Nevertheless an anaphoric relationship is possible: the adjoined maximal projection which dominates the object of sei, and which is coindexed with it, c-commands the anaphor, the precede relationship is met and the object of sei is the agent, whereas the anaphoric expression is the patient. However, in (ixb) where the precede relationship is destroyed, anaphoric relationships are no longer possible as the empty category under N j does not c-command the anaphoric expression on the one hand, and it does not precede it on the other hand. Hence (ixb) isungrammatical. 8. The difference between (35a) (in which pe and doubling do not occur) and (35b) (where both pe and doubling occur) is that in (35a) 'the girl' is perceived as being objectified in some sense. This variation is possible with few Romanian verbs, notably the verbs 'to show' a arata and 'to see' a vedea. 9. A genuine independent reflexive form (as opposed to a reflexive clitic) is not used in Romanian. Rather, the antecedent-reflexive relationship is tested here with an emphatic pronominal form which is sensitive to c-command and the binding conditions in the usual way. Thus (i) in which the antecedent 'Mary' does not c-command the emphatic pronoun is ungrammatical: (i)
*casa Marieij a ars din cauza eiinsisithe-house Mary has burned from cause her's her/emphatic 'Mary's house burned because of her/*herself
10. Note that we have not ruled out the insertion of pe, a [+P] marker, preceding a [-P] element, if such insertion takes place in PF. In this case the impossibility of interpretation is irrelevant, since pe insertion in PF does not feed into LF (I am indebted to D. Steriade for pointing this out to me). We assume, however, that whenever Case markers are inserted in the phonology, they are truly "rescuing devices." As such they are inserted only preceding [-Case] elements. Thus they will be inserted in doubling configurations but not in any other configurations. (Recall that occurrences of pe without doubling have to be instances of PI at S-structure). Thus the insertion of pe will never occur af PF preceding a [-P] element, since this element is Case marked directly by the verb: 11. Interestingly, cases of nominal pied-piping in Romanian seem to violate Kayne's generalization. Note that in (39b) pe does not appear preceding the fronted element, although this NP is doubled. The insertion o f p e preceding ale cärui rezultate 'whose results' would lead to ungrammatically as 'whose results' is not [+P] in the sense discussed in section 2 above. We would like to suggest that the only grammatical derivation of (39b) thus involves Case assignment to ale cärui rezultate in the base prior to the extraction and a defective application of clitic spell-out in the following way: (i)
[ γ V, 6 Case]
[ y V [ a number, β person, y gender] ]
The application of (i) is defective as the clitic form which is inserted by (i) does not contain Case features. These are assigned to 'whose results' in the usual way. A full, non-defective application of the clitic spell-out rule would result in absorption of the Case features of V and thus in the absence of Case marking for 'whose results' and in
Inflectional
Rules
153
a violation of the Case filter. On the other hand, a complete failure of (i) to apply, even in its defective form, would result in a violation of the semantic requirement on doubling: a non-doubled NP which is [+specific/+[a gender, β number, γ person] ]. Given this situation we expect (39b) to be a marked configuration, and indeed, it is somewhat marginal for most speakers and impossible for others. In Chapter 6 we will return to defective applications of the clitic spell-out rule. We will elaborate on the source for the number, gender and person features and on the possibility of 'Caseless' clitic.
Chapter
5
Clitics and Reanalysis in Romance
1. INTRODUCTION
In this chapter we will discuss several constructions in River Plate Spanish and in French, and show how our parametric system can account for the variation found in these constructions. We will first concentrate on "two storey" constructions: causative constructions in RP Spanish and in French and indirect-object control cases in RP Spanish. We will show that the distribution of clitics in these structures is explained naturally if we assume the analysis proposed in Chapter 2. While explaining the distribution of clitics in this way, we will be supporting a particular analysis of Case assignment and propose a process of reanalysis that can be parametrized, accounting for the distinction between French causatives and RP Spanish causative. In the previous chapters we have shown that extraction from cliticdoubling constructions is possible both in Hebrew (Chapter 2) and in Romanian (Chapter 4). Jaeggli (1982) reports that such extraction is restricted in RP Spanish. This variation will be the second topic discussed in this chapter. We will show that the difference in extraction between RP Spanish on the one hand and Hebrew and Romainian on the other hand follows from the Case marking properties of the respective dummy Case markers inserted in these languages. Some issues concerning dative clitics will be discussed in section 4. We will point out a natural way to capture the difference between dative clitic configurations in RP Spanish and dative clitic configurations in French and we will further show that some interesting properties of the inalienable possessive constructions can be captured naturally, assuming the Complement Matching Requirement. Let us start with a brief sketch of cliticization facts in RP Spanish. Clitic doubling in RP Spanish shows a great deal of similarity to clitic doubling in Romanian. As in Romanian, it has to obey certain semantic restrictions. These semantic restrictions are, however, somewhat different. Moreover, although clitic doubling is preferred in environments which satisfy the semantic requirements, it is not obligatory. The following is
156
Parametric
Syntax
a description of the semantic environment of clitic doubling in direct and indirect objects following Jaeggli (1982): 1 (1)
Indirect Objects a. Goal i.o. preferred b.Poss.i.o. obligatory c. Pronominal obligatory Direct Objects d. Inanimate impossible e. Animate, specific preferred f. Pronominal obligatory
(2)
a. (= la)
b.
c.
d.
e. f.
Miguelito lej regalo un caramelo a Mafaldaj Miguelito h e r - d a t gave a candy to Mafalda 'Miguelito gave a piece of candy to Mafalda' (= l b ) lej duele la cabeza a Mafaldaj h e r - d a t hurts the head to Mafalda 'Mafalda has a headache' (= lc) lej entregue la carta a el, h i m - d a t delivered-I the letter to him Ί delivered the letter to him' (=1 d) i. vimos la casa de Mafalda we saw the house of Mafalda ii. *la vimos la casa de Mafalda (= le) loj vimos a J u a n j him-acc saw-we to Juan (= 1 f ) loi vi a el, him-acc I saw to him
Jaeggli observes that t h e environments in which doubling occurs in RP Spanish are a subset of the environments in which the preposition a is available (note, however, that this entailment works only in one direction: it does not hold that clitic doubling is possible whenever a is present, as is exemplified by (6)): (3)
a. vimos a Juan saw-we to Juan 'we saw Juan' b. *vimosJuan c. lo] vimos a Juan him saw-we to Juan 'we saw Juan' d. *loj vimos Juan,
Clitics and Reanalysis in Romance (4)
157
a. vimos una camisa saw-we a shirt 'we saw a shirt' b. *vimos a una camisa to c. *la vimos una camisa d. *la vimos a una camisa
(5)
a. vimos a la camisa to the b. vimos la camisa c. *laj vimos a la camisaj d. *laj vimos la camisaj
(6)
a. yo vi a alguien I saw to someone b. *yo vi alguien c. *yo loj vi a alguienj d. *yo loj vi alguien.
Note that this behavior is in accordance with Kayne's generalization.
2. CLITIC GOVERNMENT AND REANALYSIS CONSTRUCTIONS
In this section we will consider the government relationship between the clitic and the coindexed NP position. It will be shown that this relationship plays a crucial role in determining the clitic distribution in " t w o storey" constructions in RP Spanish. We will further propose a parametrized reanalysis process which will distinguish between different possible causative constructions in French and in River Plate Spanish.
2.1. Clitic Government in Causative Constructions Recall that the phrases which serve as a crucial test for the government requirement in the construct state in Modern Hebrew have the structure in (7):
158
Parametric
(7)
*
Syntax Ni
N2 + c l j In (7), N 2 and the clitic attached to it cannot govern N 4 , and consequently coindexation is impossible. Can it be argued that when clitics are attached to verbs rather than nouns they exhibit the same government properties? The structure equivalent to (7) in the Romance languages would be as in (8): (8)
V,
C l j + V2
N2
The structure in (8) is not attested in the Romance languages. It could be aigued, however, that the derived structure of causative constructions and other "two-storey" constructions in Romance languages exhibits the government properties which are relevant to our claim. The sentences we have in mind in RP Spanish are as in (9) (the data from RP Spanish in this section are from Rivas, 1977): (9)
a. Maria (lej) hizo tocar la flauta a Josej Maria him-dat made play the flute to Jose 'Maria made Jose play the flute' b. Maria (Ιο;) hizo venir a Josej Maria him-acc made come to Jose 'Maria made Jose come'
Let us first clarify some of the properties of (9a-b). Note that in (9a) the clitic which corresponds to the subject of tocar 'play' is dative, whereas in (9b) the clitic corresponding to the subject of venir 'come' is accusative. Since both clitics appear in an environment in which clitic doubling is not obligatory, they are optional. Note that, although both in (9a)
Clitics and Reanalysis in Romance
159
and (9b) Jose is preceded by a, these a's are quite different. In (9a) the a is the regular dative a (and hence the corresponding clitic is dative); in (9b) the a is the object marker appearing in doubling constructions. We will return to the distinction between these two a's in section 3. It has been suggested by many scholars that the derivation of causative constructions in Romance involves the fronting of elements from a subordinate clause (to name only a few: Kayne, 1969, 1975; Aissen, 1974; Quicoli, 1976; Rivas, 1977; Rouveret and Vergnaud, 1980; Zubizarreta, 1979a, b; Buizio, 1981). Following these proposals, we will take the underlying structure of sentences such as (9a-b) to be roughly as in (10): 2
tplay > (come) However these scholars vary with respect to the nature of the fronting which takes place in causatives. Rivas (1977) shows that in RP Spanish, whenever fronting takes place, the verb must be fronted along with the complements which it strictly subcategories for. This situation is illustrated by the following paradigm: (11)
a. Maria lej hizo tocar la flauta a Jose 'Maria made Jose play the flute' (accusative complement) b. Maria lej hizo escribirlesj (una carta) a los chicosj a Jose 3 'Maria made Jose write a letter to the children' (dative complement and optional accusative complement) c. Maria le · hizo arrojar papeles en el cesto a Jose·
160
Parametric
Syntax
'Maria made Jose throw papers into the basket' (accusative and dative complements) (12)
a. Maria (loj) hizo salir a Josej de la habitation 'Maria made Jose leave the room' (non-strictly subcategorized complement) b. Maria (lej) hizo tocar la flauta a Josej en la cocina 'Maria made Jose play the flute in the kitchen' (strictly subcategorized complement vs. non-strictly subcategorized complement).
Recall that we argued that the domain of complementation is the domain of government by the head. It follows that whenever the verb strictly subcategorizes for a complement, whether accusative or dative, it has to govern it. Given our assumption that the government requirement as well as the Complement Matching Requirement are checked at S-structure, it follows that if the verb in causative constructions is fronted without any of its strictly subcategorized complements, the sentence will result in ungrammatically. 4 In this fashion we would like to capture the obligatoriness of the fronting of all subcategorized complements attested in ( l l a - c ) . (Note that this account will not explain the unavailability of fronting for non-strictly subcategorized complements. We will return to this matter below.) With respect to the derivation of causative constructions, we will make the following assumptions: first, we will assume with Rouveret and Vergnaud (1980) that any projection of V 2 in (10) may be fronted. Following Rivas (1977), we will adjoin the fronted V^ projection to V i , i.e. to the V projection of the matrix verb. Since the only requirement for subcategorized complements is the government requirement, we will assume that subcategorized PP's can be generated either under V or under V. This latter assumption will be motivated below. In this respect it is worth mentioning Jaeggli's (1982) argument that in Spanish PP's are generated under V rather than under V. In this way Jaeggli seeks to explain the fact that PP's are fronted in that language along with the direct object, although fronting is of the V constituent only. We accept this as a possible structure but we would like to claim that PP's can be generated under V as well. In this case their fronting in causative constructions will be blocked by independent factors to which we will return. Given the government requirement, it follows that although the fronting of any projection of V 2 is possible, only those fronting operations which will not split the verb and its complements will result in grammatically. Thus for ( l i b ) , which is base-generated as (13a) or as (13b), the only grammatical derivations are those in which both the direct ob-
161
Clitics and Reanalysis in Romance
ject and the indirect q_bject are fronted. Since PP can be dominated at D-structure either by V or by V, there are two possible derivations: the one in which V is fronted and the one in which V is fronted and it contains the PP. These two derivations are given in (14a-b). (14c) is an example of an ungrammatical derivation: (13)
a. Maria [y^ (lej) hizo [g Josej [y^ [y^ [y^ escribirlesj] una carta] y ] [pa los chicosj] y 2 ] ] ] b. Maria ( l e ^ i z o [ s Jose; [ y 2 [ y escribirles·] una cartal] ^ a los chicos ] y f ] ]
(14)
a.
( c l ^ + Vx (lej) hizo
(14)
V2 escribirlesj una carta
a los chicosj
b.
(clj) + Vj (lej) hizo (14)
escribirlesj
una carta
a los chicosj
c.
(clj) + V! I (lej) + hizo
escribirlesj una carta
(a) Jose'· V2 PP I ^ [e] a los chicosj
162
Parametric
Syntax
(14c) is ungrammatical because V 2 no longer governs its dative complement. Similarly, if only V 2 is fronted the derivation is ungrammatical as V 2 no longer governs its accusative complement and its dative complement. Thus the only two possible derivations are those in which V 2 still governs both its complements: the accusative complement and the dative complement. It will be shown below that (14a) is ungrammatical as well. We will also return to the dative preposition a preceding the subordinate subject in (14a-b). 2.2. The Case Tier Let us now turn to the distribution of the clitics which correspond to the subject of the subordinate clause. As can be seen from the examples in (9)-(12), that clitic is dative whenever V 2 is immediately followed by a complement. Thus we have dative clitics in (9a), ( l l a - c ) and (12b), but an accusative clitic in (9b) and (12a), where V 2 is not immediately followed by any complement. Rather, the NP which immediately follows V 2 in these cases is the subordinate subject itself. Interestingly, a similar paradigm is attested with other verbs in RP Spanish which take both an accusative and a dative complement. This is illustrated by (15)-( 16): 5 (15)
a. Pedero (lej) sirvio la c o m i d a a J u a n j Pedro (dat) served the food to Juan b. Pedro sej laj sirvio [e]j a Juanj dat acc c. *Pedro (loj) sirvio la comida a Juanj acc
(16)
a. Pedro (loj) sirvio a Juanj Pedro (acc) served to Juan' b. *Pedro lej sirvio a Juanj 6 dat
The generalization behind the data in (15)-(16) is quite clear: we have a verb that subcategorized obligatorily for a complement assigned the Θ-role of a goal. This complement appears both in (15) and in (16). Furthermore this verb optionally subcategorizes for an additional complement - a theme - which appears in (15) but not in (16). When this additional complement appears it is between the verb itself and the goal complement, and it is assigned accusative Case by the tentative rule suggested in Chapter 1. This rule is repeated here as (17):
163
Clitics and Reanalysis in Romance (17)
Accusative Case Assignment [ γ V, accusative]
V
NP [+accusative]
Rule (17), as formulated, is "blind" to the thematic role of the adjacent NP. Thus we expect (17) to apply to the theme complement when it is present and to the goal complement when the theme is not present. When both the goal complement and the theme are present, the former complement is assigned dative. Intuitively speaking, then, it is clear that in some sense the verb acquires the ability to assign dative to its complement. Note, however, that the actual assignment of Case to the NP in question cannot be done directly by the verb, since the verb is not adjacent to the goal complement in (15). Thus we will assume that the actual dative marking is achieved by the insertion of a dative marker, more or less along the lines suggested in Bordelois (1974). What is the nature of the Case assignment rules in these cases? We have already pointed out that the assignment of Case is divorced from the assignment of thematic roles. Rather, the first complement to the right of the verb will be accusative regardless of its thematic role. In order to capture this observation we will perceive of the selection of complements and the assignment of Case as two separate tiers which are linked with each other. 7 In fact, let us assume that the nature of the Case assigned to a particular argument is determined by the nature of the linking. Accordingly we will reformulate rule (17) as (18): (18)
Accusative Case Assignment (revised): Case • [+accusative] / V NP
(18) is intended to capture the generalization that an NP will be assigned accusative when it is adjacent to the verb. Rather than perceiving accusative Case as a property of a verb which requires adjacency to be realized, we are now regarding the crucial property of the verb as being [+Case assigner]. The property of assigning accusative is a result of the linking configuration in (18). Returning to verbs like 'serve' in Spanish, we now assume that they have the lexical representation of (19), entailing that they subcategorize for two Θ-marked complements in particular positions and allow for the assignment of two unspecified and unordered Cases: (19)
Serve : NP, (theme); NP2 (goal) (NP t is optional) Case; Case.
164
Parametric
Syntax
If Case is associated with NPj in (19), it will be assigned accusative by rule (18). NP2 will then be associated with the second Case slot of the verb. We will take a "second Case slot" to be spelled out as dative. In other words, given a verb with two Case slots, the one which is linked to a nonadjacent complement will be realized as dative (this is clearly reminiscent of Rouveret and Vergnaud's (1980) contention that non-adjacent Case is dative). The relevant configuration for dative assignment is thus roughly as in (20): (20)
Dative Assignment Case • [+dative] / NPj
ΝPk
Case;
]
[
(20) (21):
as formulated is clearly too strong. Thus consider the sentence in
(21)
a bookj was given ej to John
Strict application of (20) will result in the impossibility of assigning Case to John in (21). As it is not adjacent to given it would not be assigned accusative, and since there is no intervening Case slot it could not be assigned dative. Thus the realization of Case as dative is clearly not dependent solely on the presence of an intervening Case slot. Rather, we will assume that any lexical, sub categorized material intervening between the Case assigner and an additional NP will suffice to trigger the formation of dative Case. Thus (20) is now reformulated as (22): (22)
Dative Assignment (revised) Case • [+dative] / X NP
[
1
]
where X is a lexical subcategorized property (Case, Θ-role etc.) In the absence of an adjacent Case assigner for the non-adjacent complement, a dummy preposition must be inserted, serving as a direct Case assigner (or perhaps as the Case itself. We will return to this point in section 4). Crucially, however, we will take the assignment of dative Case to the goal complement in sentences such as (15a) to be a property of the verb, which is checked against the Case assigned to the non-adjacent complement by an inserted dummy. It is clear that the operations illustrated by rules (18) and (22) cannot be operations on a string of phrases. Rather, they are operations on a lexical entry of a Case assigner: they specify which morphological set of properties will be associated with which Case slot. This specification
Clitics and Reanalysis in Romance
165
is in turn dependent on the linking between the available Case slots and the subcategorized complements which exist in the lexical entry as well. The system of Case assignment proposed here makes an interesting prediction: it predicts that when a verb takes two NP complements, the dative complement will not be the adjacent one because adjacent Case slots will always be accusative. We will return to this point in section 4 below. 8 Note that since the verb 'serve' in Spanish allows for an optional theme, as is demonstrated by (16), we expect a linking of NP 2 in (19) with Case which will meet the configuration in (18) when the theme argument is not generated. In this case NP 2 will be assigned accusative Case (see (16b)). No additional mechanism is necessary to ensure the correct assignment in these cases.
2.3. Reanalysis Let us now return to causative constructions. We will assume that the interpretation of causatives requires a reanalysis which applies to structures such as (14a-b) above. Following this reanalysis, the subordinate verb becomes 'transparent.' Concretely, the verb hacer 'to cause' is perceived as taking as its complements the set of arguments which are in its government domain, the set of strictly subcategorized arguments which are in the domain of government of the subordinate verb and the subordinate subject. (For other proposals of reanalysis, similar in spirit to ours and dealing both with causative and with restructuring cases, see Rizzi, 1978; Rouveret and Vergnaud, 1980;Zubizarreta, 1979a,b). Assuming further that the reanalysis can only take place if the lower, subordinate verb is adjacent to hacer, it follows that some projection of V 2 must be fronted. Furthermore, it must be fronted precisely to a position in which it will be adjacent to hacer. Thus it has to be adjoined to V i . We can easily show that this must be the case. An adjunction to Vj would not result in adjacency, since V | would then follow the subordinate clause. Adjunction to V , , on the other hand, would imply that VJ (including the subcategorized complements) becomes part of the matrix verb, since both V x and would then be dominated by the same terminal node. It follows that the only possibility is an adjunction to V j . In effect, the process of reananysis implies that, rather than stipulating that fronting of some projection of V 2 has to take place, or stipulating that it has to be adjoined to V , , we can assume that the fronting of some projection of V 2 is subsumed under "Move a " and that the adjunction is optional at any level. Note, however, that a failure to move some projection of V 2 will fail to create adjacency and thus the causative interpretation will be blocked. On the other hand, a failure to adjoin the moved
166
Parametric
Syntax
projection to V! will result in ungrammatically as well, since only such an adjunction will create the desired adjacency. As we will see below, the reanalysis process combined with the Case assignment properties of causative constructions also predicts that only V 2 can be fronted in some cases, rather than any other projection of V 2 . Let us now turn to (14a) and (14b) above. Recall that we required that a complement be governed by the head which it is a complement of. In (14b) this requirement is met: the subordinate subject Jose is in the domain of government of hacer. The strictly sub categorized complements which are in the domain of government of the subordinate, fronted verb, are una carta 'letter' and α los chicos '(to) the children'. Since hacer is now perceived as taking all these arguments as complements as well, it follows that it has to govern these complements as indeed it does in (14b). This is due to the fact that in (14b), V was fronted rather than V. V is not a maximal projection, and therefore hacer governs its complements (and see clause (ii) of the definition of government in Chapter 2, (67) above). Thus, given the reanalysis process, the derivation given in (14b) above is well-formed. Let us now consider (14a). Following the reanalysis, hacer takes as complement the subordinate subject, Jose, which it governs. However, it also takes as its complements the complements of the subordinate, fronted verb._But in this case hacer cannot govern una carta and a los chicos, since V, a maximal projection, was fronted. It follows that (14a) is ungrammatical. Thus the reanalysis effectively forces fronting of V, rather than any other projection of V, in causative constructions in RP Spanish. The fronting of V, leaving behind its complements, will result in ungrammatically as hacerjxiW not govern the complements. On the other hand, the fronting of V will block the government of the fronted complements by hacer, thus violating the requirement that heads govern their complements. We are now equipped with an explanation for the impossibility of fronting complements that are not strictly subcategorized. Under the natural assumption that these complements are dominated by V (see Quicoli, 1980; Rivas, 1977; Jaeggli, 1982), their fronting will result in the fronting of a maximal projection. Consequently hacer will be blocked from governing its 'newly acquired' arguments and the derivation will be ruled out. The impossibility of fronting non-strictly subcategorized complements has led many scholars to argue that the rule of verb fronting (first suggested in Kayne 1969) must be stated as applying to the V projection. Within the system proposed here this fact need not be stipulated. Rather, it follows from independent principles. If the subordinate verb does not strictly subcategorize for any complements, our argument concerning government by hacer would not hold
Clitics and Reanalysis in Romance
167
and we expect fronting of V, V and V to be grammatical. Since the lower verb has no complements, the question of the projection fronted is irrelevant. Thus we expect (23a-c) to be grammatical derivations of (9b): (23)
a.
However, (24), in which a non-strictly subcategorized PP has been fronted (corresponding to the structure in (23a)), has a "scrambling" reading, equivalent to cases in which a direct object has been postposed: (24)
Maria loj hizo salir de la habitacion a Pedroj (compare with (12a) above)
We would like to suggest that (23a) is an impossible structure for (9b); the fronting of a non-subcategorized PP as in (24) is the result of a
168
Parametric
Syntax
scrambling rule which either postposes Pedroj or fronts the PP at a late stage of the derivation. We will return to the explanation of this proposal below in our discussion of Case assignment to the subordinate subjects. Let us now consider the assignment of Case in causative constructions. Recall that when the subordinate verb is fronted along with a strictly subcategorized complement the subordinate subject is assigned dative Case. However, when no complement is fronted (in other words, when the fronted verb does not have any strictly subcategorized complements) the subordinate subject is assigned accusative Case. Since we are assuming that both the subordinate subject and the complements of the subordinate verb are reanalyzed as the complements of hacer, we now have a situation similar to that of the verb servir 'to serve' above. We are considering a verb, hacer, which can take either one complement or two complements. When the two are present, the second one receives dative Case marking by rule (22). When only one is present, rule (18) should be applicable. Consider how this will work for (11a). Unlike servir, causative constructions contain two Case assigners: hacer, which we will assume to contain one Case slot, and the subordinate fronted verb, in this instance tocar 'play', which contains one Case slot. Following reanalysis, we will assume these two slots to be properties of hacer. Or, concretely, we will assume all the arguments governed by hacer to be in the government domain of these Case slots. Assuming further that the subordinate verb is 'transparent' for the purposes of the adjacency requirement imposed on accusative Case assignment, we have the configuration in (25): (25)
hacer (tocar)
NP!
NP2
I
I
Case,
Case;
NPi (in this case la flauta) will be assigned accusative Case by (18). NP 2 (in this case the subordinate subject Jose) will be assigned dative Case by (22). Now consider a slightly more complex case, as in (1 lb). Here, the subordinate verb, escribir 'write', takes two arguments, resulting after reanalysis in the configuration in (26): (26)
hacer (escribir)
NP,
I
Case;
NP 2
I
Case;
NP 3
I
Case
While NP! will be assigned accusative by (18), both NP 2 and NP 3 will be assigned dative by rule (22), resulting in the grammatical sentence in (lib).
Clitics and Reanalysis in Romance
169
If, on the other hand, the subordinate verb does not take any complements, as in (9b) and (12a) (and see corresponding structures in (23)), we have the following configuration: (27)
hacer (venir)
NP, I Case
NPi will be assigned accusative in a straightforward fashion, following (18): the absence of an intervening lexical slot between the Case assigner hacer and NP!, in this case the subordinate subject, will make this assignment possible. Returning now to (24), note that Pedro is accusative, as is attested by the corresponding accusative clitic. Rule Q8), however, requires adjacency in order to assign accusative. If in (24) Vis fronted along with the nonstrictly subcategorized PP, this adjacency condition is not met and accusative Case cannot be assigned. On the other hand, since PP is not a complement of hacer (it is neither strictly subcategorized by the subordinate verb, nor is it governed by hacer, nor is it assigned Case by it), it cannot serve as the slot in (22). Thus the environment for dative Case assignment in (22) is not met and Pedro cannot_be assigned dative Case. Therefore, if the PP is fronted along with the V2 in (24), the subordinate subject cannot receive Case and the derivation is ruled out. However, when V2 does not contain a non-strictly subcategorized PP, its fronting will not lead to ungrammaticality. Following the fronting and the reanalysis, hacer is still adjacent to the subordinate subject and accusative Case is possible in accordance with (18). 9 Since in (24) accusative Case is assigned to the subordinate subject, we conclude that at the point at which (18) applies the adjacency condition is met and that the intervening PP appears following salir as a result of a late scrambling rule. Corresponding to dative or accusative subordinate subjects we have dative or accusative clitics attached to hacer. Note that we no longer assume that verbs are inherently marked for assigning a particular Case. Rather, they are marked as assigning simply Case. Rules (18) and (22), however, specify the precise nature of these Case features. Thus we will assume that the clitics absorb those unspecified Case features. A simple checking mechanism will then ensure the correlation between the right clitic form (dative/accusative) and the linking configurations in (18) and (22). Interestingly, the distribution of the embedded subject clitics on hacer corresponds exactly to our expectation: when the subordinate verb takes complements we expect a dative subject clitic, which is indeed the case in ( l l a - c ) and (12b). When the subordinate verb does not subcategorize for any complements we expect an accusative clitic, which shows up in (9b) and (12a).
170
Parametric
Syntax
Let us now turn to the location of the clitics with respect to hacer and the subordinate verb (V2). Rivas indicates that the following paradigm holds: A. The clitic corresponding to the subordinate subject is always attached to V j {hacer): (28)
a. *Maria hizo tocarlej la flauta (a Josej) play-him (dat) b. Maria lej hizo tocar la flauta (a Josej)
(29)
a. *Maria hizo venirloj (a Jose,) him (acc) Maria loj hizo venir (a Josej)
B. The clitic corresponding to the complement of V 2 can be attached either to V 2 or to V j : 1 0 (30)
a. Maria hizo escribirla 'Maria made X write it' b. Maria la hizo escribir
((31)
a. Maria hizo escribirle 'Maria made X write to him' b. Maria le hizo escribir 'Maria made X write to him'
(32)
a. Maria hizo escribirsela 'Maria made X write it to him' b. Maria se la hizo escribir
Recal that our analysis of clitics requires that the clitic govern the NP position which is coindexed with it. In view of this requirement, and given the definition of government we assume above, let us now look at the structure of the sentences in (28)-(29) (irrelevant details omitted):
Clitics and Reanalysis in Romance (33)
171
((=28a),(29a))
In (33), the clitic attached to V 2 does not govern the NPj position, since, according to the definition of c-command in Chapter 2, (56) above, the V 2 position to which the clitic is attached does not c-command NPj. This follows from the fact that the V 2 projection does not have the same head as the projection Vj which dominates NPj. Since the clitic and the verb which it is attached to do not govern NPj, NPj cannot satisfy the subcategorization requirements of V 2 . In (29a) the verb does not strictly subcategorize for a complement and it does not have a Case slot; thus there is no possible source for the clitic and the sentence is ungrammatical. In (28a), on the other hand, the argument which does meet the government requirement, /a flauta, NP:, contains an index which is different from that of the clitic attached to V 2 ; thus the sentence is ruled out as a violation of the Complement Matching Requirement. (See Chapter 2, subsection 1.2 for discussion.) (Note that this system does not rule out a situation in which both the subordinate subject in (33), Jose, and the subordinate object, la flauta, contain index i. This situation, however, will be ruled out by the binding conditions, since the NPj position in (33) c-commands the NPj position; the indentity of indices in this case will result in la flauta being bound in its minimal governing category.) Let us now look at the grammatical (28b) and (29b). Their structure is given in (34):
172
Parametric
Syntax
In (34) the subject clitic is attached to hacer. Given the definition of government assumed here, the clitic attached to hacer governs the coindexed subject and corresponds to a Case slot. Thus the sentences are grammatical as expected. Note that the clitic attached to hacer in (34) governs both NPj and NPj. However, in order to be understood as coreferential with Jose it must be coindexed with it. If we were to assign the same index to la flauta (in other words, if i=j), the clitic coindexed with la flauta will govern it, as required. However, the coindexation of la flauta and Jose will result in a violation of the binding conditions. Now let us turn to sentences ( 3 0 ) - ( 3 2 ) . In these cases the clitic which is coindexed with the complement of V 2 can be attached either t o V 2 or to V , . This is illustrated for direct objects in (35) (irrelevant details omitted): (35)
escribirlaj ι Since both V ! and V 2 govern the NPj position it follows that the clitic can adjoin to either verb. For indirect objects the same government relationship holds in a similar fashion. The relevant diagram for (32), in which two clitics can be attached to either verb, is (36): 1 1
173
Clitics and Reanalysis in Romance (36)
V
s
V V c l d , c l a + Va
V2 \
NP d
V2 + c l a V
V2+cld,cla
[e]
se la hizo |le hizo hizo
escribir escribirla escribirsela
NP a = NP accusative NP d = NP dative
[e]
Let us summarize our discussion thus far. The distribution of subject and object clitics in causative constructions in RP Spanish follows directly from the requirement that clitics govern their coindexed NP position. The subordinate subject position is only governed from the matrix subject, and thus we expect the corresponding clitic to appear only there. The complements of the subordinate verb, on the other hand, are governed both by the matrix verb and by the subordinate verb and thus we expect a governing clitic to be attached to either one, a prediction which is borne out. One may legitimately question the cliticization of subordinate complements to V 2 , the subordinate verb. Recall that we argued that all Case slots are transferred to the matrix verb, hacer. Assuming that clitics are spell-outs of Case features attached to that element which has these features, why is the cliticization to V 2 ever allowed? Why don't we find exclusive cliticization to hacerl In order to answer this question, let us observe the behavior of causative constructions in another Romance language, which differs minimally from RP Spanish in that respect i.e. French. This sketchy comparison between French and RP Spanish will point towards an interesting parametric variation, feeding off the nature of reanalysis in causative constructions. 2.4. Parametric Variations in Causative Constructions There are two interesting differences between RP Spanish causatives and French causatives. The first relates to the distribution of clitics and the second to the fronting of subcategorized complements. Let us first consider the cliticization facts. Whereas in RP Spanish
174
Parametric
Syntax
the objects of the subordinate clause can be cliticized to the subordinate verb, as in (30a), (31a) and (32a), in French, all clitics must be attached to the matrix verb .faire. The attachment to V 2 of either the subordinate subject clitic or of the objects of V 2 results in ungrammaticality: (37)
a. *Jean a fait lui manger les bananes 'Jean made her eat the bananas' b. *Jean a fait les manger ä Marie 'Jean made Marie eat them'
It is interesting to note, however, that (37a) is considerably worse than (37b). Note that only the latter, but not the former, is possible in RP Spanish. As we will see below, the ungrammaticality of (37b) derives from the violation of a language particular condition on reanalysis. The ungrammaticality of (37a), on the other hand, derives from the requirement that a head must govern its complement which has a universal status. Now consider a second difference between French causatives and RP Spanish causatives. Kayne (1969, 1975) notes that in French direct objects are fronted with the verb in causative constructions, but indirect objects are not. Thus, compare the grammatically of (38) with the ungrammaticality of (39): (38)
Marie a fait telephoner Jean ä ses parents 'Marie made Jean telephone his parents'
(39)
*Marie a fait telephoner ä ses parents ä Jean
It is likely, however, that the ungrammaticality of (39) derives from a constraint against the occurrence of two adjacent a phrases, which is independently argued for by Kayne. Thus in cases in which the indirect object is not an ά phrase, the fronting of a strictly subcategorized indirect object is grammatical, as is illustrated by (40a-c): (40)
a. j'ai fait rever de Marie ä Jean Ί made Jean dream of Marie' b. j'ai fait mettre les bananes sur la table ä Pierre Ί made Pierre put the bananas on the table' c. j'ai fait parier de Marie ä Jean Ί made Jean talk about Marie'
Furthermore, as noted by Ruwet (1972), the constraint on two adjacent ά phrases may be violated by some speakers in cases such as (41). For these speakers, the preferred interpretation is the one in which the last
Clitics and Reanalysis in Romance
175
ά phrase is the subordinate subject, thus indicating that in these cases the indirect object is fronted along with the subordinate verb: (41)
j'ai fait ecrire la lettre ä Marie ä Jean Ί made Jean write the letter to Marie'
Further evidence for the availability of indirect object fronting in French is the fact that the dative clitic corresponds to the subordinate subject in sentences such as (42a-b). (Note that if the indirect object in (42) is not fronted, we expect the clitic corresponding to the subject t o be accusative): 1 2 (42)
a. Marie lui a fait telephoner ä ses parents Marie made him telephone his parents' b. je lui ai fait rever de Marie Ί made him dream of Marie'
Thus we conclude that the fronting of strictly subcategorized indirect objects is in fact possible. Note further t h a t , as in RP Spanish, the fronting of a PP which is not strictly subcategorized will not result in dative assignment to the subordinate subject. Rather, it will have a "postposed object" reading: (43)
Jean a fait venir de Paris Marie 'Jean made Marie return from Paris'
Thus the fronted PP's in (40) and the dative clitics in (42) pattern with the fronted complements in RP Spanish and not with the "postposed reading" constructions. There remains, however, a crucial difference between the two languages: in French the fronting of an indirect object is o f t e n optional, even when it is strictly subcategorized. Thus both (38) above and (44) below are grammatical: (44)
J'ai fait parier Jean de Marie (compare with (40c))
The grammatically of (38) and (44) would seem to present a problem for our analysis: note that in both sentences the indirect object cannot be governed by faire if it is not fronted. Clearly reanalysis in the sense discussed above cannot apply in these cases, since faire does not govern the n o n - f r o n t e d indirect object and the government requirement would be violated if reanalysis were to apply. Let us now try to account for these two differences between French
176
Parametric
Syntax
and RP Spanish. Recall that we assumed a reanalysis process which entailed the merger of the arguments and the merger of Case slots of the causative verb (hacer and faire, respectively) and the subordinate verb. Let us now assume that this reanalysis process is composed of two separate subparts. One part of the reanalysis process consists of the merger of all Θ-arguments. The other part of the reanalysis process consists of the merger of all Case slots. We will now assume that each of the mergers can occur independently of the occurrence of the other merger. Assuming that all possibilities are realized, we expect the following types: (45) Type Type Type Type
A Β C D
merger of Θ-arguments + + -
merger of Case slots + +
We can discard type D immediately. Assuming that the causative interpretation requires some sort of reanalysis, type D, where no reanalysis occurrs, will not be interpreted as a causative construction. We will now demonstrate that types A-C are attested in RP Spanish and in French and that the typology in (45) successfully covers the variations found in causative constructions. The reader should bear in mind that we are still assuming that the adjacency of the causative verb and the subordinate verb are a pre-condition for both levels of reanalysis, thus making the movement discussed earlier necessary for types A-C. Let us first consider type A. In that case the causative verb, hacer or faire, 'acquires' both Case slots and Θ-aiguments of the subordinate verb. Following the reanalysis it must govern all its complements, including the 'newly acquired' ones. Two properties of type A causatives result. First we expect the clitics corresponding to the arguments of the subordinate verb to appear only on the higher verb, the causative verb, as that verb now has all the Case slots of both verbs. Secondly, we expect all the Θ-arguments of the subordinate verb to be in the government domain of the causative verb. This is exactly the situation in RP Spanish, in those derivations in which the clitic is attached to hacer, as in (30b), (31b) and (32b). In type A cases, all the Θ-arguments of the subordinate verb are fronted so as to be in the government domain of hacer, and clitics attach to the verb which has the Case slots. This is also the situation in the French sentences (40)-(42), where there is evidence that the indirect object has been fronted along with the direct object (if one is present). Now let us consider type Β causatives. Here we have a merger of Θ-arguments, resulting in the requirement that the causative verb must govern
177
Clitics and Reanalysis in Romance
its 'new' complements. The Case slots, however, do not merge, thus the subordinate verb itself assigns Case to its complemwnts. In type Β causatives we expect the fronting of direct and indirect subcategorized complements, but the cliticization'of the subordinate arguments will be to the subordinate verb, which has the relevant Case slots. Note that this is exactly the situation in the RP Spanish sentences which allow for cliticization to the lower verb, as in (30a), (31a) and (32a): although all complements are fronted, clitics attach to the subordinate verb. Note that the assignment of dative Case to the subordinate subject in type Β causative follows from rule (22) without further machinery, as is illustrated by the diagram in (46): (46)
hacer
escribir Case I
NPj
NPj^
Case
The NPj complement (in this case a theme, say 'letter') is assigned accusative by escribir in accordance with (18). NP^ (the subordinate subject) is linked with the Case slot of hacer and is assigned dative in accordance with (22). Note that although NPj is not assigned Case by hacer, it is its Θ-complement, since a merger of ©-arguments has taken place. Thus the configuration in (46) is appropriate for dative Case assignment. Now consider type C. In this case no merger of ©-arguments has taken place, so we do not require that the causative verb would govern the complements of the lower verb. However, Case merger has taken place, so we expect all the arguments which are dependent on the subordinate verb for their Case assignment to be located in the Case domain of the causative verb, since it is the Case assigner following the reanalysis. Thus we expect the fronting of all accusative and 'true' dative complements (we will return to the status of dative complements below). However, the fronting of PP's, whether subcategorized or not, is not required. Case is assigned by the preposition and not by the verb, and since there is no Θ-arguments merger they do not have to be governed by the causative verb. What about the clitics corresponding to the arguments of the subordinate verb? Those which are fronted along with the verb we expect to show up as clitics on the causative verbs. A clitic corresponding to a non-fronted ©-complement, on the other hand, cannot appear attached to either the causative verb or the subordinate verb, since in both cases it would fail to govern the coindexed empty category. This is exactly the situation illustrated by the dialect of French described in Kayne (1975): direct objects must be fronted, PP complements are not fronted and cannot be cliticized; other clitics are always attached to faire. Few issues should be clarified with respect to type C causatives. First
178
Parametric
Syntax
note that the subordinate verb after its fronting no longer governs its PP complement directly. However, its trace does. Thus the government requirement must be interpreted so as to be satisfied either by a head or by its trace. Note that such a modification is independently plausible in order to allow the fronting of the V node without its complements in several other languages, notably, verb fronting in SOV languages (such as German and Dutch) and V fronting in Spanish (see Torrego 1981 for a discussion). It is worth pointing out that the fronting of arguments in type A and type Β causatives does not stem from the need for the subordinate verb to govern its complements, but rather from the need to have these complements governed by the causative verb following the reanalysis. Further note that in type C causatives, too, the assignment of dative Case to the subordinate subject is non-problematic as the configuration in (22) is met. Lexical, subcategorized material intervenes between the verb faire and the subordinate subject; in this case, the accusative Case assigned to the direct object of the embedded verb. This Case is X in (22), which makes the formation of dative possible. Note that the subordinate subject is not, in this case, an argument of faire, as there was no merger of arguments. In this sense the assignment of dative to the subordinate subject in these cases is similar to exceptional Case marking (and see fn. 8 for some discussion). As a last point, note that our analysis forces us to assume that the case assigned to ses parents in (38) is not assigned directly via the verb. We will return to this point in section 4 below. An interesting question arises with respect to type Β and C causatives, in which Case is assigned by one governor and Θ-role is assigned by another governor. Note that this situation is unique to causative constructions as we described them, since it is only in that special reanalysis structure that one will find two governors satisfying definition (67) in Chapter 2 for one complement. Note, however, that when the Case assigner does not coincide with the Θ-role assigner, a clitic, spell-out of Case, has no Θ-grid to attach to. This constitutes an apparent problem for the Complement Matching Requirement. Viewed from a different angle, however, this problem dissolves. Recall that the Complement Matching Requirement is a well-formedness condition on indexing which is checked inside the Θ-grid. If we assume that Case assignment involves the encoding of a referential index in a similar fashion, it is clear that Case assignment and Θ-role assignment need not be accomplished by the same agent. Rather, we could assume that all instances of the same indexing are assembled together by an extended Matching Requirement, and then that index is checked for Case, for Θ-role assignment etc. In this sense our system clearly resembles Chomsky's (1981) chains in which the compatibility of an NP with well-formedness conditions (such as Case
179
Clitics and Reanalysis in Romance
assignment and Θ-role assignment) is not checked for an isolated occurrence, but for a chain composed of all occurrences of the same index. Let us now summarize our discussion. Following the typology in (45), we predicted three types of causative constructions. All three were attested by causative construction in River Plate Spanish and in French. However, none of the languages had all three types. Rather, we can describe the availability of causative constructions in these languages by the diagram in (47): (47) RP Spanish French
arguments merger obligatory optional
Case merger optional obligatory
The table in (47) predicts that in RP Spanish all Θ-complements will be fronted, but clitics corresponding to these arguments will be attached either to hacer or to the subordinate verb. In French, on the other hand, only accusative and dative complements will be obligatorily fronted. Subcategorized PP's will be fronted optionally. On the other hand, clitics corresponding to fronted complements will always be attached to faire and never to the subordinate verb . 13 The typology offered for causative constructions, coupled with the verified predictions it makes thus supplies further support both to our analysis of clitics and to our analysis of causative constructions. 14 2.5 Permitir - Type Verbs in RP Spanish Another "two-storey" construction in RP Spanish shows a distribution of clitics that can be easily explained under our assumptions. This is the permitir-type construction - a class of cases in which V j takes a dative object which controls the subject position of V 2 . Consider the sentences in (48) (based on Rivas, 1977): (48)
a. Maria le permitio tocar la flauta a Jose 'Maria permitted Jose to play the flute' b. Maria le permitio a Jose tocar la flauta (marked order) c. *Maria permitio tocarle la flauta a Jose
We propose that the underlying representation of (48) is as in (49):
180
Parametric
Syntax
(49)
tocar
la flauta
We will assume that permitir is a reanalysis verb in the sense discussed above, and that following the fronting it takes as its complements the arguments of V 2 ._It follows that only V can be fronted, since otherwise government into V2 is not possible. The fronting of V will result in the structure in (50). (Note, however, that in this case the fronting is altogether optional): 15 (50)
V2 V2
a Josej
NP
NP
PRO [e]
The derived structure in (50) already predicts the ungrammatically of (48c) because the clitic is adjoined to V 2 , and thus fails to c-command elements outside V 2 . Hence the clitic can neither govern nor be coindexed with PPj, a Jose. Our analysis makes some more predictions with respect to the availability of clitics in the derived structure and the underlying structure. First, it predicts that if V preposing does not apply and the structure remains as in (49), (corresponding to sentence (48b)), the object clitic cannot be attached to V t : it can only be attached to V 2 . This follows from the fact that government into V2 would be impossible from a position attached to V j . However, if V 2 has been fronted, we predict that the object clitic can be attached both to V 2 and to V ! . This is due to the optionality of Case merger in RP Spanish (cf. (47) above). These predictions are in fact correct: 16
Clitics and Reanalysis
in Romance
(51)
a. sej la permiti escribir a Juanj Ί permitted Juan to write it' b. le permiti besariz a Juan Ί permitted Juan to kiss her'
(52)
a. le permiti a Juan escribir/a b. *sej la permiti a Juan escribir
181
Once again, these facts can be explained in a natural way, assuming clitic government and the reanalysis process sketched above. To conclude this section. It has been shown that some interesting facts concerning the distribution of clitics can be explained if we assume the theory of clitics sketched earlier in this study. We have further demonstrated that the relationship which holds between a dative clitic and the position with which it is coindexed is identical to that which holds between an accusative clitic and the position with which it is coindexed. Thus in causative constructions, the dative clitic corresponding to the subordinate subject could only be cliticized to a verb which governs the subordinate subject position. While dealing with the causative constructions, we proposed a system of reanalysis which yielded itself to parametric variation. These variations then predicted in a natural fashion the range of causative constructions attested in River Plate Spanish and in French. In Chapter 1 we suggested that interlanguage variation stems from the properties of the inflectional system. We took it to be a property of inflectional morphology, that it may "displace" specified properties of lexical items as well as perform operations of inflectional copying and insertion. It is worth noting that the analysis proposed here for the "two storey" constructions in Romance is entirely compatible with this view. Consider again the parametric table in (47). Two properties, we claim, are involved in reanalysis: one is a process of merger of Case slots - essentially, the transfer of Case slots from one Case assigner to another. The other involves the merger of complements (essentially Θ-slots) from one to the other. Neither of these operations eliminate a lexical feature or insert a previously non-existing lexical feature. As such, both are compatible with our interpretation of the Projection Principle in Chapter 1. In fact, both operations fit perfectly within our description of inflectional rules. In fact, there is no reason to assume that the process of Case transfer or Θ-transfer from a Case assigner or a Θ-role assigner to an NP is different in principle from the process of transferring Case or Θ-slot from an assigner to an assigner. (For a proposal along these lines see also Manzini, 1983).
182
Parametric
Syntax
Insofar as our analysis of reanalysis constructions is compatible with our view of paramters, it clearly supplies strong evidence for it.
3. EXTRACTION FROM CLITIC-DOUBLING CONFIGURATIONS IN RP SPANISH
Jaeggli (1982) observes that extraction from clitic-doubling configurations in RP Spanish is restricted to dative constructions. Thus we have the following contrast: (53)
a. lo vimos a Juan b. *a quien lo vimos? 'who did we see?' c. a quien vimos? 'who did we see?'
(54)
a. le han regalado ese libro a Juan 'they gave this book to Juan' b. a quien le han regalado ese libro? 'to whom did they send this book?'
Jaeggli shows very convincingly that the contrast extends to all occurrences of variables in the object position of doubled constructions. Thus we find the same contrast in relative clauses: (55)
*Maria, a quien la he visto ayer. estaba muy preocupada 'Maria, who I saw yesterday, was very worried'
(56)
Maria, a quien le han regalado ese libro, estaba muy preocupada 'Maria, to whom they gave that book, was very worried'
In configurations which have WH elements in situ we find the same situation: (57)
a. *le viste a quien? b. viste a quien?
(58)
le han regalado ese libro a quien?
and the same holds for quantifiers (59)-(60) and focus (61)-(62): (59)
a. *las vi a todas las chicas
Clitics and Reanalysis in Romance
183
b . vi a tod as las chicas Ί saw all the girls' (60)
les regalaron libros a todos los chicos 'they gave books to all the boys'
(61)
a. *yo le vi a JUAN b. yo vi a JUAN Ί saw JUAN' (focus reading)
(62)
yo le regalare ese libro a JUAN 'they gave this book to Juan
Jaeggli accounts for this contrast by assuming that of the pair clitic/ doubled element, the latter is never governed. This holds b o t h for doubled direct objects and doubled indirect objects. It follows that an empty category left in the doubled position, i.e. the variable left after extraction, will never be governed. Since in Jaeggli's system [ppe] are not subject to the Empty Category Principle, it follows that, although extraction from doubled indirect object configurations will leave an empty category, this will not suffice to rule out the o u t p u t . In the direct object position, however, the variable is of the type [jsjpe], and thus must be properly governed in accordance with the ECP. Since this position is never governed (let alone properly governed), all such occurrences are ruled out by the ECP. (Parallel to the dialect of RP Spanish which does not allow for extraction from direct object doubled constructions, there is another dialect of RP Spanish which allows for such extraction. See Montalbetti, 1981; Hurtado, 1980. We will refer to the dialect which allows for extraction as RP Spanish B, and to the dialect which does not allow for such extraction as RP Spanish A.) As is obvious from the analysis of clitic-doubling configurations proposed in this study, we cannot adopt the solution proposed by Jaeggli. We have shown that the doubled position is governed by the coindexed clitic. In fact we have shown that b o t h in Romanian and Modern Hebrew the doubled position is properly governed as well, thus accounting for the availability of extraction from this position in these languages. Furthermore, in section 2 above we have shown that in RP Spanish itself the distribution of clitics in " t w o - s t o r e y " constructions can be explained if we assume that b o t h the direct object position and the indirect object position are governed by the clitic. (Note that arguing for clitic government results in arguing for clitic proper government as well. Since the conditions for proper government are government and coindexing, it follows
184
Parametric
Syntax
that whenever a clitic governs a coindexed position it automatically properly governs it as well.) More evidence against Jaeggli's proposal to account for the relevant distinctions in terms of a contrast between [ppe] and [j^pel· comes from a variation which is found in RP Spanish A, and which is referred to as the leismo dialect. In this dialect it is possible to substitute the accusative clitic in doubled constructions with a dative clitic, although the doubled element is still the direct object. Such substitution is attested in (63): 1 7 (63)
lej vimos a Juan; him (dat) 'we saw Juan'
If such substitution occurs, extraction is grammatical: (64)
a. a quien; le, vimos? (compare (53b)) b. Mariaj, a quienj le, he visto ayer, estabe muy preocupada (compare (55)) c. lej viste a quienj? (compare (57a))
Note that a proposal that extraction from doubled constructions depends crucially on the categorial nature of the extracted element clearly cannot account for the grammaticality of (64). Rather, it is more plausible to argue that a doubled element which has a corresponding dative clitic can be extracted, whereas a doubled element with a corresponding accusative clitic cannot be extracted. We would like to suggest that the parameter which distinguishes Romanian and RP Spanish Β from RP Spanish A is closely related to this fact. Recall that we have been assuming the definition of proper govenment as in (65): (65)
a properly governs β iff α governs β and: i. a i s [ + V ] ; o r ii. α is coindexed with β
In our analysis of clitic doubling we have relied mainly on clause (ii) of this definition. Now let us assume that the coindexing referred to in (ii) is well-formed only if a agrees in all its features with ß. Such agreement of features will include agreement in gender, person and number (a fact that we have been tacitly assuming in our discussion of Romanian) and also Case. We will assume that the requirement of Case agreement is only valid if β has Case. It is important to note here that the Case agreement requirement is a condition on proper government and not on coin-
Clitics and Reanalysis
in
Romance
185
dexing. Thus we do not assume that coindexed elements have to agree in Case. Rather, we assume that a coindexed governor has to agree in Case with the coindexed element in order to properly govern it. Let us now assume (contrary to Jaeggli, 1982) that in RP Spanish A the marker a has Case assignment properties identical to its prepositional counterpart: it assigns dative, not accusative Case. 18 Thus in (66) Juan is dative: (66)
loj vimos a Juanj (dat)
Note that the coindexing between lo and Juan is still well-formed, although the clitic is accusative and Juan is dative. This is due to the fact that Case agreement is not a condition on coindexing, as explained above. However, under extraction the situation is different. In sentences such as (67) the empty category is marked as dative, as is its antecedent, a quien. The clitic, however, is a spell-out of the Case features of the verb, and thus is accusative: (67)
*a quienj loj vimos [e]j? (dat) (acc) (dat)
Following the requirement that coindexed governors agree in Case with empty elements which they properly govern, we expect the ungrammaticality of (67). Thus it follows that precisely in those cases where the clitic is accusative and the doubled element is dative extraction is not possible. A different situation holds in (68). (68)
loj vimos [e]j (acc)
In this case a was never inserted. Consequently there is no reason to suppose that the empty category is dative. In fact there is no reason to assume that it is Case-marked at all. Thus it can be properly governed by the coindexed, governing clitic, and the ECP is not violated. One could argue that verbs in RP Spanish, as in Romanian, are proper governors themselves. It then follows that (67) should be grammatical, since the verb can govern the empty category. Recall, however, that we are assuming that the complex clitic + verb is one lexical unit, and that the clitic is a spell-out of a feature of the verb. We would like to argue that since this complex contains an index which is identical to the index of the governed element, clause (ii) of the definition of proper government
186
Parametric
Syntax
has to be m e t : this coindexing has to enter the proper government relation. Since coindexing cannot create a situation of proper government here due to the Case conflict, the complex as a whole cannot serve as a proper governor. In effect this means that although the verb is a proper governor, it cannot properly govern an element which contains conflicting information. Let us now turn to the grammaticality of sentences such as (53c) in RP Spanish A. In this case extraction of an a phrase has occurred, leaving, presumably, a dative empty category. In this case, however, the sentence is grammatical. Still there really is no reason why (53c) should be ungrammatical: the empty category is governed by the verb. Since the governing category does not contain an element which is coindexed with the trace, the Case requirement is irrelevant and the [+V] element can freely properly govern an empty category. The difference in extraction facts between Romanian and RP Spanish A is now explained by utilizing the properties of a d u m m y Case assigner: the marker a in RP Spanish A and pe in Romanian. Note that as in Romanian (but not Hebrew) the formative a is available in the base. This is clear from the fact that it interacts with "Move a " ; when a WH element is fronted, it is fronted with the marker a. Unlike Romanian pe, however a assigns to its complement a Case which is different from the Case features of the verb. In Romanian pe assigns accusative, just like the verb preceding it. This difference accounts for the availability of extraction in the latter and for its ungrammaticality in the former. Clearly it is strange, functionally speaking, for a marker which is essentially a semantic marker (as α is in RP Spanish A) to assign a Case differing from that assigned by the verb that actually s u b c a t e g o r i e s for the NP in question. However, when no clitic is attached to the verb, as in (69), the situation is even stranger: (69)
vimos a Juan
The marker a assigns dative Case to Juan. Consequently the accusative Case features of the verb 'to see' are not realized at all: they are never assigned to any complement, nor are they spelled out as a clitic. This situation, we would like to argue, triggers the rule of Clitic Spell-Out. The "idle" accusative features are spelled out as a clitic on the verb, resulting in clitic doubling in direct object configurations. On the other hand, a may be reanalyzed as an accusative marker. This, we believe, is the situation both in RP Spanish B, in which extraction from direct object configurations is possible despite doubling, and in Standard Spanish, in which doubling does not occur in direct object configurations.
Clitics and Reanalysis in Romance
187
4. A NOTE ON DATIVE CLITICS
Throughout the discussion in the previous sections we have been assuming that dative clitics are a spell-out of dative Case features of the head verb. This assumption was made explicit in section 2, when we discussed the assignment of dative Case to non-adjacent arguments in configurations such as (20) and (22). In essence, we argued that verbs have unspecified Case slots and that these Case slots will be spelled out as accusative Case under adjacency and as dative Case when the assigner is not adjacent to the element linked to the Case slot. The relevant configuration for a verb like sirvir in RP Spanish is given in (70): (70)
sirvir ^ ß ?
NP
teasel
teasel
acc
dat
In (70), sirvir has dative features which may, in t u r n , spell out as a dative clitic attached to the verb. Given this account, several questions come to mind immediately. First, we clearly need additional arguments to account for the assignment of dative to an adjacent argument in sentences such as (71 a - b ) : (71)
a. Jean a telephone ä ses parents Jean has telephoned to his parents 'Jean called his parents' b. Pedro lesj llamo a suspadresj Pedro them called to his parents 'Pedro called his parents'
Since in (71 a - b ) dative Case is assigned to an adjacent argument, these cases clearly cannot be collapsed with the dative Case assignment rule in (22). A second question concerns clitic doubling in dative constructions, both for sentences such as (71b) and sentences such as (72), in which dative Case is assigned in accordance with (22): (72)
(= 15a)
Pedro lej sirvio la comida a J u a n j Pedro him served the food to Juan 'Pedro served the food to Juan'
The doubling in (71b) and in (72) (RP Spanish) contrasts with the impossibility of doubling for (71 a) and for (73) (French):
188
Parametric
Syntax
(73)
Marie (*luij) a servi la soupe ä Pierrej Marie him has served the soup to Pierre 'Marie served the soup to Pierre'
If the dative clitic is the spell-out of the dative Case slot, then dative Case is absorbed by the clitic and the grammaticality of (71b) and (72) remains unexplained. If, on the other hand, the clitic does not absorb the Case, the source of the clitic remains mysterious. Furthermore, the reason for the difference between French and RP Spanish is entirely unclear. Why does one language allow doubling whereas the other does not? 4.1. Dative Case Assignment Consider first the assignment of dative Case in (71a-b). We are tacitly assuming that the argument of the verb in these sentences is an NP. Note, however, that there is another logical possibility. One might suggest that the verb subcategories for a PP rather than for an NP. Dative Case would then be assigned directly by the preposition and the verb would have no dative Case features at all. According to this last hypothesis, dative clitics could no longer be regarded as a spell-out of dative Case features of the verb. Interestingly, both in RP Spanish and in French there is some evidence that indicates that these a NP/d NP combinations are NP's rather than PP's. Vergnaud (1974) provides two tests which indicate that some ά NP combinations in French are NP's rather than PP's. Kayne (1975) observes that other ά NP combinations systematically fail some of these tests. Thus it is plausible to assume that, whereas the former are NP's which are marked as dative, the latter are genuine PP's. This contrast is exemplified in (74)-(76): (74)
conjunction of objects of prepositions: a. *ils ont parle' ä Marie et le directeur 'they talked to Mary and the director' b. ils se sont assis sur la table et les chaises 'they sat on the table and the chairs' c. ils ont pense ä Marie et le directeur 'they thought about Marie and the director'
(75)
PP's vs. NP's as heads of relative clauses. a. il a parle ä l'homme et ä la femme qui se sont recontres hier 'he talked to the man and to the woman who met yesterday'
Clitics and Reanalysis in Romance
189
b. *il a compte sur l'homme et sur la femme qui se sont rencontre's hier 'he counted on the man and on the woman who met yesterday' (76)
PP clitics vs. NP clitics a. je parle ä Jean b. je lui parle c. je vais ä Paris d. j'y vais e. je pense ä Jean f. j'y pense g. *je lui pense
In (74) we see that, unlike the real preposition sur in (74b), the marker ά cannot take a conjoined object. However, when it appears following the verb penser it can take a conjoined object. In (75) we see that dNP's may serve as heads of relative clauses, whereas PP's cannot. In (76) we observe that, while the argument of the verb parier may be a dative clitic, the argument of the verb penser cannot be a dative clitic (cf. 76g). Rather, if cliticized, it has to be a PP clitic (cf. 76f). (For a more detailed review of various distinctions among ά NP's which are real PP's, ά NP's which are NP's. and PP's see Jaeggli, 1982.) The object of ά both in the penser cases and in the parier cases is plausibly marked as dative. Nevertheless there are two differences between these two configurations. First, parier subcategorizes for an NP while penser takes a PP. Secondly, while parier takes a dative clitic, penser takes a PP clitic. If we wish to reduce these two distinctions to one it is plausible to assume that with penser, dative is assigned by the preposition itself, whereas with parier the verb has dative Case features. This account will reduce the contrast between the availability of dative clitics for parier and their unavailability for penser to the fact that penser subcategorizes for a PP, whereas parier requires a dative complement. A similar situation holds in RP Spanish, which has no PP clitics. However, a NP's behave differently from real PP's in two respects. First, the former have corresponding dative clitics whereas the latter two do not: (77)
a. lesj mandaron cartas a los padresj 'they send letters to the parents' b. Juan lesj hablo a sus padresj 'Juan spoke to his parents'
(78)
a. Juan fue a Paris 'Juan went to Paris" b. *Juan le fue
190
Parametric
Syntax
Secondly, there is, as discussed above, a stylistic constraint in RP Spanish against the occurrence of two a phrases (see footnote 3 for discussion). This constraint holds for a sequence of two a phrases when they are both indirect objects or when they consist of a direct object preceded by a and followed by an indirect object. However, the constraint does not hold for two a phrases when they are both directional PP's, nor does it hold when the first a phrase is an indirect object and the second is a directional PP: (79)
a. ?Juan presento a Pedro a Jose 'Juan presented Pedro to Jose' b. ?Juan lej hizo escribirlesj α los chicosj a Josej 'Juan made Jose write to the children' c. Juan llevo a Maria al cine a las cinco 'Juan took Maria to the movie at 5 o'clock' d. Juan lo presento a Pedro a las cinco 'Juan introduced him to Pedro at 5 o'clock' e. Juan lesj hablo a sus padres; a las cinco 'Juan spoke to his parents at 5 o'clock
The distinction between the grammaticality of (79c-e) and the maiginality of (79a-b) indicates that the restriction against two adjacent a phrases does not hold for genuine PP's. Furthermore, the marginality of both (79a) and (79b) suggests that in these sentences all the a's present are of the same type. We have shown above that the a preceding Pedro in (79a) is an inserted Case marker. This strongly suggests that the a's preceding Jose in (79a) and (79b) and the a preceding los chicos in (79b) are inserted Case markers as well. Thus it is clear that both RP Spanish and French have verbs which take dative complements which are NP's and not PP's. We will now discuss the way these NP's are assigned Case. Two plausible mechanisms may assign dative Case to an adjacent argument. The first one would involve the marking of a Case slot as [+dative], or, for concreteness, as a. Thus, assuming this mechanism, the verb telephoner in French would be marked as follows: (80)
telephoner
NP
t In essence, this proposal implies that the dative and accusative Case assignment procedures described in section 2 may be preempted by an already existing Case marking in a Case slot. These cases, however, will be more costly. In (80), a is assigned directly to the post-verbal NP, once it is linked with the Case slot of telephoner. Note that if this is in-
Clitics and Reanalysis in Romance
191
deed the correct way to assign Case to a dative NP, we do not expect doubling. A clitic would be linked with the same Case slot and will absorb its Case, making the appearance of a phonologically realized NP impossible. Now consider another possible way to assign dative to post-verbal indirect objects. According to this second system a rule of a insertion takes place, very much like the rule of sei insertion and the rule of pe insertion discussed in Chapter 4: (81)
a Insertion Φ
(AI)
*a/[vp
NP]
Some things should be noted with respect to (81). First, note that the same rule could insert a preceding a direct object and an indirect object, since the rule in (81) does not specify the type of argument which triggers the application of the rule. Furthermore, it could apply preceding the indirect object in sentences such as (77a), inserting a preceding Jose. Secondly, note that the rule in (81) is independently needed in RP Spanish in order to insert a preceding direct objects in the semantic environments specified in (1). Thus the application of (81) preceding the indirect object in Spanish cannot be blocked. Overgeneration of the rule in (81) will be ruled out by Case conflict. If a is inserted preceding a [NP, PP] which is assigned Case by the selecting preposition, the derivation will be ruled out. On the other hand, consider the insertion of a by rule (81) in sentences such as (82a-b): (82)
a. (loj) vimos a Juanj 'we saw Juan' b. * vimos a una camisa 'we saw a shirt'
In (82) the insertion of a preceding Juan is not only grammatical: its failure to apply would lead to ungrammatically (see discussion in section 1). On the other hand, insertion of a preceding una camisa would result in ungrammatically. This is due to the animate and specific interpretation that would be imposed on 'shirt', resulting in a semantic anomaly. Thus, rather than stating that accusative Case assignment is obligatory in (82b), the insertion of a in this case would result in ungrammaticality. Now let us reconsider doubling in dative constructions in RP Spanish. The insertion of a preposition, as in (81), does not supply the verb with dative Case features. Nevertheless we find dative clitics in RP Spanish, so we would like to argue that in RP Spanish, verbs such as hablar are marked as having a dative Case slot similar to telephoner in French. It
192
Parametric
Syntax
follows that in RP Spanish we have a redundancy: the verbs which select an adjacent dative complement are marked as [+dative], but the language also has a rule whose existence is independently motivated, which inserts a preposition in the environment specified by (81). It is this duality which allows for clitic doubling in dative constructions: in sentences such as (71a) and (72), the clitic is the spell-out of the dative Case slot, while the indirect object is marked as dative by the inserted preposition. Now consider French. We have no reason to believe that French has a rule of preposition insertion alongside the dative Case slot suggested in (80) for the verb telephoner. Thus the redundancy which is attested in RP Spanish never occurs in French; consequently we do not find doubling. An interesting piece of evidence for our analysis is presented by Jaeggli (1982), who points out that in RP Spanish a in indirect objects behaves like a real preposition in the following environment: 19 (83)
a. les compararon una casa a Maria y Pedro 'they bought a house for Maria and Pedro' b. Juan les hablo a Maria y Pedro 'Juan talked to Maria and Pedro'
In (83) the preposition a may take a conjoined object. This state of affairs contrasts with the ungrammaticality of the equivalent French sentence, given in (74a). The assumption that dative marking in RP Spanish is achieved by rule (81), while dative marking in French is achieved by a device of the sort illustrated in (80) can account for the grammaticality of (83) vs. the ungrammaticality of (74a). Let us consider the nature of the rule in (81). In this case, the preposition a is inserted and Chomsky-adjoined to the NP. We argued above that this preposition has dative Case-assignment features. These Case-assignment features are then assigned to each of the conjoined NP's in a fashion similar to the assignment of Case to conjoined NP's by a governing preposition in a genuine PP. Thus we expect the grammaticality of (83). On the other hand, in French, following the procedure in (80), ä is simply a morphological manifestation of the Case marking of a particular NP. Since every NP has to be Case-marked, and in (74a) there are two NP's with only one morphological manifestation of Case assignment, the sentence is ruled out. Of course, if ά is attached to both parts of the conjunction the sentence is grammatical. In this way the different properties of (80) and (81) predict the difference between (74a) and (83). Thus we find again that parametric variation between languages can be reduced to the properties of the inflectional component. The different
Clitics and Reanalysis in Romance
193
process of dative Case assignment in French and in RP Spanish accounts for the availability of doubling in one vs. the absence of doubling in the other. Lastly, we would like to reconsider sentences such as (38). In these cases we must argue that the preposition ά is a true preposition which assigns Case directly to ses parents. The fronting of the verb telephoner and the merger of the Case slots of telephoner and faire does not enable telephoner to assign Case to ses parents directly. We would like to propose that while the Case assignment device proposed in (80) is the dative Case assignment used primarily in French, French may also utilize a rule of ά insertion similar in nature to the rule of Spanish. This rule, however, is utilized only in the absence of a Case slot. Thus while RP Spanish has both the dative Case slot and the preposition insertion rule, French utilizes one of these techniques, but not both, the second one being a marked option. Note, incidentally, that the problem does not arise for dative complements which follow the Case assignment pattern in (22) above. In those cases the dative complements cannot be left behind in causative constructions, and must be fronted due to their Case dependency. (See (41) above and related discussion.) This fact further suggests that we are dealing here with a marked structure, governed by lexical idiosyncracy. 4.2. Inalienable Possession Constructions As a last point we will consider the inalienable possession construction in RP Spanish. This construction will supply interesting confirmation for our proposal that clitics are best viewed as a morphological part of the verb. Consider the sentences in (84): (84)
a. lej duele la cabeza a Juanj him hurts the head to Juan 'Juan has a headache' b. lej rompieron la pataj a la mesaj they broke the leg to the table 'they broke the table's leg'
Two important properties of the inalienable possession construction should be pointed out. First, the doubling in (84a-b) is obligatory. The failure to double results in ungrammatically, as is demonstrated by (85a-b): (85)
a. *duele la cabeza a Juan b. * rompieron la pata a la mesa
194
Parametric
Syntax
Secondly, the indirect object in (84a-b) must be construed as the possessor of one of the other arguments. In (84a) la cabeza 'the head' can only be interpreted as Juan's head. In (84b) la pata 'the leg' can only be interpreted as the table's leg. Jaeggli (1982) observes that the ungrammatically of inalienable possession sentences in which there is no doubling seems to stem from the fact that in the absence of the clitic the indirect object in (85a-b) is interpreted as a goal. Since the verb in these cases does not select a goal complement, the sentences are ruled out as semantically deviant. In view of this, Jaeggli concludes, it is natural to assume that the obligatoriness of a clitic must be derived from the Θ-criterion. We will adopt the intuition behind Jaeggli's reasoning and derive the obligatoriness of the clitics in (84a-b) from the Θ-criterion. However, we will execute this idea in a different way. Jaeggli argues that in (84a-b) the clitic can be said to bear a special Θ-role, Θρ. This Θρ is then transmitted to the NP object by a special rule which is sensitive to coindexing (but not to government or c-command. Recall that in Jaeggli's system the clitic crucially does not govern or c-command the doubled element). We would like to argue, on the other hand, that no transmission rule is necessary. Rather, the clitic, as a morphological affix of sorts, simply effects the nature of the Θ-role assigned to the indirect object. Consider how this is done. Recall that we argued that the process of Θ-role transmission involves the transference of a referential index from an argument to an empty slot in the thematic grid of the head. Recall that when a clitic is present, it is linked to one of the thematic grids of the head. It is the presence of the clitic in the Θ-grid which allows it to alter the Θ-role to Θ ρ . For concreteness, we will assume that a verb such as rompir in (84b) has two Θ-grids. One is that of the direct object, which is a patient. The second lacks a concrete Θ-specification (in particular, it is not a goal). Rather, the attached clitic which is present in that Θ-grid is specified as + Θ ρ affix. The resulting situation is given in (86): (86)
[VV,
•cli
Δ _theme_
(
Δ _ ρ-role _
The empty slots in (86) are replaced by the referential indices of the subcategorized complements. In Jaeggli's system, the 'formation' of Θρ crucially depends on the process of transference. In our system, on the other hand, it is a morpho-
Clitics and Reanalysis in Romance
195
logical process which effects the Θ-structure of a given head. Any account of morphology will have to allow for such operations. Morphological operations may obviously alter the Θ - s t r u c t u r e of the heads to which they are attached. As a simple example one could point to the relationship between an active verb and its passive participle (eat-eaten), a middle verb and a transitive verb (break-break), a causative verb and its n o n causative counterpart in languages which allow for morphological causativization (as in Hebrew 'axal - he'exil, 'ate - 'fed') etc. (for a more detailed discussion of the interaction between morphology and argument structure, see Williams, 1981, Marantz 1981 and references cited there, as well as Chapter 1 above.) Thus the dependency between the clitic and the Θ-role assigned t o the indirect object confirms in an interesting way our claim that the clitics are morphologically part of the head: they participate in a grammatical operation which is typical of morphological affixes and which is best restricted to a word-internal level. They influence the nature of a Θ-role assigned to an argument. (For a more detailed discussion of inalienable possession constructions and some discussion of the phenomenon as it appears in French, see Jaeggli 1982 and references cited there.) 2 0 Let us now summarize our discussion in this chapter. In section 2 we argued that the requirement that clitics govern their doubled elements (whether fully realized or [e]) can account in an interesting way for the distribution of clitics in " t w o - s t o r e y " constructions in RP Spanish. Further, we proposed an analysis of causative constructions that involved a rule of V^ fronting. We showed that limiting the fronting to V in most cases follows from b o t h the requirement that a head must govern its complements and from the process of reanalysis in causative constructions. We further showed that adjunction of the fronted category is itself free, but is only well-formed at the V level of the matrix verb. This fact too follows from the government requirement and from the formulation of the reanalysis process. It was shown that once the right configuration has been established, the distribution of clitics in causative constructions follows directly from the government properties of the structural configuration. Clitics consistently appear precisely in those positions which allow them to govern the doubled elements. In section 2 we proposed a particular system of Case assignment, involving the linking of two tiers: the Case tier and the complements tier. Crucially, the Case slots were unspecified and their identity was dependent on the linking with the complements tier. This system was coupled with an analysis of reanalysis constructions, the latter subject to parametrization in a particular fashion. This coupling allowed us to derive certain
196
Parametric
Syntax
distinctions between causatives in French and causatives in RP Spanish. Likewise, a distinction in the distribution of the subordinate object clitics in causative constructions in these languages was reduced to the same parameter. The parameter involved conformed crucially to our model. It involved the transferrence of features from one lexical item to another, a process allowed by the inflectional rules formulated in Chapter 1. In section 3 we argued that some differences between extraction facts in some dialects of RP Spanish and Romanian can be explained if we bear in mind the fact that a in RP Spanish is a dative marker, whereas pe in Romanian is an accusative marker. Given this distinction and a slight change in the formulation of proper government, requiring that a coindexed proper governor agree in Case with the element which it governs, we reduced the differences between Romanian and the relevant dialect of RP Spanish to the idiosyncratic properties of dummy Case markers. In section 4 we elaborated on the nature of dative clitics. We produced some evidence for the existence of dative Case assignment features in particular verbs which can in turn be spelled out as clitics. Thus we explained the nature of dative Case assignment, and argued that the assignment of dative Case in RP Spanish differs from the assignment of dative Case in French. While the former involves the insertion of a preposition, the latter involves linking an NP with a dative Case slot. This difference was then shown to account for the availability of clitic doubling in indirect objects in RP Spanish vs. its absence in French. In the last paragraphs of section 4 we showed that Jaeggli's conclusions with respect to the interpretation of inalienable possessive can be incorporated quite naturally into a system which assumes the mechanism of Θ-role assignment outlined in Chapter 2. Within such a system the obligatoriness of the clitic is captured as a morphological process effecting the assignment of Θ-roles. As such, it is on a par with the obligatoriness of passive morphology in a passive construction, effecting the assignment of Θ-role to the subject. Insofar as the parametric system proposed in this chapter accounts for the phenomena presented, they supply strong evidence for the system as well as for the theoretical assumptions which motivated it.
NOTES 1. Wc are tacitly assuming here that verbs have dative assignment features, thus accounting for the appearance o f a dative clitic attached to the verb. Note that this assumption is by no means necessary. On could, for example, assume that the verb subcategorizes for a PP. regardless o f the Case assigned to the objcct o f the preposition. Our assumption will be discussed and motivated in section 4 below.
Clitics and Reanalysis
in
Romance
197
2. Note that we assume that in causative constructions the subordinate clause is S rather than S. To the best of our knowledge there is no direct evidence against this assumption. In fact, in many investigations of causative constructions it is suggested that following the preposing rule, the lower clause is completely destroyed (Aissen (1974), Rivas (1977)). For our purposes the assumption that it is S rather t h a n S is crucial, as a maximal projection would block government of the subject position by the matrix verb (we take S to stand for INFL and S stand for INFL; see Chapter 1 for a discussion). Insofar as our analysis accounts for causative constructions in a natural way it supplies evidence for S vs. S in this position. 3. Rivas (1977) observes that a sequence of two a phrases, as in ( l i b ) , is not felicitous. Following Rivas, we take this constraint to be stylistic in nature. Thus if one of the a phrases is fronted the sentence loses its awkwardness: (i)
a Josej Maria lej hizo escribirlesj a los chicosj
(see below for some discussion of a similar constraint in French.) 4. It is possible that the requirement that a verb should govern its complement at S-structure should be relaxed to allow the trace of a fronted verb to satisfy this requirement. As will become apparent from the text, the obligatoriness of the fronting of all strictly subcategorized complements in RP Spanish will follow even if the trace, rather than the antecedent verb, can satisfy the requirement that a head should govern its complements. As will be shown below, following reanalysis, the verb hacer has to govern the subcategorized complements of V 2 , thus rendering the fronting of the subcategorized complements obligatory. 5. The proposal to account for the Case assignment properties of causative constructions by equating them with the Case assignment properties of certain verbal paradigms is from Ii. Wehrli. Wehrli notices that the same situation holds in French. Thus we have the sentences in (i) and (ii) below contrasting with those in (iii) and (iv): (i)
a. Marie a servi la soupe ä Pierre 'Marie served the soup to Pierre' b. Marie lui a servi la soupe c. Marie l'a servie ä Pierre
(ii)
a. Jean a vole lc livre ä Pierre 'Jean stole the book from Pierre' b. Jean lui a vole le livre c. Jean l'a vole ä Pierre
(iii)
a. b. c. d.
(iv)
Jean a servi Pierre *Jean a servi a Pierre Jean l'a servi *Jean lui a servi
a. Jean a vole Pierre b. *Jean a vole ä Pierre c. Jean l'a vole d. *Jean lui a vole 6. (16b) is in fact grammatical for speakers of the leismo dialect, briefly mentioned in section 3 below. For these speakers accusative clitics can be replaced freely with the dative clitic le. For non-leismo speakers, however, (16b) is ungrammatical.
198
Parametric
Syntax
7. Our system of Case assignment is obviously influenced by current theories of autosegmental phonology. For references see in particular Halle and Vergnaud (forthcoming), McCarthy (1979) and references cited there. 8. For an analysis of dative constructions in English which is compatible with this conclusion see Stowell (1981), w h o argues that t h e structure of (i) is as in (ii), where John is a clitic-like element attached to t h e verb: (i)
Mary gave J o h n a book
(ii)
Mary gave + J o h n j a book [e]j
Note that the Case assignment system proposed here has interesting implications for the so-called exceptional Case assigners. In those cases we have a Case slot, but no thematic c o m p l e m e n t . It is presumably this property which licenses the exceptional Case assignment to t h e subordinate subject. 9. We will see in section 3 below that in the dialect which we call RP Spanish A, the subordinate subject marked by a is always assigned dative Case. This, however, is not relevant to o u r point. It will be argued below t h a t , regardless of the Caseassignment properties of a, the verb in direct object configurations will come to have accusative Case assignment features by virtue of the configuration in (18). The dative marking of direct objects is a side effect of the use of the prepositon a as an object marker and of t h e fact that a has dative Case assignment properties even when it appears in an accusative environment. With respect to the causative constructions this would imply that, although hater will have an accusative Case assignment feature in the relevant sense, it may not be realized on the subordinate subject. Rather, it will be absorbed by a clitic and the subordinate subject itself will be marked as dative, due to t h e interfering effect of dative a. 10. Rivas (1977) observes that the attachment of clitics to V 2 and V, is partially determined by semantic considerations. Thus if the clitic is inanimate it is perferable to attach it to V, . On the other hand, if the subject is present and the object clitic is animate, it has to be attached to V 2 . This is exemplified by the following sentences: (i)
Maria 1c hizo escribirle 'Maria made him write to him'
(ii)
*Maria j ' ^ j le hizo escribir
(iii)
?Maria le hizo tocarla 'Maria made him play it'
(iv)
Maria sc la hizo tocar
In (i)-(ii) the complement of V 2 is animate (him) and thus it has to be attached to V 2 . In (iii)-(iv) it is inanimate (it) and thus it is preferably attached to V , . I believe that this can be explained if we bear in mind that RP Spanish is a clitic-doubling language and that in (i)-(iv) t h e subordinate subject has a corresponding clitic which is attached to V , . The availability of two animate clitics attached to V, would result in an ambiguity, but not the situation in which one of the clitics is inanimate: pragmatically speaking, the causer will usually be animate. Rivas further observes that certain clitic combinations are blocked. In particular, if a dative subject clitic
Clitics and Reanalysis
in
199
Romance
is attached to V , , then of t h e following possible c o m b i n a t i o n s , o n l y ( a ) - ( d ) are grammatical; (v)
b.
a.
d.
f. d a t , dat + V,
V2
d a t , acc + V,
V 2 + (acc), (dat)
V2 V , + dat
C o m b i n a t i o n (c) is grammatical only if the clitic which corresponds to the complem e n t of V 2 is inanimate. The explanation of this fact was discussed above. As for t h e u n g r a m m a t i c a l l y of (f), it seems to derive f r o m a constraint that blocks crossing lines between clitics and t h e gaps which correspond to t h e m , as is d e m o n s t r a t e d in (vi): (vi)
a.
*acc I
1 b. (dat
dat
V
[c| acc
V]
[e] acc
1
acc I
|e] dat
j—• |c] dat
In ( f ) the situation is as in (vii): (vii)
*|dat
acc + V , |
[V 2 + d a t ] I
[e] acc
[e] dqt
[c] dat
Since t h e fronting of t h e accusative clitic alone will create crossing lines, t h e sentence is ungrammatical. Recent work by Pcsetsky ( 1 9 8 2 ) links crossing constraints to p a t h s established b e t w e e n Α-binders and A-bindees. T o t h e extent that t h e crossing restrictions o n clitic ordering can be accounted for by t h a t system, it suggests that clitics may be viewed as Α-binders along the lines suggested by Riny H u y b r e g t s ( f o r t h c o m i n g ) . See also A o u n (1982), Kayne ( 1 9 8 2 ) and references cited there. For o t h e r accounts of the facts in (i)-(vii), involving crossing constraints or filters see Perlmutter ( 1 9 7 1 ) , Bordelois (1974), and references cited there. 11. The other logically possible c o m b i n a t i o n s are blocked by t h e constraints discussed in f o o t n o t e 10. The marginality of the | c l d + V, ] [V 2 + c l a ] c o m b i n a t i o n is explained by t h e preference for attaching inanimate clitics t o t h e higher verb discussed in f o o t n o t e 10.
200
Parametric
Syntax
12. Kayne (1975) suggests that in sentences such as (42a) in the text the clitic lui cannot be construed as the indirect object due to the specified subject condition (and see Rouveret and Vergnaud, 1980; Aoun, 1982; for different versions of the same idea). Note, however, that the specified subject condition is only relevant if the indirect object is not fronted. If it is fronted, as in (41) the specified subject condition can no longer be invoked. Note also that this restriction in Frcnch is similar in nature to the restriction in RP Spanish discussed in footnote 10 above. Further evidence against the specified subject condition as an explanation for the unavailability of an indirect object reading for the clitic in (42) is given in Wehrli (1981). Wehrli notes that the phenomenon attested in (42a) extends to the paradigm in (i) in which the cliticization of a direct object is blocked when certain clitic forms are used. In this case, the specified subject condition cannot be invoked: (i)
a. Jean a fait cmbrasser Marie ä Pierre 'Jean made Pierre kiss Marie' b. Jean I'a fait embrasscr [e] ä Pierre c. *Jean m'a fait embrasser [e| ä Pierre 'Jean made Pierre kiss me d. *Jcan vous a fait embrasser ä Pierre 'Jean made Pierre kiss you' e. Jean lui a fait embrasser Marie [c| 'Jean made him kiss Marie' f. Jean m'a fait embrasscr Marie [e] 'Jean made me kiss Marie' g. Jean vous a fait cmbrasser Marie [c] 'Jean made you kiss Marie'
The contrast between the ungrammatically of (i c - d ) and the grammaticality of (i f - g ) is similar in nature to the contrast between the ungrammaticality of (42a) with the clitic interpreted as the indirect object and the grammaticality of the reading in which it is the subordinate subject. However, the specified subject condition clearly cannot be invoked to explain the ungrammaticality of (i c-d). We thus conclude that the unavailability of an indirect object interpretation for the clitic (42a) is not related to the specified subject condition. 13. Note that our analysis crucially applies only to accusative and dative clitics in French. Thus, although wc predict that these clitics will never be attached to the lower verb, our analysis makes no claims with respect to the clitics en and y or the reflexive se. We say nothing in this study about the PP clitics en and y. Note that in many respects they differ from accusative and dative clitics: first, they cannot be described as a spell-out of Case features or as an insertion of gender, number and person features. Secondly, even in those dialects of Frcnch and Italian which allow for doubling, the latter is never attested with PP clitics. Further, it cannot be said that these clitics are subcategorizcd by the verb, hence wc do not expect the Complement Matching Requirement or the Government Requirement to be relevant for these clitics. In fact it seems that the best characterization of these clitics is indeed as pronominal elements moved and adjoined to the verbal element. An exception to this are cases when, for example, the clitic y actually corresponds to subcategorizcd elements, as in (ii): (i)
j'ai pense a Pierre
Clitics and Reanalysis (ii)
j ' y ai p e n s c
(iii)
*je lui ai p e n s c
in
Romance
201
We shall b r i e f l y r e t u r n t o t h e s e c o n s t r u c t i o n s in section 4 b e l o w . 14. T h e reader m a y wish t o raise a q u e s t i o n at this p o i n t : if w e are assuming t h a t clitics can o n l y bo a t t a c h e d to t h o s e e l e m e n t s w h i c h have t h e Case slots c o r r e s p o n d ing t o t h e m , w h y is it t h a t in R o m a n c e clitics are a t t a c h e d to t h e auxiliary verbs, r a t h e r t h a n t o t h e v e r b s themselves, as is d e m o n s t r a t e d by (i): (i)
J e a n l'a m a n g e ' J e a n ate i t '
F u r t h e r , u n d e r t h e a s s u m p t i o n t h a t t h e s t r u c t u r e of (i) is as in (ii), n o t e t h a t t h e clitic d o e s n o t govern t h e c o i n d e x e d [ e j : (ii)
J e a n lj-a [ y p m a n g e [ e ] j ]
We will r e t u r n to t h e s e s t r u c t u r e s in C h a p t e r 6, w h e r e we will argue t h a t t h e s t r u c t u r e of (i) is as in (iii) at S - s t r u c t u r e b u t as in (iv) in P F : (iii)
Jean a | y p I j - m a n g e [ e | j ]
(iv)
J e a n | y p l j - a + m a n g e [ejj]
Given t h e s e s t r u c t u r e s , t h e q u e s t i o n s a b o v e are a n s w e r e d : a t S - s t r u c t u r e t h e clitic governs t h e e m p t y p o s i t i o n a n d is a t t a c h e d t o t h e verb w h i c h has its Case f e a t u r e . Only at P F . f o l l o w i n g t h e a p p l i c a t i o n of a rule w h i c h lowers I N F L i n t o t h e VP, is t h e clitic a t t a c h e d t o t h e a u x i l i a r y . 15. N o t e t h a t in ( 5 0 ) w e a s s u m e t h a t t h e s u b o r d i n a t e clause is d o m i n a t e d b y S. Recall t h a t f o r causative c o n s t r u c t i o n s we a s s u m e d t h a t t h e s u b o r d i n a t e clause is d o m i n a t e d b y S, t h u s e n a b l i n g g o v e r n m e n t of t h e s u b o r d i n a t e s u b j e c t . T h i s s i t u a t i o n should be b l o c k e d in ( 5 0 ) , since t h e s u b j e c t p o s i t i o n is o c c u p i e d b y P R O . 16. Rivas observes that t h e same r e s t r i c t i o n s o n clitic d i s t r i b u t i o n w h i c h hold in causative c o n s t r u c t i o n s hold in ' p e r m i t ' t y p e verbs. F o r an e x p l a n a t i o n of t h e s e r e s t r i c t i o n s , see f o o t n o t e 10. 17. I a m i n d e b t e d t o M.-L. Z u b i z a r r e t a a n d Y. A o u n f o r p o i n t i n g t h e s e s e n t e n c e s out to me. 18. T h a t a is an accusative Case assigner is in f a c t a s s u m e d b y m o s t of t h e r e f e r e n c e s cited a b o v e o n R P S p a n i s h . H o w e v e r , o n l y Jacggli a s s u m e s t h i s f o r R P Spanish A; t h u s o n l y w i t h respect t o his s t u d y is t h i s a s s u m p t i o n relevant t o o u r discussion. As will be s h o w n b e l o w , we will a d o p t t h e a s s u m p t i o n t h a t a is an accusative m a r k e r f o r all dialects of S p a n i s h e x c e p t f o r R P S p a n i s h A . 19. Jacggli ( 1 9 8 2 ) argues that indirect o b j e c t s in R P Spanish are in fact PP's, a n d t h a t , as s u c h , t h e y d i f f e r f r o m indirect o b j e c t s in F r e n c h . Jacggli s u p p o r t s his claim w i t h t w o k i n d s of evidence. First, he i n d i c a t e s t h a t c o n j o i n e d N P ' s can serve as an o b j e c t of a in RP Spanish b u t n o t as o b j e c t s of ά in F r e n c h (of. ( 7 4 a ) a b o v e ) . As will b c c o m e clear b e l o w , we o f f e r a n o t h e r e x p l a n a t i o n f o r this f a c t . T h e s e c o n d k i n d of evidence is t h e fact t h a t in causative c o n s t r u c t i o n s s u b c a t e g o r i z e d PP's are f r o n t e d with t h e f r o n t e d V c o n s t i t u e n t in jRP Spanish, b u t n o t in F r e n c h . Assuming t h a t PP c o m p l e m e n t s are g e n e r a t e d u n d e r V , b u t t h a t i n d i r e c t N P c o m p l e m e n t s are g e n e r a t e d
202
Parametric
Syntax
under V, and further assuming that V-fronting in causatives always moves V, it follows that indirect objects are fronted in Spanish but not in French. Note, however, that in French PP's can be left behind: (iii)
je faisais parier Jean de Marie Ί made Jean talk about Marie'
Further recall that we had shown above that the differences between causative constructions in French and in RP Spanish do not depend on the categorical status of the indirect object. Jaeggli argues that the availability of doubling in indirect objects in RP Spanish but not in French follows from the different categorical nature of the indirect object in these two languages. Although our attempt to account for the same phenomenon is inspired by this idea, wc reject the conclusion that the categorical status of the indirect objects differs. Rather, the nature of the dative Case assignment is different. In RP Spanish it is assigned by a real P, whereas in French it is not. Rather than assume that indirect objects are PP's in RP Spanish, we assume that they are marked as dative by an adjoined preposition. 20. Wc are leaving open in this study the question of whether morphological affixes cffect Θ-structures only through attachment with Θ-grids. Note, however, that this would be a very interesting theoretical result, which we hope to pursue in future work.
Chapter 6
Parameters for INFL
1. INTRODUCTION
This chapter contains a discussion of some properties of the INFL node. Moreover, it will be shown that some of these properties can be parametrized to yield an interesting range of grammatical constructions in different languages. In the previous chapters we argued for a particular analysis of clitics. In this chapter we will extend this analysis to another type of clitics which are attached to the INFL node rather than to a phrasal head such as N, V, or P. The evidence will be from existential sentences in Modern Hebrew. It will be shown that in those constructions the clitic exhibits an interesting interaction with the pro-drop phenomena. Section 2 of this chapter contains a presentation of existential sentences in Hebrew, indicating that clitics in these constructions behave both as agreement markers and as object clitics. Section 3 contains a parametrized analysis of the pro-drop phenomena inspired in part by analyses of Jaeggli (1982) and Chomsky (1981). Sections 4 and 5 include the nominative and accusative derivations of existential sentences in Hebrew, elaborating on the behavior of the clitics in these constructions and pointing out the consequences that these constructions have for the theory of grammar in general and for the analyses of pro-drop and clitics in particular.
2. EXISTENTIAL SENTENCES - PRESENTATION
Existential sentences in Modern Hebrew are formed by using the particle yes, roughly 'exist', to assert existence and the particle 'eyn, roughly 'exist not' to negate existence: (1)
yes" slosa xatulim ba-gan exist three cats in-the-garden 'there are three cats in the garden'
204
Parametric
Syntax
(2)
'eyn slosa xatulim ba-gan exist-not three cats in-the-garden 'there aren't three cats in the garden'
Sentences ( l ) - ( 2 ) exhibit a typical behavior with regard to existential sentences. The NP whose existence is being asserted, the understood subject, cannot be definite and does not appear in the [NP, S] position. The definiteness is demonstrated in (3)-(4) (we adopt here Milsark's 1974 account for the definiteness restriction in existentials): (3)
*yes sloset ha-xatulim ba-gan 'there are the three cats in the garden'
(4)
*'eyn sloset ha-xatulim ba-gan 'there aren't the three cats in the garden'
The sentences in ( l ) - ( 4 ) have counterparts in which the subject appears in the regular [NP, S] position. In these cases the subject may be definite or indefinite as expected (and compare with the English counterparts). Interestingly, when the subject appears in the [NP, S] position, we find that a clitic is obligatorily attached to the particles yes and 'eyn. The clitic agrees with the subject in gender, number and person: 1 (5)
a. slosa xatulim yes-nam ba-gan three cats exist-they in-the-garden 'three cats are in the garden' b. slosa xatulim 'eyn-am ba-gan 'three cats are not in the garden' c. sloset ha-xatulim yes-nam ba-gan 'the three cats are in the garden' d. sloset ha-xatulim 'eyn-aw ba-gan 'the three cats are not in the garden'
The requirement that the clitic agrees with the subject in number, gender and person is demonstrated in (6): (6)
a. 'anij'eyn-enij/*no/*nxa ba-gan I exist-not-I/*he/*you in-the-garden Ί am not in the garden' b. ha-yaldaj yes-naj/*no ba-gan the girl exist-she/*he in-the-garden 'the girl is in the garden'
Parameters for 1NFL
205
A failure to attach a clitic to the particles yes and 'eyn when they appear following the subject results in ungrammaticality: (7)
a. *sloset ha-xatulim b. *'ani 'eyn ba-gan
yes
ba-gan
(Compare with (5c)) 2 (Compare with (6a))
Similarly, the attachment of a clitic when the subject is not fronted is ungrammatical: (8)
a. *yes-namj slosa xatulimj b a - g a n 3 (Compare with (1)) b. *'eyn-aj yaldaj ba-gan exist-not-her girl in-the-garden 'there is not a girl in the garden'
At first sight it seems that the contrast between the grammatically of (1), (2) and (5) on the one hand, and the ungrammaticality of (7) and (8) on the other can be explained b y assuming the analysis of clitics illustrated previously. According to that analysis, the clitic would be a spellout of the Case features of the particle. The particle itself would not be a proper governor. Hence only when the clitic was present could extraction occur, allowing the coindexed clitic to properly govern the empty position. Thus (7) would be ungrammatical because the extraction site following the particle was not properly governed. (8) would be ungrammatical because the clitic absorbed the Case features of the particle, hence the NP slosa xatulim 'three cats' would not be assigned Case and would violate the Case filter. Some independent evidence that particles are not proper governors comes from cases of quantifier raising. Thus wide-scope interpretation of many is impossible in (9a) but possible in (9b): (9)
a. 'eyn harba yeladim ba-gan 'there aren't many children in the garden' b. harbe yeladim ' e y n - a m ba-gan 'Many children are not in the garden'
A wide scope interpretation for many in (9a) would result in an empty category following 'eyn, as in (10a). Since 'eyn is not a proper governor, this reading is ruled out and the only possible reading of (9a) is the one in which the negation marker 'eyn has scope over many. (9b), on the other hand, is given the representation in (10b). (10)
a. for χ , χ a child, 'eyn many χ in the garden b. for χ , χ a child, many χ 'eyn+ c l , [e]j in the garden
206
Parametric
Syntax
However, if we extend the analysis proposed in the previous chapters to the clitics on the existential particles discussed above, a few questions immediately arise. What is the status of the [NP, S] position in sentences such as (l)-(2)? How does the fronted element in (5a-b) receive Case? How is Case assigned to the post-particle subject? The Case assigned to the fronted subject in sentences (5a-b) is nominative, as is evident from the appearance of nominative pronominal forms in that position, as in (6a) above. Assuming that the rule of nominative Case assignment is roughly as in (11) (but see below for a more precise formulation), we conclude that the value of AGR in (5a-b) is [+]: (11)
Assign nominative to NP if it is governed by [+AGR]
On the other hand, there is some evidence that the Case which is assigned to the post-particle position is nominative as well. Thus consider the sentences in (12): (12)
a. 'eyn'ani yoda'at'et ha-tsuva neg I know acc the-answer Ί don't know the answer' b. 'eyn hi h a - r o f a neg she the-doctor 'she is not the doctor'
The sentences in (12) are illustrations of another environment in Hebrew in which the particle 'eyn is used. It may serve as a marker of negation in present tense and in nominal sentences. In such cases the post-particle subject is, of course, not constrained by the definiteness requirement. As we can see, a nominative pronominal element appears in that position. As in the case of the negative existential 'eyn, the placement of the subject preceding 'eyn will result in obligatory cliticization: (13)
a. 'ani'eny-eni yoda'at'et ha-tsuva I neg+cl know accthe-answer b. hi 'eyn-ena r o f a she neg+cl doctor
(14)
a. *'ani 'eyn yoda'at 'et ha-tsuva b. *hi 'eyn r o f a
Such cliticization is impossible in the pre-subject position: (15)
a. *'eyn-eni 'ani yoda'at 'et ha-tsuva b. *'eyn-ena hi r o f a
Parameters for
INFL
207
Adopting as a working hypothesis the assumption that the negative existential 'eyn has the same properties as the 'eyn occurring in (12)-(15), the post-particle subject in ( l ) - ( 2 ) may be assigned nominative (note that in these cases a nominative pronominal is never possible, due to the definiteness restriction). 4 Following this conclusion, we seem to be confronting a problem. If we wish to maintain the assumption that the clitics in (5) and in (13) are spell-outs of the Case features of the particles, we would have to conclude that existential sentences and sentences such as (12)-(13) contain two nominative markers: the AGR node and the particle itself. Apart from the general unattractiveness of such a result, note that it will also rob us of an explanation for the ungrammatically of (8): although the Case features of the particle itself were absorbed by the clitic, we would still expect it to be possible for AGR to assign nominative Case to the post-particle NP, since nominative Case assignment by AGR to post-verbal positions is otherwise possible in Modern Hebrew, as is demonstrated by (16a-b): (16)
a. huku slosa yeladim bagan were-beaten three children in-the-garden 'three children were beaten in the garden' b. nis'me'a yilelat xatul was-heard wailing cat 'the wailing of a cat was heard'
Clearly in (16a-b) the post-verbal subject is assigned nominative, since it agrees with the verb in number, gender and person, a property characteristic to nominative Case. The verbs in (16a-b) are ergative verbs (or unaccusative verbs in the sense of Burzio 1981, and Perlmutter 1978, respectively). The single argument of the verb is base-generated in the post-verbal position and is assigned Case directly there. (We will return to these constructions below. See also Borer, 1980). Another curious property of the particle + clitic complex in sentences such as (5) and (13) is that it behaves exactly like a fully inflected verb with respect to pro-drop. Pro-drop is the name given to a cluster of phenomena attested in languages such as, among others, Italian, Spanish, Romanian, Arabic and Hebrew. In these languages a pronominal subject is optional. In Hebrew this phenomenon has an unusual distribution: it seems to be attested only in the past and future tenses, and then only in the first and second person (but see below for some exceptions). This situation is exemplified by the following paradigm:
208
Parametric
Syntax
(17)
a. ( ' a n i ) ' a x a l t i ' e t ha-banana (I) ate acc the-banana b . 'ani 'oxelet 'et ha-banana I eat acc the-banana c. *'oxelet 'et ha-banana eat acc the-banana d. ('ani)'oxal 'et ha-banana will-eat
(18)
a. ('atem) 'axaltem 'et ha-banana (you-pl)ate acc t h e - b a n a n a b . 'atem 'oxlim 'et h a - b a n a n a eat c. *'oxlim 'et ha-banana d . ('atem) toxlu 'et ha-banana will-eat
(19)
a. hu 'axal 'et ha-banana he ate acc t h e - b a n a n a b. * ' a x a l ' e t ha-banana c. h u ' o x e l ' e t ha-banana eats d. * ' o x e l ' e t ha-banana e. hu yoxal 'et ha-banana will-eat f. *yoxal 'et ha-banana
As has been noted before (see Borer, 1980), the availability of p r o - d r o p in Hebrew seems to be related to the availability of person markers on the AGR node. In the present tense AGR in Hebrew is defective: it contains markers only for gender and number but not for person. The third person in the other tenses is the unmarked person of AGR in Hebrew. It would thus be plausible to assume that the person marker in these forms is not sufficiently specific and thus cannot "trigger" p r o - d r o p (we will return to the formulation of this generalization below). Returning now to the particle + clitic complex in (5) and (13), it is interesting to note that it behaves as a fully inflected verb (namely, a non-present tense verb containing a person marker) with respect t o p r o drop. Thus the sentences in (20) and (21) are grammatical but those in (22) and (23) are n o t : (20)
a. ('ani) 'eyn-eni ba-gan Ί am not in the garden'
Parameters
(21)
for
INFL
209
b. ('ata) 'eyn-xa bagan 'you are not in the garden' c. ('atem) 'eyn-xem ba-gan 'you (pi) are not in the garden' d. hu 'eyn-enu ba-gan 'he is not in the garden' e. hem 'eyn-am ba-gan 'they are not in the garden' a. ('ani) 'eyn-eni yoda'at 'et ha-t§uva Ί don't know the answer' b. ('atem) 'eyn-xem yod'im 'et ha-tsiiva 'you (pi) don't know the answer'
(22)
a. *'eyn-enu ba-gan b . *'eny-am ba-gan
(Compare with (20d)) (Compare with (20e))
(23)
*'eyn-ena r o f a
(Compare with (13b))
The grammatically of the pro-drop versions of (20a-c) and (21a-b) vs. the ungrammatically of (22) and (23) is related to the degree of specificity on the clitic: when it is specified for first or second person, p r o - d r o p is possible, as it is possible in (17a, d) and (18a, d). However in (22) and (23), where the person marker is third person, pro-drop is blocked, as it is blocked in (19b, f). Thus the clitic on the particles patterns with agreement markers on verbs. These facts seem to suggest that the clitics on the particle should be viewed as AGR markers rather than as a spell-out of Case features of the particles themselves. Such a move, however, will lead to new problems. Rizzi (1982) argues very convincingly that AGR is neither a proper governor in the pro-drop languages nor in the n o n - p r o - d r o p languages. Rizzi's argument, constructed in the On Binding framework, is based on showing that Nominative Island Condition (NIC) effects are attested in the [NP, S] position in pro-drop languages just as they are attested in n o n - p r o - d r o p languages. Compare the following facts from French (see Kayne 1980b) and Italian (Rizzi, 1982): (24)
a. Je exige qu'ils n'arretest personne Ί demanded that they will arrest no one' b . Je n'exige qu'ils arretest personne c. Je exige que personne ne soit pas arrete Ί demanded that no one will be arrested' d. *Je n'exige que personne soit arrete
210
Parametric
Syntax
(25)
a. non voglio che tu parli con nessuno not want that you talk with anyone (for no χ , I want that you speak with x) b. *non voglio che nessuno venga (for no χ, I want that χ come) c. voglio che nessuno venga (I want that for no χ, χ come)
Kayne (1980b) observes that the ungrammatically of (24d) can be explained if we assume that the negative polarity item ne in French functions as a scope marker. When it is in the embedded clause, the quantifier personne will have scope only over the embedded clause, or, formulated in terms of Quantifier Raising, the quantifier personne will be raised and adjoined to the embedded S node. On the other hand, when ne is in the matrix, the quantifier will be raised and adjoined to the matrix S, having wide scope. Note that such an operation will result in grammaticality in (24a-c). The trace left by the quantifier will be a nominative anaphor only in (24c), where, in turn, it will be bound by its antecedent,personne, having narrow scope. (24d) is ungrammatical, according to this reasoning, since the nominative anaphor in the embedded clause, the trace of the (wide scope) personne is free. In a framework in which NIC effects are accounted for by the ECP, (24a-b) are grammatical since the trace of personne is properly governed by the verb. In (24c) it is properly governed by its antecedent. If AGR were a proper governor in French, we would expect (24d) to be grammatical. Its ungrammatically leads us to the conclusion that AGR is not a proper governor in French. Rizzi (1982) observes that a similar contrast exists in Italian, as is demonstrated by (25a-c). In Italian too a wide scope quantification from the [NP, S] position of an embedded clause results in ungrammaticality. Thus we must conclude that AGR is not a proper governor in Italian either. 5 Similar facts can be shown to hold in Hebrew. Thus the quantifiers klüm and sum davar (both 'nothing') and 'af'exad 'nobody' exhibit a similar distribution with respect to the negative marker lo 'no' in (26a-d). Note that, as in French, (26b) is marginal. There is, however, a sharp contrast between this configuration and the completely ungrammatical (26d): (26)
a. 'amarti se-hem lo hevi'u 'af'exad/sum davar said-I that-they not brought no one/nothing Ί said that they brought nobody/nothing b. ?lo 'amarti &-hem hev'iu 'af'exad/sum-davar no no one/nothing Ί didn't say that they brought anybody/nothing
Parameters for
INFL
211
c. 'amarti se-'af 'exad lo ba/se-sum davar lo naxuc said-I that-no one not came/that nothing not necessary Ί said that nobody came/that nothing is necessary' d. *lo'amarti se-'af 'exad ba/se-sum davar naxuc no no one came/that nothing necessary Summarizing this short digression, we conclude that in Hebrew (as in Italian and French), AGR is not a proper governor. Now consider the clitics on yes and 'eyn. If we were to suggest that they are agreement markers, we would have to conclude that they are not proper governors as well. Since the particles themselves are not proper governors (see (9)-(10) and related discussion), we would expect every extraction from a post-particle position to result in a violation of the ECP: the empty category left following the extraction cannot be governed by the clitic, as the clitic is an agreement marker which does not function as a proper governor. It cannot be properly governed by the particle either, since the particle is not a proper governor. The availability of extraction from the post-particle position thus poses a problem for the assumption that the clitics on the particles are agreement markers. Further consider the difficulties that we would confront if we wished to argue that the clitics are agreement markers. We would have no ready explanation for the peculiar restrictions on their distribution: why would agreement markers show up only when preceded by the subject and not when followed by it? Note that other agreement markers on verbs show up regardless of the location of the subject, as is demonstrated by (27): (27)
a. nehergu slosa yeladim were-killed three children 'three children were killed' b. slosa yeladim ne'hergu three children were-killed c. *ne'herag slosa yeladim was-killed three children
Thus we reach the contradictory result that the clitics on existential particles function as AGR markers with respect to the pro-drop phenomena, but as clitics (i.e., as proper governors) for the purposes of extraction from the post-particle position. In order to solve this apparent contradiction it would be necessary to digress and review at some length the analysis proposed for the prodrop phenomena. We will show that a proper analysis of these phenomena will solve the contradiction above in a natural way, accounting both for the clitic-like and the agreement-like properties of the clitics in question.
212
Parametric
Syntax
3. PRO-DROP: ANALYSIS
The following cluster of properties is traditionally associated with the pro-drop phenomena (cf. Perlmutter 1971, Rizzi 1982): (28)
a. b. c. d.
empty pronominal subjects. free inversion of the subject. long movement of the subject. violations of the [that e] filter (in the sense of Chomsky and Lasnik 1977).
The properties in (28a-d) are demonstrated for Hebrew in (29a-d) (see Rizzi, 1982 for the relevant examples in Italian, and Jaeggli, 1982 for the relevant examples in Spanish): (29)
a. 'axalti 'et ha-tapua'x ate-1st acc the apple Ί ate the apple' b. kafcu min ha-matos Ran ve-Dan jumped from the-plane Ran and Dan 'Dan and Ran jumped from the plane' c. ha-'isj se-hayiti roca la-da'at 'eyfo e· haya the-man that-would-I want to-know where was bizman ha-milxama during the-war d. mij xaiavta se ej 'exer la-mesiba? who thought-you that-late-was to-the-party? 'Who did you think (that) was late to the party?'
In essence, the properties in (28a-d) are a list of environments in which an empty category is licensed in the [NP, S] position in pro-drop languages but is barred from that position in non-pro-drop languages. Rizzi (1982) argues that a crucial property distinguishing pro-drop languages from non-pro-drop languages is (28b). Properties (28c) and (28d) can be reduced to (28b) in his account. For concreteness, when extraction of the subject takes place in pro-drop languages, it is not from the [NP, S] position but from the post-verbal position. Thus, Rizzi observes, we can contrast the ungrammatical (25b) above, in which extraction (in Logical Form) has taken place from the [NP, S] position, with the grammatical (30), in which the subject is extracted from the post-verbal position: (30)
no voglio che venga nessuno (for no χ . I want that χ come)
Parameters for 1NFL
213
Again, a parallel situation holds in Hebrew: (31)
'ani lo roca se-yikre sum davar (for no χ , I want that χ will happen)
Chomsky (1981), accepting Rizzi's observations and following some proposals of Jaeggli (1982), suggests that the presence of the cluster of properties in (28) in some languages vs. its absence in others can be explained if we assume that the rule which attaches AGR markers t o the verb (essentially the rule of Affix Hopping of Chomsky, 1957) is a local rule (in the sense of Emonds, 1976). This rule, which may always apply in the phonological component, may also apply in the syntactic component if the grammar of a particular language allows it t o do so. The rule ( R ) is given in (32). In p r o - d r o p languages, it is specified as applying optionally in the syntax, as is stated in (33):
(32)
Affix Hopping (R): AGR V
(33)
>
[ γ V,AGR]
(It is assumed that AGR does not leave a trace.)
R may apply in the syntax
Following our terminology, we will assume that (32) is an inflectional rule (see Chapters 1 and 4), and we will restate (33) as (34), a restriction applicable to n o n - p r o - d r o p languages: (34)
R may not apply at S-structure
(The reader is asked to recall that the operation of inflectional rules at S-structure is, in fact, mapping from the o u t p u t of 'Move a ' to S structure proper. It crucially takes place prior to the operation of various checking mechanisms which are relevant at S-structure, e . g . , t h e binding conditions and nominative Case assignment. For a detailed discussion of this proposal, see Chapter 1.) Let us now see how (32) combined with (34) can derive the distribution of p r o - d r o p . Consider the following S-structure representations of sentence (17a) above: (35)
a. b. c. d.
I + A G R [ y p ate the banana] I [ y p ate + AGR the banana] [e] +AGR [ y p ate the banana] [e] [ y p ate + AGR the banana]
R R R R
has has has has
not applied applied not applied applied
214
Parametric
Syntax
Representation (35a) is well-formed: the AGR node governs the subject position at S-structure, and thus it can assign nominative Case. This derivation gives rise to the grammatical non-pro-drop version of (17a). Representation (35b) is ungrammatical: the subject position is not governed by AGR; hence the lexical NP / cannot receive Case. 6 Therefore this derivation violates the Case filter. Now consider the representation in (35c). Chomsky (1981) suggests that empty categories (so-called NP-trace, WH-trace, quantifier trace and PRO) are all tokens of the same type. The particular function of any given empty category is determined by its context. Thus [e] is a variable (WH-trace and quantifier trace) if it is in an Α-position and is locally A-bound. [e] in an Α-position that is not a variable is an anaphor. Since pronominals (which are not interpreted as bound variables) are either free or locally Α-bound by an antecedent with an independent Θ-role, an [e] which is not a variable is a pronominal if it is free or localy Abound by an antecedent with an independent Θ-role. It follows that the [e] usually referred to as PRO is a pronominal anaphor (see Chapter 3 for a more detailed discussion). Now consider the empty category in (35c). As it is not ϊ - b o u n d it cannot be a variable. If we consider it to be an anaphor, it must be bound in its minimal governing category (see Chapter 1). However, it is free in its minimal governing category. Thus if [e] in (35c) stands for an anaphor, the sentence is ruled out. Can [e] in (35c) stand for PRO? Note that the position is governed by the AGR node, and thus PRO in this case would have a governing category. A pronominal anaphor cannot have a governing category (in which it would have to be both free and bound). Therefore if [e] stands for PRO the sentence would be ruled out as well. So there is not grammatical representation for the [e] in (35c) and the representation is ungrammatical. Now consider (35d): it is well-formed. The [e] cannot be a variable and if it is an anaphor it would be free, but it could be PRO: lacking a governing category in that position it would be free to appear there. However, the representation in (35d) is only possible in pro-drop languages. In non-pro-drop languages AGR will always govern the subject position at S-structure, and all cases of empty categories in the [NP, S] position will have the ungrammatical structure in (35c). Only in prodrop languages will the configuration in (35d) arise, allowing for a missing subject. 3.1. Pleonastic Subjects and Nominative Case Assignment Chomsky's analysis can account for other pro-drop phenomena: empty pleonastic subjects and free inversion. These constructions in Hebrew
Parameters for INFL
215
and their counterparts in non-pro-drop languages are given in (36)-(38): (36)
Pleonastic elements in "raising" and extraposition configurations: a. it seems that John is late again b. nir'e se-Itamar suv me'axer seems that-Itamar again late c. it annoys me that John is always late d. maigiz 'oti se-Itamar tamid me'axer annoys me that-Itamar always late
(37)
Pleonastic elements in "ergative" configurations (in the sense of Buizio, 1981): a. il est arrive un gar^on 'there arrived a boy' b. nisma cilcul pa'amon was-heard ring bell 'bell-ringing was heard'
(38)
Subject inversion: a. *ate the apples three men/John and Mary b. ??there ate the apples three men/*John and Mary c. 'axlu'et ha-tapuxim slosa 'anasim/Raxel ve-Dan ate acc the-apples three men Rachel and Dan
Some points remain to be clarified with respect to the examples in (36)(38). Note that in (36b,d) the subject position in Hebrew is empty, although the AGR node does not contain a person marker. We will return to this point below when we formulate more precisely the interaction between the person marker on the verb and the pro-drop phenomena. The ergative configurations in (37) are, in essence, configurations in which the subject is base-generated in the post-verbal position and is assigned its θ-role there. It can subsequently be preposed to the regular subject position or stay in its post-verbal position, providing it can receive Case there. In this, we follow the essentials of Burzio (1981) and Borer (1980). A similar analysis for the selection of auxiliaries in Italian was first proposed by Perlmutter (1978) (the "unaccusative hypothesis"). 7 We can raise several questions with respect to the pro-drop variants of (36)-(38). First, are pleonastic elements restricted to non-pro-drop languages? Can we find them in pro-drop languages? What will determine their distribution? Another type of question concerns the device which allows the post-verbal NP to be assigned nominative Case in (37b) and (38c). Turning first to the distribution of pleonastic elements, note that their
216
Parametric
Syntax
availability in non-pro-drop languages follows directly from the existence of the restriction in (34). As the subject position may never be empty, a lexical NP must be available to appear in that position, even when it is not a Θ-position. In the latter case the lexical NPmust be non-referential so as to avoid a violation of the Θ-criterion. In other words, the lexical element must be pleonastic. On the other hand, (34) does not predict the absence of pleonastic elements in pro-drop languages. We expect them to show up in those derivations in which R applies in the phonology. A partial explanation for the rarity of pleonastic elements in pro-drop languages would be available if we invoked the Avoid Pronoun principle of Chomsky (1981). This principle, if adopted, would effectively force R to apply at S-structure in constructions such as (36b,d) and (37b), allowing for the empty category PRO to appear in the [NP, S] position. Adopting this principle, however, immediately leads to new questions: why would this principle apply obligatorily to sentences with ηοη-Θ-subject, but optionally to sentences with a Θ-subject (e.g., (17a) above)? We would expect a pleonastic element whose distribution parallels that of 'ani, T , not an obligatory absence of such a pronominal element. Interestingly, there are sentences corresponding to (36b,d) in substandard Hebrew which utilize a pleonastic element: (39)
a. Ize nir'e se-Itamar suv me'axer 'it seems that Itamar is late again' b. ze margiz 'oti se-Itamar tamid me'axer 'it annoys me that Itamar is always late'
However, the counterparts of (37b) and (38c) with the pleonastic element ze are completely ungrammatical: (40)
a. *ze nisma cilcul pa'amon b. *ze 'axlu 'et ha-tapuxim slosa 'anaSim
As we will see below there are independent reasons to believe that precisely in these cases R is forced to apply at S-structure in order to form a grammatical derivation. Thus, assuming that (34) is indeed the parameter distinguishing pro-drop languages form non-pro-drop languages, it will correctly predict the distribution of pleonastic elements in substandard Hebrew. Let us now turn to the assignment of nominative Case in (37b) and (38c). We know that nominative rather than any other Case is assigned to the post-verbal subject since the verb has to agree with the subject, whether preposed or postposed. Thus (41a-b) are ungrammatical:
Parameters for INFL (41)
a. *nisma
217 cilculey pa'amon
was/is-heard ringings bell 'ringings of a bell are/were heard' b. *'axal 'et ha-tapuxim slosa 'anasim ate(sg) acc the apples three men In order to capture the assignment of nominative Case to post-verbal subjects, Chomsky suggests that a D-structure rule co-superscripts the [NP, S] position and AGR. He further suggests that the rule for nominative Case assignment is then re-stated as in (42): (42)
Assign nominative Case to an NP which is b o t h governed by AGR and co-superscripted with it. (Applies at S-structure.)
Nominative Case is now assigned to (38c) in the following way: at D structure, following the co-superscripting, the representation of (38c) is as in (43): (43)
NPi AGR' [ V p
]
Following the postposing of the subject, the possible configurations at S-structure are as in ( 4 4 a - b ) : (44)
a. [e]' AGR1 [ v p [VP ] NP 1 ] 1 b. PRO' [ V P [ V P V+AGR . . .] NP']
R did not apply R applied
(44a) is ungrammatical as nominative Case cannot be assigned to the post-verbal NP: although it is co-superscripted with AGR, it is not governed by it and hence (42) cannot apply. (44b), on the other hand, is grammatical. AGR is attached to the verb and now it governs the cosuperscripted NP in the post-verbal position (recall that we are assuming that a head can govern into adjoined structures. See the definition of c-command and government in Chapter 2 above). 8 From the application of R at S-structure, it follows that no lexical NP, and hence no lexical pleonastic element, can appear in the subject position, quite independently of the Avoid Pronoun principle. The appearance of a lexical NP in the subject position would be ruled out since that NP cannot receive Case. The ungrammaticality of (40b) above follows. In order to account for the assignment of nominative Case in "ergative constructions", some more machinery is necessary. Assuming, following Burzio (1981) and Borer (1980), that in sentences such as (37b) the post-verbal subject is base-generated in the VP, we cannot appeal to superscripting at D-structure as a way to co-superscript AGR and the post-verbal NP.
218
Parametric
Syntax
We will assume that every NP can be freely assigned a superscript. Only if that superscript matches that of AGR will the NP in question be assigned nominative Case in accordance with (42). Note that at this point we deviate from Chomsky's (1981) analysis. Chomsky assumes that in ergative constructions the post-verbal nominative Case assignment is accomplished by means of the insertion of an expletive PRO in the [NP, S] position. By convention this PRO is cosuperscripted with the post-verbal subject, and by transitivity of indexing the post-verbal subject must be co-superscripted with AGR. Following the application of R, the assignment of nominative Case in the postverbal position proceeds in the regular fashion. At first glance it seems that the altered system which we propose here, although clearly simpler, would fail to block the assignment of nominative Case to the direct objects of transitive verbs. Since superscripting is free and since AGR in pro-drop languages may move into the VP at S-structure, it should be possible to have the situation in (45a), in which AGR is co-superscripted with the direct object: (45)
a. PRO1 [ y p V + AGR1 NP1 ] nom *PRO 'axal-ti ha-tapuax ate-1 the apple Ί ate the apple' b. PRO1 [ y p V + AGRi NP) ] PRO 'axal-ti 'et ha-tapuax acc
In (45a) the direct object was assigned the same superscript as that of AGR and consequently it was assigned nominative Case. However, the sentence is ungrammatical. The grammatical sentence is as in (45b), in which accusative Case is assigned to the direct object. The ungrammatically of (45a), however, is due to other factors. Burzio (1981) and Chomsky (1981) observe that verbs which assign accusative Case also assign a Θ-role to their subjects ("Burzio's generalization"). Note further that verbs which assign accusative Case also assign a Θ-role to their object. It follows that for every transitive verb, there are two referential expressions which occupy the two relevant Θ-positions. 9 Given these generalizations, the ungrammatically of (45a) follows immediately. Recall that a Θ-role is assigned to an Α-chain and that Achains are defined on superscripting as well as on co-indexing (see Chapter 3 for discussion). Since the sequence PRO1 - The apple1 in (45a) has the same superscript assigned to each of its members, it constitutes an Achain and only one Θ-role can be assigned to it. It follows that one of the
Parameters for INFL
219
Θ-roles which correspond to the verb eat in (45a) is not assigned. There is only one A-chain, but two Θ-positions and two Θ-roles. Hence (45a) is ruled out as a violation of the Θ-criterion. Our account makes a clear prediction: if one could find verbs that violate "Burzio's generalization", i.e. verbs that assign accusative Case but do not assign a Θ-role to their subject, we would expect some freedom with respect to the assignment of Case to the complements of these verbs. We expect them to be either accusative or nominative, depending on the superscript picked by the post-verbal NP. If that superscript matches that of AGR, we expect nominative Case. Otherwise we expect accusative Case. As we will see in section 4 below, this prediction is verified, thus supplying strong evidence for the free superscripting of post-verbal NP's. We will now return to the contrast between the grammatical (39a-b), and the ungrammatical ( 4 0 a - b ) . While the grammaticality of ( 3 9 a - b ) need not be explained, the ungrammatically of ( 4 0 a - b ) now follows from the fact that once R applied at S-structure, the pleonastic element in the [NP, S] position can no longer be assigned Case. It is no longer governed by AGR at S-structure. On the other hand, the failure of R to apply at S-structure will make it impossible to assign Case to the post-verbal NP. This situation does not arise in (37b) and (38c): in these sentences the [NP, S] position is occupied by PRO, a state of affairs that frees R to apply at S-structure and leave the [NP, S] position ungoverned.
3.2. The Person Marker We will now turn to the function of the person marker in delimiting the p r o - d r o p phenomenon. Let us assume that the person feature of AGR contains a referential index. Naturally, this referential index is present only when the person feature is present. Thus in present tense in Hebrew there is no such index. The third person marker, we assume, is different in that its index is not referential. Thus the only person markers which contain referential indices are first and second person markers. We further assume that in configurations such as (46) the person index - whether referential or not - is transmitted to the [NP, S] position as part of the superscripting process (outlined above) between AGR and the subject position: AGR 1 gender number person:
220
Parametric
Syntax
A few things should be noted with respect to our proposal. First, it is important to point out that the lack of reference of third person AGR does not mean that third person AGR is incompatible with referential reading. Rather, it cannot itself assign referentiality to an empty category. In essence, then, our proposal entails that although empty categories are specified as having φ features (see Chapters 1 and 3 for discussion), they nevertheless lack an inherent reference. Such reference index may be inherited from a co-superscripted element, as in the case of AGR. A trace inherits a referential index from its antecedent. Yet another way to acquire a referential index is through a controller. Some of these ways to inherit a referential index will be discussed further in the following sections. An interesting question arises with respect to empty categories in clj + [e]j combinations. Note that in cases of extraction, the empty category acquires reference from its antecedent. On the other hand, when no extraction occurs, we assumed that the clj + [e]j combination is reanalyzed as a pronoun (see Chapter 3 for discussion). It is this pronoun form which supplies reference to the empty category. It is worth mentioning at this point that a referential index will not be transferred to the [NP,S] position if the position is Θ. This follows from the Θ-criterion and the Projection Principle. Following the Projection Principle, a "genuine" empty category with φ features is base-generated only in the subject position of a "regular" pro-drop clause, since the position is a Θ-position and must be filled at D-structure. [NP,S] positions which are not Θ-positions are truly null at D-structure. The crucial distinction here is between a position filled by a category which does not have a phonological representation and a position which is simply not filled. In the remainder of this chapter we will be using the empty-bracket notation - [j^p ] - to indicate a null category which is generated in the base and is filled in the course of the derivation either by a moved element, or by a pleonastic element or expletive PRO inserted at S-structure. The reader will bear in mind that this null category differs from Ijsjpe] in that it does not contain 0-features (see Chapters 1 and 3 for a discussion). In drawing this distinction we follow Chomsky (1981). 10 We can now account for the facts of pro-drop in Hebrew, as illustrated by (17)-(19) above. We assumed that whenever the subject is not realized phonologically it may be occupied only by PRO, since whenever an [e] appears in the [NP, S] position the only well-formed derivation is one in which R has applied at S-structure. The following paradigm results:
221
Parameters for INFL (47)
a. PRO·
'axal-ti
'et ha-banana
[ y p ate + AGR'
acc the-banana
[
1 st person "I
+referentialjj
Ί ate the banana' 'et ha-banana
toxl-u
b. PRO·
[ V p will-eat + AGR'
acc the-banana
T2nd person"] PiI +referential I; 'et ha-banana
'oxel *PRO'
[ V P eat + A G R '
acc the-banana
[ - person] Ί/you/he eats the banana' d.
'et ha-banana
'axal *PRO]
[ V P ate + A G R '
[
acc the-banana
3rd person Ί -referentialjj
'he ate the banana'
Note that the PRO receives a referential index from A G R only in ( 4 7 a - b ) . Since the subject position o f 'eat' is a 0 - p o s i t i o n , a referential expression has to appear in this position. Since in ( 4 7 c - d ) the PRO does not receive a referential index from A G R , the sentences are ungrammatical. Now consider how this process will apply to 'ergative' verbs, exemplified by ( 4 8 ) : (48)
a. nisbarti bi-zman ha-bxina broke-I during the-test Ί broke down during the test' b. nixnasu be-sa'a xames entered-we in-hour five 'we entered at five o'clock'
The D-structure o f ( 4 8 a ) is given in (49). Note that no person-index is transferred to the [NP, S ] position at D-structure, since it is not a ©-position and, in accordance with the Projection Principle, it is null at D structure.
222
Parametric
(49)
[
Syntax
]' AGR 1 [1st person ]j
[ v broke PRO]
A free superscript is now assigned to PRO. However, in its post verbal position the PRO is governed. The representation in (49) will be ruled out at S-structure unless two things happen: PRO must move to the [NP, S] position in order to escape the government of the verb and AGR must move to the VP in order to prevent government of PRO in the [NP, S] position following its movement. PRO may move into the [NP, S] position only if its superscript matches that of AGR as the [NP, S] position and AGR are co-superscripted by a D-structure rule, as pointed out above. Now recall that the referential index of the person-marker in AGR is transferred to all referential elements co-indexed or co-superscripted with AGR. It follows that PRO in (49) must have the referential index j in order for the representation to the grammatical. Thus the S-structure representation of (49) is as in (50): (50)
PROj [ v broke + AGR' [1st person]j
[e]j]
Thus, although the D-structure of (48) is quite different from the Dstructure of (17) above their S-structures are identical in all relevant respects (and compare (50) with (47a)). Similarly, if the 'ergative' verb is not inflected in the first or second person, PRO will not inherit a referential index from AGR and pro-drop will be blocked. This statement, as well as the account for the ungrammatically of (47c-d) above makes a prediction: if there is another way to assign a referential index to the PRO in (47c-d), we would expect these constructions to be grammatical. There are two other ways to assign a referential index to a PRO. First, it can be assigned a referential index by a controller and secondly, it can have arbitrary reference. And, indeed, (47c-d) can be "salvaged" in these situations. In (51a-b) a control situation is illustrated. In (52a-b) the PRO has arbitrary reference: (51)
a. Talilaj ma'amina se-PROj hiclixa ba-bxina Talilaj believes that PROj succeeded in-the-test 'Talila believes that she passed the test' b. Dani bikes me-Talilaj se PROjtavo 1 1 Dani asked from-Talilaj that PROj will-come 'Dani asked Talila to come'
Parameters for 1NFL (52)
a. 'amru
'et ze ba-radio
223 'etmol
said(pl) it i n - t h e - r a d i o yesterday 'it was said on the radio yesterday' b. 'omrim se-Rina lo hiclixa ba-bxina12 say(pl) t h a t - R i n a not succeeded i n - t h e - t e s t 'it is said that Rina did not pass the test' When no referential index is assigned to PRO it is a non-referential PRO, a pleonastic one. Since it is not a referential expression, its distribution is restricted by the Θ-criterion. Thus we find this PRO in "raising"-type constructions, in "extraposition"-like constructions and in the subject position of ergative verbs. This situation was illustrated by ( 3 6 ) - ( 3 7 ) above. The verbs in these constructions typically appear in the present tense and in the third person, failing to transmit a referential index. Some interesting evidence for the presence of PRO in the subject position of p r o - d r o p constructions comes from the following sentences: (53)
a. Dan 'amar l e - R a n PRO la-vo Dan said to Ran to come 'Dan told Ran to come' b. Dan 'amar le-Ran se PRO yavo Dan said to Ran that will come i. Dan told Ran to come ii. Dan told Ran that he will come
While the sentence in (53a), a regular control case, allows only for object control, we would not expect such a strict restriction if the embedded [NP,S] position in tensed clauses is not a PRO. Nevertheless the sentence in (53b) is ambiguous. Its preferred reading entails that Dan commanded Ran to come, and is synonymous with (53b). Like (53a), it allows only for object control. The second reading of (53b), which some speakers reject altogether, allows the subject position to be controlled freely by the subject or the object of the matrix and entails that Dan asserted that either he, himself, ox Ran will come. If we wished to claim that the empty category in the [NP,S] position of (53b) is not a PRO, we would have to complicate the theory of control so as to allow it to apply in an identical way to n o n PRO empty categories.
3.3. Summary Let us summarize our discussion of the p r o - d r o p phenomena and its analysis. We proposed, following Chomsky (1981) that the clustering of properties associated with the p r o - d r o p languages can be accounted for in a unified fashion if we assume a parametrized rule, moving AGR to the
224
Parametric
Syntax
VP. The availability of the configuration [V+AGR] at S-structure for some languages resulted both in the availability of nominative Case assignment in the VP and in the licensing of the empty category PRO in that position. While accepting the essence of the nominative Case assignment system proposed in Chomsky, we deviated from his analysis by assuming a process of free indexing to replace a rule co-superscripting an inserted expletive PRO with a post-verbal subject. While using this system as a starting point, we elaborated a range of phenomena in Modern Hebrew: the distribution of pleonastic elements in pro-drop languages; the distribution of pro-drop in Hebrew as following from the restrictive distribution of person markers; and the availability of control and arbitrary readings in environments which do not allow for referential pro-drop. In section 5 we will return to the question of pleonastic elements: discussing their distribution in pro-drop languages and in non-pro-drop languages. We will further supply some evidence for the process of free indexing proposed above.
4. INFLECTIONAL CLITICS
We will now return to the discussion of existential sentences in Hebrew. Recall that we are seeking to explain the grammaticality of (54a-d) vs. the ungrammatically of (55a-d): (54)
a. 'eyn slosa xatulim ba-gan 'there aren't three cats in the garden' b. 'eyn 'aniyoda'at 'et ha-tsuva Ί don't know the answer' c. slosa xatulim 'eyn-am ba-gan d. 'ani 'eyn-eni yoda'at 'et ha-tsuva
(55)
a. b. c. d.
*'eyn-am slosa xatulim ba-gan *'eyn-eni 'ani yoda'at 'et ha-tsuva *slosa xatulim 'eyn ba-gan *'ani 'eyn yoda'at 'et ha-tsuva
Earlier we concluded that the clitic attached to 'eyn in (54c, d) functions both as a proper governor for the extraction site, and as a "trigger" for pro-drop (see (20)-(23)). Given our account of pro-drop in Hebrew, it would be desirable to conclude that, when the clitic is present, both a superscript and a referential index can be transmitted to the [NP, S]
Parameters for INFL
225
position of sentences such as (56), accounting for the grammaticality of (56) vs. the ungrammatically of (57): (56)
a. 'eyn-eni ba-gan Ί am not in the garden' b. 'eyn-eni yoda'at 'et ha-tsuva Ί don't know the answer'
(57)
a. * ' e y n - e n u ba-gan 'he is not in the garden' b. *'eyn-enu yode'a 'et ha-tsuva 'he doesn't know the answer'
We would like to show that only in (56) the PRO is assigned a referential index by the person marker in the clitic. On the other hand, the sentences in (57) are ruled out as the clitic is third person and hence it cannot transfer a referential index. It would be desirable to argue that at least in this respect the clitics on 'eyn and on _yes"behave as agreement markers. Recall that on the other hand these clitics function as proper governors for extraction from the position immediately following the particle. When a clitic is absent, such extraction leads to ungrammatically. We pointed out before, that the particle 'eyn has two functions. In sentences such as (2) and (9) above it clearly functions as the head of a predicate. in a role which parallels that of the particle yes. On the other hand, in sentences such as (12) and (13) it functions as a sentential negation marker. The structure of each of these constructions must be different. In what follows we will analyze each of these constructions separately, nevertheless showing, that with respect to the source of the clitic and its behavior, the two pattern together, regardless of their different structure.
4.1. Ergative particles It is easy to see that in its predicative role, 'eyn behaves very much like ergative verbs. Unlike them, however, it is not a proper governor. Given this minimal difference we will assume the following: A. Particles such as 'eyn and .yes are 'ergative' in nature. By this we mean that their sole argument is generated as their complement, following them. Like other ergative elements, they do not assign Case to that complement. It is assigned Case by AGR once AGR moves into the predicate and is adjoined to the particle (Pt). Unlike ergative verbs, however, particles are not proper govenors. Thus an empty category following the particle is
226
Parametric
Syntax
not properly governed unless a coindexed clitic is available to properly govern it. B. In spite of the fact that particles do not have Case features, the rule of Clitic spell-out may apply optionally in particle constructions resulting in a defective feature matrix, given in (58): (58)
[p t Pt]
» [ptPt, [a gender, β number, γ person]]
The clitic formed by (58) is defective in that it does not contain the feature [δ Case]. Let us assume that the output of (58) is ruled out by a phonological filter, unless the feature [δ Case] is added to the matrix of features of the clitic at some point of the derivation to yield the grammatical clitic representation in (59): (59)
[a gender, β number, γ person, δ Case]
Intuitively speaking, our proposal implies that although the representation in (58) counts as a clitic for the purposes of proper govenment and the interpretative component, it cannot be regarded as a well-formed clitic in PF unless it contains the feature [δ Case.] 13 Thus once (58) has applied the derivation is ruled out unless there is a way to add the missing Case feature to the clitic derived by (58). Let us briefly digress and examine the theoretical motivation for (58). We have argued previously in this study that clitics are a spell-out of Case features, and that they attach to the heads which possess these features. It has become clear, however, that the availability of Case features, while a necessary condition, is by no means sufficient. Thus complementation plays a crucial role in licensing clitics, through the notion of Θ-grid discussed in Chapters 2 and 5. It has further become evident in our discussion of causative constructions that the Case features in question may be supplied by a different Case marker from that which subcategorized for the cliticized complement. This, in fact, is our characterization of reanalysis: a situation in which Case is not assigned by the lexical selector of an item but rather, by another Case assigner. A similar situation exists in the existential sentence constructions in Hebrew: while the cliticized complement is a complement of yes or 'eyn, the missing Case feature is supplied by another element. One may ask why (58) may apply to the existential particles but not to ergative verbs in general. In explaining this fact, we would like to invoke the similarity between the particles yes and 'eyn and prepositions. Not only is the clitic paradigm similar (see fn. 1), but, like prepositions and unlike verbs 'eyn and yes are not proper governors. Furthermore, they do not have inflectional paradigm, unlike verbal elements and similar to
Parameters for INFL
227
prepositions. Thus, it seems to us, the best characterization of the particles is as Caseless prepositions. What is the relationship between prepositions and their complements? It is not clear that they always assign Θ-role to their complements. Thus we would like to argue that in sentences such as (60), the Θ-role locative is assigned primarily by the verb put and only secondarily - if at all - by the preposition under. In fact it would seem plausible to assume that the Θ - r o l e is assigned to the entire PP, and that the relationship between under and table, although intuitively clear, cannot be characterized in terms of Θ-role assignment. Certainly under does not qualify table semantically in any way. Rather, it states a spatial relation between the act of
putting and the table: (60)
I put the litterbox under the table
Yet another possibility would be to argue that Θ-role is assigned to the objects of prepositions compositionally by the verb and the preposition. Thus table in (60) would be assigned Θ-role by put + under. Note, however, that neither version allows us to state a unique relationship between a preposition and its object. We would like to suggest that the unique relationship between prepositions and their objects is expressed by means of co-superscripting, similar in nature to the co-superscripting which holds between the I N F L node and the [NP, S] position. In the absence of direct Θ-relations, such superscripting will characterize both the relationship 'subject o f and the relationship 'object of.' Let us return to the rule in (58), now assuming that the particles in Hebrew are defective prepositions (they do not assign Case) but that nevertheless they are co-superscripted with their complements. It is this superscript, we claim, which licenses the operation of the rule in (58): the superscript, i.e., the index of the complement, includes the features number, gender and person, which, in turn, can supply the first three features in the clitic matrix. Let us now considere the derivation (54a,c), (55a,c) (56a) and (57a). The derivation of (54a) is straightforward. R applies at S-structure and the optional (58) does not apply. At S-structure the result is the configuration in (61): (61)
PROJ [ p t p P t i + AGRJ N P M ] [+nominative]
The NP in (61) is co-superscripted with 'eyn, being its complement, and must be co-superscripted with AGR in order to be assigned nomina-
228
Parametric
Syntax
tive Case. Therefore the only grammatical derivation is the one in which i=jLet us now turn th the counterpart of (61) in which/? does not apply at S-structure. In this case the derivation is as in (62) (here, as in all the following examples, AGR and the particle must assign the same superscript. We omit the particle superscript for the sake of simplicity): (62)
•PRCy
AGR*
[p^PtNp·]
The post-verbal subject in the derivation in (62) cannot be assigned nominative Case. In (62) AGR* does not govern NP*, since government into maximal projections is blocked. It follows that NP1 cannot assigned Case and the derivation is ungrammatical. Note further that PRCP in (62) is governed. Thus the sentence is ruled out twice. Now let us consider the derivation in which the post-verbal subject has been fronted to the subject position, but (58) has not applied. The output is as in (63): (63)
(=5 5c) *NPj AGR> [ P t P Pt
[e]j]
In (63) NP| was moved to the subject position, leaving behind a coindexed trace. Although NP] can be assigned nominative Case by the governing, co-superscripted AGR, the derivation is nevertheless ungrammatical, as the empty category left following the fronting of the post-verbal NP is not properly governed (recall that the particle itself is not a proper governor). Now consider a derivation in which (58) has applied, R has applied at S-structure and the post-verbal subject has «or been fronted. This derivation is given in (64): (64)
(=55a)
*PR0"
[ p t p Pt + clj + AGR1
Npj]
In (64), (58) has applied, resulting in a clitic, attached to the particle, which is coindexed with the complement NP}. (64) is ungrammatical. Recall that after application of (58) above the clitic is defective; unless a Case feature is added to it, it will be ruled out. In (64) the addition of the missing Case feature is possible: the clitic can absorb the nominative Case feature of AGR which is now attached to the particle. However, after the absorption of this feature the assignment of nominative Case to the post-verbal NP) is no longer possible. As NP3 cannot receive Case in any other way the derivation is ungrammatical. Let us make this absorption process more concrete. We assumed that clitics are morphologically part of the lexical head to which they are
Parameters for INFL
229
attached. As such, a clitic attached to the particle in (64) will govern the complement NP. Clearly, in order to govern that NP following the application of R. AGR must be part of the particle as well. Since the particle is not a verb, it is natural to assume that the particle does not have a morphological slot for AGR. Thus when AGR is attached to the particle, it is attached in the only slot which is available to it: the missing Case slot of the clitic. Thus the clitic does not absorb nominative Case as such: rather, it absorbs the entire AGR node, the assigner of nominative. Since nominative Case as such is not absorbed (and, in fact, we will argue below that nominative Case is not itself a feature, and hence it cannot be absorbed unless its assigner is absorbed) it would seem that AGR could still assign nominative Case following rule R. This is blocked, however, since AGR is now part o f the clitic and clitics are not Case assigners. If, on the other hand, there is no post-verbal NP which has to be assigned nominative Case in cases such as (64), we expect the derivation to be grammatical. Thus corresponding to (64) and (55a) we have the grammatical sentence in (56a) - the case of pro-drop: (65)
(=56a)
PROj [ P t P P t + c l j + AGR)
[e]j]
Let us consider in detail the derivation o f (65). At D-structure, the structure of (65) is as in (66a): (66)
a. [ N P j i
AGR) 1 st person referential
[ P t P Pt PRO]
[
Two operations apply to (66a): first, the post-verbal PRO is co-superscripted with the particle. At this stage, however, there is no reason to assume that this superscript ( / ) which is assigned to PRO is identical to /', the superscript of AGR and of the null category in the subject position. Secondly (58) applies, resulting in a clitic which is coindexed with the complement PRO. These two operations result in the structure in (66b): (66)
b. [ N P p
AGR)
[ P t P Pt + c l k
PRO£]
[1st person]j Again, several processes apply to (66b). First, by the application of "Move a " , the post-verbal PRO is moved to the subject position, leaving behind a coindexed empty category. (Note that the failure of "Move a " to apply in this case would result in a governed PRO at S-structure, and hence in ungrammaticality.) If PROj, is moved to [ n P }) and l ^ j , the derivation is ruled out. Hence the only grammatical derivation in this
230
Parametric
Syntax
respect is that in which l=j. Note that PRO now carries the same superscript as AGR. Let us assume that, as such, it also has to have the same subscript as AGR, if AGR contains a subscript. Since in (66) it does, we conclude, that k=i, and that the correct representation of PRO in (66) is as PROj. Secondly in (66) R applies at S-structure. Again, the failure of R to apply at S-structure will result in a governed PRO in subject position, thus placing the sentence in violation of the binding conditions. The application of "Move a " and R results in the representation in (66c): (66)c.
PROj [ P t P
[Pt + clj + AGR]] [e]j]
Following the attachment of AGR to the particle, nominative Case is absorbed by the clitic, through the absorption of the AGR node. Thus the missing feature is supplied, rendering the clitic well-formed in accordance with (59). (66c) contains an empty category, which is properly governed by the coindexed clitic. The clitic, an output of rule (58), is well-formed since nominative Case has been absorbed, in accordance with (59). AGR has moved into the PtP at S-structure, resulting in an ungoverned PRO in subject position. Consequently there remains no reason to rule out the sentence in (56a) - and, indeed, it is grammatical. Let us now turn to the last case, the one illustrated by (54c). In this case R applies in the phonology; thus, nominative Case can be assigned to the fronted NP in the subject position. On the other hand, (58) applies as well, resulting in a clitic which governs the empty category left by the preposed subject, and which is coindexed with it. Thus at S-structure the structure of (54c) is as in (67): (67)
(= 54c) NP] AGRJ
[PtP
[ P t Pt +
[e]j]
The clitic in (67) is defective in that it does not contain a Case feature. Recall, however, that we assume that the well-formedness condition for clitics applies in the phonological component. Thus, for the LF component, (67) is well-formed: NP] is assigned nominative by AGR, and the empty category is properly governed by a coindexed, governing clitic. In the phonological component, R applies, attaching AGR to the particle. As such it supplies the missing feature for the clitic: nominative Case. Hence in the phonology, once AGR has been added, the derivation is well-formed, and the clitic adheres to the description in (59) above. Note that we are crucially assuming that in the phonological component, once AGR is moved to the PtP and is attached to the particle,
Parameters for INFL
231
it still contains nominative Case features. It can thus still supply the missing feature to the defective clitic in (67), rendering the clitic well-formed. Nominative Case is still part of AGR, although nominative Case has been assigned at S-structure to NPj in (67). The fact that AGR still contains nominative Case after nominative Case has been assigned supplies evidence for the difference between the assignment of nominative Case and the assignment of other Cases. Recall that inflectional rules of Case assignment require adjacency. Nominative Case does not require adjacency, as is attested by the assignment of nominative Case to postposed subjects. Inflectional rules of Case assignment do n o t require superscripting, whereas nominative Case assignment does. We thus conclude that the assignment of nominative Case is not an inflectional rule, and that unlike inflectional rules of Case assignment it does not involve the linking of an NP with an inflectional feature. Instead we assume that nominative Case is a default Case, assigned in the presence of a co-superscripted, governing AGR node. This AGR feature which licences the assignment of nominative Case is still present in the phonological component as part of AGR, and it can supply the missing component for the clitic in (67).
4.1.1. A Note on Clitics and PF Interestingly, we can spot another case in which the surface well-formedness of clitics is determined after S-structure and in the phonological component. In Chapter 5, while discussing the location of clitics in Romance, we relegated to a footnote (no. 14) the surface location of clitics when auxiliary verbs are present. This location, always preceding the auxiliary verb, is illustrated by (68): (68)
Jean l j - a mange [e]j 'Jean ate it'
The sentence in (68) appears to be problematic in two respects. First, note that the clitic in (68) does not seem to govern its corresponding gap, since it is attached to the auxiliary node, presumably dominated by INFL, whereas the gap is in the VP. Secondly, the clitic is attached to the auxiliary avoir, although we claimed that it is the Case spell-out of manger. Both these problems immediately disappear if we assume that the S-structure representation of (68) is as in (69a), while its PF representation, following the application of Λ in the phonological component, is as in (69b): (69)
a. Jean a [ y p l j - m a n g e [e]j] b. Jean [ y p l j-a-mange [e]j]
232
Parametric
Syntax
The first problem disappears, since given the representations in (69), the clitic governs the gap at S-structure (and subsequently at LF). In fact it governs it at PF as well. The second problem disappears too: following the inflectional rule of affix hopping, the clitic is attached to the V + auxiliary complex, now reanalyzed as one morphological unit. 4.2. The Negation marker 'eyn as an Exceptional Preposition While in its function as a predicate head 'eyn clearly heads a phrase, this is not the situation in the case of the negative marker 'eyn. Thus in a sentence such as (12) above repeated here as (70b), it is hard to see in what sense 'eyn takes a complement or heads any phrase at all: (70)
a. 'ani yoda'at 'et ha-ts'uva I know acc the-answer b. 'eyn 'ani yoda'at 'et ha-tsuva η eg Ί don't know the answer'
Rather, in (70b) 'eyn functions as a negation marker on the clause, and it would be plausible to assume that it is located in the COMP position. Thus we propose the following structure for the clause in (70a): (71)
tg 'eyn [g 'ani yoda'at 'et ha-ts'uva] ]
The structure of the S constituent in (71) would be identical to the structure of the S constituent in sentence (70a), thus capturing the obvious similarities between (70a) and (70b): in both cases, 'ani, Ί ' is the subject occupying the [NP, S] position and in both cases it is assigned Case by INFL. Furthermore, the assignment of Θ-role is identical in both cases. The addition of 'eyn simply negates the entire S. Now consider the sentence in (72) (identical to (13)): (72)
'ani 'eyn-eni yoda'at 'et ha-tsuva I neg-I know acc the-answer Ί don't know the answer'
In (72), we claim, the subject is simply topicalized and appears in a preCOMP position, yielding, at the first step, the structure in (73): 14 (73)
[g 'anij fg 'eyn [g ej yoda'at 'et ha-tsvuva]]]
Now consider the structure in (73), which is ungrammatical. It contains
Parameters for INFL
233
a topic and an element in COMP, followed by an empty category coindexed with the topic. While the sequence [topic-filled COMP] is usually permitted in Hebrew, as is illustrated by (74a), the empty category in the [NP, S] position in such constructions typically exhibits 'NIC' effects, as is demonstrated by the contrast between (74a) and (74b): (74)
a. 'et Rinaj mij [e]j pagas [e]j ? acc Rina who met 'who met Rina?' b. *Rinaj,'et mij [e]j pag& [e]j ? Rina acc who met 'Rina who did she meet?'
Note that a similar contrast is evident in English (although superiority effects seem to bar the literal translation of (74a)): (75)
a. ?this apple, who ate? b. *this man, what apple ate?
Returning to the representation in (73), its ungrammatically follows immediately. The empty category in the [NP, S] position cannot be properly governed by its antecedent, since the intervening COMP is filled by 'eyn. As 'eyn itself is not a proper governor (as demonstrated by (9)-(10) above), the representation in (73) cannot be salvaged. Recall now that we assume that 'eyn is preposition-like in that it is co-superscripted with its object. We will now make an additional assumption about the nature of 'eyn: we will assume that in its function as a negation marker it resembles in many respects the preposition for in English. Consider the preposition for. In most of its occurrences, it is simply the head of a PP, subcategorized for a complement NP to which it also assigns Case. However, in its function as an exceptional Case marker, illustrated by (76), it assigns Case to an NP which is not its subcategorized complement: (76)
fg for[gJohn to come would be irresponsible]]
In (76) John is assigned Θ-role by the predicate headed by come and it is an argument of come. Nevertheless it is assigned Case by the preposition for. Let us assume that the exceptional property of prepositions such as for is that they are allowed to be co-superscripted with an NP which they are not subcategorized for. We can now re-write (76) as (77): (77)
[for 1 [John 1 to come would be irresponsible]]
234
Parametric
Syntax
Within such a system the property of exceptional Case assignment derives from the exceptional assignment of superscripts. Once an NP is co-superscripted with a Case assigner, it will be assigned Case by it. The truly exceptional property is the co-superscripting. 15 Note that the property of assigning superscript by itself is not exceptional: for has this property by virtue of being a preposition. Now consider the negation particle 'eyn. We suggested that it is a defective preposition in that it does not assign Case. However, it does assign a superscript. If we assume that the exceptional property of prepositions such as for is the assignment of a superscript to a non-complement, it is clear that this property can be extended to 'eyn although it is not a Case marker. Thus in a representation such as (73), 'eyn will be exceptionally co-superscripted with the [NP, S] position although that position is not its complement and is not assigned Case by it. Note that the [NP, S] position already bears a superscript: it was cosuperscripted with AGR by the D-structure rule discussed in section 3. Since it clearly cannot bear two conflicting superscripts, we must assume that the only grammatical representation is that in which AGR and 'eyn bear the same superscript. This representation is given in (78): (78)
'anij fjj'eyn 1 [ s ej AGR 1 yoda'at 'et ha-tsuva]]] I neg know acc the-answer
Recall that the presence of a superscript on the particle 'eyn licensed the application of a clitic formation (see (58) above). The clitic, defective in that it does not contain a Case feature, can then serve as a proper governor for a coindexed empty category. Note, in particular, that it can properly govern the empty category in the subject position. Thus the representation in (79) is well-formed at S-structure and in LF (and see the parallel derivation for the ergative particle illustrated by (67)): 079)
[§ fc ' e y n
+clJ
i Is ^i A G R *
k n o w the
answer]]]
At PF, however, the derivation in (79) will be ruled out since the clitic on 'eyn is not well-formed: it does not contain a Case feature. In order to rescue the derivation in (79) we will assume (with Stowell 1982, Pesetsky 1982) that the INFL node may move into the COMP position. We will take this movement to be a syntactic movement, i.e., not an inflectional rule. (We will deviate from the respective analyses in the references, however, in assuming that this movement is not conditioned by Case assignment, scope of tense or any other factor. It is entirely optional.) (79) can now be rescued if the INFL node moves into COMP. Following the same procedure described above for the example in (67),
Parameters for INFL
235
the nominative Case assignment features of INFL will be absorbed by the clitic on 'eyn, supplying the missing Case feature and yielding a wellformed clitic at PF. We have now accounted for the grammaticality of (70b) and (72), while correctly excluding the ungrammatical (73) on the grounds of the lack of proper government for the empty category in the [NP, S] position. Now consider the pro-drop case, illustrated by (80): (80)
'eyn-eni yoda'at 'et ha-tsuvat neg I know acc the-answer
At D-structure, (80) is represented as in (81): (81)
fg 'eyn) [ S PROJ AGR) [know the answer]]]
Both 'eyn and the AGR node assign a superscript to PRO following the procedure described above. In the same way the identity of the superscript assigned by 'eyn and by AGR follows. AGR will further assign a person index to the PRO (see section 3 for discussion), in this case, a first person singular. If the representation in (81) is allowed to reach S-structure, however, it will be ruled out: the PRO in the [NP, S] position is governed both by the AGR node and by 'eyn, resulting in ungrammatically. Note that if rule R applies at S-structure it will eliminate government by AGR, but not government by 'eyn. The structure can be rescued if PRO is topicalized, resulting in (82): (82)
[ | PRoj [ s 'eyn* [ s ej AGR* [know the answer]]]]
The representation in (82) does not contain a governed PRO, but it does contain an empty category which is not properly governed in the [NP, S] position. The proper government of that position is achieved if the superscript on 'eyn is spelled-out as an indexed clitic. This clitic is in turn rendered well-formed by the promotion of AGR into the COMP node and the merging of the clitic features with the Case feature of AGR. Note that in cases (72) and (80) we crucially require the movement of INFL into COMP. We have assumed above that this movement is a syntactic movement, i.e., it must happen prior to S-structure. Excluding this type of movement from the class of inflectional rules is natural, since it may move over an intervening element (typically the [NP, S] position, regardless of its "heaviness"). Following the movement of the INFL node into COMP, and given the presence of 'eyn in that COMP, the following are possible outputs:
236
Parametric
Syntax
(83)
a.
COMP 'eyn + c l j
b. INFL
COMP 'eyn + c l j + lNFL
Whereas the structure in (83a) branches, the structure in (83b) represents a state of affairs in which INFL is absorbed by the clitic matrix on 'eyn, resulting in a non-branching structure. We will follow standard analyses in assuming that in (83a), INFL may not govern the [NP, S] position, since it does not c-command it. In this respect INFL government of the [NP, S] position from COMP is blocked in (83b) in the same fashion that the government by the trace in COMP is blocked in sentences such as *who did you think [e,· that [e, saw John]]? (cf. Rizzi 1982, based on Pesetsky 1982). Now consider the structure in (83b): INFL does govern the [NP, S] position, but as part of a clitic. Clitics, however, are not Case assigners. Thus INFL in (83b) cannot assign Case to the [NP, S] position. Returning to (72) and (80), in which INFL moves into COMP, we now see that the [NP, S] position cannot be assigned nominative Case. Since the trace is properly governed, due to its coindexation with the governing clitic this is not a problem for (72) and (80). Consider, however, the remaining case, the ungrammatical sentence in (84): 16 (84)
*'eyn-enij 'anij yoda'at 'et ha-tsuva neg-I I know acc the-answer
A well-formed clitic in (84) can only be derived by moving INFL to COMP, but such movement will result in the structure in (85): (85)
fg 'eyn+ cl + INFL [ s I [know the answer]]]
In (85) either INFL is absorbed by the clitic, in which case it can no longer assign nominative Case, or, it creates a branching structure, destroying the government relationship between INFL and the [NP, S] position. Thus the derivation in (84) is ungrammatical. To conclude this section, it is clear that the clitic on the negative marker 'eyn is identical to the clitic on the ergative particle 'eyn. In both cases it is a defective clitic matrix, supplemented by the Case features of INFL. In both cases, it can function as a proper governor, rescuing otherwise ungrammatical extractions. Furthermore, we have given a unified account of the particle 'eyn, assuming that it is a Caseless preposition. Utilizing a superscripting device, we have shown that 'eyn is best characterized as a preposition which takes complements, to which it sometimes assigns a Θ-role (the ergative 'eyn) and sometimes does not. In
Parameters for INFL
237
the latter case, it is best characterized as an exceptional superscript assigner, on a par with the preposition for in English. Lastly, we wish to direct the reader's attention to scope differences which correspond to whether or not a clitic is attached to the particle 'eyn. Thus consider the following sentences: (86)
'eyn sney gvarim 'ohavim 'ota 'isa neg two men love same woman 'It is not the case that (any) two men love the same w o m a n ' 'It is not the case that there is a woman such that two men love her'
(87)
sney gvarim ' e y n - a m 'ohavim 'ota 'isa two men n e g - t h e m love same woman 'there are two men such that they don't love the same woman'
'two m e n ' cannot take wide scope over the negation marker in (86). It does have a wide scope in (87). Note that a wide scope reading in (86) would result in an ECP violation (see (9) and (10) above and related discussion). Wide scope in (87) will not result in a similar violation. Here, again, we find a striking similarity between the properties of negative 'eyn and ergative 'eyn and the properties of the clitics attached to them.
5. EXISTENTIAL SENTENCES - THE ACCUSATIVE DERIVATION
Interestingly, some occurrences of the existential particles in Modern Hebrew are currently undergoing a process of reanalysis. This is particularly true of all the uses of particles in which n o true existential meaning is expressed. Thus alongside the sentences in (54) and (55) above we have the following: 1 7 (88)
'eyn 'et ha-sefer h a - z e ba-sifriya exist-not acc t h e - b o o k t h e - t h i s in-the-library 'this book is not in the library'
Clearly (88) indicates that the particle 'eyn in cases such as (88) functions as an accusative-assigning particle. Furthermore, when the particles function as accusative assigners, they allow for extraction from the postparticle position without the presence of a coindexed clitic. Thus (89) and (90) are possible;
238
Parametric
Syntax
(89)
'et maj yes" [ej] ba-sifriya? acc what exists in-the-library 'what is there in the library?'
(90)
yes" 'et sloset ha-sfarim se-xipasta ba-sifriya exist acc three the-books that-searched-you in-the-library 'the three books that you were looking for are in the library'
In (89) extraction takes place with the accusative marker 'et, leaving an empty category in the position following the particle. In (90), 'three books' receives wide scope interpretation. Interestingly, fronting of the object of the particle into the subject position once accusative Case is assigned is impossible. Thus (91a) has only a topicalized reading, and (91b) is ungrammatical: (91)
a. sifrei yeladim yes ba-sifriya, sifrey mevugarim, lo books children exist in-the-library books adults no 'there are children's books in the library, but not adults' books' b. *sifrey ha-yeladim yes ba-sifriya etc. books the-children 'the children's books are in the library etc.
We believe that the sentences in (88)-(91) can be explained if we assume that particles in Modern Hebrew are reanalyzed as accusative Case assigners. The assignment of accusative Case is optional. However, when the particle does assign accusative, it completely assimilates to the verbal class. In particular, it can function as a proper governor without the presence of an attached clitic coindexed with the empty category. Thus the representation given in (92) is grammatical in the accusative derivation: (92)
[ p t p Pt [+acc]
[e]
]
The configuration in (92) is the representation of both (89) and (90) at the stage at which ECP is relevant. The availability of proper government in (92) thus renders (89) and (90) grammatical. 18 Now recall that we argued that the superscripting of the [NP, VP] in the VP in "ergative" constructions is random. Only if the superscript assigned to the [NP, VP] (or in our case, [NP, PtP]) agrees with the superscripting of AGR can the post-verbal NP receive nominative Case. This occurs when R applies at S-structure and AGR both governs the post-verbal NP and is co-superscripted with it. Recall that such a system would pre-
Parameters for INFL
239
diet nominative Case assignment to direct objects, under the random assignment of supercripting. However, such assignment is blocked by exploiting "Bureio's generalization" (see section 2 for a discussion). However, the reanalyzed particles in Hebrew seem to violate "Burzio's generalization". Although they have accusative Case assignment features, their subject is not in a Θ-position. Consequently, precisely in the case of these Hebrew particles we expect two possible derivations, depending on the superscript assigned to the post-particle NP. If the post-particle NP is assigned the same superscript as the A G R element, we expect nominative Case for the [NP, PtP], and we do not expect the particle to function as a proper governor. This derivation is the nominative derivation outlined in detail in section 4. If, on the other hand, the superscript assigned to the post-particle NP does not agree with that of AGR, nominative Case cannot be assigned. Consequently the derivation can only be salvaged if accusative Case is assigned by the particle. In that case the particle becomes a proper governor and we predict the grammaticality of (89) and (90) above. In the accusative derivation a clitic on the particle would be reanalyzed as well: it would no longer be a composition of the AGR node attached to the particle with gender, number and person features inserted by the rule (58) above. Rather, it would be the regular clitic, incorporating the accusative Case features of the particle itself, having the structure in (93): (93)
[ p t Pt, [a gender, β number, δ person, + accusative]
Consequently we would not expect the sentence in (94) to be grammatical in the accusative derivation (recall that it was ungrammatical in the nominative derivation as well, b u t due to different reasons. See (64) above and related discussion): (94)
* ' e y n - e n u j 'et h a - x a t u l j ba-gan exist-not acc the-cat in-the-garden 'the cat is not in the garden'
In (94) the clitic absorbs the accusative Case and hence 'the cat' cannot be assigned Case. 19 Note, however, that we are dealing here with a hypothetical situation, since Modern Hebrew no longer has accusative clitics. Now note that, given these two possible derivations, the non-topicalized reading of (91a) and the sentence in (91b) are still ruled out. First consider (91b). The accusative Case marker 'et is not present; hence we know that the definite NP sifrey ha-yeladim 'the children's books' is not marked as accusative. However, if the particle does not assign accusative, it cannot function as a proper governor. It follows that
240
Parametric
Syntax
(91b) contains an empty category which is not properly governed in the post-particle position. Of course that empty category could be properly governed if (58) above applied, resulting in a clitic spell-out. This situation would yield the grammatical (95) corresponding to (54c) above (and see derivation (67) above): (95)
sifrey ha-yeladim yes-nam^ [e]j ba-sifriya 'the children's books are in the library'
Now consider (91a). There is no overt accusative Case marker on indefinite direct objects (see Chapter 2 section 4.2 for a discussion of 'et)·, hence we do not know whether 'children's books' was assigned accusative or not. If no accusative Case was assigned, the non-topicalized reading of (91a) is ruled out in the same way that (91b) is ruled out. Now let us consider the possibility that accusative Case is assigned. When the NP is moved to the subject position, it moves into a null category that is cosuperscripted with AGR in the base (see above for discussion). If the superscript of the moved element conflicts with that of the null category, the conflict will lead to ungrammaticality. If it agrees with that of the null category, the moved NP carries the same superscript as AGR and thus will be assigned nominative Case. If we assume that accusative Case is assigned to the moved NP prior to its preposing, such assignment will result in a Case conflict and hence in ungrammaticality. If accusative Case is not assigned prior to movement, it will be assigned to the coindexed trace. However, since the antecedent of that trace carries the same superscript as AGR, so does the trace. Thus this assignment will result in the situation in (96): (96)
NPJ AGRJ [ p t p P t [nominative]
[e}i] [accusative]
In (96) NPJ forms an Α-chain with its co-superscripted trace. This chain is thus assigned two distinct Cases; hence this situation results in Case conflict and in ungrammaticality. Now consider the topicalization reading. In this derivation the NP is moved to a non-Case position and hence the trace left behind can be accusative. This fact does not cause any conflict in Case assignment to chains. Thus the topicalized reading is the only possible reading of (91a). It has been observed by Shoshani (1980) that the process of reanalyzing post-verbal subjects as direct objects is more general, and applies to other ergative configurations as well. Thus, in substandard Hebrew, (97c), (98c) and (99c) are grammatical:
Parameters for INFL
241
(97)
a. hayta ktuva yedi'a xasuva ba-'iton was-f written-f new(s)-f important in-the-paper 'an important piece of news was written in the paper' b. ha-yedi'a ha-zot hayta ktuva ba-'iton the-new(s)-f the-this-f was-f written-f in-the-paper c. Iiaya katuv 'et ha-yedi'a ha-zot ba-'iton was-m written-m acc the-new(s)-f the-this-f in-the-paper
(98)
a. meforatim harbe dvarim ba-karoz ha-ze specified-m-pl many things-m-pl in-the-leaflet the-this 'many things are specified in this leaflet' b. harbe devarim meforatim ba-karoz ha-ze many things-m-pl specified-m-pl in-this leaflet c. meforat 'et ha-dvarim ba-karoz ha-ze specified-m-sg acc t h e - t h i n g s - m - p l in this leaflet
(99)
a. karta 1-i te'una xamura ba-derex happened-f t o - m e accident-f serious-f in-the-road Ί had a serious accident on the road' b. te'una xamura karta 1-i ba-derex accident-f serious-f happened-f t o - m e in-the-road c. kara 1-i kvar 'et ha-te'una ha-zot kodem happened-m to-me already acc the-accident-f the-this-f before
The verbs in (97)-(99) are ergative verbs which were reanalyzed to have accusative Case assigning features similar to the reanalysis which applied to yes and 'eyn.20 If the post-verbal NP in these cases is assigned the same superscript as AGR, once R applies at S-structure, the post-verbal NP is assigned nominative. Alternatively, the post-verbal NP can move to the [NP, S] position, in which case R does not apply at S-structure and nominative Case is assigned in the usual way. These cases are demonstrated by sentences (a) and (b) of (97)-(99). When the post-verbal NP is assigned nominative, we expect full agreement between the verbal inflection and the subject, and indeed, we find such full agreement in the (a) and (b) cases. In the (c) cases, on the other hand, the post-verbal NP is assigned a superscript which differs from that of AGR. Hence it is assigned accusative by the verb itself. In these cases we do not expect agreement between the verbal inflection and the accusative NP. As shown by the (c) cases above, we find in these cases that the verb is inflected in the third person masc. sing., the unmarked form, regardless of the gender and number of the post-verbal NP. Note that both in (97c), (98c) and (99c) and in (88)-(90) we nevertheless have to assume that R applies at S-structure so as to permit a
242
Parametric
Syntax
(pleonastic) PRO to appear in the [NP, S] position. If AGR does not move into the VP at S-structure, the subject position is governed but not properly governed, blocking the occurrence of any empty element, PRO or [e]. Thus both the accusative derivation of existential sentences and the accusative derivation of some ergative verbs in Modern Hebrew supply interesting evidence for our claim that there is no rule which co-superscripts the inserted PRO in ergative configurations and the post-verbal subject. Rather, NP's are superscripted at random. If they agree in superscripting with AGR they are assigned nominative. But if they do not, and if there is no other way to assign Case to them, the configuration is ruled out. On the other hand, if the post-verbal NP's (or post-particle NP's) can be assigned accusative by the verb or by the particle, we expect both derivations to be possible and grammatical, and indeed they are. 5.1. Free Superscripting and there Insertion An interesting confirmation of the assumption that superscripts are assigned at random to post-verbal NP's is found in English. Both (100a) and (100b) are grammatical, although the latter is considered substandard English: (100) a. there are at least seven people in the garden b. there's at le^st seven people in the garden The rule suggested in Chomsky (1981) which co-superscripts expletive inserted PRO's with post-verbal subjects in ergative constructions is also utilized to co-superscript the pleonastic element there in (100) with the post-verbal subject. In this case the co-superscripting is utilized to form an Α-chain which consists of the pleonastic element inserted at S-structure and the post-verbal NP. This Α-chain is then assigned Case by AGR. Note that since in English R cannot apply at S-structure, the formation of an Α-chain linking the pleonastic element to the post-verbal NP is crucial. In this way the assignment of nominative Case to the pleonastic element enables the post-Z>