Features of Person: From the Inventory of Persons to Their Morphological Realization 0262535610, 9780262535618

A proposal that person features do not have inherent content but are used to navigate a “person space” at the heart of e

332 98 3MB

English Pages 382 Year 2018

Report DMCA / Copyright

DOWNLOAD FILE

Polecaj historie

Features of Person: From the Inventory of Persons to Their Morphological Realization
 0262535610, 9780262535618

Table of contents :
Contents
Series Foreword
Preface
Abbreviations
1. Person: A To-Do List
2. Person Features: Deriving the Inventory of Persons
3. Interactions between Person and Number
4. Impersonal Pronouns
5. The Symmetric Nature of Agreement
6. Syntactic Restrictions on Agreement
7. The Morphological Realization of Person
8. Conclusion
Notes
References
Index

Citation preview

Features of Person

Linguistic Inquiry Monographs Samuel Jay Keyser, general editor A complete list of books published in the Linguistic Inquiry Monographs series appears at the back of this book.

Features of Person From the Inventory of Persons to Their Morphological Realization

Peter Ackema and Ad Neeleman

The MIT Press Cambridge, Massachusetts London, England

© 2018 Massachusetts Institute of Technology All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information storage and retrieval) without permission in writing from the publisher. This book was set in Times LT Std by Toppan Best-set Premedia Limited. Printed and bound in the United States of America. Library of Congress Cataloging-in-Publication Data Names: Ackema, Peter, 1965- author. | Neeleman, Ad, author. Title: Features of person : from the inventory of persons to their morphological realization / Peter Ackema and Ad Neeleman. Description: Cambridge, MA : The MIT Press, [2018] | Series: Linguistic inquiry monographs ; 78 | Includes bibliographical references and index. Identifiers: LCCN 2017057053| ISBN 9780262038195 (hardcover : alk. paper) | ISBN 9780262535618 (softcover : alk. paper) Subjects: LCSH: Grammar, Comparative and general--Person. | Grammar, Comparative and general--Pronoun. | Grammar, Comparative and general--Pronomials. | Grammar, Comparative and general--Morphosyntax. | Generative grammar. 10 9 8 7 6 5 4 3 2 1

Contents

Series Foreword  vii Preface  ix Abbreviations  xi 1

Person: A To-Do List  1 1.1 Introduction  1 1.2 The Inventory of Persons and Their Interpretation  1.3 Agreement for Person  10 1.4 Syntactic Restrictions on Agreement  13 1.5 The Realization of Person Features  15 1.6 Maximal Encoding  16 1.7 Summary  19

4

2

Person Features: Deriving the Inventory of Persons  21 2.1 Introduction  21 2.2 The Basic Proposal  22 2.3 The Interpretation of Person in Discourse  30 2.4 The Third Person as Default  37 2.5 Effects of Maximal Encoding  40 2.6 Honorifics  42 2.7 Consequences for the Morphosyntax of prs  50 2.8 Bound Pronouns  62 2.9 Conclusion  68

3

Interactions between Person and Number  71 3.1 Introduction  71 3.2 Two Types of Number System  72 3.3 Number in Pronouns and Associative Plurals  3.4 Conclusion  104

4

Impersonal Pronouns  105 4.1 Introduction  105 4.2 Overview  106

84

vi Contents

4.3 Generic and Arbitrary Readings of Plural Pronouns  116 4.4 Dedicated Impersonal Pronouns: imp-2  120 4.5 Generic Second Person Singular Pronouns  123 4.6 Dedicated Impersonal Pronouns: imp-1  127 4.7 Why Inclusive Impersonal Pronouns Are human  130 4.8 Conclusion  132 5

The Symmetric Nature of Agreement  133 5.1 Agreement: Symmetric or Asymmetric?  133 5.2 Feature Spreading  135 5.3 Referential Unagreement  146 5.4 Quantificational Unagreement  159 5.5 The Link between Unagreement and Pro Drop  5.6 Conclusion  167

163

6

Syntactic Restrictions on Agreement  169 6.1 Introduction  169 6.2 Agreement and the Theory of Syntactic Dependencies  171 6.3 Syntactic Agreement versus Restrictor Agreement  176 6.4 Agreement: Upward and Downward  188 6.5 The Domain of Feature Spreading  214 6.6 Intervention Effects  234 6.7 Conclusion  239

7

The Morphological Realization of Person  241 7.1 Introduction  241 7.2 Patterns of Syncretism  242 7.3 Accounting for Person Syncretisms  246 7.4 Shifts in Person Syncretism in Dutch  259 7.5 Conclusion  281

8 Conclusion  283

Notes  287 References  323 Index  351

Series Foreword

We are pleased to present the seventy-eighth volume in the series Linguistic Inquiry Monographs. These monographs present new and original research beyond the scope of the article. We hope they will benefit our field by bringing to it perspectives that will stimulate further research and insight. Originally published in limited edition, the Linguistic Inquiry Monographs are now more widely available. This change is due to the great interest engendered by the series and by the needs of a growing readership. The editors thank the readers for their support and welcome suggestions about future directions for the series. Samuel Jay Keyser for the Editorial Board

Preface

This book is the result of a long-standing collaboration, dating back to our time as undergraduate students at Utrecht University in the 1980s. We cannot recall when we started talking about person in particular, but it is so long ago that it is not possible to thank everyone who helped us develop our ideas. Apologizing, therefore, for any oversights, we would like to thank the following people. For helpful discussion or other forms of assistance during the writing of this book, we thank Suzanne Aalberse, Klaus Abels, Sam Ackema, Matthew Baerman, Amela Camdzic, Greville Corbett, Jan Don, Antonio Fábregas, Jane Grimshaw, Daniel Harbour, Georg Höhn, Olaf Koeneman, Hans van de Koot, Jeffrey Lidz, Marc Lowenthal, Anne Mark, Klaartje Neeleman, Tessa Neeleman, Nadja Rajgelj, Henk van Riemsdijk, Maaike Schoorlemmer, Jolijn Sonnaert, Yasutada Sudo, Kriszta Szendrői, Misako Tanaka, Jochen Trommer, Edwin Williams, Anthony Zannino, Ziren Zhou, and Wim Zonneveld. Some people carefully read a first draft of this book, and their comments and objections significantly helped shape the final result. They include Caroline Heycock, Andrew Nevins, Omer Preminger, and three anonymous reviewers for MIT Press. We also received very useful feedback from participants in courses where we presented this material. Apart from students at the University of Edinburgh and University College London, we would like to mention the audience at the University of the Basque Country in 2013 and at the LOT Summer School at Utrecht University in 2016. Aspects of the research we report on were presented in talks at the following institutions: the University of Verona (Words Don’t Come Easy Workshop, 2008), the University of York (North East Syntax Seminar, 2010), the City University of Hong Kong (2011), the Radboud University Nijmegen (Morphology Days, 2011), the University of Tromsø (2011), the Catholic University of Leuven (BCGL 7, 2012, and BCGL 10, 2017), the Goethe University at

x Preface

Frankfurt am Main (2012), the University of Amsterdam (2012), the University of Edinburgh (2012), the University of California at Santa Cruz (2013), the University of Olomouc (2013), the School of Oriental and African Studies (2014), the University of Manchester (2014), the University of Zadar (Agreement without Borders conference, 2015), and the University of Leipzig (2017). The questions raised at these occasions made us rethink various issues. Our work was greatly aided by Leverhulme Research Fellowships to Ackema (RF-2013-181) and Neeleman (RF-2013-144). These made it possible to finish a first draft of the manuscript within a reasonable amount of time. Parts of this book are based on earlier publications. Chapter 5 is a substantially revised version of Ackema and Neeleman 2013b; chapter 7 and sections of chapter 2 are reworked from Ackema and Neeleman 2013a.

Abbreviations

1 first person 2 second person 3 third person a aspect abs absolutive acc accusative act active voice af agent focus agr agreement ap applicative arb arbitrary asp aspectual marker assoc associative aug augmented auth author aux auxiliary clf classifier cmp comparative com completive aspect comn common gender comp complementizer conj conjunct coor coordination ctr selectional requirement for agreement controller dat dative dec declarative def definite deic deictic

dim diminutive dir direct dist distal du dual emph emphatic erg ergative excl exclusive expl expletive fem feminine foc focus fut future fv final vowel gen genitive gnr generic hon honorific imp impersonal imper imperfective inan inanimate incl inclusive inf infinitive inv inverse lnk linker loc locative masc masculine min minimal nmb number neg negation neut neuter nom nominative

xii Abbreviations

O object obj object agreement obv obviative OP operator par participant part participle agreement perf perfective pl plural poss possessive pret preterit prog progressive prox proximal prs person

prt particle pst past refl reflexive sbj subjunctive sg singular sr selectional requirement ta transitive verb with animate object trg selectional requirement for agreement target vta transitive verb with animate agent and animate patient

1

Person: A To-Do List

1.1 Introduction

Φ-features provide a unique window into the language faculty. They have semantic import, they are involved in the remarkable syntactic relation of agreement, and their morphological realization displays intricate patterns of syncretism and allomorphy. In this book, we will develop a theory of one type of φ-feature—namely, person—aiming to elucidate the journey of person features in grammar, from the semantics, through the syntax, to the system of morphological realization. As a strategy, this approach can be contrasted with studies of the behavior of a larger set of features in a single module of grammar. Both approaches are necessary. The value of an in-depth cross-modular study of a single φ-feature is that it forces us to develop a theory in which assumptions made about the behavior of that feature in one module bear on possible assumptions about its behavior in other modules. In the case at hand, those cross-modular implications mainly concern the inventory of person features, which must guide the semantic interpretation of pronouns and at the same time must provide insight into attested patterns of morphological realization. In the course of analyzing the grammatical behavior of person, we will defend a number of theoretical choices that are either new or not generally accepted. To give an idea of what lies ahead, we will mention a few of these choices, starting with a claim pertaining to the semantics of person. In many theories, person features are assumed to directly refer to discourse participants (which we will henceforth simply refer to as participants). Such direct reference is an assumption in traditional grammar, where first and second person are understood to include reference to speaker and addressee, respectively. The idea is also present in the influential proposals by Halle (1997) and Harley and Ritter (2002), who adopt features such as speaker (or author), addressee, and participant.

2

Chapter 1

By contrast, we will argue that person features are functions that operate on a domain of discourse referents, but do not pick out fixed referents. Rather, they either select or discard part of their input. Combinations of these features then result in specific subsets of the discourse domain that correspond to traditional persons. We will argue that this allows for a more insightful account of the crosslinguistic inventory of persons. As for syntax, person is centrally important to this module mainly because of its involvement in agreement. In person agreement, the feature makeup of a verbal element appears to depend on the feature makeup of a nominal argument. There is an undeniable semantic asymmetry between these two sets of features, in that person contributes to the interpretation of nominal arguments, but not to the interpretation of verbs. This semantic asymmetry is reflected in the terms controller and target of agreement (see, for instance, Corbett 1983, 2006). Probably because of this asymmetry, many theories assume that there is also a syntactic asymmetry between controller and target, in two respects. First, the features of the target are taken to be dependent on the features of the controller: if the controller does not carry a particular feature in syntax, then the target cannot carry that feature either. Second, controller and target are assumed to always stand in the same c-command relation with respect to each other. A clear example of such an asymmetric theory is Chomsky’s (2001) probe-goal system. In this system, the target (the probe) is required to value its φ-features and can only do so if it c-commands a controller (the goal) that carries an interpretable version of these features. We will argue in favor of an alternative line of analysis according to which the semantic asymmetry between target and controller does not translate into a syntactic asymmetry. There is no difference in the syntactic status of φ-features in controller and target. This implies, on the one hand, that the target can have more features in syntax than the controller and, on the other hand, that the structural relation between controller and target need not always show the same direction of c-command. At the same time, however, agreement relations need to be encoded syntactically, and we will show that as a consequence specific instances of agreement are structurally asymmetric. In some cases, the target needs to c-command the controller; in others, the controller needs to c-command the target.1 There are no cases, however, where both options are available (resulting in variability in the direction of agreement between two elements). The claim that there is no difference in the syntactic status of φ-features in controller and target is shared by analyses of agreement in terms of unification. Examples of this idea can be found in theories such as Head-Driven Phrase Structure Grammar and Lexical-Functional Grammar (see, for instance,

Person 3

Gazdar et al. 1985, Shieber 1986, Barlow 1992, and Pollard and Sag 1994). In Minimalism, a similar line is adopted in Brody 1997. Turning now to the morphological realization of person features, we will follow the general line of models adopting “late insertion.” That is to say, the syntax contains only syntactic information, which includes person features. It does not contain the morphological forms that are associated with these features. Rather, a language specifies a set of spell-out rules and possibly also a set of rules of impoverishment, which operate at the interface between morphosyntax and morphophonology. A novel aspect of our proposal is that these rules are sensitive to the way that the syntax represents the order of function application, with features that represent functions applied later “outside” of features that represent functions applied first. At the point of morphological realization, the presence of an outer feature blocks spell-out rules and impoverishment rules that exclusively target inner features. This effect will be crucial in explaining patterns of syncretism. It will be clear that, in a project of this type, principles that operate at the interfaces between the various modules of grammar play a crucial role. A classic principle that codetermines which morphological realization is chosen, given a syntactic input, is the Paninian principle of blocking, introduced into  modern grammar as the Elsewhere Principle (see Kiparsky 1973). The same idea occurs in various other guises, including the Subset Principle of Distributed Morphology (see Halle and Marantz 1993). According to the Paninian  principle, a morphological form is blocked if it expresses a subset of the features expressed by another available form. We will argue for a generalized version of this principle that restricts not only the interface between syntax and morphophonology, but also the interface between semantics and syntax. This principle, which we will dub Maximal Encoding, requires that in the trajectory from semantics to phonology each subsequent representation encode as much of the information contained in the previous representation as possible. Just as the morphophonology is a maximal encoding of syntax, so the syntax itself is a maximal encoding of semantics (for a related proposal, see Bobaljik and Wurmbrand 2012). In the remainder of this introductory chapter, we will outline some of the problems that motivate the theoretical outlook just sketched. The organization of the chapter mirrors that of the book: we will discuss the interpretation of person features, then their role in syntax, and finally their morphological realization.

4

Chapter 1

1.2  The Inventory of Persons and Their Interpretation

At first sight, the semantics of person seems fairly trivial. People (and objects) can be classified according to their role in discourse: there is an I (the speaker), a you (the addressee), and others, which we represent by the symbols i, u, and o. Let us assume, for now, that i and u are unique, while there can be multiple tokens of o. (We will come back to this assumption later.) A reasonably adequate typology of grammatical persons can then be generated by simply combining these three primitives into sets, as in (1). (We use o+ to indicate inclusion of one or more o’s. Thus, in the first and second person plural, the denotation of the pronoun can include one or more o’s. In the third person plural, it contains at least two o’s.)2 (1) First person singular: Second person singular: Third person singular: First person plural inclusive: First person plural exclusive: Second person plural: Third person plural:

{i} {u} {o} {i, u}, {i, u, o+} {i, o+} {u, o+} {o, o+}

This theory seems to work well. At the very least, it does not seem to present an obvious problem of overgeneration; that is, we cannot construct sets that fail to correspond to person specifications attested in natural language. However, one of the assumptions underpinning this a priori attractive proposal is problematic. It is not true that everyone other than speaker and addressee is equal. Rather, in a first person exclusive plural pronoun the others must all be associated with the speaker, while in a second person plural pronoun the others must be associated with the addressee. In a third person pronoun, the others do not have an association with either speaker or addressee. (We will discuss this associative effect in more detail later, but for now we should note that association is not a constant relationship; rather, it is context-dependent.) To illustrate this, we consider a situation in which Peter and Ad are going to a football game, with Peter supporting Vitesse and Ad supporting Feyenoord. If Ad is coming back after collecting the obligatory cans of lager and asks Peter what happened, Peter might utter (2a) or (2b). (2) a. We scored. b. You scored.

Person 5

It is clear that the reference of we in (2a) is not Peter and some random collection of other people. It must be Peter and those he is associated with in the context at hand, that is, the players of Vitesse (and other supporters of that club). Similarly, you in (2b) must refer to Ad and the players of Feyenoord (and other supporters). The opposite association would suggest that Peter has switched sides (which is unlikely to happen). This means that the others subsumed under we are different from those subsumed under you, which are yet different from the group that they may refer to (some third party, possibly the officials in the situation described). This kind of observation is not new. A number of authors have noted that there seems to be a semantic parallel between so-called associative plurals and the plurals of first and second person pronouns (see Jespersen 1924, Barulin 1980, Corbett 2000, Moravcsik 2003). Associative plurals of regular R-expressions are found in a variety of languages. As Corbett (2000:102) notes, Hungarian has two plural forms of proper names like János. One, Jánosok, refers to more than one person called János. The other, János-ék, refers to János and his associates. A clear difference between the regular plural and the associative plural is that the latter is nonhomogeneous: the elements in the plurality are not all tokens of the same type. Such nonhomogeneity is of course typical of the plural of first and second person pronouns. A first person plural pronoun can refer to a group containing the speaker and others; a second person plural can refer to one or more addressees and possibly others. In principle, then, one could try to attribute the interpretations in (2) to the number system rather than the person system. On this view, the “associative effect” in the first and second person plural would be a consequence of pronouns only accepting associative plural marking. While this is indeed a logical possibility, we will argue in chapter 3 that it faces various problems when compared to a theory that attributes the interpretations in (2) to the person system. Another possibility is that the associative effect derives from a general requirement of “cohesion” that restricts the reference of pronouns. This requirement would state that all individuals referred to by a pronoun must be associated with each other through some contextual criterion. This cannot work, however, because there are cases that fail to meet this more general criterion. One example is the first person plural inclusive pronoun. In the denotation of such a pronoun, one may find others that are associated with the speaker, as well as others that are associated with the addressee. However, these two types of associated others need not themselves be associated. For example, if

6

Chapter 1

speakers A and B discuss the possibility of their families meeting, speaker A might utter the following: (3)  We should really go out for a meal some time. Here, we includes A and A’s family, as well as B and B’s family. The use of we is felicitous even if the members of A’s family have no connection with the members of B’s family, or are even unaware of their existence. Impersonal pronouns are a still stronger counterexample to a general notion of cohesion. Their reference frequently includes nonassociated individuals. Consider, for example, the generic use of you in (4). The interpretation of you here is ‘people in general’, and it seems unlikely that all people are associated with each other through some contextual criterion. (4)  You should not give matches to young children. We conclude that the associative effect observed in plural pronouns cannot be the result of a general principle of cohesion requiring that all individuals in the denotation of a pronoun must be associates of each other. What the above discussion shows is that, while there may be three primitives i, u, and o, the status of o’s varies: there are o’s that are genuine others, there are o’s that are associated with the addressee, and there are o’s that are associated with the speaker. The above observations about the associative effect can be captured if we divide the person domain into three distinct layers, one containing i and others, one containing u and others, and one containing only others. The generalization then is that others that occur in the same layer as a participant must be associated with that participant; beyond this, there is no demand for cohesion. To distinguish the various types of others, we will label the others associated with i as ai, the others associated with u as au, and the nonassociate “true” others as o. From here on, we will use the term others exclusively for the latter. This leads to the following organization of the person space, to be explored in more detail in the next chapter: (5)  i (ai+) u (au+) (o+) Of course, this particular organization of the person space is at this point no more than a suggestion. However, it is not difficult to demonstrate that some kind of system governing the status of different type of o’s is necessary. Not

Person 7

developing such an organizing system is tantamount to working with five rather than three primitives (namely, i, ai, u, au, o). In a system with five primitives, there are many more potential pronouns, since free combination of these five primitives gives rise to 30 distinguishable nonempty sets, even when we abstract away from the possibility of multiple tokens of ai, au, and o. These are listed in (6). (6) {i, ai, u, au, o}, {i, ai, u, au}, {i, ai, u, o}, {i, ai, au, o}, {i, u, au, o}, {ai, u, au, o}, {i, ai, u}, {i, ai, au}, {i, ai, o}, {i, u, au}, {i, u, o}, {i, au, o}, {ai, u, au}, {ai, u, o}, {ai, au, o}, {u, au, o}, {i, ai}, {i, u}, {i, au}, {i, o}, {ai, u}, {ai, au}, {ai, o}, {u, au}, {u, o}, {au, o}, {i}, {ai}, {u}, {au}, {o} We can divide the collection of sets in (6) into two groups: one group that contains sets that correspond to some attested reading of a pronoun and another containing sets that do not. These two groups are given in (7). (7) a. Sets that correspond to attested readings {i, ai, u, au, o}, {i, ai, u, au}, {i, ai, u, o}, {i, u, au, o},{i, ai, u}, {i, u, au}, {i, u, o}, {i, ai}, {i, u}, {u, au},{i},{u},{o} b. Sets that do not correspond to attested readings {i, ai, au, o}, {ai, u, au, o}, {i, ai, au}, {i, ai, o}, {i, au, o}, {ai, u, au}, {ai, u, o}, {ai, au, o}, {u, au, o}, {i, au}, {i, o}, {ai, u}, {ai, au}, {ai, o}, {u, o}, {au, o}, {ai}, {au} The following two generalizations describe when a set expresses an attested reading. (i) No such set may contain associates of α but not α itself. (ii) No such set may contain a nonassociate other (o) and exactly one of i and u. The first generalization would be falsified if a language contained, for example, a pronoun that refers to the speaker and the associates of the addressee (i and au+), but not the addressee (u). Consider again the situation of Peter and Ad attending a football match. Suppose that Vitesse (Peter’s team) has scored; then Ad might want to express that he and Peter’s associates are happy (Ad is happy for Peter, and Peter’s associates are happy that Vitesse finally scored). However, there is no pronoun that can refer to this group. If Ad uses we, he refers to himself and his own associates; if he uses you, he refers to Peter and his associates; if he uses they, he excludes himself. To the best of our knowledge, no pronoun of the type described has ever been identified. The second generalization can be illustrated by considering the set {i, ai, o}. At first sight, one might think that this corresponds to the exclusive reading of the first person plural. However, this cannot be correct. Consider once more the situation of the football match, and in particular We scored as uttered by Peter when Vitesse scores. Clearly, the use of the pronoun in this example is

8

Chapter 1

exclusive: it refers to the speaker and his associates, but crucially not to the addressee and his associates. On reflection, it will be clear that the reference of the pronoun in this context must also exclude any third parties, such as the officials (unless the speaker thinks the referee is a closet Vitesse supporter); hence, no o may be included in the reference of this pronoun. The two generalizations (i) and (ii) separate the attested and unattested readings of pronouns. This is not enough to derive the inventory of possible persons, however. A further question to be addressed is whether a given reading can ever have a separate form in a pronominal paradigm. If this is not the case, it is unlikely that that reading constitutes a distinct person. Our assessment of the typological situation is that the attested readings are distributed over three persons with a singular interpretation, four persons with a plural interpretation, and a subsystem of impersonal pronouns. The distribution of readings across forms is given in (8). (8) First person singular: Second person singular: Third person singular: First person plural inclusive: First person plural exclusive: Second person plural: Third person plural: Impersonal:

{i} {u} {o} {i, ai, u, au}, {i, ai, u}, {i, u, au}, {i, u} {i, ai} {u, au} {o, o+} {i, ai, u, au, o+}, {i, ai, u, o+}, {i, u, au, o+}, {i, u, o+}

There is one clear generalization that captures the ambiguity of the first person plural inclusive and impersonal pronouns, namely, that pluralities containing participants but not their associates are never formally distinguished from pluralities containing participants as well as their associates (compare Bobaljik 2008a). The hypothesis that the associative effect is inherent in the person system and not an effect of the number system does not imply that no interaction between person and number exists. In fact, it is well-known that crosslinguistically there are two distinct number systems that differ in how the first person inclusive behaves in case it refers to a group of exactly two (speaker and addressee but no associates). In some languages, such a pronoun looks like a plural pronoun, while in other languages, it appears to be singular. The person system must therefore be paired with a number system that can explain this variation. To summarize, any theory of person features has to account for the generalizations in (9) about possible readings and their expression. We will develop such a theory, based on the structure in (5), in chapters 2–4.

Person 9

(9) a. There is no person that contains associates of a participant but not the participant itself. b. There is no person that consists of others and only speaker(s) or of others and only addressee(s). c. No person system distinguishes pluralities containing multiple participants but not their associates from pluralities containing one or more participants as well as their associates. d. First and second person plural pronouns show an associative effect: any element contained in them other than i or u must be ai or au and cannot be o. As already suggested, impersonal pronouns form their own subsystem. The main issue here concerns the possible readings that such pronouns can have. Roughly speaking, we can distinguish two main interpretations, a generic one and an arbitrary one. The former is illustrated by the pronoun you in an example like You should not eat with your hands. Here, the interpretation of you can be paraphrased as ‘people in general’. The latter reading is illustrated by the pronoun they in an example like They said on the radio that the prime minister is on holiday again. The interpretation of they here can be paraphrased as ‘someone or other’. Not every impersonal pronoun permits both a generic and an arbitrary reading. Impersonal you, for example, does not permit an arbitrary reading. The question is why. More generally, a theory of person must explain how the generic and arbitrary readings distribute over particular forms, both within a language and across languages. Related to this, it must explain why some impersonal pronouns share their form with a referential pronoun (such as you), while others do not (such as one). There are a number of generalizations to be captured. In particular, we will show that the following hold: (10) a. Singular pronouns do not permit generic or arbitrary readings, with two exceptions: second person singular pronouns permit generic readings, and dedicated impersonal pronouns are singular. b. Impersonally used second person singular pronouns differ from dedicated impersonal pronouns in two ways: (i) second person singular pronouns require a generic reading, while dedicated impersonal pronouns come in two types, one of which allows an additional arbitrary reading; (ii) dedicated impersonal pronouns systematically trigger third person agreement. c. Plural pronouns can all be generic, but an additional arbitrary reading is available only for the third person. These generalizations are the main topic of chapter 4.

10

Chapter 1

1.3  Agreement for Person

It is generally assumed that φ-features can only be interpreted when located in nominal elements. In order to understand this restriction, we must know what the unique semantic contribution is that nominal elements make. This issue is discussed by Geach (1962), Gupta (1980), Larson and Segal (1995), and Baker (2003). These authors suggest that all lexical categories have a criterion of application. This means that lexical categories define a criterion that determines which elements are part of the set they denote and which elements are excluded. For example, man stands for a set whose members are male humans, tall stands for a set whose members are tall, and walk stands for a set whose members walk. What is unique to nominal categories is that they have an additional criterion of identity, which allows access to individual members of the set that the lexical category refers to. This unique property of nominal categories makes it possible, for example, to determine whether one member of the set is identical to another, something that is not possible for nonnominal categories. As a result, it is possible to say things like That is the same man as he saw yesterday, while there is no similar nonnominal construction (see Baker 2003:101ff. for discussion). The criterion of identity is a precondition for quantification. It is not possible to talk about quantities unless one has access to individual members of a set. If the criterion of identity is a defining property of nominal categories, it follows that only nominal categories permit quantification, as noted already in Geach 1962 and Gupta 1980. Likewise, the interpretation of φ-features is such that it requires access to individuals. For example, number is like quantification in that it requires countability (which is not possible without individuation). Person, too, relies on identification of distinct members of the set—namely, as being the speaker or the addressee or neither. Again, this requires access to individuals. The implication is that φ-features must be interpreted in nominal categories. In view of this conclusion, it is a remarkable fact that φ-features can be expressed by categories other than nouns. It is very common, for example, for a verb to agree with a particular nominal in person and number features. This fact raises various questions. A central one is whether the semantic asymmetry between φ-features on nominal categories and φ-features on other categories translates into a parallel syntactic asymmetry. It is quite a common view that this is the case. In the probe-goal system proposed by Chomsky (2001), for example, “uninterpretable” (that is, nonnominal) φ-features are valued by “interpretable” (nominal) φ-features in syntax. The syntactic presence of nonnominal φ-features is therefore licensed only in the presence of nominal

Person 11

φ-features, while the presence of the nominal φ-features themselves need not be syntactically licensed. The same asymmetry is present even more clearly in the theory of Bobaljik (2008b), who suggests that agreement consists of the copying of nominal φ-features onto the agreeing head. Thus, φ-features on verbs or adjectives cannot exist in the absence of their original on a nominal category. Although the semantic asymmetry is well-motivated and seems to be uncontroversial, there is no logical necessity for it to translate into a syntactic asymmetry. Indeed, there is an alternative view according to which agreement is established by a form of unification. The idea is that the φ-feature set of the nominal element is unified with the φ-feature set of the nonnominal category, with the resulting set bearing on the interpretation of the nominal category only. On this view, the presence of φ-features on a verb or an adjective does not depend on the presence of identical features on a nominal category (although they can only be interpreted if in the end these features are associated with a nominal category). Unification-based theories are proposed in Shieber 1986 and Barlow 1992, and in models such as Head-Driven Phrase Structure Grammar (Pollard and Sag 1994) and Lexical-Functional Grammar (Bresnan 2000). There is an important difference between the asymmetric and symmetric views of syntactic agreement. In both approaches, situations in which the nominal category has the same φ-features as or more φ-features than the verbal or adjectival one (as in (11a,b)) are permitted. However, only under the symmetric theory is it possible for the verbal or adjectival category to have more φ-features in syntax than the nominal one (as in (11c)). The additional φ-features will be interpreted in the nominal category after unification.3 (11) a. [DP F1 F2] … [V/A F1 F2] b. [DP F1 F2] … [V/A F1] c. [DP F1] … [V/A F1 F2] There are some cases of agreement that, at least at face value, seem to instantiate the distribution of features in (11c). Pollard and Sag (1994:62) mention gender marking on adjectives in French in the context of a first person singular subject. Although first person pronouns in French do not show any gender distinction, adjectives can be inflected for gender in their presence. The resulting interpretation is such that the speaker is identified as male or female.4 (12) a. Je suis heureux. I am happy.masc ‘I am happy.’

(spoken by a man)

12

Chapter 1

b. Je suis heureuse. I am happy.fem ‘I am happy.’

(spoken by a woman)

Such data can be made compatible with the asymmetric view on syntactic agreement if the pronoun can be argued to carry a gender feature that has no phonological reflex. Pollard and Sag argue that this is an unattractive assumption, because it requires unwarranted redundancy in the French lexicon: one must assume that the language has two first person singular pronouns, one masculine and the other feminine, which happen to be homophonous. Since this point holds not just for this particular case, but for numerous comparable cases in a variety of languages, the result would be massive redundancy. Notice that this is avoided if the agreement relation is one of unification, as unification allows “transfer” of the feature of the predicate to the subject. Thus, French need only have one first person singular pronoun, which is unspecified for gender. Although persuasive at first sight, this argument is only valid under a view of morphology in which lexical items are fixed bundles of phonological, syntactic, and semantic features. Such a view has been challenged by a number of researchers, who argue for a separation of the phonological properties of an overt morpheme on the one hand, and the formal features present in a syntactic position on the other. The core idea is that overt forms are associated with the syntax through correspondence rules that refer to syntactic features, but that do not need to stand in a one-to-one relationship with any specific feature bundle. An example of a framework based on this idea is Distributed Morphology (Halle and Marantz 1993), which implements it through “late lexical insertion,” but there are other varieties of the general approach as well (see, for instance, Ackema and Don 1992, Anderson 1992, Beard 1995, and Ackema and Neeleman 2004). How does the existence of such theories affect the analysis of data like (12)? In any “separationist” model, it is possible to formulate a single correspondence rule that associates first person singular with the overt form je in French. This rule simply does not mention gender, and hence can be applied to syntactic feature sets that are marked as first person singular, but vary in gender features. (Alternatively, the gender feature may be deleted before spell-out as the result of a rule of impoverishment.) Therefore, it is not a coincidence that the same form, introduced by the rule in (13), is used to spell out the syntactic inputs in (14a) and (14b). (For now, we will simply describe the relevant person and gender features in traditional terms. In the interest of readability, we will use the usual orthographic representation of overt lexical items here and below, rather than phonetically faithful ones; the slashes at the right-hand

Person 13

side in (13) are simply meant to indicate that we are dealing with an overt form inserted in the PF component of grammar.) (13) {1 sg} ⇔ /je/ (14) a. [1 sg masc] b. [1 sg fem] One could refer to an analysis of this type as a “hidden-feature” analysis: there is a feature present in syntax that does not find an expression in the overt realization of the relevant terminal node. Thus, the separationist view of morphology appears to undermine any argument based on lexical redundancy against an asymmetric view of agreement (and hence in favor of unification). However, it is possible to make a very similar argument that is valid within a separationist model of morphology in case the feature not mentioned by the spell-out rule is not overtly realized on the relevant category in any language. Suppose, for example, that not just French, but every single natural language, fails to spell out gender on first person pronouns. This situation could only be captured under the asymmetric theory of agreement if it happens to be the case that no language has a spell-out rule that mentions both first person and gender. However, there is nothing in the theory that prevents the existence of such a spell-out rule, so we would be left with a conspiracy theory of the data. In the case of gender in first person pronouns, it is likely that an argument of this form cannot be made, simply because there do seem to be languages that express gender in first person pronouns (see Corbett 1991). Examples are Thai (Cooke 1968), Ngala (Laycock 1965, Siewierska 2011), and Korana (Meinhof 1930, Siewierska 2011). However, in chapter 5 we will argue that there is at least one case in which the controller of agreement really must have fewer features than the target (see also Ackema and Neeleman 2013b). A hidden-feature analysis for this case fails because the relevant features universally fail to be realized on the controller. In chapter 5, we will therefore conclude that agreement must be modeled as a form of unification. 1.4  Syntactic Restrictions on Agreement

If there is no inherent syntactic asymmetry between target and controller, as implied by the assumption that agreement is a form of unification, then a certain degree of flexibility is introduced into the syntax: in principle, the target of agreement could be located structurally higher or lower than the controller. In chapter 6, we explore whether this flexibility is in fact attested. An initial assessment must be mixed. There are apparent cases of upward agreement as

14

Chapter 1

well as apparent cases of downward agreement, as expected. However, in any individual case of agreement the relationship must be either upward or downward; in this respect, there does not seem to be flexibility. Thus, in general Universal Grammar (UG) seems to allow for variation in the direction of agreement, but in specific cases a choice of direction is forced. In general, this may be understood as follows. Unification in itself is not syntactically asymmetric. Yet, every instance of agreement constitutes a syntactic dependency and should therefore fit into a general theory of such dependencies. One hallmark of any syntactic dependency is a distinction between the dependent category and its antecedent. For example, in predication, the dependent is the predicate and the antecedent is the subject; in binding, the dependent is the anaphoric category and the antecedent is (unsurprisingly) the antecedent; in movement, the dependent is the trace and the antecedent is the moved category. A number of generalizations over syntactic dependencies must be formulated in terms of the distinction between antecedent and dependent. Koster (1987) refers to this set of generalizations as the “configurational matrix.” Perhaps the best-known aspect of the configurational matrix is that the dependent must be c-commanded by the antecedent. Thus, subjects c-command their predicates, antecedents c-command anaphors, and moved categories c-command their traces. Hence, if agreement relations are to fit into the configurational matrix, a distinction between the two elements in the relation must be made such that one counts as the antecedent and the other as the dependent. This is where the symmetric nature of unification comes in. If there is no inherent syntactic asymmetry between target and controller, then, in principle, the grammar has the choice in any specific instance of agreement to identify either the target or the controller as the dependent (and the other as the antecedent). This suggestion may resolve a theoretical stalemate. Some authors argue that agreement is uniformly upward (see, for instance, Wurmbrand 2012 and Zeijlstra 2012); others argue that it is uniformly downward (see, for instance, Chomsky 2000 and Preminger 2013). Each provides convincing cases in support. We can now take these cases at face value, without allowing optionality in specific instances of agreement. In sum, a symmetric theory of agreement predicts that there is no inherent direction in this relationship. However, for each of type of agreement in each language a direction must be fixed: either the controller must invariably c-command the target, or vice versa. There is no optionality in individual cases. This is a highly testable theory, and part of the work we will do in chapter 6 is to confront it with the data.

Person 15

A second prediction is that, where a specific type of agreement seems to have a universal direction, this must be due to an identifiable independent factor. An example may be person agreement between a complementizer and a subject (see, for instance, Bennis and Haegeman 1984). This seems to be systematically downward (an apparent counterexample is analyzed in chapter 6). If this kind of universal asymmetry cannot be reduced to a universal asymmetry in the agreement relation itself, it must be the result of the interaction between the theory of agreement proper and an external factor. Exploring cases of this type is the second aim of chapter 6. 1.5  The Realization of Person Features

In the previous sections we have traversed the grammar, starting with the challenges the person system poses for the interpretation and inventory of persons and continuing with an exploration of some syntactic issues associated with person agreement. This leaves the realization of person features at the interface between syntax and phonology. One could have the impression that this is simply a matter of finding the right set of spell-out rules for individual languages. While this is true, there is a paradoxical situation that needs to be addressed, as we will now explain. There seems to be a lot of variation in the way person is realized crosslinguistically. Even within a single language, there are shifts in syncretism depending on various factors, such as the context in which an agreeing verb finds itself, whether the agreeing verb belongs to a particular lexical class or not, and others. Dutch, for example, has two verbal person agreement endings in the singular, -t and ∅. Depending on the kinds of factors just mentioned, these arrange themselves into a regular 2-3 syncretism (∅ for first person, -t for second and third person), a 1-2 syncretism (∅-∅-t) under subject-verb inversion, or a 1-3 syncretism (∅-t-∅) with modal verbs. In fact, even a 1-2-3 syncretism occurs (∅-∅-∅), as an alternative option with modal verbs. If this kind of variation is attested even within a single language, the spellout system must be quite powerful. Nevertheless, and this is what is paradoxical, there are reasons to think that the spell-out system must be constrained quite seriously. If there were no restrictions, one would expect that anything goes, meaning that there should be random variation in the patterns of syncretism that are attested crosslinguistically. But this is not the case. There are a number of generalizations concerning the realization of person. A very striking one is the following crosslinguistic regularity: (15)  1-2 and 2-3 syncretisms are far more common than 1-3 syncretisms.

16

Chapter 1

Evidence for this generalization comes from work by Baerman and colleagues (see Baerman, Brown, and Corbett 2005:59 and Baerman and Brown 2011). It should be clear that a theory of person that simply adopts “1,” “2,” and “3” as atomic features will find it hard, if not impossible, to capture the generalization in (15) in a principled way. What is required is a theory of φ-features in which each person is characterized by a unique feature specification, while a particular feature may nonetheless be shared by two persons. The basis for an explanation of (15) could then be that second person shares a feature with both first and third person, while first and third person do not share any feature. Thus, the feature system should be of the general form in (16) (compare Kerstens 1993 and Halle 1997). Of course, as mentioned in sections 1.1 and 1.2, the same system must account for the crosslinguistic inventory of possible persons and their semantics, as expressed by the generalizations in (9). Indeed, the feature system we will develop in chapter 2 has the general shape of (16). (16) 1st person [F1]

2nd person

3rd person

[F1 F2]

[F2]

Although it will form the basis for an explanation of (15), the architecture of the syntactic person feature system is not, in itself, enough to account for this generalization. After all, the generalization concerns overt forms; hence, the spell-out system that operates on the syntactic input must play a role in any account of syncretism. Moreover, (15) does not state that there are no 1-3 syncretisms, so there must be ways in which the spell-out system can associate first person and third person with the same phonological form, leaving a distinct form for second person. Nonetheless, there must be something about such an association that leads to a reduction in its crosslinguistic distribution. The aim of chapter 7 is to develop a theory of the realization of person features that, on the one hand, captures crosslinguistic generalizations like (15) and, on the other, allows an insightful analysis of language-internal patterns of the type found in Dutch. 1.6  Maximal Encoding

So far, we have mentioned the issues that the category of person gives rise to in various modules of grammar. However, we have not discussed interface issues that may bear on the grammar of person. There are two relevant interfaces, between syntax and phonology, and between syntax and semantics. Using common terminology, we will refer to these as PF and LF, respectively.

Person 17

A well-known constraint that applies at the PF interface is the Elsewhere Principle (Kiparsky 1973), which regulates mappings between morphosyntactic information and its morphophonological spell-out. (Different frameworks incorporate this principle in different ways; examples are the Subset Principle of Distributed Morphology (see Halle and Marantz 1993) and the Superset Principle of Nanosyntax (see, for instance, Caha 2009).) In general, the Elsewhere Principle says that if a more specific rule can be applied, this blocks the application of a less specific rule. A rule counts as more specific if its context of application is a proper subset of the context of application of the less specific rule. Of interest here are the consequences of the Elsewhere Principle for spell-out. Suppose there are three spell-out rules that could apply to a syntactic feature bundle [F1 F2], as in (17). In that case, the Elsewhere Principle requires that the spell-out rule with the richest structural description—(17c)—be applied, because if a spell-out rule mentions a superset of features, its context of application will be a subset of the contexts of application of competing rules. Application of either (17a) or (17b) is blocked, because they leave one feature unencoded that (17c) does encode. (17) a. {F1} ⇔ /a/ b. {F2} ⇔ /b/ c. {F1 F2} ⇔ /c/ In the absence of a rule with a richer structural description, the Elsewhere Principle permits application of a rule mentioning fewer features. For the abstract example just given, this implies that absence of (17c) would make it possible to apply either (17a) or (17b). Note that, in this kind of situation, the Elsewhere Principle does not decide which of these two rules will be applied. We will discuss situations of this type in chapter 7. Turning now to the LF interface, we assume that here, too, the association of syntactic structures and semantic interpretations is partly regulated through competition. This competition is based on the idea that syntax can be considered an encoding of semantics (see Bobaljik and Wurmbrand 2012; see also Hendriks and De Hoop 2001 and other literature on Optimality Theory semantics). At work is a principle, analogous to the Elsewhere Principle, which requires that given a particular semantic representation, a feature structure is licensed in syntax only if there is no alternative feature structure that encodes the relevant semantics more explicitly. The easiest way to illustrate this is to consider a series of syntactic representations that can all be associated with the same interpretation: {, , …, }. Competition has the effect that given that interpretation, the syntactic representation chosen must be the one with the richest feature specification among {R1, R2, …, Rn}. For example, if a noun does not carry a feature pl (for plural) in a language that

18

Chapter 1

does make this feature available, it will be interpreted as singular. By contrast, the interpretive effect of the absence of a feature is different in case there is no competing structure that includes the feature in question. Hence, in languages that lack all means of marking plurality in nouns (see Haspelmath 2011) or even pronouns (see Daniel 2013), a nominal that does not carry a pl feature may nevertheless receive a plural reference, as there is no more explicit syntactic representative for that interpretation.5 Given the evident parallel between how the Elsewhere Principle operates at the PF interface and how competition works at the LF interface, we will adopt a unified principle of competition, operating at both interfaces. It regulates the mapping between any two levels of representation, as follows: (18)  Maximal Encoding A mapping R → R* is licit only if R* is the maximal expression of R at the relevant level of representation. We should be explicit about the conditions under which R* counts as the maximal expression of R, and we therefore adopt the definition in (19). (19)  R* expresses R maximally if there is no alternative R′ that encodes more properties of R or encodes these properties in more locations. In the mapping between syntax and phonology, the morphosyntax functions as R in (19) and the morphophonology as R*. This is presumably uncontroversial. In the mapping between syntax and semantics, however, the semantics functions as R and the syntax as R*. This seems to imply a directionality of mapping that runs counter to the directionality assumed in the standard Y-model of generative grammar, where derivations start with selection from the lexicon and end with representations presented to the interfaces with semantics and phonology (see Chomsky 1965, 1981, 1995 for various incarnations of this model). This incompatibility only holds if grammar operates in a unidirectional fashion. We do not think that this is the case. There are alternative models of grammar that have no directionality built into them. In Ackema and Neeleman 2004, for example, an argument is made for a “parallel architecture” in the sense of Jackendoff 1997. In a model along these lines, the semantic, syntactic, and phonological modules are independent generative systems, whose outputs are associated through a system of mapping rules: (20) Semantics ↔ LF ↔ Syntax ↔ PF ↔ Phonology However, the nondirectionality of the model as a whole does not imply that the principles that define well-formed mappings cannot be asymmetric in that

Person 19

they take one representation as a given and then impose a particular restriction on a second representation. Rather, we expect both mapping rules that operate “from left to right” in the model in (20) and mapping rules that operate “from right to left.” A clear example of a principle that operates from right to left is compositionality. After all, compositionality takes the LF representation as given and maps it onto semantics in such a way that the interpretation of a syntactic category is restricted by the material it contains and the way this material is combined. An example of a principle that operates from left to right is the Minimal Link Condition as interpreted by Fanselow (2004) and others. The starting point of Fanselow’s discussion is the observation that Superiority violations are tolerated when they result in a meaning that cannot be expressed by alternative syntactic structures (see also Golan 1993, Kitahara 1994, Sternefeld 1997, and Reinhart 2006). Fanselow concludes from this that the Minimal Link Condition must be a principle that is sensitive to interpretation. Given a particular semantics, a syntactic representation must be chosen that has the shortest possible movement links. Maximal Encoding in (18) is simply another principle that operates from left to right. 1.7 Summary

In sum, this is our to-do list: • Explain why the crosslinguistic inventory of persons is the way it is (in chapter 2). •  Explain the interactions between person and number (in chapter 3). • Explain the possible readings and forms of impersonal pronouns (in chapter 4). •  Find out whether person agreement should be dealt with through a syntactically symmetric or asymmetric theory (in chapter 5). •  Integrate the theory of syntactic agreement with the general theory of syntactic dependencies, and use the result to account for syntactic restrictions on agreement relations (in chapter 6). • Account for crosslinguistic patterns of syncretism and language-internal variation in the realization of person with one and the same theory of person features and spell-out (in chapter 7).

2

Person Features: Deriving the Inventory of Persons

2.1 Introduction

One long-standing aim of research into person has been to achieve a deeper understanding of the inventory of personal pronouns by decomposing them into a limited set of person features. The main aim of this chapter is to make a contribution to this enterprise. We will argue that two privative person features suffice to derive the full inventory of attested pronouns and their interpretations, without generating nonattested pronouns. In chapter 1, we showed that the crosslinguistic inventory of attested persons is a small subset of the set of theoretically possible persons. The theoretically possible persons are those that can be formed by freely combining the speaker (for which we use the symbol i), associates of the speaker (ai), the addressee (u), associates of the addressee (au), and others (o). The generalizations that distinguish between the attested and nonattested persons are the following (repeated from (9) in chapter 1): (1) a. There is no person that contains associates of a participant but not the participant itself. b. There is no person that consists of others and only speaker(s) or of others and only addressee(s). c. No person system distinguishes pluralities containing multiple participants but not their associates from pluralities containing one or more participants as well as their associates. d. First and second person plural pronouns show an associative effect: any element contained in them other than i or u must be ai or au and cannot be o. This leaves us with three persons with a singular interpretation and four persons with a plural interpretation. Any theory of person features and their semantics must explain why this, rather than some other subset of all logical possibilities, is the inventory of attested persons.

22

Chapter 2

There is limited room for maneuvering, as a number of observations about person should be derivable from the proposed system. To mention just one, third person has an apparent “default” status. Thus, third person pronouns, as opposed to first and second person ones, can be used in situations where they do not seem to have any referent. Examples are expletive pronouns that can appear in subject position in certain languages so as to avoid an EPP violation, or pronouns that appear with “weather” verbs. (2) a. It seems that Mary has left for Paris. b. It is raining again in Edinburgh. Similarly, some languages allow use of third person verb forms in the absence of an agreeing subject. Observations of this type have led to the assumption that the third person in fact has no person features (see Cysouw 2003 and references cited there), in contrast to what we will argue. We are therefore committed to show how the system to be developed can explain this behavior of the third person. In section 2.2, we outline our basic proposal. We will define two privative person features that operate on a person space with a universal structure, and show how these features generate exactly three singular persons and four plural ones. The rest of the chapter is taken up by a discussion of how this feature system sheds light on the workings of the person system in various domains. Issues to be addressed include the use of third person as a default, the typology of honorific pronouns, the lack of person features in R-expressions, the location of person within the extended pronominal projection, and the use of pronouns as bound variables. 2.2  The Basic Proposal

As noted in chapter 1, verbal agreement frequently shows 1-2 and 2-3 syncretisms, while 1-3 syncretisms are relatively rare. This pattern suggests that second person has a feature in common with both first and third person, while first and third person share no features (see (3)). (3) 1st person [F1]

2nd person

3rd person

[F1 F2]

[F2]

Proposals that assume this distribution of features can be found in Kerstens 1993 and Halle 1997. Kerstens uses binary features (the features that correspond to F1 and F2 in (3) are underlined): first person is characterized as [+utterance +speaker], second person as [+utterance –speaker], and

Person Features 23

third person as [–utterance –speaker]. The fourth logical possibility, [–utterance +speaker], is ruled out as contradictory. Halle proposes a similar system, built on the features ±participant (in the speech event) and ±author.1 The system we propose is similar, but based on privative features. The use of privative features is familiar from a number of publications, including Harley and Ritter 2002 and McGinnis 2005, and was discussed in phonology as early as Trubetzkoy 1939. The features we will employ are prox for proximate and dist for distal. The names for these features are of course not crucial to the analysis below, but are chosen to reflect the intuition that there is a parallel between the person system and the system of demonstratives (for relevant discussion, see Diessel 1999, Lander and Haegeman to appear, and references mentioned there; see also note 10 below). We interpret these features as functions, following insights in Harbour 2011a,b,c, 2016. Both operate on an input set to deliver a subset as output. The initial input set for the person system represents all potential referents in a given context (Si+u+o in (4)). The input set has a fixed structure, already introduced informally in chapter 1. In particular, it contains a subset Si+u, which itself contains a subset Si. Si has the speaker (i) as an obligatory member; its other members, if there are any, are associates of the speaker and/ or further individuals identified as speaker.2 Si+u has one addressee (u) as an obligatory member, in addition to all members of Si; its other members, if there are any, are associates of the addressee and/or further individuals addressed by the speaker.3 Si+u+o contains all members of Si+u; its remaining members, if there are any, are neither associates of the speaker nor associates of the addressee. (The idea that the input to the person system is a set of nested structures on which person features act is borrowed from Harbour 2011a. Note, however, that we treat these nested structures as sets of atoms, rather than as power sets.) (4)  Si

i u

Si+u Si+u+o

We assume that there is a feature Π that encodes Si+u+o, and a dedicated category NΠ that bears this feature. By definition, pronouns are projected from NΠ. The person features are introduced in a separate node that we will label

24

Chapter 2

prs. In section 2.7, we will discuss the status of the prs node in more detail, but for now it is sufficient to assume that it combines with NΠ and delivers Si+u+o to the person features that it is specified for. We now define these person features. The feature prox introduces a function that operates on an input set and discards its outermost “layer.” Applied to Si+u+o, it delivers Si+u. By contrast, dist introduces a function that selects the outermost layer of its input set. Applied to Si+u+o, it delivers Si+u+o − Si+u. We assume that the various sets in (4) are ordered such that Si is the predecessor of Si+u, while Si+u is the predecessor of Si+u+o (we will abbreviate predecessor as Pred). (5) a. Pred(Si+u) = Si b. Pred(Si+u+o) = Si+u Given (5), characterization of prox and dist is simple. The definitions in (6) express that prox discards, while dist selects, those elements that are part of the outermost layer of the input set. (6) a. prox(S) = Pred(S) b. dist(S) = S − Pred(S) We now consider how first, second, and third person readings are derived, starting with the singular (which for now we take to impose a requirement that the output set cannot contain more than one member; for more discussion, see below and chapter 3, where singular inclusive forms are analyzed). The specification of the third person singular is straightforward: it should be dist, as this feature will derive Si+u+o − Si+u, a set that excludes the speaker and any addressees. A second person singular reading can be generated by applying both prox and dist. Notice that there is only one order of application that yields an interpretation. If prox is applied first, Si+u is selected, a set containing the speaker (and any of the speaker’s associates) and individuals that the speaker addresses (and any of their associates). Applying dist to this set removes Si, leaving only any addressees (and any associates) as potential members—the required result (see (7)). In the singular, this will yield a pronoun that refers to exactly one addressee. (7) [prs prox–dist] = dist(prox(prs)) = dist(prox(Si+u+o)) = dist(Si+u) = Si+u − Si = Su

(by definition) (by (6a)) (by (6b))

Person Features 25

The opposite order of function application is not coherent. Dist applied to Si+u+o yields Si+u+o − Si+u (a set that includes neither the speaker nor any addressees). But this set is not layered (that is, Pred(S) is not defined for this set). Therefore, prox cannot apply to it. We assume that the order of function application is reflected in syntax. The notation we use to represent this is borrowed from theories according to which φ-features occur in a so-called feature geometry (see Gazdar and Pullum 1982 and Harley and Ritter 2002, among others): features representing functions applied later are dominated by features representing functions applied earlier. Thus, our representation of the second person singular is as in (9b) below. Here, prox is the first function to be applied, and dist represents the function applied subsequently.4 In what follows, we will need to refer to the relative relationships of prs and the various features. To this end, we will use the terms host and dependent. Thus, in the second person, prs is the host of prox and prox is a dependent of prs. Similarly, prox and dist stand in a host-dependent relationship in the second person. Consider finally the first person. Notice that in the singular, just applying prox to Si+u+o will not do. This is because the output it delivers, Si+u, is a set with two obligatory members: the speaker and an addressee. Such a set obviously cannot be construed as (semantically) singular.5 Therefore, at least in the singular, a first person reading requires that prox be applied to the output of prox. As prox discards the outermost layer of its input set, this will deliver Si, a set whose only obligatory member is the speaker and which therefore can be interpreted as singular. (8) [prs prox–prox] = prox(prox(prs)) = prox(prox(Si+u+o)) = prox(Si+u) = Si

(by definition) (by (6a)) (by (6a))

The feature structures for singular persons are summarized in (9). We use boxes to make clear that the person features are contained in the prs node. (9) 

a. 1st person PRS PRS Singular

b. 2nd person

c. 3rd person

PRS NΠ

PRS

PROX

PROX

PROX

DIST



PRS PRS DIST



26

Chapter 2

We now turn to plural pronouns. We take number to be encoded through a nmb node, merged after prs (see subsection 2.7.3). This node can contain a feature that marks plurality. For now, we will label this feature pl and assume that it requires that the cardinality of the output set of the person system be larger than one. (For a more detailed discussion, including a refinement of this proposal, see chapter 3.) Singular is the absence of the pl feature in nmb. There is, indeed, evidence that plural is marked and singular unmarked. First, Greenberg (1963:94) observes that “there is no language in which the plural does not have some nonzero allomorph whereas there are languages in which the singular is expressed only by zero.” Second, plural is both a target for morphological impoverishment rules and a context that triggers such rules. This behavior is typical of marked features (see Aalberse and Don 2009, 2011 and Nevins 2011). As in the singular, the interpretation of third person plural forms is straightforward. Dist delivers a set that contains neither speaker nor addressees; adding pl simply adds the requirement that the cardinality of this set be two or more. For second person plural forms, our proposal implies that the set selected contains at least one addressee and one or more other members. The latter can be additional addressees and/or associates of the addressee (compare also Daniel 2005). First person plural pronouns are more complex, as here languages can make an inclusive-exclusive distinction. Inclusive first person pronouns refer to a group containing one or more speakers and one or more addressees (as well as possible associates). Exclusive first person pronouns disallow reference to any addressee (or their associates). That a language may have distinct pronouns expressing these readings follows without further stipulations. Suppose that the plural feature is simply added to the singular first person form, in which prox applies twice (see above). Repeated application of prox delivers Si, a set containing the speaker and any contextually given associates of the speaker, but no addressees. Pl indicates that the cardinality of this set is two or more, so the pronoun in this case receives an exclusive reading: it refers to one or more speakers and possible associates, but excludes reference to any addressees. However, this is not the only possibility. Recall that in singular first person pronouns prox has to apply twice because the output after one application contains two obligatory members, the speaker and an addressee, which is incompatible with a singular interpretation. This consideration is obviously irrelevant in the plural. If prox applies only once, its output will be Si+u. In

Person Features 27

combination with pl, this allows for a range of inclusive readings. Reference can be to any group that contains the speaker and at least one addressee; in addition, more speakers (if these exist; see note 2), addressees, or associates of either can be present. These are indeed the remaining readings that a first person plural pronoun can have. Thus, the first person exclusive plural has the same person feature structure as the first person singular, whereas the first person inclusive plural has a person feature structure distinct from this. This chimes well with an observation by Cysouw (2005a:74), according to whom “the exclusive is regularly homophonous with the first person singular [while] the inclusive is hardly found to be homophonous with the first person singular.” (For related discussion, see Moskal 2014 and subsection 7.3.6.) We thus arrive at the following inventory of plural forms: (10) 

a. 1st person inclusive

b. 1st person exclusive

NMB PRS

NMB Plural PL

NMB

PRS

PRS

NMB NΠ

PL

PRS



PROX

PROX

PROX c. 2nd person

d. 3rd person

NMB

NMB PRS

NMB Plural PL

PRS PROX

PRS

NMB NΠ

PL

PRS



DIST

DIST The introduction of the number node implies that in the singular structures, too, prs must be analyzed as being combined with nmb. The only difference with the plural structures is that in this case this node does not contain the feature pl.

28

Chapter 2

(11) 

a. 1st person

Singular

b. 2nd person

c. 3rd person

NMB

NMB

NMB

NMB PRS

NMB PRS

NMB PRS

PRS



PRS

PROX

PROX

PROX

DIST



PRS



DIST

In the system just outlined, the first person (inclusive or exclusive) does not form a natural class with the third person to the exclusion of the second person. Similarly, the first person inclusive does not form a natural class with the second person to the exclusion of the first person exclusive. This is relevant in view of the results of a large-scale study reported in Harbour 2016. Harbour looked at which systematic patterns of syncretism are attested crosslinguistically, where a systematic pattern of syncretism is a syncretism characteristic of all paradigms of a given language. He found that no language had a systematic syncretism for first and third person, or for first person inclusive and second person. On the assumption that the distribution of systematic syncretisms reflects the underlying distribution of features, this shows that no set of features is shared uniquely by the relevant combinations of persons. The absence of systematic syncretisms of first and third person does not imply that there are no grammars that deliver such syncretisms in particular paradigms. In chapter 7, we discuss what kinds of morphological rules are necessary to generate such a syncretism and we show that grammars that contain such rules are relatively hard to acquire, accounting for the fact that such syncretisms are relatively rare. The absence of systematic syncretisms for first person inclusive and second person is in line with a typological generalization discussed by Zwicky (1977). Zwicky argues that in languages that lack the distinction between inclusive and exclusive first person pronouns, the inclusive reading is systematically expressed by the first person rather than the second person plural pronoun— this despite the fact that the inclusive reading covers both speaker and addressee. An account for this observation would be impossible if first person inclusive and second person did form a natural class to the exclusion of first person exclusive.6 As it stands, the person specifications in (9) and (10) are the only ones that can be generated by the system we propose. Recall that both prox and dist require a layered set as their input (see (4)–(6)). Given that Si+u+o has only three

Person Features 29

layers, the number of possible feature combinations is restricted. (i) dist always delivers an unstructured set as output; hence, neither prox nor dist can be applied after dist has applied. (ii) If applied to Si+u+o, prox delivers a layered set. This leaves open three possibilities: after prox has applied, either (a) prox applies again, or (b) dist applies, or (c) neither prox nor dist applies. The first two options yield an unstructured set and hence make any further application of prox or dist impossible. Option (c) delivers a set with two obligatory members, hence is only feasible in the plural. Consequently, our proposal allows exactly the personal pronouns that are attested and no others. (In chapter 4, we will show that the same person features suffice to analyze impersonal pronouns.) This is not as trivial as it may seem, because there are distinct interpretations of pronouns that do not ever seem to be expressed by a distinct form. In (1c), it was already stated that no person system distinguishes pluralities containing participants but not their associates from pluralities containing participants as well as their associates. This generalization covers two universals defended in detail by Bobaljik (2008a) (his Universals 1 and 2). Bobaljik also defends the validity of a third universal. We give the full set in (12), adapted to our terminology regarding the person space.7 (12) Universal 1: No language distinguishes [i+i] from [i+ai]. Universal 2: No language distinguishes [u+u] from [u+au].8 Universal 3: No language distinguishes among [i+i+u], [i+u+u], and [i+u+ai/u]. Universal 1 states that no language has different forms for reference to a multitude of speakers (if this “choral we” exists in the first place; see note 2) and reference to one speaker and one or more associates. Similarly, by Universal 2, no language has different forms for reference to a multitude of addressees and reference to one addressee and associates. Universal 3 states that no language has different forms for inclusive first person pronouns that vary in their reference according to the number of speakers, addressees, and associates included. Our system allows all these readings, but the readings grouped together in each of the universals are not distinguished in terms of feature content. Rather, they are distinguished in terms of the population of the input set. For example, the readings mentioned in Universal 2 are all specified as [prox–dist]; they only differ in what elements are present in Si+u – Si, other than the one obligatory instance of u. As a consequence, the spell-out system cannot distinguish between the readings mentioned in Universal 2.

30

Chapter 2

An issue not addressed yet is what parameter distinguishes languages that have the inclusive-exclusive distinction from languages that do not. It is inherent in our proposal that this must be a matter of spell-out. All languages allow the inclusive and exclusive readings associated with (10a) and (10b) and must therefore have these feature structures available in their syntax.9 However, not all languages have two spell-out rules for first person plural pronouns, one that mentions a single instance of prox in its input and one that mentions two instances of this feature. If there are two such spell-out rules, the inclusive and exclusive pronouns will have different forms. However, if there is only a spellout rule that mentions a single instance of prox (besides pl), this rule will be applied to both (10a) and (10b). (In general, if no more specific spell-out rule is available for a particular input, a less specified one may be applied, given Maximal Encoding; see also section 2.5.)10,11 2.3  The Interpretation of Person in Discourse

So far, we have not said anything about the relation between Si+u+o and the discourse context. There are three issues to consider here. First, given that the set Si+u+o has mostly optional members (only i and u are obligatorily present), a question presents itself: what factors decide the population of this input set in actual discourse contexts? How much of the discourse context is represented in a given instance of Si+u+o? Second, during a discourse the roles of speaker and addressee can be reassigned and the question is at what points in the linguistic structure this is possible. Third, it is not clear whether the various discourse roles are mutually exclusive as a consequence of the logic of their semantics, or whether the appearance of exclusivity is a consequence of a pragmatic convention that can be overridden in certain circumstances. 2.3.1  Linking Person to Discourse

Let us start with the first issue mentioned above: what determines the population of Si+u+o? We propose that Si+u+o is a purely formal object, used to construct the semantics of person. Its initial population is arbitrary from the point of view of discourse. This is not to say that the output of the person system is not connected to the actual discourse. The output set must be compatible with the context of the discourse in which a pronoun is used, which means that for every element in the output set there must be a corresponding individual in the domain of discourse. For example, use of an exclusive first person plural pronoun (which has a nonsingleton Si as its output set) requires a discourse that makes available associates of the speaker. However, the correspondence goes only one way: the discourse does not require that all individuals in the context be represented in the output set delivered by a given pronoun. For

Person Features 31

example, use of a first person singular pronoun (which implies that the output set Si is a singleton set) is perfectly admissible in a discourse context where the speaker has associates. 2.3.2  The Values of i and u

We now turn to the second issue mentioned at the outset of this section, which concerns the value of the obligatory elements i and u. At which points in the discourse can these variables be assigned a (possibly new) value? An initial guess might be that this happens at the onset of a new utterance. The speaker of this utterance would then be the value for i and its addressee the value for u. There is evidence, however, that this is not correct. In a number of languages, certain matrix verbs allow a shift in the value of i and u in their complement clause, with the consequence that an example like ‘John said that I will leave’ means ‘John said that he (John) will leave’. Here, the embedded i refers to the same individual as the matrix subject, while the matrix i refers to the speaker of the utterance. (For discussion of this phenomenon, see Schlenker 2003, Anand and Nevins 2004, and Sudo 2012, among others.) Some examples from Amharic are given in (13) (these are taken from Schlenker 2003:68; Schlenker attributes (13a) to D. Petros and quotes (13b) from Leslau 1995). (13) a. ǰon ǰəgna nə-ññ yɨl-all John hero be.perf-1sgO 3masc.say-aux.3masc ‘John says that he is a hero.’ b. mɨn amt’-a ɨnd-al-ə-ññ what bring.imper-2masc comp-say.perf-3masc-1sgO al-səmma-hu-mm neg-hear.perf-1sg-neg ‘I didn’t hear what he told me to bring.’ In (13a), the embedded first person subject refers not to the speaker but to the subject of the main verb ‘say’ (the embedded subject is a null pronoun in this example, but we know that it is first person because of the first person agreement on ‘be’). In (13b), the second person subject of the imperative verb ‘bring’ refers not to the addressee but to the object of the matrix verb ‘say’ (which happens to refer to the speaker of the entire utterance). Schlenker shows that these are not cases of direct quotation; rather, they require shifted values for what we call i and u, a shift imposed by certain matrix verbs. Anand and Nevins (2004) show that there is what they call a “shift-together” effect in cases like (13a,b), not just in Amharic but in a range of languages (but see note 13 for possible exceptions). That is to say, if the matrix verb triggers

32

Chapter 2

a shift in the value of i, then all first person pronouns in its complement must be interpreted in accordance with this shift. It is not possible for some first person pronouns to shift while others still refer to the value of i established in the matrix clause. The same is true, mutatis mutandis, for shifts in the value of u, or indeed the value of other indexicals such as certain place and time adverbials. Anand and Nevins further show that besides situations in which all indexicals shift in value, there are situations in which, for example, the first person shifts but the second person does not. Restrictions of this type appear to be governed by the matrix predicate. Similarly, whether a shift is obligatory, optional, or disallowed depends on the choice of matrix predicate. Anand and Nevins argue that the various shifts are implemented via a context-shifting operator, which in effect assigns new values to i, u, and/or time and place variables.12 The shift-together phenomenon can be accounted for if these operators are merged somewhere high in the left periphery of a clause, and there is a locality condition such that all instances of i and u in a clause can only take the local operator as their binder (a minimality effect).13 We adopt Anand and Nevins’s analysis, with one extension. The reason for this extension has to do with the nature of the locality constraint mentioned above. Consider a case in which first person pronouns shift, but second person pronouns do not. On the analysis proposed by Anand and Nevins, this implies that first person pronouns in the relevant clause are locally bound by a contextshifting operator. However, second person pronouns must be bound nonlocally by a matrix context-setting operator. This is schematically represented in (14a). While it will be technically possible to formulate relativized minimality conditions in such a way that this type of nonlocal binding is possible, it seems simpler to us to assume that the context-shifting operator also binds second person pronouns, but that the value it assigns to u is inherited by default from the matrix operator, as in (14b). (14) a. [CP OP … verb … [CP OP …1sgc … 2sgb …]] b. [CP OP … verb … [CP OP …1sgc … 2sgb …]] If there is a default rule that assigns embedded context variables the same value as their matrix counterparts, we can make the straightforward assumption that every CP comes with a context operator.14 The highest in a sequence of operators acquires its values from the discourse context. Operators in embedded clauses look upward to acquire values. In the usual case, these are inherited by default from the higher operator, but in some languages certain embedding predicates may also determine values, yielding possible shifts. That this is on the right track is confirmed by the fact that even in languages that do not allow shifted readings of first and second person pronouns, shifted readings of other indexicals can occur (see Schlenker 2003 and references

Person Features 33

cited there). For example, the interpretation of Dutch vandaag ‘today’ in the embedded clause in (15) can be anchored to the matrix context time (that is, the time of utterance) or to the matrix event time (that is, the time at which Jan said something). The first reading is the only one available in (15a). The second reading is predominant in (15b). Schlenker analyzes such phenomena as involving indexical shift pertaining to matrix and embedded context time. Note that, as with other indexical shifts, the matrix verb seems to drive the process. Thus, ‘say’ allows a shift in (15b), but factual verbs like ‘forget’ do not, as illustrated in (15c). In our implementation, the embedded context operator can either inherit its temporal value from the matrix context operator by default, or shift to the value of the matrix event time. The interpretation of the adverbial is then regulated via local binding by the embedded operator. (15) a. [OP Jan zei gisteren [OP dat ie vandaag Jan said yesterday that he today jarig is] yeary is ‘Jan said yesterday that it is his birthday today.’ b. [OP Jan zei gisteren [OP dat ie vandaag jarig Jan said yesterday that he today yeary was] was ‘Jan said yesterday that it was his birthday today.’ c. *[OPt Jan was gisteren vergeten [OP dat ie Jan was yesterday forgotten that he vandaag jarig was] today yeary was ‘Jan had forgotten yesterday that it was his birthday today.’ This is not a peculiarity of vandaag ‘today’. As shown in (16), it also holds for other adverbials anchored to a context time. (16) a. [OP Jan zei gisteren [OPv dat ie overmorgen Jan said yesterday that he over.tomorrow jarig is] yeary is ‘Jan said yesterday that it is his birthday the day after tomorrow.’ b. [OP Jan zei gisteren [OP dat ie overmorgen Jan said yesterday that he over.tomorrow jarig was] yeary was ‘Jan said yesterday that it was his birthday the day after tomorrow.’

34

Chapter 2

It is striking that in these examples the tense of the embedded verb seems to be sensitive to whether or not a shift takes place. There is an extensive literature on this phenomenon in a variety of languages (see, among others, Enç 1987, Abusch 1988, Kratzer 1998, Schlenker 1999, von Stechow 2003, and Khomitsevich 2007). We cannot discuss this literature here, but simply conclude that the generalization for Dutch is that any shift must be signaled by past tense, even if the time referred to is in the future with respect to both the matrix context time and the matrix event time, as in (16b). The absence of a shift is compatible with both past and present tenses (which then receive the same range of interpretations as in root contexts, because the matrix context time is inherited by default by the embedded context operator). We close this section with a brief discussion of what might appear to be a clause-internal shift in the value of u. When addressing a group, it is possible to use multiple instances of a second person singular pronoun to refer to different members of that group. A Dutch example of this type is given in (17). (17) Jij begeleidt jou en jij begeleidt jou. you accompany you and you accompany you ‘You1 will accompany you2 and you3 will accompany you4.’ At first sight, this phenomenon is unexpected, as the proposal outlined above rules out such clause-internal shifts. However, arguably the example does not, in fact, require a shift. Rather, the context operator assigns multiple values to the addressee variable (a possibility required independently to deal with homogeneous readings of plural second person pronouns; see note 3). The use of a singular pronoun in this context requires that one of these values be selected to the exclusion of the others. This explains the necessity of contrastive focus on second person pronouns used in this way, as exclusion of alternatives is exactly the semantic effect of contrastive focus. Indeed, the strong pronouns in (17) cannot be replaced by their weak counterparts, as weak pronouns cannot be focused. (18) *Je begeleidt je en je begeleidt je. you accompany you and you accompany you ‘You1 will accompany you2 and you3 will accompany you4.’ Other consequences of the assumption that every CP contains a context operator will be discussed in chapter 6. 2.3.3  Coincidence of Discourse Roles

The final issue to be discussed in this section again concerns the values of the various discourse roles. In the usual case, these roles seem to be mutually

Person Features 35

exclusive, in that the person functioning as speaker will not also be the addressee, and so on. The question we are interested in here is whether this is a matter of discourse conventions, or a consequence of the semantics of the person system. Within the theory outlined above, this effect must be a matter of discourse convention, because there is nothing inherent in the notions i, u, and o that makes it impossible for a single individual to be associated with more than one of these roles. This is not true, however, of all theories of person. Consider the well-known feature system proposed by Halle (1997), mentioned in section 2.2. (19) 

PARTICIPANT

+ AUTHOR –

+ 1 2

– * 3

This feature system predicts that, within the domain of a single context operator, it should not be possible for a first and a second person pronoun to refer to the same individual. Neither should it be possible for a third person pronoun to refer to the same individual referred to by either a first or a second person pronoun. This is because the feature system explicitly contains negative values. For instance, if something is specified as –author (as both second and third person pronouns are), it cannot refer to an individual that acts as author. Several observations indicate that in certain situations or styles the same individual can be referred to by pronouns that have different person specifications. One example is self-talk. Here, the same individual that acts as i also acts as u, as in (20) (for examples of this kind, see also Holmberg 2010:186). With respect to (20b), it is important to note that English does not permit shifts in the values of i and u between the matrix clause and an embedded clause (see above). (20) a. (Context: John discovers he has a winning lottery ticket.) IJohn can’t believe yourJohn luck! b. (Context: Mary, an athlete, is at the start of an important race.) Come on, IMary know youMary can do it! Of course, there are many facts about self-talk that require further research. As Holmberg shows, the first and second person pronouns are not in free variation when you are talking to yourself.15 This may suggest a possible solution for proponents of a system like the one in (19). One might argue that what is happening in self-talk is that the speaker imagines there to be two individuals

36

Chapter 2

in the domain of discourse, a speaker and a distinct addressee, the latter being a particular guise of the former. The problem with this assumption is that it predicts that it should be possible in self-talk to use a first person plural pronoun to refer to these two individuals. For many speakers, however, this is not possible.16 (21) a. (Context: John discovers he has a winning lottery ticket.) IJohn can’t believe yourJohn luck! *If we hurry, we can still collect the money today. b. (Context: Mary, an athlete, is at the start of an important race.) Come on, IMary know youMary can do it! *If we win, we’ll have a glass of champagne afterwards. It is also possible for the same individual to function as both i and o. This can happen, for instance, when two people are speaking about a third person, who overhears them. (22) (Context: Marie and Piet are discussing Mien’s ability to take care of herself, not knowing that Mien is in the next room.) a. Marie to Piet: Mien moet eigenlijk naar een verzorgingstehuis, denk je ook niet? Mien must really to a care.home think you also not ‘Mien should really go to a care home, don’t you think?’ b. Mien: IkMien denk dat zeMien dat niet wil! I think that she that not wants ‘I think she won’t like that!’ A similar context allows for the same individual to function as both u and o. (23) (Context: Mummy says that someone should tie Sam’s shoelaces.) a. Sam to Daddy: Hoorde je dat? Iemand moet Sam’s schoenveters strikken! heard you that someone must Sam’s shoelaces tie ‘Did you hear that? Someone should tie Sam’s shoelaces!’ b. Daddy to Sam: Ja, jijSam moet z’nSam schoenveters strikken! yes you must his shoelaces tie ‘Yes, you should tie his shoelaces!’ These effects, then, favor an analysis of person in terms of privative features.

Person Features 37

2.4  The Third Person as Default

A striking characteristic of the person system proposed above is that the third person is not a “nonperson,” but is characterized by a specific feature structure. While this is not unique to our proposal (see, for instance, Nevins 2007), it goes against a traditional idea according to which only first and second person have a feature specification while “third person” is the absence of any person information (see, for instance, Forchheimer 1953, Harley and Ritter 2002, Béjar and Rezac 2003, Cysouw 2003, Kayne 2003, Anagnostopoulou 2005, Adger and Harbour 2007). One important argument that has been given for this view is that the third person behaves as a default. This expresses itself in various ways. One is that expletive pronouns are invariably third person. The English data in (24) are typical in this respect. (24) a. It seems/*seem that Vitesse won. b. *I/*You seem that Vitesse won. In the same vein, so-called default agreement, which is the inflection showing up on a verb in the absence of an agreeing argument, must be third person singular. In Dutch, an example of default agreement can be found in the impersonal passive. (25) Nog jaren is/*ben/*bent/*zijn naar een oplossing gezocht. still years be.3sg/be.1sg/be.2sg/be.pl for a solution searched ‘People searched for a solution for many years.’ The fact that in our system the third person does have a feature gives rise to two questions. First, we need to explain why the third person behaves as a default. Second, especially in view of our use of privative features, we would expect some element in the system to be characterized by the absence of person features. If this is not the third person, we may wonder what the relevant pronoun is. In this section, we will focus primarily on the issue of defaults. We will also touch on the issue of featureless items, but this is explored in more detail in chapter 4. We first consider the fact that expletive pronouns are consistently third person forms. Any explanation must be rooted in the observation that such pronouns do not have a reference. We therefore assume that a pronoun can only function as an expletive if its φ-feature specification can deliver an empty set. The definition of dist, the feature that characterizes third person pronouns, indeed makes this possible. Recall that dist selects the outer layer in (4) (Si+u+o − Si+u). But given that the only obligatory members of Si+u+o are one speaker and one addressee, which are contained in the rejected inner layers of (4), dist may deliver an empty set.

38

Chapter 2

Precisely because both inner layers of (4) contain an obligatory member, prox, which discards the outer layer of its input set, can never yield an empty set. Therefore, first and second person pronouns cannot be used as dummies. The logic behind this proposal—namely, that expletive pronouns must deliver an empty set—also explains why they must be singular. A plural specification would require the cardinality of the pronoun’s output set to be greater than one (see subsection 3.2.4). This is obviously not true of the empty set. We have analyzed singular as the absence of a plural specification. This allows the pronoun’s output set to be a singleton set (in the normal case) or the empty set (in the case of expletive pronouns). Our analysis of expletive pronouns extends to default agreement. As we have illustrated in (25), agreement morphology that shows up on the verb in the absence of an agreeing argument must be third person singular. It is generally assumed that φ-features cannot be interpreted in verbs (compare the distinction between interpretable and uninterpretable features in Chomsky 1995). The rationale for this is that features that force a nominal reference are incompatible with verbal semantics. There are two ways to deal with a verb inflected for φ-features. The normal procedure is that the features are identified with features of the agreeing nominal argument (compare checking theory; also see chapter 5 for more detailed discussion). But if there is no such argument, the structure can still be licit as long as the φ-features present in the verb need not be given any reference at all. As just noted, the only form that allows this absence of reference is the third person singular.17 (A different issue is why in structures like (25) the verb should carry φ-features to begin with. We assume that it is simply a morphological property of a language whether or not finite verbs carry φ-features. If they do, this is the case in any context.) We conclude that the default status of third person singular does not militate against its carrying a person feature. In fact, there are arguments supporting the view that third person does have feature content. A first argument is that, as we will show in chapter 7, crosslinguistic patterns of syncretism require that second person share a feature with first person and a (different) feature with third person. This can, of course, only be the case if third person has feature content to begin with. A number of further arguments are discussed in Nevins 2007. One involves the “spurious se” phenomenon in Spanish (see Perlmutter 1971 and Bonet 1995, among others). In Spanish, a sequence of a third person dative clitic and a third person accusative clitic does not surface as the expected le lo. Rather, the dative clitic (le) is replaced by se, which in other contexts functions as a reflexive or impersonal clitic. An example illustrating this is given in (26) (adapted from Bonet 1995).

Person Features 39

(26) a. *A Pedro, el premio, le lo dieron ayer. to Pedro the prize 3.dat 3.acc gave.pl yesterday b. A Pedro, el premio, se lo dieron ayer. to Pedro the prize se 3.acc gave.pl yesterday ‘As for Pedro and as for the prize, they gave it to him yesterday.’ The spurious se phenomenon is likely to be an example of dissimilation of adjacent clitics (see, for instance, Grimshaw 1997a). Nevins shows that this dissimilation cannot be purely phonological in nature, but must be based on avoidance of a sequence of adjacent morphemes with identical feature specifications. He demonstrates that the simplest formulation of the relevant rule must make reference to a third person feature. (27)  Delete/Alter the features corresponding to third person on a dative when it precedes another third person. Formulations that try to avoid making reference to a third person feature either are inadequate empirically or say the same thing as (27) in artificially complex terms (see Nevins 2007 for detailed discussion). Of course, there is nothing wrong as such with a pronoun lacking feature content. Ruling out this possibility would require an additional stipulation. Therefore, if third person does have feature content, we would expect there to be some other type of pronoun that is characterized by a lack of person features. Such a pronoun would refer to the entire input set Si+u+o. This yields a pronoun with a very broad reference. We will argue in chapter 4 that this type of pronoun indeed exists. In particular, certain impersonal pronouns, exemplified by Dutch men, are best analyzed as featureless (see also Egerland 2003b). If the hypothesis that certain impersonal pronouns are featureless is correct, we can understand why it is se, rather than some other clitic, that acts as the replacement form for le in (26). One of the uses of se in Spanish is as an impersonal pronoun, which implies on our view that it can be analyzed as lacking feature content. Indeed, this is exactly what Grimshaw (1997a) argues. This means that a simple rule of feature deletion (or impoverishment, as this process is generally known; see (28)) suffices to account for the le/se alternation. The third person clitic pronoun carries the feature dist in syntax. This must be the case in order for the correct interpretation to obtain. However, before the pronoun is phonologically realized, it undergoes the rule in (28). (28)  dist → ∅ / [clitic __ ] [clitic [dist]] The result is that the form that realizes the clitic in the context of another third person clitic is the one specified as realizing the empty feature set (that is, se).

40

Chapter 2

A variant of the rule in (28) is found in two varieties of Catalan discussed in Walkow 2012. Here, it is the rightmost clitic that undergoes impoverishment of its third person feature when preceded by another third person clitic. The two varieties differ in the order within the cluster: in one, the indirect object clitic is rightmost; and in the other, the direct object clitic is rightmost. For both, the rule in (29) holds. (Note that in Catalan the clitic that remains is not entirely featureless, but spells out case; hence, what surfaces is not the impersonal clitic found in Spanish.) (29)  dist → ∅ / [clitic [dist]] [clitic __ ] Note that in the theory proposed here a pronoun that lacks person features must refer to the input set Si+u+o, as no selection from this set is made. Given that the input set contains two obligatory elements (i and u), the pronoun necessarily has referential properties. The proposal therefore predicts that featureless pronouns cannot act as dummies. Indeed, the Dutch impersonal pronoun men cannot be used as an expletive. (30) *Men schijnt dat men regent. one seems that one rains ‘It seems that it rains.’ 2.5  Effects of Maximal Encoding

In the previous sections, we have developed our basic proposal for the featural decomposition of persons and shown how this accounts for the crosslinguistic inventory of pronouns and their interpretation. When we consider individual languages, however, the interpretation associated with a pronominal form depends not only on the interpretation of the pronoun’s own φ-features, but also on which competing pronominal forms are present. This is because of Maximal Encoding, a principle introduced in section 1.6. The relevant definitions are repeated here for convenience: (31)  Maximal Encoding A mapping R → R* is licit only if R* is the maximal expression of R at the relevant level of representation. (32)  R* expresses R maximally if there is no alternative R′ that encodes more properties of R or encodes these properties in more locations. Consider the phonological realization of subject pronouns in Dutch. The following spell-out rules are relevant:18

Person Features 41

(33) a. {D prox} ⇔ /ik/ b. {D prox dist} ⇔ /jij/ c. {D dist} ⇔ /hij/zij/het/

d. {D prox pl} ⇔ /wij/ e. {D prox dist pl} ⇔ /jullie/ f. {D dist pl} ⇔ /zij/

As noted in chapter 1, the general effect of Maximal Encoding for spell-out is that where two or more rules may realize a syntactic feature structure, the rule that realizes the most features must be used. For example, if the syntactic input is the second person singular (see (9b)), then any of the rules in (33a–c) may in principle be applied, as the structural description of each of these rules is compatible with the input. (The rules in (33d–f) cannot be applied, as they mention the feature pl, which is absent in (9b). Hence, the structural description of these rules is not met.) The choice among (33a–c) is determined by Maximal Encoding: (33b) realizes a superset of the features realized by (33a) and (33c). Therefore, in the realization of a second person singular pronoun, the spell-out rule in (33b) blocks the spell-out rules in (33a) and (33c). It goes without saying that in the realization of, say, a first person singular pronoun, (33b) will not have this blocking effect, as it cannot spell out (9a). Dutch does not have separate forms that express the inclusive and exclusive readings of the first person plural pronoun, although the distinction must exist in the syntax on the assumptions made above (see note 9). As noted, this situation can be described as an underspecification of the set of spell-out rules available in the language. Rule (33d) mentions [prox pl]. Since there is no alternative form that marks repeated application of prox (that is, no rule of the form in (34a), as opposed to (34b)), (33d) will be used to realize both the inclusive and the exclusive first person pronoun. (34) a. {D prox prox pl} ⇔ /a/ b. {D prox pl} ⇔ /b/ In languages that do make the distinction between inclusive and exclusive first person plural pronouns, Maximal Encoding will have the effect that rules of the form in (34a) will block rules of the form in (34b) when the feature structure present in syntax contains two instances of prox (that is, in the exclusive reading).19 Hence, in such languages, rules of the form in (34b) can only be used for inclusive first person plural pronouns. Conversely, rules of the form in (34a) cannot be used for such pronouns, as their structural description (which mentions two instances of prox) in that case does not match the pronominal feature structure (which has only one instance of prox). Maximal Encoding is relevant not only to the realization of the various objects in (9) and (10), but also to their interpretation. In chapter 1, we already noted that the principle explains why the interpretive effect of the absence of a feature depends on what alternative feature structures are available in the

42

Chapter 2

language. Thus, a pronoun not specified as plural in a language that does have plural pronouns will typically be interpreted as singular (in principle, the cardinality of the output set could also be zero, an issue to which we turn in chapter 3). In a language that lacks plural pronouns, an unspecified pronoun can receive either a singular or a plural interpretation. A further example of the interpretive effect of Maximal Encoding is mentioned in the next section. It is interesting to note that on the proposal developed here, there cannot be LF competition (governed by Maximal Encoding) between different persons in the same way that there can be between different numbers. The reason for this is that person representations at LF are not compatible with more than one person interpretation, even when they have fewer features. For example, it may seem that a feature structure [dist] competes with the more fully specified [prox–dist]. However, the former cannot possibly be interpreted as referring to a set containing u and is therefore not in competition with the latter for encoding such a set. This is a general property of the person system proposed here, which can be traced back to the fact that the input set has a specific structure and population. Even the most underspecified pronoun, not containing any person feature, is incompatible with an interpretation equal to any specific person. Rather, as already suggested, it must be interpreted as a particular type of impersonal (generic or arbitrary) pronoun. The implication is that any underspecification for person in the pronominal system of a language must be a matter of PF. In syntax, the person distinctions discussed so far are universal. 2.6 Honorifics

Above, we have given an account of the inventory of persons as it is conceived of in much of the literature. However, there is reason to think that this does not exhaust all possible feature structures grouped under the prs node. In particular, there is evidence for a further feature in the person system, one that encodes politeness. Typologically, a distinction exists between “referent” honorifics and “addressee” honorifics (see, for instance, Brown and Levinson 1987). Referent honorifics are markers of honorificity that can pertain to the referent of a nominal expression. They are not limited to any particular person in principle, but in many languages they pertain to second persons only. By contrast, addressee honorifics convey respect to the addressee, without necessarily referring to them. The honorifics we are interested in here are those that are part of the person system, that is, polite alternatives to regular second person pronouns.

Person Features 43

Such honorifics can take different forms. Some languages have dedicated polite pronouns, that is, pronouns that do not occur elsewhere in the pronominal paradigm. The example we will consider below is Dutch u (not to be confused with the symbol u we use to reference the addressee element in Si+u+o). Another common option is to recruit an independently existing pronoun as a polite form. In French, for example, vous is a plural second person form, which can be used as a polite form of address in the singular. Similarly, in German the third person plural pronoun sie is recruited as an honorific. We will discuss below which independently available forms can be used in this way, and why, but first we consider the feature structure of honorific addressee pronouns. Our basic proposal is quite straightforward: we analyze all such honorifics as involving a feature hon. Hon is different from the person features discussed so far in two ways. First, for obvious functional reasons, pronouns that carry hon must be formally distinct from the informal addressee forms in the language (it makes no sense to be polite if this is indetectable to the addressee). Second, hon does not affect the reference of the pronoun in which it is present: it delivers the same set that it receives (see (35b)). However, it adds the information that relevant members in that set are honorable. In languages with addressee-oriented referent honorifics, only second person pronouns have a polite form, so hon must come with an appropriate selectional restriction (see (35a)). The semantic effect of hon is not that it makes a selection out of its input set, but rather that all addressees contained in its input set are identified as honorable (see (35b,c)). (35) a. hon selects [prox dist] b. hon(S) = S c. If x ∈ hon(S) ∧ x = u, then honorable(x) The interpretation of pronouns in systems that encode honorificity is partly governed by Maximal Encoding. In languages that have distinct honorific pronouns, use of the general second person pronoun implies familiarity. In languages that lack honorific pronouns, use of the general second person pronoun has no such effect (compare English you, for example). On the definition in (35), associates of the addressee(s) are not marked as honorific through the application of hon. This accounts for the following observation by Daniel Harbour (personal communication). If one addresses a friend and uses a second person plural pronoun to refer to that friend and his honorable but absent parents, the familiar form will be used, without this implying any familiarity toward the parents. However, if one addresses the parents, using a second person plural pronoun to refer to them and their

44

Chapter 2

son, then the polite form must be used if honorificity is intended, without this implying any formality toward the son. We illustrate this with the Dutch examples in (36) (see below on the singular agreement on the verb when u is used). (36) a. (Context: I am addressing the son.) Gaa-n jullie / #Gaa-t u nog wel eens samen op go-pl you.pl go-2sg you.hon yet indeed sometimes together on vakantie? holiday ‘Do you still sometimes go on holiday together?’ b. (Context: I am addressing the parents.) #Gaa-n jullie / Gaa-t u nog wel eens samen op go-pl you.pl go-2sg you.hon yet indeed sometimes together on vakantie? holiday ‘Do you still sometimes go on holiday together?’ Notice that this observation necessitates that Si+u − Si can contain associates of u. If all members of this set were addressees, it would be impossible to explain why use of the polite pronoun does not necessarily mark all of them as honorific. With (35) in place, we can add the pronominal structures in (37) to the crosslinguistic inventory. (37) 

a. Singular

b. Plural

NMB 2nd person polite

NMB PRS

NMB PRS

PRS

NMB NΠ

PL

PRS

PROX

PROX

DIST

DIST

HON

HON



We now look in some more detail at the various ways in which politeness is marked overtly in pronouns. The simplest way languages do this is by having a dedicated spell-out rule for polite forms. Dutch, for example, has the spell-out rule in (38), which specifically mentions hon in its structural description.

Person Features 45

(38) {D prox dist hon} ⇔ /u/ The form u can have both singular and plural referents, although it systematically triggers singular agreement on the verb. This is illustrated in (39). (39) a. Heb-t u in uw eentje deze hele kamer have-2sg you.hon in your.hon one.dim this whole room geschilderd? painted ‘Have you painted this whole room on your own?’ b. Of heb-t u elkaar geholpen bij het schilderen? or have-2sg you.hon each.other helped by the painting ‘Or have you helped each other paint?’ c. *Of hebb-en u elkaar geholpen bij het schilderen? or have-pl you.hon each.other helped by the painting There are two possible accounts for this behavior, which make use of underspecification and impoverishment, respectively. The first account is that Dutch does not morphosyntactically specify number in polite pronouns. In other words, the polite pronoun u is specified as in (37a) and does not have a plural counterpart. As a result, Maximal Encoding does not block a plural interpretation (see the discussion in section 2.5). Note that number agreement on the verb will be the default singular in the absence of number information in the subject. The second account assumes that there is a singular-plural distinction in polite pronouns in syntax, which is, however, obliterated by an impoverishment rule targeting both the pronoun and the verb before spell-out. In Ackema and Neeleman 2013a, we opted for the second account on the basis of data that we now think are inconclusive.20 Given that both accounts are compatible with the facts of Dutch, we remain agnostic, but will for the sake of concreteness adopt the account based on syntactic underspecification (though nothing hinges on this in what follows). As noted, instead of using a dedicated pronoun to express politeness, a language may recruit independently existing pronouns and assign them a polite second person interpretation. The question is which pronouns allow this, and why. We can rule out the second person singular pronoun on functional grounds. If it were selected as a polite pronoun, it would not be possible to mark politeness towards individual addressees, which largely defeats the purpose of a system of politeness marking. We can also rule out the first person singular, on the grounds of learnability. In order to acquire a polite pronoun that shares its form with a regular pronoun, there need to be contexts in which the polite pronoun is used and the reference of its regular counterpart is excluded. For example, a context in which there is a single addressee and vous is used in French reveals that the plural feature in the pronoun is not

46

Chapter 2

interpreted. However, there are no contexts in which an utterance does not have a single speaker. Hence, the first person singular pronoun can always be interpreted as referring to the speaker, which makes its acquisition as a polite form of address difficult if not impossible. Our general proposal is that a personal pronoun can be recruited as a polite second person pronoun in case the set of features present in syntax can be successfully connected to the right semantics by rules of impoverishment that apply at LF and/or successfully connected to a phonological form by rules of impoverishment that apply at PF. The starting point for this process is a syntactic feature specification that must be independently available in the language. The correct semantics for a second person polite pronoun can only be delivered by the feature structure [prox dist hon], and therefore this must be the feature structure presented to the mapping rules that connect syntax and semantics. We first consider the possibility of LF impoverishment as a means of recruiting existing pronouns as polite forms of address; the possibility of PF impoverishment is discussed later. LF impoverishment involves the deletion of features after syntax and before interpretation. Although the term may be unfamiliar, the idea is hardly new. It is used, for example, in accounts of why verbs recruited as support verbs typically have less specific semantics than verbs that are not recruited in this function, the idea being that support verbs are derived by deletion of semantic content and that such deletion must be minimized (see, for instance, Grimshaw 1997a,b). Nevins (2008), who calls the process “deprivation,” explicitly discusses it as a counterpart to the more familiar notion of impoverishment at PF (after syntax and before spell-out).21 On the assumption that there is no feature insertion at LF, but only feature deletion, the syntactic source for recruited polite pronouns must contain the features [prox dist hon]. This means that pronouns specified in syntax as first or third person can never give rise to an interpretation as a second person polite pronoun via LF impoverishment. What is possible, however, is that a pronoun specified as [prox dist hon pl] undergoes LF impoverishment of its number feature. This in effect delivers a pronoun underspecified for number, which is therefore compatible with both a singular and a plural polite second person reading (compare the discussion above of the dedicated polite pronoun u in Dutch). After all, Maximal Encoding only applies in the presence of an opposition in feature content, and this is obliterated by the rule in (40). (40) pl → ∅ / [ __ hon]

(at LF)

It is common for languages to use second person plural pronouns to mark politeness. An example of this pattern is found in French, where vous can have both a polite singular and a polite plural interpretation (alongside the regular plural interpretation) (examples from Wechsler 2011:1016).

Person Features 47

(41) a. Vous êtes loyal. you.pl be.2pl loyal.masc.sg ‘You (one male addressee, polite) are loyal.’ b. Vous êtes loyaux. you.pl be.2pl loyal.pl ‘You (addressee + associate(s), potentially polite) are loyal.’ Such pronouns are syntactically specified as plural, but undergo the LF impoverishment rule in (40). The input delivered to the spell-out system is the feature structure present in overt syntax. Given that French does not have a designated spell-out rule for pronouns specified for hon, this input will be realized by the same form that realizes the regular second person plural pronoun.22 As expected, polite vous allows both a singular and a plural construal. At the same time, the presence of the pl feature in syntax ensures that the agreeing verb will systematically show up in its plural form. Hence, (41a) does not have the variant in (42). (42) *Vous es loyal. you.pl be.2sg loyal.masc.sg ‘You (one male addressee, polite) are loyal.’ An interesting twist in the French data is that, in contrast to the finite verb, the predicative adjective in (41) shows singular agreement when polite vous has a singular construal. The contrast between verbal and adjectival agreement in examples like (41) is sometimes referred to as “mixed agreement”; see Corbett 2006, Wechsler 2011, and Wechsler and Hahm 2011. Our take on this is that adjectival agreement, at least in the case at hand, is an instance of socalled semantic agreement (see also Corbett 2006:231 on similar cases from Bulgarian). We will come back to the analysis of semantic agreement in chapter 6. For now, we can summarize our position as follows. Semantic agreement is not, in fact, a form of agreement at all (see also Ackema 2014). Rather, in the case at hand, it consists of information in the adjective that encodes the presupposition that the adjective’s external argument has a particular cardinality. This presupposition must match the semantic properties of the subject in order for the sentence to make interpretive sense. Thus, in (41a) the adjective comes with the presupposition that its external argument is singular, which is compatible only with a singular construal of vous.23 The second means of recruiting an existing pronominal form as a polite pronoun is through PF impoverishment. Any rule of PF impoverishment will have to operate on the required syntactic input [prox dist hon (pl)] if the pronoun is to be assigned the right interpretation at LF. Reduction of this input through feature deletion can yield either [prox (pl)] or [dist (pl)].24 These feature bundles will be spelled out by the form that otherwise realizes the

48

Chapter 2

first person inclusive or the third person (singular or plural). Notice that for the plural forms, there will also have to be LF impoverishment for number.25 An example of a language in which third person forms have been recruited as polite pronouns is German. In current standard German, the third person plural form sie is used as such. Simon (2003) discusses the history of the German polite address system and shows that in earlier stages of the language third person singular forms could be used in the same way. As just discussed, such forms must involve PF impoverishment, and in particular the rules in (43) (rule (43a) pertains to present-day German; (43b) is the historical rule). (43) a. prox → ∅ / [ __ dist hon pl] b. prox → ∅ / [ __ dist hon]

(at PF) (at PF)

The rules in (43) are underspecified for syntactic category. As a result, they also apply to any verb that syntactically agrees with a polite pronoun. Indeed, such verbs must show third person agreement, rather than second person. (44) Hab-en/*Hab-t/*Ha-st Sie das Geld? have-3pl/have-2pl/have-2sg you.hon (lit:they) the money ‘Do you (polite) have the money?’ It is not surprising that the verbal agreement undergoes the same rule of impoverishment as the polite pronoun. After all, the source for recruitment is an existing structure with a third person pronoun, including the agreement that such a pronoun triggers. Integral to our account of third person pronouns used as polite forms of address is that they are syntactically distinct from true third person pronouns. The surface similarity is only established at PF. We therefore expect to see some reflexes of this divergent syntactic specification. Such reflexes indeed exist. For German, Simon (2003) points out a number of phenomena that show that polite Sie behaves, syntactically, like a second rather than a third person pronoun. One clear syntactic difference between polite Sie and third person sie is that polite Sie can show up in structures of close apposition, like second person pronouns but unlike the plural third person sie (examples adapted from Simon 2003:101). (45) a. ihr Finnen you.pl Finns ‘you Finns’ b. *sie Finnen they Finns ‘them Finns’ c. Sie Finnen you.pl.hon Finns ‘you (polite) Finns’

Person Features 49

A further difference concerns relative clauses. Subject relative clauses that take a second person pronoun as antecedent repeat this pronoun in the relative clause, while relative clauses with a third person antecedent do not show this behavior. Again, polite Sie patterns with second person pronouns in this respect, rather than third person (examples adapted from Simon 2003:102). (46) a. Ihr, die *(ihr) Bücher sammelt, besitzt fünf Ausgaben you.pl who you.pl books collect own five editions der Bibel. the.gen bible ‘You (plural),who collect books, own five editions of the bible.’ b. Sie, die (*sie) Bücher sammeln, besitzen fünf Ausgaben they who they books collect own five editions der Bibel. the.gen bible ‘They, who collect books, own five editions of the bible.’ c. Sie, die *(Sie) Bücher sammeln, besitzen fünf you.pl.hon who you.pl.hon books collect own five Ausgaben der Bibel. editions the.gen bible ‘You (polite), who collect books, own five editions of the bible.’ The final pronoun that can be recruited as a polite second person pronoun according to the proposal above is the first person inclusive. While this seems to be rare, there are examples, discussed in detail in Cysouw 2005b and references cited there. A case in point is Muna, as described by Van den Berg (1989). In this language, the inclusive pronoun intaidi is used as a polite form of address, uninflected in the singular and affixed with the marker -imu in the plural (data from Van den Berg 1989:51). (47) 1sg 2sg 2sg hon 3sg 1excl 1incl dual 1incl pl 2pl 2pl hon 3pl

Pronoun

Verb form

Gloss

inodi ihintu intaidi anoa insaidi intaidi intaidi-imu ihintu-umu intaidi-imu andoa

a-kala o-kala to-kala no-kala ta-kala do-kala do-kala-amu o-kala-amu to-kala-amu do-kala

‘I go’ ‘you (singular) go’ ‘you (singular polite) go’ ‘he/she/it goes’ ‘we (exclusive) go’ ‘you and I go’ ‘we (inclusive) go’ ‘you (plural) go’ ‘you (plural polite) go’ ‘they go’

Notice that the inclusive pronoun is formally distinct from both the exclusive pronoun and the second person familiar pronoun.

50

Chapter 2

Similar to the case of the third person just discussed, our analysis of the use of inclusive pronouns as polite forms of address is based on PF impoverishment (here, of the feature dist), and therefore we predict that, syntactically, these forms should behave differently from their formally identical true inclusive counterparts. There is a distinct paucity of evidence in this area, simply because only a few languages have been reported to use inclusive pronouns as polite forms of address in the first place. Note, however, that at least in Muna the polite pronoun intaidi triggers an agreeing form of the verb not found elsewhere in the paradigm, in particular, not with the actual inclusive pronoun intaidi (nor with the informal second person form ihintu). Van den Berg describes the relevant verb form as containing a specific honorific agreement marker to-. In our terms, this indicates that the feature hon is present in the polite version of intaidi, but absent in its inclusive incarnation. In sum, we have shown that through LF or PF impoverishment a language may recruit as a polite form of address the second person plural, third person singular or plural, or first person inclusive pronoun. The one pronoun that cannot be recruited as a polite form of address through either form of impoverishment is the first person exclusive plural. LF impoverishment cannot deliver the feature structure required for a second person interpretation if given [prox prox pl] as input, while PF impoverishment cannot derive the feature structure [prox prox pl] from a [prox dist hon pl] input.26 This appears to be an adequate description of the typology of honorific pronouns in languages that do not have a dedicated polite pronoun. Cysouw (2005b:225) notes that the use of exclusive pronouns as the polite addressee form is not attested in any known language. To conclude, the typology of polite forms of address follows from the feature structures proposed here. Their analysis merely requires the introduction of the feature hon and the option of impoverishment at LF and/or PF. 2.7  Consequences for the Morphosyntax of

prs

2.7.1 Introduction

In the previous subsections, we have outlined the system of person features and their semantics. Our proposal has consequences for the syntactic context in which prs, the node hosting person features, can find itself. We will show that these consequences are the following: (i) R-expressions cannot be specified for person; (ii) nmb cannot be merged prior to prs; (iii) categories specified for person cannot be combined with an intersective (or subsective) modifier.

Person Features 51

We have argued that prs takes an input set (represented by Π) and delivers it to the person system. This implies that the semantics of prs is an identity function over sets. (48)  prs: λP.P As a consequence, [prs–prox–dist] has the semantics λP.dist(prox(P)). Given that the interpretation of prox and dist is dependent on the input set Si+u+o, which has the layered structure described in section 2.2, prs must always combine with a node containing Π in the nominal projection in which the person features are interpreted. We therefore expect that pronouns can be built from an invariant stem expressing NΠ that combines with distinct person markers. It is common for NΠ to remain silent, which is not surprising given that its presence is predictable from the presence of person features. However, there may be cases in which NΠ and prs are spelled out separately. It has been observed that in a number of languages strong pronouns consist of a person/ number morpheme and a separate root. As Siewierska (2004:19) notes, “Etymologically the generic pronominal root is often the word for person, body, self or the verb ‘to be’ or ‘exist’. ” There is some discussion in the literature about the synchronic function of such roots; see, for instance, Keegan 1997, Jelinek 1998, Cardinaletti and Starke 1999, and Siewierska 2004. At least in some cases, we take these roots to spell out NΠ. Consider the Shuswap pronominal paradigm in (49) (from Déchaine and Wiltschko 2002:414, adapted from Kuipers 1974 and Lai 1998). (49) Shuswap independent pronouns Singular 1 n-tsets-we7 (1sg-emph-deic) 2 7-enwi7 (2sg-emph) 3 newi7-s (emph-3)

Plural wll-enwi7-kt (pl-emph-1pl) wll-enwi7-s-kucw (pl-emph-3-excl) wll-enwi7-mp (pl-emph-2pl) wll-enwi7-s (pl-emph-3)

The relevant morpheme is glossed by Déchaine and Wiltschko as emph (for emphatic), but we suggest it is a realization of NΠ.27 Notice that the relevant root is located between morphemes that express person and number and must therefore have a position low in the pronoun’s extended projection, in line with this suggestion. (50) [pl [NΠ prs]]

52

Chapter 2

2.7.2  R-Expressions Do Not Have Person

If it is correct that person features are defined in terms of the function Pred, and that Pred in turn is defined in terms of the sets Si, Si+u, and Si+u+o (section 2.2), we expect that prs cannot combine with a regular R-expression, as R-expressions do not deliver Si+u+o: they differ from pronouns in that their head does not have the Π-feature that introduces this set. Hence, we predict that R-expressions cannot be specified for person (but see chapter 5 for discussion of an apparent exception).28 This explains why there are no first or second person R-expressions. A first person R-expression, for instance, would refer to the speaker and would obligatorily trigger first person agreement. It is certainly possible to use an R-expression to refer to the speaker or addressee (see, for instance, Collins and Postal’s (2012) discussion of what they term “imposters”). However, this is never accompanied by obligatory first or second person agreement, where obligatory agreement is understood to be syntactic agreement triggered as a lexical property of the R-expression. As far as we know, no such R-expressions exist. For instance, the Dutch and English examples in (51) are possible in certain registers, with the DP subject referring to the speaker. Nonetheless, these DPs cannot license first person agreement, let alone obligatorily require it. (51) a. Deze jongen gaa-t/*ga morgen fijn vissen. this boy go-3sg/go.1sg tomorrow nicely fish ‘I’m happy that I’m going out fishing tomorrow.’ b. The present author thinks/*think that this is not justifiable. Further evidence that R-expressions like those in (51) are not specified as first person comes from the observation that in discourse they can easily be used ironically to refer to the addressee, as well as the speaker. (52) A: Yours Truly has been awarded a knighthood. (Yours Truly = speaker) B: Well, then Yours Truly must be absolutely (Yours Truly = thrilled. addressee) (53) A: Deze jongen gaat dit weekend lekker (deze jongen = speaker) this boy goes this weekend nicely vissen. fish ‘I’m going out fishing this weekend!’ B: Maar weet deze jongen wel dat het gaat (deze jongen = but knows this boy well that it goes addressee) regenen? rain ‘But do you realize it’s going to rain?’

Person Features 53

It goes without saying that the equivalent is not possible with pronouns, because pronouns have a fixed association with i, u, or o, as outlined earlier. The following is impossible, for example (no matter how ironic B’s reply is intended to be): (54) A: I have been awarded a knighthood. B: #Well, then I must be absolutely delighted.

(I = speaker) (I = addressee)

(55) A: Ik ga dit weekend lekker vissen. I go this weekend nicely fish ‘I’m going fishing this weekend!’ B: #Maar weet ik wel dat het gaat regenen? but know I well that it goes rain ‘But do you realize it’s going to rain?’

(ik = speaker)

(ik = addressee)

Similar observations can be made for R-expressions that refer to the addressee. Our proposal predicts that R-expressions cannot carry a third person feature either. At first sight, this seems unlikely, given that R-expressions trigger what appears to be third person agreement. However, since default agreement necessarily takes the form of third person agreement (see section 2.4), this is not a particularly compelling argument. If R-expressions indeed do not have person features, they would be expected to trigger default person agreement. There is evidence that R-expressions are not like third person pronouns. Their reference can contain speaker or addressee, as already illustrated in (51)–(53), and as corroborated by the examples in (56), in which a first or second person pronoun refers back to an R-expression (underlining is used to indicate intended coreference).29 By contrast, a third person pronoun cannot be the antecedent for a first or second person pronoun, as illustrated in (57). This follows if third person pronouns are specified as dist, while R-expressions are not. (Note that in (56) the Dutch can be replaced by a proper name like the Millers without the judgments being affected, indicating that these do not depend on the antecedent being interpreted generically.) (56) a. Anyone who knows the Dutch realizes they no longer wear wooden shoes. b. Anyone who knows the Dutch realizes we no longer wear wooden shoes. c. Anyone who knows the Dutch realizes you no longer wear wooden shoes.

54

Chapter 2

(57) a. Anyone who knows them realizes they no longer wear wooden shoes. b. *Anyone who knows them realizes we no longer wear wooden shoes. c. *Anyone who knows them realizes you no longer wear wooden shoes. The absence of person features in R-expressions means that such expressions cannot impose restrictions dictating which persons are included in their reference. This makes it possible to construe the set that an R-expression refers to as one containing speaker or addressee. A third person pronoun, however, explicitly encodes (via dist) that it selects a part of the person space that excludes speaker and addressee, and therefore coreference with a first or second person pronoun is impossible. We will come back to this difference between R-expressions and third person pronouns in chapter 5.30 Examples similar to the ones in (56), but with a singular R-expression as antecedent, are not always felicitous (for instance, Collins and Postal (2012: 20–21) report considerable speaker variation in this respect). However, they are not impossible. For a start, it is relatively unproblematic for a first or second person pronoun to refer back to an R-expression across a sentence boundary. (58) a. Deze jongen gaat van het weekend lekker vissen. this boy goes of the weekend nicely fish Ik heb er echt zin in. I have there really appetite in ‘I’ll go for some great fishing this weekend. I’m really looking forward to it.’ b. Yours Truly has been nominated for the Royal Society Medal. Not that I care, of course. (59) a. (Context: The doctor is visiting Kobus. Kobus has served coffee, and asks the doctor the following.) Wil de dokter ook een koekje? Of vindt u dat niet gezond? wants the doctor also a biscuit or find you that not healthy ‘Would the doctor like a biscuit as well? Or do you think that’s not healthy?’ b. (Context: A member of Parliament is addressing the Speaker of the House.) Could the Speaker allow more time for this debate? You allowed this when we discussed the first draft of this bill. Even within a single sentence, coreference between a singular R-expression and a first or second person pronoun is not excluded.

Person Features 55

(60) a. Hoewel deze jongen in het weekend liever vist, although this boy in the weekend rather fishes moet ik zondag de voordeur verven. must I Sunday the front.door paint ‘Although I prefer to go fishing on the weekend, I have to paint the front door this Sunday.’ b. As for Yours Truly, I was in full-on glutton mode. (Jim Finnerty, A Pleasant Stroll to Everest, p. 339) Nevertheless, there appears to be a contrast between singular and plural R-expressions in that the latter permit this kind of coreference more easily. The examples in (56b,c) are entirely unobjectionable, while their singular counterparts in (61b,c) are degraded. (61) a.

Anyone who knows this Dutchman realizes he no longer wears wooden shoes. b. ??Anyone who knows this Dutchman realizes I no longer wear wooden shoes. c. ??Anyone who knows this Dutchman realizes you no longer wear wooden shoes.

Given the examples in (59)–(60), this cannot be a matter of grammar. We assume, instead, that the use of an R-expression normally triggers the presupposition that its denotation is not exhaustively identical to i and/or u. That is, whenever an R-expression includes i or u in its denotation, it also includes some other individuals. In the singular, this presupposition cannot be met if the R-expression refers to i or u. There are circumstances in which this presupposition can be suppressed, however. For example, the Dutch case in (59a) represents the use by certain speakers of R-expressions as polite forms of address. In other cases (for instance, (58)–(60)), suppression of the presupposition is tied to a particular “jolly” style of speech. To our mind, plural DPs that refer exhaustively to i’s or u’s trigger exactly the same kind of stylistic effect as found in the singular (in contrast to expressions like the Dutch in (56)). (62) a. Wat de huidige auteurs betreft, wij denken niet dat dit voorstel what the current authors concerns we think not that this proposal werkt. works ‘As far as the current authors are concerned, we don’t think this proposal works.’

56

Chapter 2

b. Ik wil voor de trouwe luisteraars graag een plaatje draaien, I want for the faithful listeners gladly a disc turn als dank voor al jullie berichten. as thanks for all your messages ‘I would like to play a disc for all my loyal listeners now, as a thank-you for all your messages.’ We conclude that R-expressions do not have person features that determine their reference. They never obligatorily trigger first or second person agreement, and they can be coreferential with any pronoun.31 2.7.3  Number above Person

Our proposal about the nature of prs has a number of other consequences, which become apparent once we assume that UG requires that grammars be designed so as to preserve information encoded by φ-features in the interpretation of a DP. We formulate this “design principle” as in (63). It is not intended to be a principle that is part of individual grammars. Rather, it requires that grammars contain conditions that imply that no violations of (63) can ever occur. (63) Preservation of Φ Any operation applied in a DP must be information-preserving with regard to the semantic contribution of φ-features. An operation is information-preserving if it comes with the guarantee that it will not destroy the information encoded by φ-features. The exact effects of Preservation of Φ depend on the nature of the operations involved and on the nature of the φ-features present in the DP. A first result of (63), in conjunction with our characterization of the person system, is that number must be interpreted after person. Number specifies the cardinality of its input set, but does not change this set in any way. Hence, information contributed by person features is preserved when number is added. (See chapter 3 for more discussion on the nature of the number system and its interaction with person.) Suppose, however, that number were interpreted before person. Number would then specify the cardinality of Si+u+o (which is delivered by NΠ). This is necessarily plural, as this set contains at least two obligatory members, i and u. Interpreting person next may destroy this number information, since the person features may select a singular set (for example, if prs contains proxprox and Si does not contain any associates). Recall that an operation is information-preserving only if it comes with the guarantee that it will not destroy the information encoded by φ-features. Therefore, application of

Person Features 57

person is not information-preserving with respect to number, but application of number is information-preserving with respect to person. We assume that in this case Preservation of Φ is implemented in grammars through a restriction on the interpretation of number. In particular, we assume that plural nmb is specified as follows (this will be revised in our discussion of the number system in chapter 3, but the restriction on the input set of the number feature will stay in place): (64) a. pl(S) is defined iff |S| > 1 ∧ S ≠ Si+u+o b. pl(S) = S if defined The effect of (64) is that the number feature cannot be directly applied to NΠ, but can be applied to a set delivered by the person system. This is compatible with a DP-internal structure in which nmb c-commands prs or in which nmb and prs are combined into a single node (on such conflation of functional structure, see Platzack 1983, Bayer, Schmid, and Bader 2005, and Citko 2008, among others; and see chapter 3). In the latter situation, the interpretive system can simply choose to interpret person first, in line with our proposed principle. What is excluded, however, is a structure in which prs c-commands nmb. In such a structure, compositionality requires that person be interpreted after number, which implies that (64) will be violated. The data suggest that this is the right result (see also Harbour 2016). There are many languages whose pronouns fuse person and number. There are also languages in which pronouns are built from a person stem that is inflected for number. A straightforward example is Japanese, where, for instance, ‘we’ is built from the word for ‘I’ (watashi) and a plural suffix (see chapter 3). The evidence that the person information constitutes the base is twofold. First, the plural suffix also attaches to regular nouns. Second, inflectional affixes in Japanese are generally suffixes, not prefixes. Some varieties of English show the same pattern in the second person, as these have a distinct plural form built from the singular form you by addition of a number marker; examples are youse and y’all (see Hickey 2003 and references cited there). Other cases are given in Daniel 2013, a chapter in the World Atlas of Language Structures on the expression of number in pronouns. There are 42 languages in Daniel’s sample that show this morphological pattern. Significantly, Daniel does not mention a category of pronoun that is built by attaching a person affix to a number stem. To the best of our knowledge, such pronouns do not exist, confirming that prs may not c-command nmb.

58

Chapter 2

2.7.4  Pronouns and Modification

A second result of (63) is that intersective modifiers can never be merged with a pronoun, as we will now explain. For a start, such modifiers must occur below nmb if (63) is correct. If they were to merge above nmb, they could destroy the information encoded by the number features. For example, the plural form old men could express a singular. Even though nmb encodes that its input set has a cardinality larger than one, subsequent intersection with the set of old things might result in a singleton set. Precisely this is forbidden by (63). Therefore, if a language is to have an operative number system, intersective modifiers must select a category lower than nmb. For the same reason, intersective modifiers should be located below prs. Suppose such a modifier were merged after prs. The person features deliver a particular subset of Si+u+o as their output. Preservation of Φ demands, then, that no operation can be applied that could yield a different set as output. But this is exactly what intersective modification does: its output is a proper subset of its input set. Hence, attaching an intersective modifier after person has applied is not information-preserving. We assume that in this case Preservation of Φ is again implemented through a selectional restriction, this time on intersective modifiers. (65)  Intersective modifiers select NP. The implication is that in pronouns, intersective modifiers must attach below prs. However, this leads to a different problem. prs needs as its input the structured set Si+u+o delivered by NΠ, since the person features are defined only in relation to this set. However, when applied to NΠ, an intersective modifier delivers a set whose members both meet the description of the modifier and are elements in Si+u+o. However, if the modifier is truly intersective, this set will not equal Si+u+o, as some elements of the latter will be discarded. The consequence of this is that the predecessor relations that structure Si+u+o (see section 2.2) no longer apply, so that the set resulting from intersection is not a suitable input to the person features.32 If intersective modifiers can be located neither above nor below prs, they must be in complementary distribution with person information. Indeed, R-expressions, which we argued lack person information, can be combined with both intersective and nonintersective modifiers. Pronouns, on the other hand, only combine with nonintersective modifiers. We illustrate this using relative clauses and PPs. Consider the Dutch examples of modification of pronouns in (66).

Person Features 59

(66) a. Ik met mijn grote mond stond zo weer buiten. I with my big mouth stood right again outside ‘I, being such a loudmouth, found myself in the street again right away.’ b. Wij met onze warme jassen aan hadden geen last van de kou. we with our warm coats on had no bother of the cold ‘We, with our warm coats, were not bothered by the cold.’ c. Jullie die in tranen voor ons staan zijn hierbij vergeven. you.pl that in tears before us stand are hereby forgiven ‘You, who are standing before us in tears, are hereby forgiven.’ These examples only permit a reading in which the modifier is not intersective, but applies to every individual in the denotation of the pronoun. It will be clear that the modifier in (66a) must be nonintersective. After all, the pronoun denotes a singleton set containing just the speaker. The sentence in (66b) cannot mean that only those among us who had warm coats were not bothered by the cold; rather, it implies that all of us had warm coats and hence were not bothered by the cold. Similarly, (66c) implies that all of you were standing before us in tears. At first sight, things are less clear when third person pronouns are modified. There are clear cases where a modifier is interpreted nonintersectively with respect to a third person pronoun. An example is the following: (67) De mensen uit Groningen zeiden dat zij, die nu eenmaal al the people from Groningen said that they that now once already twee uur in de trein hadden gezeten, als eerste naar binnen mochten. two hours in the train had sat as first to inside may.pst ‘The people from Groningen said that they, already having been on the train for two hours, were allowed to enter first.’ However, such examples exist alongside cases in which the modifier seems to have an intersective flavor. (68) Zij die hun kaartje nog hebben mogen nu naar binnen, maar zij they that their ticket still have may now to inside but they die hun kaartje kwijt zijn moeten achter aansluiten. that their ticket lost are must behind join ‘The people who still have their ticket can go inside now, but the people who have lost their ticket must join the back of the queue.’ However, it is not difficult to show that examples of this type do not, in fact, have an intersective interpretation. If they had, it should be possible for the pronoun to be coreferential with some R-expression in the context, with the

60

Chapter 2

modifier selecting a subset of the set introduced by this R-expression. This is clearly impossible, though. The example in (69) is ungrammatical on the reading indicated by the translation. (Here and below, underlining indicates intended coreference; binding is indicated by coindexation.) It is grammatical on a different reading in which the people from Amersfoort claimed that all and only those who still have their ticket should be allowed entry. (The core difference between the two readings concerns the status of people from outside Amersfoort who still have a ticket; they would be refused entry on the intended, but ungrammatical, reading.) (69) *De mensen uit Amersfoort beweerden dat alleen zij die hun the people from Amersfoort claimed that only they that their kaartje nog hebben naar binnen mogen. ticket still have to inside may.pst ‘The people from Amersfoort claimed that only those among them who still had their ticket are allowed to go in.’ What seems to be going on instead is that the relative clause identifies the reference of the pronoun. That is to say, the denotation of the pronoun consists of all and only those that meet the description given in the relative clause.33 The disjunction in (68) conspires to suggest a restrictive (intersective) reading of the relative clauses, but this is only the result of the mistaken assumption that the pronoun zij ‘they’ refers to the same set of people in the two clauses. In reality, two different groups of people are contrasted, with the relative clause merely helping to identify the reference of the pronouns. That this is possible is demonstrated by the grammaticality of (70). Here, the different denotations of the pronouns are determined deictically. (70) Zij mogen nu naar binnen, maar zij moeten achter aansluiten. they may now to inside but they must behind join ‘They are allowed to go in now, but they must join the back of the queue.’ Note that similar cases also exist with singular pronouns. (71) Hij die het laatste kaartje heeft gekocht mag nu naar binnen. he that the last ticket has bought may now to inside ‘The person who bought the last ticket may now enter.’ This confirms that the relative clause cannot be intersective in these cases, since no proper subset can be selected from a singleton set (other than the empty set, which is clearly not relevant here). The hypothesis that relative clauses can identify the reference of a pronoun implies that the whole constituent is like an R-expression in that its denotation

Person Features 61

is determined internally to it.34 This may shed some light on the fact that pronouns thus modified behave like R-expressions syntactically: they can neither be bound by, nor be coreferential with, a c-commanding R-expression (compare Condition C of the binding theory). Even on the otherwise permitted identificational reading, examples like the following are ungrammatical: (72) a. *[Elke groep reizigers]i beweerde dat [zij die hun kaartje nog every group travelers claimed that they that their ticket still hadden]i naar binnen mochten. had to inside may.pst ‘Every group of travelers claimed that they were allowed to go inside as they still had their tickets.’ b. *De mensen uit Groningen zeiden dat zij die uit het verre the people from Groningen said that they that from the far noorden kwamen als eerste naar binnen mochten. north came as first to inside may.pst ‘The people from Groningen said that they were allowed to enter first, as they came from the far north.’ The example in (72b) can be distinguished from (67) on the basis of its intonation. The latter requires focus on the pronoun and intonational breaks before and after the relative clause. None of these properties hold of “identificational” relative clauses like the one in (68). Coreference in (67) is allowed because here the relative clause does not identify the reference of the pronoun zij. Whereas the DP zij die hun kaartje nog hebben ‘they who still have their ticket’ in (68) refers to exactly those people who still have their ticket, a DP like zij, die hun kaartje nog hebben ‘they, who still have their ticket’ (comparable to the relevant DP in (67)) refers to a group of people whose identity is determined by the pronoun’s antecedent. All the relative clause tells us in this case is that the people in question happen to still have their ticket, but it is entirely possible that there are other people not contained in the denotation of the pronoun who still have their ticket as well. In conclusion, the principle in (63) seems well-motivated. In combination with the semantics proposed for person features, it correctly predicts that pronouns do not allow intersective modification.35 It also has the consequence that if nmb and prs are not conflated, then nmb must c-command prs, which appears to be in line with the data. We will argue in chapter 4 that (63) has other desirable consequences, in particular with regard to the interpretation of generic and arbitrary pronouns.

62

Chapter 2

2.8  Bound Pronouns

So far, we have concentrated on the referential use of pronouns. However, it is well-known that pronouns have a second use as bound variables. Our theory of person should of course at least be compatible with the bound variable interpretation of pronouns. In this section, we will argue that no adjustment of our theory of person features is required to deal with the distinction between referential and bound pronouns. However, specific assumptions must be made about the way φ-features function in bound pronouns. Our core hypothesis is that bound pronouns are variables and that the φ-features they contain impose restrictions on this variable and thereby on the element that binds it. The φ-features can impose such restrictions in one of two ways. Either the features of the pronoun are used to impose limitations on the φ-feature content of the antecedent (via agreement), or they are used to limit its interpretive domain (see also Heim 2008 and references mentioned there). In the former case, the relevant φ-feature must be present in the antecedent. In the latter case, the φ-feature in the pronoun is used to construct a semantic restriction on the variable, and the choice of antecedent must be compatible with that restriction. (This is comparable to the standard treatment of φ-features in bound variable pronouns as introducing presuppositions on the range of the variable; see Cooper 1979 and Heim and Kratzer 1998). The general idea that φ-features in bound pronouns are either interpreted or licensed under agreement, in which case they are not interpreted, is very similar to the proposals in von Stechow 2003 and Kratzer 2009. Kratzer’s implementation of this idea differs from von Stechow’s in that Kratzer assumes that features licensed under agreement are not generated in the pronoun but copied onto it from the antecedent in the PF branch in the grammar. Essentially following von Stechow, we assume instead that agreement provides a license for noninterpretation of features in the pronoun. However, the choice of implementation has no empirical consequences, as far as we can tell. It is easiest to demonstrate the two ways in which φ-features can contribute to the properties of a bound pronoun by considering gender features first, as these are sometimes interpreted and sometimes licensed under agreement. We will use Dutch to illustrate these options. In R-expressions, standard Dutch has a two-way distinction in formal gender, with nouns specified as neuter or nonneuter (the latter are usually said to have common gender). We assume, essentially following Toebosch (2011), that neuter corresponds to the absence of gender, while common gender nouns are specified as [gender].36 The pronominal system makes a three-way gender distinction in the singular (though not the plural): there are neuter, masculine, and feminine singular pronouns, which we respectively analyze as having no gender features, being specified

Person Features 63

as [gender], and being specified as [gender feminine]. Note that according to Maximal Encoding, use of an underspecified pronoun where a more highly specified pronoun is available has the effect of negation of the missing features. The choice of a pronoun under coreference works as follows. If a pronoun is selected on the basis of the formal features of the antecedent, then a masculine pronoun must take a common gender antecedent and a neuter pronoun must take a neuter antecedent. (A feminine pronoun cannot be used in this case, given that there is no formal feminine gender in R-expressions.) If the pronoun is chosen on the basis of the semantics of the antecedent, then a masculine pronoun is used for a male referent, a feminine pronoun for a female referent, and a neuter pronoun for nonindividuated referents, such as the referent of mass nouns (for a detailed description of the system, see Audring 2009). These alternatives are on a par in principle, but their use is subject to certain rules. In particular, if a pronoun has the gender feature and its antecedent has biological gender, then the pronoun must be chosen on the basis of the semantics of the antecedent. Thus, if the antecedent is a common gender noun that refers to a female (for example, de vrouw ‘the woman’), use of the default masculine pronoun hij is excluded, even though this pronoun could in principle be licensed through syntactic agreement. Instead, the feminine pronoun zij, which is licensed semantically, must be used. By contrast, if the antecedent is a neuter noun that refers to a female (for example, het meisje ‘the girl’), use of the neuter pronoun het is permissible, alongside use of the feminine pronoun zij. As a neuter pronoun does not encode [gender], the above rule does not apply, and therefore it is sufficient for pronoun and antecedent to match in formal feature content. (For related discussion, see Corbett 1991:228ff.) This system is also at work when a pronoun functions as a bound variable. The gender features in a bound pronoun can impose restrictions either on the formal feature content of the antecedent quantifier via agreement or on the individuals in its domain via the imposition of a semantic restriction on the variable. This is evident when the quantifier has a head noun that is formally neuter, but whose denotation contains female individuals. In that case, a bound pronoun can be feminine or neuter, but not masculine. (73) a. [Elk meisje]i denkt dat zei gaat winnen. each girl thinks that she goes win ‘Each girl thinks that she will win.’ b. [Elk meisje]i denkt dat heti gaat winnen. each girl thinks that it goes win ‘Each girl thinks that she will win.’

64

Chapter 2

c. *[Elk meisje]i denkt dat-iei gaat winnen. each girl thinks that-he goes win ‘Each girl thinks that she will win.’ The example in (73a) illustrates the possibility of the pronoun imposing a semantic restriction; (73b) illustrates the use of gender to impose a formal restriction; (73c) is unacceptable, because the gender features of the pronoun cannot be used to impose either a semantic or a formal restriction. When the antecedent is a quantifier headed by a common gender noun that refers to humans whose gender is left unspecified, both the masculine and the feminine pronoun can act as a bound variable, but with contrasting effects. Since the masculine pronoun is specified as [gender], the domain of the antecedent quantifier must contain either male referents only or both male and female referents, as in (74a). Use of the female pronoun, which is specified as [gender feminine], implies that the domain of the antecedent quantifier contains exclusively female referents, as in (74b). (74) a. [Elke student]i denkt dat-iei gaat winnen. each student thinks that-he goes win ‘Each student thinks that they will win.’ b. [Elke student]i denkt dat zei gaat winnen. each student thinks that she goes win ‘Each student thinks that she will win.’ In sum, gender features in bound pronouns may impose either formal or semantic restrictions on their binder. We now turn to number. Number in bound pronouns can impose only a formal restriction on the antecedent, not a semantic one. In order to see why, consider what the semantic contribution of number would be. In general, semantically interpreted φ-features in bound pronouns say something about the elements in the domain of the quantifier by imposing a semantic restriction on the bound variable. Hence, number would encode whether or not there is a plurality of such elements. Thus, if a bound pronoun is plural, and pl is interpreted, then the quantifier should range over a nonsingleton set. This is unproblematic. If the pronoun is singular, then the quantifier should range over a singleton set. Crucially, this is problematic. Arguably, quantification over a singleton set is not permitted (see also Hinterwimmer 2008; for more discussion, see section 4.3). This means that absence of pl in bound pronouns may not be construed as the negation of plurality. Therefore, plural is not semantically contrastive in bound pronouns, which leaves as the only option that its presence is licensed because it imposes a formal restriction on its binder.

Person Features 65

This accounts for the data below, which follow straightforwardly if the number specification of a bound pronoun reflects the feature content of its binder. (Notice that (75a) is grammatical on a reading in which each boy thinks that the group of boys will win. This is not a bound variable reading.) (75) a. *[Elke jongen]i denkt dat zei gaan winnen. each boy thinks that they go win ‘Each boy thinks that he will win.’ b. [Elke jongen]i denkt dat hiji gaat winnen. each boy thinks that he go win ‘Each boy thinks that he will win.’ (76) a. *[Alle jongens]i denken dat-iei gaat winnen. all boys think that-he go win ‘All boys think that they will win.’ b. [Alle jongens]i denken dat zei gaan winnen. all boys think that they go win ‘All boys think that they will win.’ We finally turn to person. Here, a great deal hinges on whether the binder is an R-expression or has person features itself. We begin by considering pronouns bound by quantifiers like every boy or each boy, which, being R-expressions, do not have person features. With such quantifiers, the situation is the reverse of what we have discussed for number: person features on bound pronouns cannot be licensed by agreement; they must therefore impose a semantic restriction on the variable, and thus an indirect restriction on the domain of the quantifier. The consequence is that only third person pronouns can act as variables bound by a regular quantifier, as we will now argue. Consider an example with a bound third person pronoun, such as (75b). The pronoun contains the feature dist, whereas its binder elke jongen ‘each boy’ does not contain any person features. Therefore, dist imposes a semantic restriction, not a formal one. In particular, the variable introduced by the pronoun must range over the set Si+u+o – Si+u, which contains only o’s. (77)  ∀x, x=boy, [x thinks that x, x ∈ Si+u+o – Si+u, will win] This interpretation is coherent, because the boys in the set under discussion can be construed as others.37 If person features in bound pronouns must have an interpretive effect, then first and second person bound pronouns should lead to a semantic representation in which the variable introduced by the pronoun must range over Si or Si+u in the first person and Si+u – Si in the second. The result is that first and second person pronouns cannot, in fact, be bound by regular quantifiers. We

66

Chapter 2

show this using the examples in (78a) and (78b), neither of which allows the bound variable reading described in (78c).38 (78) a. *[Elke jongen]i denkt dat iki ga winnen. each boy thinks that I go win b. *[Alle jongens]i denken dat wei gaan winnen. all boys think that we go.pl win c. ∀x, x=boy, [x think that x, x ∈ Si, will win] In these examples, the pronoun must be translated as a variable plus a restriction (x, x ∈ Si), while the quantifier is translated as an operator plus a restriction (∀x, x=boy). Binding has the consequence that for every boy an element in Si must be chosen that is identical to that boy. The problem is that there is no guarantee that the speaker will be included in the subset of elements taken from Si under variable binding. For example, the speaker might be female, in which case all the elements from Si paired with the boys over which the quantifier ranges must be associates, and cannot be the speaker. This means that a bound variable construal of a first person pronoun runs afoul of Preservation of Φ (63), the “design principle” that states that the preservation of the semantic contribution of φ-features must be guaranteed at the level of the pronominal DP (see subsection 2.7.3). Earlier instantiations of Preservation of Φ restricted operations that apply within a pronominal DP. In the case of bound variable pronouns, however, what needs to be constrained is the relation between the QP and the restriction on the bound variable that originates in the pronoun’s φ-features. We propose, then, that Preservation of Φ is implemented in this domain through the grammatical condition in (79). (79)  Let Rφ be a set over which a variable ranges (i.e., the set mentioned in its restriction), and S be the set of elements selected from Rφ under variable binding. If inclusion of i in Rφ is encoded, it must be guaranteed that S contains i, and if inclusion of u in Rφ is encoded, it must be guaranteed that S contains u. Condition (79) expresses that those elements obligatorily included in the set that forms the restriction on the variable introduced by the pronoun must be guaranteed to be contained in the range of the quantifier. Thus, prox in the first person pronoun in (78a) encodes the inclusion of the speaker in the restrictor of the variable, but as explained there is no guarantee that the speaker is included in the set of elements in the range of the quantifier.39 By the same logic, second person pronouns cannot be bound by regular quantifiers, as there is no guarantee that the addressee will be included in the individuals selected from Si+u – Si.

Person Features 67

(80) a. *[Elke jongen]i denkt dat jei gaat winnen. each boy thinks that you.sg go win ‘Each boy thinks that he will win.’ b. ∀x, x=boy, [x think that x, x ∈ Si+u – Si, will win] Since the reasoning above depends solely on the quantificational element not having person features, we expect that other R-expression quantifiers cannot bind first or second person pronouns either. This seems correct.40 (81) a. *Er denkt [een jongen]i dat iki ga winnen. there thinks a boy that I go win ‘A boy thinks that he will win (namely, me).’ b. ∃x, x=boy, [x think that x, x ∈ Si, will win] Our account predicts that there are situations in which a first or second person pronoun can act as a bound variable. To begin with, first or second person bound plural pronouns can occur if the variable is construed as ranging over elements that are systematically paired with speaker or addressee. This is because on such a construal, neither i nor u is contained in the set over which the variable ranges (although of course they help define this set), and therefore they are not part of its restriction. However, if neither i nor u is part of the variable’s restriction, (79) is not applicable. Examples of this type, which go by the name of partial binding, indeed exist, as observed by Partee (1989) and further discussed by Rullmann (2004), among others. Rullmann provides the following examples:41 (82) a. [Every woman I date]i wants usi to get married. ∀x, x=woman that I date, [x wants i and x, {i x} ⊂ Si, to get married] b. [Every woman you date]i wants youi to get married. ∀x, x=woman that you date, [x wants u and x, {u x} ⊂ Si+u – Si, to get married] First or second person bound pronouns can also occur if the binder is itself a pronoun. As pronouns are of course specified for person features, the complications we noted when the binder is a personless R-expression do not arise. Consequently, examples like the following are unobjectionable on a bound variable reading:42 (83) a. Wiji denken elk dat wei de ander gaan verslaan. we think each that we the other go beat ‘We each think that we will beat the other.’

68

Chapter 2

b. Julliei denken elk dat julliei de ander gaan verslaan. you.pl think each that you.pl the other go beat ‘You each think that you will beat the other.’ When the antecedent is a pronoun, singular first and second person pronouns also permit a bound variable interpretation. This is exemplified by sentences like the ones below, noted in Heim 2002, 2008 and references mentioned there. In (84a) and (85a), my and your must be bound variables in order to account for the fact that the elided VP permits a sloppy reading (see Reinhart 1983). In (86), binding is necessary in order to account for the fact that the sentence permits an interpretation on which the speaker did their homework, but no other person did theirs. (84) a. Ii did myi homework and Johnj did eVP , too. b. i λx [x did x’s homework] and John λx [x did x’s homework] (85) a. Youi did youri homework and Johnj did eVP , too. b. u λx [x did x’s homework] and John λx [x did x’s homework] (86) a. Only Ii did myi homework. b. i λx [x did x’s homework] c. No x, x ∈ {Bill, Susan, …}, [x did x’s homework] In sum, no changes are necessary to the proposed system of φ-features in general and person features in particular in order to deal with bound variable interpretations of pronouns. The φ-features of a bound pronoun always impose restrictions—either formal or interpretive—on their binder. Which of these two options is available in practice depends on various factors, including the nature of the feature makeup of the binder. This means that different φ-features will behave in different ways, even though they belong to the same system.43 2.9 Conclusion

The feature system described in this chapter exhausts our typology of person.44 There are two privative person features, prox and dist. These features represent functions that operate on an input set and deliver an output set. The initial input set is an abstract representation of the domain of discourse. The output set is a subset of this. The system delivers three persons with a singular interpretation and four persons with a plural interpretation. This feature system is different from a number of alternative proposals in that the third person has feature content. Nonetheless, it predicts that only the third person can act as an expletive. It can be shown that only the person feature that delivers third person interpretation can also deliver an empty set

Person Features 69

as output. (There are featureless pronouns as well, but these are a type of impersonal pronoun; see the discussion in chapter 4.) The theory has a number of other consequences. First, it predicts that R-expressions cannot be specified for person. We put forward several pieces of evidence in support of this result. Second, the proposal can be used to generate a typology of polite pronouns that is descriptively adequate. Third, in combination with a Preservation of Φ principle, it requires that nmb be merged higher in the structure than prs. This appears to be in line with crosslinguistic observations regarding the morphological structure of pronouns. Preservation of Φ has the further consequence that modification of pronouns must be nonrestrictive. Finally, our person feature system is compatible with a fairly standard treatment of bound variable pronouns. In particular, in contrast to gender and number, person is never interpreted in bound variable pronouns, again as a consequence of Preservation of Φ.

3

Interactions between Person and Number

3.1 Introduction

While this book is about person, it is necessary to discuss the role that number plays in the grammar of pronouns as well. This is because some questions about person are intertwined with questions about number. We will consider two such questions here. The first question concerns the distinction between number systems that mark plural and number systems that mark augmented numbers. In the latter, the effects of number marking crucially depend on information from the person system. We will explore what is behind the distinction between the two types of number system. Our proposal is that all languages make use of a feature aug, for augmented; however, only some languages allow the number system to access person information, and this has an effect on the number that aug delivers. Whether access is available depends on whether nmb and prs are separate heads or conflated. The second question relates to the status of others versus the associates of speaker and addressee when considering the interpretation of plural pronouns. We have claimed that plural first and second person pronouns can contain associates (ai and au) but not others (o), something we call the associative effect. Other authors have ascribed this effect to the number system instead, claiming that the plural of first and second person pronouns is an associative plural, something also found with regular nouns in a number of languages. We will argue that the latter view is problematic. The overall conclusion of the chapter is that the person system as developed previously does not need to be modified to account for interactions between person and number.

72

Chapter 3

3.2  Two Types of Number System 3.2.1  Plural in the First Person Inclusive

Perhaps the best-known interaction between person and number has to do with the first person inclusive (in languages that make a distinction between inclusive and exclusive forms). In order to explain the issue, let us assume for the time being that the number system is built on two features, pl (for plural) and min (for minimal) and that the first of these simply encodes that the cardinality of the set referred to, which we will indicate by n, exceeds 1 (n > 1). The second feature operates on the output of the first and selects the minimal plural (so n = 2). Maximal Encoding requires that number distinctions be marked where possible. Therefore, we expect that pronominal systems with a threeway distinction among singular, plural, and dual will be organized as in (1), which is the Maori paradigm (see Cysouw 2003:191). Note that on the analysis under discussion, what is usually referred to as the singular is really the absence of the pl and min features (see also subsection 3.2.4; see Mattens 1970 for an early discussion of the general idea that there is no singular feature, but that singular readings arise from competition with forms marked for plural). (1) Maori pronouns (singular-plural-dual system)

1incl 1excl 2 3

Singular

Plural

Dual

_ au koe ia

tā-ua ā-ua kōr-ua rā-ua

tā-tou mā-tou kou-tou rā-tou

In languages that make the inclusive-exclusive distinction but lack the min feature (and therefore the dual), one would expect pronominal paradigms of the type found in Mandarin (see Cysouw 2003:258; there is considerable dialectal variation—the paradigm in (2) is found in what Cysouw calls “Northern modern Chinese”). (2) Mandarin pronouns (singular-plural system)

1incl 1excl 2 3

Singular

Plural

_ wǒ nǐ tā

zán-men wǒ-men nǐ-men tā-men

Interactions between Person and Number 73

Notice that in neither type of system can there be a singular form of the first person inclusive pronoun, as this would violate Maximal Encoding. As the first person inclusive contains both speaker and addressee, it is inherently plural, and therefore plural marking (or dual marking) is obligatory where available. A number of authors have observed that certain pronominal systems resist analysis in terms of the singular-plural opposition. Ilocano is a language often discussed in this context (see Cysouw 2003:85–90 for an overview). It appears to have a dual form of the first person inclusive, but of no other person. The resulting paradigm looks awkward on a singular-plural-dual analysis. (3) Ilocano pronouns (on a singular-plural-dual analysis)

1incl 1excl 2 3

Singular

Plural

Dual

tayo mi yo da

ta

co mo na

The conclusion widely accepted in the literature is that languages like Ilocano should not be analyzed using a feature like pl. Instead, Ilocano has a number system built on the feature aug. (This take on languages like Ilocano goes back to Thomas 1955, Conklin 1962, and McKay 1978; see Bobaljik 2008a and Cysouw 2011 for modern incarnations.) The interpretation of aug can be paraphrased as ‘n is larger than the minimal cardinality allowed by the person system’ (see below for a formal implementation). In all persons except the first person inclusive, the minimal cardinality allowed by the person system is 1, and so aug delivers n > 1 (just like pl). In the first person inclusive, however, the minimal cardinality allowed by the person system is 2 (as the inclusive contains i and u). In this context, aug delivers n > 2. On this analysis, the Ilocano pronominal paradigm looks much more elegant—namely, as in (4), where there is one series of pronouns that carries the aug feature, and another that is not marked for number (just like the singular in (1) and (2)). (4) Ilocano pronouns (nonaugmented-augmented analysis)

1incl 1excl 2 3

Nonaugmented

Augmented

ta co mo na

tayo mi yo da

74

Chapter 3

Systems based on aug can have a three-way system in which what looks like a dual marker triggers a trial interpretation in the case of the first person inclusive (but nowhere else). An example is Rembarrnga (see (5), from Cysouw 2003:233). The first person exclusive series (ngunu – yarru – yarr-bbarrah), like the second and third person series, receives the interpretations singularplural-dual, but the first person inclusive series (yukku – ngakorru – ngakorrbbarrah) is interpreted as dual-plural-trial. Such languages can be analyzed using the aug feature and the min feature introduced above. If these features are combined, the resulting interpretation is ‘n is the smallest number larger than the minimal cardinality allowed by the person system’ (see below for details). This interpretation is often referred to as “unit-augmented.” (5) Rembarrnga pronouns

1incl 1excl 2 3

Nonaugmented

Augmented

Unit-augmented

yukku ngunu ku nawu/ngadu

ngakorru yarru nakorru barru

ngakorr-bbarrah yarr-bbarrah nakorr-bbarrah barr-bbarrah

Striking confirmation for the existence of aug comes from Sierra Popoluca (see Cysouw 2003:147–149). This language has the following “subject prefixes” that mark person in the verb: (6) Sierra Popoluca subject prefixes 1incl 1excl 2 3

taʔami∅-

These subject prefixes do not overtly encode number. However, verbs do mark number using a suffix, táʔm, whose distribution corresponds perfectly with the hypothesis that it realizes aug (under agreement). It shows up in all plurals, but is absent when a first person inclusive subject receives a dual interpretation. (7) a. ʔa-moŋ-táʔm-pa 1excl-sleep-aug-imper ‘We (exclusive) are sleeping.’ b. mi-moŋ-táʔm-pa 2-sleep-aug-imper ‘You (plural) are sleeping.’

Interactions between Person and Number 75

c. ta-moŋ-pa 1incl-sleep-imper ‘We (inclusive dual) are sleeping.’ d. ta-moŋ-táʔm-pa 1incl-sleep-aug-imper ‘We (inclusive plural) are sleeping.’ Thus, at first sight it appears that we need to distinguish the two types of languages on the basis of the number features they employ: some number systems are based on pl and others on aug. Both can be enriched with an additional feature min. 3.2.2 Aug as a Universal Number Feature

While the existence of two distinct number systems seems undeniable, an analysis in terms of crosslinguistic variation in feature types seems unattractive to us. This is partly because the interpretations of pl and aug overlap in almost all cases, making unification desirable, and partly because in other cases of parameterization in the number system the relevant difference concerns not a choice between different features but one between more or fewer features from a constant hierarchy; at least, that is what is suggested by Greenberg’s (1963) Universal 34, according to which no language has a trial number unless it has a dual and no language has a dual unless it has a plural. A similar observation holds of parametric variation in other feature systems, such as those for case (see, for instance, Caha 2013). The alternative we will develop here is based on the hypothesis that there is no pl feature. Rather, even in singular-plural systems plurality is encoded by aug. The difference between the two types of number system discussed above has to do with whether or not number can access person information. If it can, the result is a system of the type in (4) or (5); if not, we are dealing with a system of the type in (1) or (2). We will refer to systems in which number has access to person information as relative number systems, and to systems in which number does not have such access as absolute number systems. Whether access is granted depends on the morphosyntactic structure of pronouns, which is partially revealed by their phonological form. We begin by making our assumptions explicit. We assign the nmb node the following semantics, where C is the set of possible cardinalities determined by the features in nmb (see below): (8) a. nmb(S) is defined iff |S| ∈ C ∧ S ≠ Si+u+o b. nmb(S) = S if defined

76

Chapter 3

What these definitions express is that nmb passes on any set delivered by its sister, as long as the cardinality of this set falls within the range determined by the number features. (The condition S ≠ Si+u+o was motivated in subsection 2.7.3.) In the case of pronouns, the sister of nmb is prs, which delivers a particular set as determined by the person features (see chapter 2). Number then simply reflects the cardinality of this set. In the case of R-expressions, however, the sister of nmb is an NP. It is commonly assumed that nonpronominal NPs provide the number system with a set of sets, structured as a join semilattice (see Link 1983 and subsequent literature). If so, nmb applies to each of these sets, filtering out all that have a cardinality that falls outside of C. For our current purposes, the core question is how the features in nmb determine which cardinalities C contains. We assume that C is generated by the number features on the basis of the input set N, the set of natural numbers (which we take to include 0). For concreteness’ sake, we assume this input set is represented by a feature # contained in the nmb node. The semantics of aug is given in (9a). According to this definition, aug selects a subset of N that includes all numbers larger than a reference number nR. This reference number itself is preferentially determined by the person specification of a DP. If there is access to that information, nR corresponds to the lowest possible positive number compatible with the DP’s person specification. If there is no access to the DP’s person specification, or if the DP has no person specification, nR assumes a default value of 1.1 The assumptions that yield this result are stated in (9b–d). The definitions in (9c,d) formalize the notion of “lowest possible number compatible with a DP’s person specification.” Sperson is the output set delivered by the person system; the function strip removes all elements that are not i or u; nperson is defined as the cardinality of this reduced set. (9) a. b. c. d.

aug(S) = S′, S′ ⊆ S, n ∈ S′ ⇔ n > nR nR = nperson iff nperson is accessible and nperson > 0; otherwise nR = 1 nperson = |strip(Sperson)| strip(Sperson) = S′, S′ ⊆ Sperson, p ∈ S′ if and only if p ∈ {i, u}

A second feature required for an analysis of the number system is min, whose semantics is defined as follows:2,3 (10)  min(S) = S′, S′ ⊆ S, n ∈ S′ ⇔ n > 0 ∧ ∄n′, n′ ∈ S ∧ n′ < n This definition characterizes min as selecting a subset whose sole member is the smallest element in the input set (on the condition that this is a positive integer).

Interactions between Person and Number 77

Except for the notion of accessibility of person information and the associated default value of nR, there is nothing in (9) that is new. The semantics of aug and the way the value of nR is determined are part of every theory that has an aug feature (although they may be stated differently). Obviously, it is crucial that we develop a theory of access to person information. We will turn to this issue below, but first we demonstrate how the proposal works on the assumption that there is indeed variation in access to person information. Absolute number systems receive a simple analysis. In such systems, aug has no access to person information. Therefore, for all persons, the following holds: (11)  Absolute number system •  nR = 1 (by default) •  nmb–aug: n > 1 •  nmb–aug–min: n = 2 In relative number systems, aug does have access to person information. Therefore, the value of nR will vary depending on the composition of the set delivered by the person system. In the first person exclusive (see (12b)) and the second person (see (12c)), the cardinality of the “stripped” version of the person set (nperson) is 1 (see (9)). This means that nR = 1, yielding a regular plural interpretation when aug is applied. In the third person (see (12d)), nperson = 0. As nR must be a positive integer, it cannot be equal to nperson and will therefore assume its default value of 1, again yielding a regular plural interpretation. The first person inclusive is different from the other persons in that nperson and therefore nR equals 2 (see (12a)). This implies that aug will encode n > 2, while the combination of aug and min will lead to the interpretation n = 3. (12)  Relative number system a.  First person inclusive: •  nperson = |strip({i, ai+, u, au+})| = |{i, u}| = 2 •  nR = nperson = 2 •  nmb–aug: n > 2 •  nmb–aug–min: n = 3 b.  First person exclusive: •  nperson = |strip({i, ai+})| = |{i}| = 1 •  nR = nperson = 1 •  nmb–aug: n > 1 •  nmb–aug–min: n = 2

78

Chapter 3

c.  Second person: •  nperson = |strip({u, au+})| = |{u}| = 1 •  nR = nperson = 1 •  nmb–aug: n > 1 •  nmb–aug–min: n = 2 d.  Third person: •  nperson = |strip({o+})| = |{ }| = 0 •  nR = 1 (by default) •  nmb–aug: n > 1 •  nmb–aug–min: n = 2 Thus, if it is possible to make sense of the notion of access to person, a unification of absolute and relative number systems is within reach, at least where the features used are concerned. 3.2.3 Access

We postulate that whether or not aug has access to person information depends on the morphosyntactic structure in which it finds itself. Recall from subsection 2.7.3 that Preservation of Φ has the consequence that number must be interpreted after person. This is compatible with a structure in which nmb c-commands prs, but not with a structure in which prs c-commands nmb. By compositionality, the latter configuration would force number to be interpreted before person. What we have suggested is possible, however, is for number and person to be introduced in the same node (conflation; see chapter 2). In that case, compositionality does not force a particular order of interpretation, which means that it is possible to interpret person before number, as required. Thus, there are two relevant licit structures for a plural inclusive first person pronoun:4 (13)

a. Nonconflated NMB/PRS NMB

  1st person inclusive

AUG

NMB/PRS PRS

NMB PRS

b. Conflated NMB/PRS

NMB PRS NΠ



AUG PROX

PROX We propose that number has access to person information if and only if nmb and prs are contained in the same node. The rationale for this proposal lies in the assumptions that the semantics of a node must be computed independently

Interactions between Person and Number 79

of the semantics of its sister and that information must be preserved under inheritance. Arguably, these assumptions follow from compositionality. The effect is that number in (13a) has no access to person, in the sense that the value of nR must be determined in ignorance of the output of the person system—in particular, in ignorance of the value of nperson. In (13b), number does have such access, because number and person information are contained in the same node. Of course, in (13a) the number information and the person information are both inherited by the node that dominates nmb and prs (the higher nmb node), but given the second assumption above, tinkering with the value of nR is no longer allowed at that stage. Let us look in detail at how this derives the difference between absolute number systems and relative number systems described above. Recall that these differ in the first person plural inclusive; here, the value of nR is 1 in absolute number systems (as it is everywhere else), but 2 in relative number systems. The relevant structures are given in (14) and (15). (14) Inclusives in an absolute number system Unmarked inclusive *

Augmented inclusive

NMB

NMB PRS

NMB PRS



PROX

Minimal augmented inclusive NMB

AUG

PRS PRS

AUG

PRS PROX

nperson = 2 (*Maximal Encoding)

NMB

PRS

NMB



MIN PROX nperson = 2; nR = 1 ⇒ n = 2

nperson = 2; nR = 1 ⇒ n > 1



80

Chapter 3

(15) Inclusives in a relative number system Unmarked inclusive NMB/PRS NMB PRS

Augmented inclusive NMB/PRS



NMB PRS



AUG PROX

PROX

nR = nperson = 2 ⇒ n > 2

nperson = 2 ⇒ n = 2 Minimal augmented inclusive NMB/PRS NMB PRS



AUG PROX MIN nR = nperson = 2 ⇒ n = 3 When the semantics of number in (15) is computed, the value of nperson is accessible to aug, because prs is part of the same terminal node. Since applying prox once delivers a set with i and u as obligatory members, nR = nperson = 2, so that aug in this case requires that n > 2. When the semantics of the terminal containing aug in the structures in (14) is computed, however, the value of nperson is not accessible, because [prs–prox] is generated inside a sister node. This means that nR assumes its default value of 1, so that aug now requires that n > 1. One thing worth noting about absolute number systems is that they do not permit inclusive pronouns that do not bear number information (see (14a)). In such forms, nperson is 2, while nR in an absolute system is always 1. Consequently, the cardinality of the set referred to by the pronoun exceeds nR. As this information is exactly what is encoded by aug, Maximal Encoding requires that this number feature be present. Our analysis makes a crucial prediction about the morphological form of pronominal number. In absolute systems, plural can be either agglutinative or fusional. If the terminals introducing person and number are spelled out separately, an agglutinative number paradigm will emerge; if spell-out targets a string of terminals or a nonterminal node (on a par with {go past} ⇔ went), the number morphology will be fused with the person morphology. If person and number are introduced in the same terminal, however, as is the case in

Interactions between Person and Number 81

relative systems, they must be fusional (there is no position in which a distinct number morpheme could be anchored).5 We therefore predict that if number marking is agglutinative in pronouns, then the number system must be of the absolute type. If we go back to the paradigms in subsection 3.2.1, it turns out that the two languages with a relative number system that we discussed are indeed fusional for person and aug (see (4) for Ilocano and (5) for Rembarrnga). Remarkably, Rembarrnga is agglutinative for min—something our proposal allows for relative as well as absolute systems. The two languages that are agglutinative for aug indeed have an absolute system (see (1) for Maori and (2) for Mandarin). This seems to be the general pattern. Cysouw (2003:89, 263) notes that languages that have a relative number system and are agglutinative for aug are extremely rare, if they exist at all (see also Greenberg 1988). He remarks that in those languages with a relative number system that do seem to have a separate morpheme for aug, this morpheme is never the sole expression of number. Rather, different stem forms already encode the distinction between augmented and nonaugmented numbers.6 Note that languages with a relative number system do allow transparent number marking on the verb (see, for example, the Sierra Popoluca data in (7)). But this arguably involves agreement and features that, by general consent, remain uninterpreted (see chapters 5 and 6 for discussion). In sum, the aug feature appears to be shared by all number systems, but its interpretive effect depends on whether or not it has access to information delivered by the person features, which in turn depends on the syntactic structure of pronouns. 3.2.4  Can Plurals Mean ‘One or More’?

The proposal outlined in the previous subsections is partly based on the assumption that the default value of nR is 1. This implies that elements marked with aug must at least refer to sets containing more than one member.7 This is an intuitive result, but one that has been labeled as naive in some of the semantic literature, mainly because there are contexts in which bare plurals seem to allow an interpretation as ‘one or more’. Some examples are given in (16) (from Farkas and de Swart 2010:3). (16) a. Have you ever seen horses in this meadow? b. If you have ever seen horses in this meadow, you should call us. c. Sam has never seen horses in this meadow. The question in (16a), for example, will be answered positively if the addressee has seen a single horse in the meadow. Similar considerations hold for the

82

Chapter 3

other examples. At the same time, this interpretation is not generally available. The plural in an example like (17) must be interpreted as ‘more than one’ rather than ‘one or more’. (17)  I have seen horses in this meadow. The generalization that has been established in the literature is that the weaker reading of the plural is only available in downward-entailing environments (see Sauerland, Anderssen, and Yatsushiro 2005 and Farkas and de Swart 2010). We hypothesize that plurals always mean ‘more than one’ and that bare plurals can be accompanied by a silent quantifier equivalent to any. Thus, Have you seen horses? is interpreted on a par with Have you seen any horses? At first sight, this may seem to shift the problem, since now the question becomes how plurals are interpreted in the context of any, and how this interpretation differs from both a singular accompanying any and a plural in a referring expression. Recall that a nonpronominal NP provides the number system with a set of sets, structured as a join semilattice. In languages with an absolute number system, aug selects out of this set all sets with a cardinality larger than 1. The absence of aug gets its interpretation through competition and will hence normally correspond to a set of singleton sets. In a definite DP, the determiner selects a unique set from the output of the number system; hence, the boy refers to a unique singleton set while the boys refers to a unique nonsingleton set. In the context of quantifiers, including any, the contribution of number in the restrictor should not differ from its contribution in referring expressions. Hence, a plural NP still delivers a set of nonsingleton sets, while a singular NP still delivers a set of singleton sets. However, we assume that any ranges over atoms contained in the set of sets delivered by NP. (18)  ⟦any NP⟧ = any x, x ∈ S ∧ S ∈ ⟦NP⟧ This implies that expressions like any boy and any boys do not differ in the elements that the quantifier ranges over; rather, they differ only in the structure of the set from which these elements are selected. Thus, both Have you seen any horse? and Have you seen any horses? can be answered in the affirmative if the addressee has seen either a single horse or more than one horse. However, the way that the domain of the quantifier is structured does have an effect. After all, this domain is contextually relevant, and therefore it matters whether it is structured as a set of singleton sets or as a set of nonsingleton sets. In particular, the use of the plural in any horses implies that in the context of the

Interactions between Person and Number 83

utterance horses usually come in groups, whereas the singular in any horse has no such implication. Rather, in competition with the plural, the singular generates the expectation that horses will appear on their own. On the assumption that bare plurals may be accompanied by a silent version of any, it follows that a question like (16a) can also be answered in the affirmative if the addressee has seen only a single horse, while at the same time it is predicted to come with the expectation that horses (in the context of this question) usually show up in groups. We think that this is correct. For one thing, the existence of this expectation accounts for the contrast in felicity between (19a) and (19b), on the assumption that the use of well marks that an expectation is defeated. (19) Q: Are there (any) horses in this meadow? a. A: Yes, there is one. / Well, there is one. b. A: Yes, there are five. / #Well, there are five. Notice that with the singular the expectation is the reverse, and therefore the judgments shift. (20) Q: Is there any horse in this meadow? a. A: Yes, there is one. / #Well, there is one. b. A: Yes, there are five. / Well, there are five. If the above analysis is correct, it can be maintained that aug always marks a cardinality larger than 1. The hypothesis that the interpretation of the examples in (16) relies on the presence of a silent any predicts that only in those contexts in which any can appear is it possible for the plural to receive an apparent ‘one or more’ interpretation. This means that plural pronouns should never allow this interpretation, regardless of whether they occur in a downward-entailing environment or not. After all, they cannot be the restrictor of any: *any them is ungrammatical and any of them presumably involves an empty nominal head (equivalent to ‘one’) in the complement of any. Plural pronouns indeed do not allow a ‘one or more’ interpretation. Consider (21). (21) Q: Have you seen them in the meadow? a. A: #Yes, I have seen one. / Well, I have seen one. b. A: Yes, I have seen them. / #Well, I have seen them. The question in (21) can be contrasted with Have you seen any of them in the meadow?, which can, of course, be answered affirmatively if only one of them was spotted. The infelicity of the affirmative answer in (21a) indicates that the plural pronoun does not permit a ‘one or more’ reading, in accordance with the analysis of number in pronouns outlined in this section.

84

Chapter 3

3.3  Number in Pronouns and Associative Plurals

We now turn to a second issue regarding the interaction between person and number: whether, in the sets denoted by first and second person pronouns, members other than i and u must be associates of i and u, and, if so, what is behind this associative effect. We have assumed above that a first person plural pronoun selects a part of the person space that contains the speaker(s) and one or more associates. Similarly, a second person plural pronoun selects a part of the person space that contains one or more addressees with the possible addition of associates. The parts of the person space selected by these pronouns do not contain what we have termed others (o’s), however. We have referred to the resulting impossibility of including nonassociates in the reference of these pronouns as the associative effect. In our account, the associative effect is an inherent property of the person system. It follows from the way in which the person space is structured and populated, in combination with how the functions expressed by the person features operate. There are two challenges to this view. Some authors have argued that the associative effect is not real, in that the non–speaker/addressee individuals that can be included in the reference of a first or second person plural pronoun need not be associates, but can be others. Other authors have argued that, while the associative effect is real, it is a property of the number system rather than the person system. We will discuss these challenges in turn. 3.3.1  Is the Associative Effect Real?

The first challenge is perhaps most clearly expressed by Daniel (2000), according to whom the assumption that heterogeneous readings of first and second person plural personal pronouns are always associative is false. Daniel argues that, in addition to speaker or addressee and their associates (our i, u, and ai/u), such pronouns can include others (our o) in their reference. To illustrate this, he gives the following examples: (22) a. She called me and we talked until dawn. b. Look, we are both wearing striped socks, me and that girl over there. In evaluating examples like (22a,b), it is important to be clear about the notion of associate that is assumed. In our view, associates in the person system do not entertain some fixed relationship with speaker or addressee. For instance, “associates” cannot be equated to something like “family and friends.” The notion as operative in the person system is entirely context-dependent. Who is to be regarded as an associate of i or u depends on the context of the

Interactions between Person and Number 85

utterance and/or the preceding discourse; this information can be updated as the discourse proceeds. Such updating is exactly what happens in both (22a) and (22b). The speaker—say, Bill—signals that there is a particular relationship between himself and an o, with the effect that this o can now also be regarded as an associate of his, permitting this ai to be included in the reference of we. In (22a), this relationship is established in the first conjunct. In (22b), the relationship seems to be established through the coordination of me and that girl over there. Somewhat surprisingly, we is used before the addressee encounters this coordination. Notice, however, that (22b) involves rightdislocation of the coordination in question. Right-dislocation is known to be felicitous under specific discourse conditions only. In particular, Lambrecht (2001) observes that the right-dislocated category must be given; moreover, it must have been established earlier in the discourse as a topic. (This identification of the category in question may have been contextual; it need not have been established by linguistic means.) Lambrecht cites the French examples in (23) as evidence of this, but the effect has been found to be very general crosslinguistically (see Gundel 1988, Fretheim 2001, and Gundel and Fretheim 2004; relevant English data can be found in Ziv and Grosz 1994 and Ward and Birner 1996). (23) (Context: Husband and wife at dinner table; husband looks at food on his plate.) H: Ça n’a pas de goût, ce poulet. that neg.has not of taste this chicken ‘It has no taste, this chicken.’ W: Le veau, c’est pire. the veal it.is worse ‘The veal is worse.’ Lambrecht notes that it is felicitous for ce poulet ‘this chicken’ to be rightdislocated in the husband’s remark, because of its prominence in the situational context. This suffices to establish it as a given topic. In contrast, le veau ‘veal’ in the wife’s reply is left-dislocated, left-dislocation being a structure used to introduce new discourse topics. Lambrecht explicitly remarks that it would be highly infelicitous for the wife to use right-dislocation, exactly because le veau is not contextually given. Given this function of right-dislocation, me and that girl over there in (22b) must have been established in the context as a topic. This is indeed plausible. Notice that the speaker—again, Bill—starts the utterance with the imperative look, which must be accompanied by some form of pointing. The mere fact that Bill performs this action is sufficient to establish a contextual

86

Chapter 3

relationship between himself and the girl over there. This in turn licenses the inclusion of the referent of that girl over there as an associate of Bill’s in the construal of we. There are, in fact, good empirical reasons to distinguish between instances of o on the one hand and associates included in the reference of first and second person plurals on the other. As we will demonstrate, what is an o at a particular point in the discourse cannot be included in the reference of a first or second person plural pronoun without first being turned into an associate in some way. The other side of the coin is that an associate cannot be treated as an o without first being introduced as such. We consider each of these points in turn. We can use the following context to show that an o cannot be included in the reference of a first person plural pronoun. Suppose that Ad and Peter are having coffee in some café. Across the street, they see a famous singer passing by. Peter could then address Ad by uttering (24a) or (24b), but (24c) is infelicitous.8 (24) a. Do you see that Neil Young is wearing the same coat as me? b. Do you see that he’s wearing the same coat as me? c. #Do you see that we’re wearing the same coat? Of course, following an utterance like (24a) or (24b), use of we will be licensed, since these sentences establish a relationship of same-coat-wearing, so that the singer on the other side of the street can now be regarded as an associate of Peter. Thus, Peter can now continue by uttering (25), for example. (25)  We must have bought it in the same shop. As a second example, consider a situation in which Peter and Ad are watching a film starring George Clooney. In this situation, if Peter asks Ad the question in (26a), the reply in (26b) has a strangely comical effect, which can be explained if we can only refer to the speaker and the speaker’s associates, not to the speaker and others, regardless of how prominent the o is in the given context. In other words, by uttering (26b) Ad suggests that in some sense or other George Clooney is one of his associates (and that Peter is supposed to know this). This effect is absent in (26c). The latter example does not involve any boasting on Ad’s part regarding his connections with famous film stars, even though there may be boasting regarding his taste in coffee. (26) a. (Peter:) Do you know whether George Clooney likes good coffee? b. (Ad:) #Yes, we both drink Illy. c. (Ad:) Yes, he drinks Illy, just like me. Notice that in contexts where the non-i element included in we is indeed an associate of the speaker, the same example is of course fine. Thus, if

Interactions between Person and Number 87

Peter inquires about Ad’s partner Nadja, the following is a perfectly natural discourse: (27) a. (Peter:) Is Nadja keen on good coffee? b. (Ad:) Yes, we both drink Illy. The effect in (26) is somewhat subtle. It is confirmed, however, by a pilot experiment reported in Sonnaert to appear. In this experiment, 32 speakers rated examples like (26b) and (26c); (26b) was judged significantly worse. In a different context that established a relationship between George Clooney and the speaker, the example improved significantly. Thus, it is indeed the case that the reference of a first person plural pronoun cannot include any o’s. The same argument can be made for second person plural pronouns. For instance, in the context given for (26), it would be strange for Ad to reply Yes, you both drink Illy (compared to Yes, he drinks Illy, just like you).9 We now turn to the reverse situation, which shows that an associate cannot be treated as an o unless explicitly reintroduced as such. Consider the following scenario. While Peter and Ad are working, Ad receives a phone call from Nadja, his partner. Peter knows that the caller is Nadja, and Ad knows that Peter knows this. After Ad hangs up, he may turn to Peter and utter (28a), but it is decidedly odd for him to utter (28b). (28) a. Good news! We’re going to the opera on Friday. Nadja has just bought tickets. b. Good news! We’re going to the opera on Friday. #She’s just bought tickets. On our view, (28b) is degraded because she must refer to an o, while its intended referent is an associate of i. Notice that if Nadja is reintroduced separately in the discourse, as in the last sentence in (28a), she can be construed as an o, so that subsequent use of she is fine. For example, (28a) could be followed by (29). (29)  She says we have good seats. Notice that this continuation also shows that the reintroduction of the associate as an o does not mean it loses its status as associate; after all, Nadja is also still included in the reference of we in (29). This means we must allow an element that is not i or u to be represented both as an o and as an associate in the person space. This is not surprising, however. In section 2.3, we have already shown that the same individual can fulfill multiple discourse roles even within a single sentence.

88

Chapter 3

An alternative analysis of the effect in (28) might be based on the assumption that the reference of a third person singular pronoun must be introduced separately, rather than as part of the reference of a plural pronoun. This assumption is incorrect, however. Consider a variant of the situation discussed above. Suppose Ad receives a phone call that is about a couple who are acquaintances of both his and Peter’s. Let us assume that Ad knows that Peter knows what the call is about and who is calling. Ad may then utter (30), without this being particularly awkward. (30)  Good news! They’ve just bought a new home. She’s very happy about it. Taken together, the data discussed above show that it is necessary to distinguish between associates and o’s, and that what we call the associative effect is always present when a first or second person plural pronoun is used to refer to a plural group that does not consist exclusively of speakers and/or addressees. 3.3.2  Associative Plurals

As noted above, the associative effect is built into our analysis of the person system. However, there is an alternative view in the literature, which attributes the effect to the number system instead. This view is based on the existence of dedicated associative plurals, which occur with some regular R-expressions in a number of languages. Hungarian, for example, has two different plural forms of a proper name like János, as in (31) (from Corbett 2000:102). The plural in (31a) is the regular plural, leading to an interpretation in which there are multiple persons called János. The plural in (31b) is the associative plural, leading to a reference to János and his associates. (31) a. János-ok John-pl ‘Johns’ (more than one person called John) b. János-ék John-assoc.pl ‘John and associates’, ‘John and his group’, ‘John ’n them’ Given the existence of associative plurals, one could hypothesize that the associative effect observed with first and second person plural pronouns comes about as a result of these pronouns taking the associative plural (for discussion, see Nakanishi and Tomioka 2004, Vassilieva 2005, Nakanishi and Ritter 2009, and Kiparsky and Tonhauser 2012, among others). On this view, the person system itself can be reduced to just the notions i, u, and o, without needing

Interactions between Person and Number 89

the notion of associates of i and u. In its strongest version, this hypothesis would require first and second person pronouns to take an associative plural; a weaker version would allow pronouns to have either an associative or a regular plural. In this subsection, we will argue against this general idea. We maintain that plural in pronouns is the regular plural, and that the associative effect is induced by the person system. We begin with a short overview of how associative plurals can be formed from nouns, based on Corbett 2000 and Daniel and Moravcsik 2011. The following options exist. (i) There is a dedicated associative morpheme, without there being any other marker of plurality. (ii) There is a dedicated associative morpheme, combined with the regular plural marker. (iii) The regular plural marker allows for an associative interpretation. (iv) There is no overt number marking on the noun, while an agreeing predicate carries the regular plural marker. Option (i) was already illustrated by the Hungarian example in (31b). Option (ii) is found in Central Alaskan Yup’ik, as shown by the paradigm in (32) (adapted from Corbett 2000:109), which shows the regular number inflection of qayaq ‘kayak’ and the associative number inflection for the proper name cuna ‘Chuna’. The associative marker is -nku, while the regular plural marker is -t. Note that the associative marker can also be combined with the dual marker -k. (32) Associative and number morphology in Central Alaskan Yup’ik

Regular Associative

Plural

Dual

qaya-t cuna-nku-t

qaya-k cuna-nku-k

Option (iii) is illustrated by Turkish. Although Turkish uses the same morpheme -lar/-ler to mark regular and associative plurals, the two can be distinguished not only semantically but also through their morphosyntactic positioning. In a possessive construction, a regular plural reading results when the plural marker precedes the possessive suffix. When the plural marker follows the possessive suffix, an associative reading is triggered (see Görgülü 2011:72–73). (33) a. teyze-ler-im aunt-pl-1sg ‘my aunts’ b. abi-ler-im brother-pl-1sg ‘my brothers’

90

Chapter 3

(34) a. teyze-m-ler aunt-1sg-pl ‘my aunt and her family/associates/friends’ b. abi-m-ler brother-1sg-pl ‘my brother and his family/associates/friends’ Finally, option (iv) is instantiated by the Talitsk Russian example in (35a) (Bogdanov 1968, cited here from Corbett 2006:155). This example can be analyzed in two ways: it may be that the subject is syntactically singular, or that it carries a zero associative morpheme that in turn triggers a zero allomorph of the plural marker on the noun. The latter analysis may be the right one for Talitsk Russian, since an unmarked noun can also receive an associative plural reading when it is not the controller of plural agreement on the verb; this is demonstrated by the example in (35b), in which a plural pronoun refers back to the proper name Kuz’mu that occurs in the sentence-initial adjunct PP (Bogdanov 1968:71, adapted here from Corbett 2006:210). (35) a. Moj brat tam toža žy-ľ-i. my.masc.sg brother(masc.sg) there also live-pst-pl ‘My brother and his family also lived there.’ b. Pra Kuz’mu my šypka ab’is’n’it’ toža n’e možym, paš’imu about Kuz’m.acc we much explain.inf also not can.1pl why on’i n’e p’išut vam. they not write.3pl you.dat ‘About Kuz’ma and his associates we also can’t explain much, why they don’t write to you.’ Following Corbett and Mithun (1996) and Corbett (2000), we take option (ii) to be indicative of a need to separate associativity from number. What the Yup’ik data in (32) suggest is that an associative plural is created by first adding a distinct associative morpheme to a nominal phrase, and subsequently attaching a regular plural marker, as in (36). (36)   NMB

NMB ASSOC ASSOC

N

The associative morpheme generates a set consisting of the reference of the nominal, plus the associates of this referent. The number markers express the cardinality of this set in the usual way. Options (i), (iii), and (iv) can be seen as

Interactions between Person and Number 91

underlyingly identical to option (ii), but with a different morphophonological realization of the various morphemes. Option (i) may be a case of conflation or fusion of the associative and plural morphemes, or a case of zero spell-out of the plural morpheme in the context of the associative morpheme; option (iii) can be seen as a case of zero spell-out of the associative marker; finally, option (iv) could be a case of zero spell-out of both the associative and plural morphemes (as noted for Talitsk Russian above).10 On our characterization of the person system, one would not expect first and second person plural pronouns to be formed through addition of an associative marker to their singular counterparts (possibly followed by addition of a regular plural marker). The reason for this is that the person system itself accounts for the associative effect in such pronouns. In contrast, if the associative effect has its origin not in the person system but in the number system, we would expect first and second person plural pronouns to show signs that they can be composed in the same way as associative plurals of regular nouns. Moreover, to the extent that first and second person plural pronouns allow a homogeneous reading in which their reference contains only speakers or only addressees, we would expect that there is variation in how plural first and second person pronouns are formed. In particular, in that case the second person plural pronoun should have one form when it refers to multiple addressees only and another when associates are included; the same is true, mutatis mutandis, for first person plural pronouns, if they can indeed refer to a multiplicity of speakers (see chapter 2). We will now argue that the hypothesis that the associative effect is an inherent property of the person system yields the better empirical results. 3.3.3  Plurals of Pronouns Are Not Associative Plurals

To begin with, there are languages that have associative morphology, which nonetheless is not used in the plural of first and second person pronouns. An example is Hungarian.11 Recall that in Hungarian there is a dedicated marker for the associative plural (-ék) that is in complementary distribution with the marker used in the regular plural (-ak/-ok/-ek/-ök/-k).12 Eva Csato, cited by Corbett (2000:103), observes that, at least in the oblique cases, plural pronouns contain a distinct number marker that is identical to the regular plural marker, rather than to the associative one. The morphology of Hungarian pronouns is somewhat complicated, but it seems reasonable to assume that oblique pronouns are built up according to the template case + person + number, where the person forms seem to vary with number as well, possibly as the result of context-dependent allomorphy. For a detailed description of the system, see Spencer and Stump 2013 and the references given there. The thing to note in

92

Chapter 3

the present context is that the number marker that appears in the plural is the regular one, not the associative marker -ék. Consider, for example, the paradigm for dative pronouns in (37). (37) Hungarian dative pronouns

1 2 3

Singular

Plural

nek-em nek-ed nek-i

nek-ün-k nek-te-k nek-i-k

Crucially, the first person plural pronoun does not have a form that includes the associative marker. Both forms in (38) are ungrammatical. The first is built using the regular case marker, the plural first person allomorph, and the associative plural marker; the second is built using the case marker, the singular first person allomorph, and the associative marker. Similar considerations apply to the form of the second person plural pronoun. (38) a. *nek-ün-ék dat-1pl-assoc.pl b. *nek-em-ék dat-1-assoc.pl The argument probably extends beyond oblique forms to nominative and accusative pronouns. In the accusative, the case morpheme follows, rather than precedes, the person and number morphemes (person + number + case). Although there are some irregularities that could be dealt with by using allomorphy rules, it is clear that the number morpheme is best analyzed as the regular plural marker -(V)k. Nowhere in these forms is there a trace of the associative marker -ék. The nominative first and second person forms show no transparent number marking in the standard language. However, there is a substandard first person form min-k that again contains the regular rather than the associative plural morpheme. The relevant data are given in (39). (39) Hungarian nominative and accusative pronouns Nominative

1 2 3

Accusative

Singular

Plural

Singular

Plural

én te ő

mi/min-k ti ő-k

engem(-et) téged(-et) ő-t

min-k-et ti-tek-et ő-k-et

Interactions between Person and Number 93

In conclusion, the associative plural marker that Hungarian possesses does not show up in any plural pronoun, whereas the regular plural marker does. The second pattern of associative plurals mentioned above involves the combination of a distinct associative marker and regular number markers, as found in Central Alaskan Yup’ik. As before, if the associative effect in first and second person plural pronouns is a consequence of the use of the associative plural, we would expect this pattern to be reflected in the pronominal morphology. The relevant part of the pronominal paradigm of Central Alaskan Yup’ik is given in (40) (from Miyaoka 2012:383). (40) Central Alaskan Yup’ik first and second person pronouns

1 2

Singular

Plural

Dual

wii/wiiŋa iɬpit

waŋkuta iɬpici

waŋkuk iɬpitik

As shown in (32), the associative marker in Central Alaskan Yup’ik is -nku. In the second person plural and dual, there clearly is no evidence for the presence of this marker. The first person pronouns contain a sequence ŋku that would appear to resemble this marker. However, there is a plausible historical explanation for this. The first person pronoun in Central Alaskan Yup’ik is probably related—at least diachronically—to a demonstrative meaning something like ‘this one here’ (see Corbett and Mithun 1996:15 and Miyaoka 2012:383). The plural of demonstratives is built by first combining the stem with a “nonsingular expander” (Miyaoka 2012:384), followed by merger of the regular plural. This nonsingular expander takes the form -ku for demonstratives. Therefore, the proper segmentation of waŋkuta is likely to be waŋ-ku-ta. Notice that on this analysis the -nku marker is absent. It is not a coincidence that the associative marker shares part of its form with the nonsingular expander. Associative markers are often diachronically related to regular plural markers (sometimes combined with another element; see note 12), and indeed Miyaoka (2012:713) suggests that -nku contains -ku, at least historically (see also the appendix in Corbett and Mithun 1996). It is possible, even, that the apparent associative marker is still synchronically complex and consists of an actual associative marker -n followed by a nonsingular expander -ku. However that may be, there is no convincing evidence for the presence of the associative marker in the first person plural (and dual) pronouns of Central Alaskan Yup’ik.13 Next, consider Turkish. In this language, there is a semantic restriction on the use of the associative plural: it only allows collective readings, whereas

94

Chapter 3

the regular plural is compatible with both distributive and collective readings. Recall that Turkish exhibits no morphemic distinction between the associative and the regular plural, but the two can be distinguished on the basis of the position they occupy with respect to possessive markers; see (33) and (34). Consider now the following data from Görgülü 2011:74. The examples in (41a) and (42a) show the regular plural reading, while those in (41b) and (42b) illustrate the associative plural reading. The regular plural is compatible with a distributive reading. For instance, (41a) can describe a situation in which each of the speaker’s aunts left at a different time. By contrast, the associative plural does not allow a distributive reading. Thus, (41b) cannot describe a situation in which each individual member of the group consisting of the speaker’s aunt and her associates left at a different time. While this restriction is not a universal property of associative plurals, Görgülü (2011) shows that it holds consistently of associative plurals in Turkish. (41) a. Teyze-ler-im çık-tı. aunt-pl-1sg.poss leave-pst ‘My aunts left.’ i. My aunts left together. ii. My aunts left at different times. b. Teyze-m-ler çık-tı. aunt-1sg.poss-pl leave-pst ‘My aunt and her family/friends/associates left.’ i. My aunt and her associates left together. ii. *My aunt and her associates left at different times.

(collective) (distributive)

(collective) (distributive)

(42) a. Abi-ler-im Ankara-ya git-ti. brother-pl-1sg.poss Ankara-dat go-pst ‘My brothers went to Ankara.’ i. My brothers went to Ankara together. (collective) ii. My brothers went to Ankara at different times. (distributive) b. Abi-m-ler Ankara-ya git-ti. brother-1sg.poss-pl Ankara-dat go-pst ‘My brother and his family/friends/associates went to Ankara.’ i. My brother and his associates went to Ankara (collective) together. ii. *My brother and his associates went to Ankara (distributive) at different times. With this in mind, consider the possible readings of first and second person plural pronouns in Turkish. Crucially, these do allow distributive readings, as shown by the examples in (43) (from Balkız Öztürk, personal communication).

Interactions between Person and Number 95

(43) a. Biz farklı zaman-lar-da gel-di-k. we different time-pl-loc come-pst-1pl ‘We came at different times.’ b. Siz farklı zaman-lar-da var-dı-nız. you.pl different time-pl-loc arrive-pst-2pl ‘You arrived at different times.’ This contrast between associative plurals and plurals of first and second person pronouns militates against a unification of the two. Consider finally Talitsk Russian, a language with associative plurals that are expressed by absence of number marking on the noun in the presence of a plural marker on the predicate in case this predicate agrees with the noun (Corbett 2006:155; see (35) above). If the associative effect in first and second person plural pronouns is the result of attaching an associative plural, we would expect the pattern in (35) to carry over to such pronouns. This is not the case, however. While we have not been able to consult native speakers of Talitsk Russian, we have consulted Russian speakers who accept associative plurals of the type in (35). However, these same speakers do not accept pluralization of a first or second person pronoun by using the singular form of the pronoun in combination with a plural verb. Again, then, the supposed parallel between associative plurals and plurals of first and second person pronouns breaks down. To summarize our findings so far, in none of the languages that we have used to illustrate the various patterns of associative plural formation is there any evidence that plural first and second person pronouns follow an extension of this pattern. In fact, there is evidence to the contrary, as in some of these languages these pronouns are built using the regular plural. There is an additional argument against the hypothesis that first and second person pronouns typically have associative plurals. As noted previously, such pronouns permit two construals. The first is a heterogeneous one, in which the group referred to includes associates of speaker(s) or addressee(s). The second is homogeneous: it exclusively consists of a multiplicity of speakers or addressees. If the associative effect is a result of using an associative plural, then the homogeneous readings could show a form that is different from the form used for the heterogeneous readings, as only the former can be derived using a regular plural. This means that in principle we might expect there to be two different forms of first and second person plural pronouns (at least in languages that have an associative plural distinct from the regular plural), linked to their heterogeneous and homogeneous readings. In contrast, the theory outlined in chapter 2 does not make it possible to formally distinguish these readings, because they are characterized by exactly

96

Chapter 3

the same person feature or features. The different readings result from the optionality in the way the various subsets of the person space are populated. Si must contain one instance of i, but may additionally contain associates and/ or further instances of i. Likewise, Si+u – Si must contain one instance of u, but may additionally contain associates and/or further instances of u. As a consequence, the same feature specification that selects Si permits heterogeneous and homogeneous construals in the plural; the same holds for the feature specification that selects Si+u – Si. The crosslinguistic data appear to be in line with our proposal. The relevant universals, defended by Bobaljik (2008a), were given in chapter 2 and are repeated here for convenience. (44)  Universal 1: No language distinguishes [i+i] from [i+ai]. Universal 2: No language distinguishes [u+u] from [u+au]. Universal 1 states that what we have called a homogeneous reading for first person plural pronouns is never formally distinguished from what we have called a heterogeneous reading.14 Universal 2 says the same about second person plural pronouns. These observations are unexpected if the associative effect in pronouns is the result of applying a special associative plural. We conclude that the associative effect observed with plural first and second person pronouns does not result from attachment of an associative marker. This does not exclude the possibility that a language, in addition to having regular plural pronouns, can form an associative plural on the basis of first and second person singular pronouns. However, the result of this derivation should be formally distinct from the regular plural and is likely to also be distinct in its semantics and/or pragmatics, given that associative plurals typically involve close, permanent associates. The point can be illustrated on the basis of Dutch. While this language does not have the morphological means to express associative plurals, there is a specialized syntactic form that can be used instead. As shown in (45), the structure in question is a coordination whose second conjunct consists of the definite determiner and a plural possessive pronoun. It is available for all human DPs. We find some of the associative forms in (45) somewhat marginal, but all are attested (as a Google search confirms). However this may be, they are very clearly different in form from the regular plural pronouns in the right-hand column. (45) a.

Associative ik en de mijnen I and the mine.pl

Regular wij we

Interactions between Person and Number 97

b.

jij en de jouwen you and the yours.pl c. u en de uwen you.hon and the yours.hon.pl d. hij en de zijnen he and the his.pl e. % zij en de haren she and the hers.pl f. wij en de onzen we and the ours.pl g. % jullie en de jouwen you.pl and the yours.pl h. % zij en de hunnen they and the theirs.pl i. Piet en de zijnen Piet and the his.pl j. Marie en de haren Marie and the hers.pl

jullie you.pl u you.hon.pl zij they zij they — — — Pieten Piet.pl Maries Marie.pl

In addition to the requirement that the left conjunct be human, the associative construction in (45) is subject to a restriction familiar from associative plurals in other languages: the persons denoted by the right conjunct must be family or other permanent close associates of the person denoted by the left conjunct. In contrast, as we have shown in section 3.1, the associates contained in a first/ second person plural pronoun do not have to be permanent; they are contextdependent. For example, imagine a situation in which Jan is joining a group of colleagues that happens to consist of Kees, Marie, and Anna. In that situation, it is perfectly natural for Kees to say (46a) to Jan, with the first person plural pronoun referring to himself, Marie, and Anna. It is not possible for him to use (46b) if he intends to refer to the same group (unless Marie and Anna, in addition to being colleagues, happen to be his wife and daughter). (46) a.

Wij gaan naar de film; heb je zin om mee te gaan? we go to the film have you appetite for along to go ‘We are going to the movies; do you want to come along?’ b. # Ik en de mijnen gaan naar de film; heb je zin om I and the mine go to the film have you appetite for mee te gaan? along to go ‘Me and mine are going to the movies; do you want to come along?’

98

Chapter 3

A very similar situation can be observed in Haruai, a language that is like Talitsk Russian in that associative plurals are not overtly marked on DPs, but trigger plural agreement. Probably as opposed to Talitsk Russian (see above), Haruai can generate an associative plural on the basis of the first person singular pronoun. Note, however, that this form is distinct from the regular first person plural pronoun. This is illustrated in (47a) (from Corbett 2000:192, attributed there to Bernard Comrie) and (47b) (from Comrie 1988:146). (Corbett gives the translation in (47a), but notes in the text around the example that the others must be permanent associates of the speaker.) (47) a. N dy-n-ŋ-a. I go-fut-1pl-dec ‘I and some others will go.’ b. An hön pay-n-ŋ-a. we pig hit-fut-1pl-dec ‘We will hit the pig.’ Of course, such clear distinctions between the actual associative plurals of singular pronouns and the regular plural pronouns of a language further confirm that the associative effect with regular plural pronouns should not be attributed to associative plural marking. 3.3.4  Plural Pronouns in Mandarin and Japanese

The conclusion of the previous subsection goes against claims that have been made in the literature on Japanese and Mandarin Chinese plural marking. In these languages, plural first and second person pronouns (and third person pronouns in Mandarin) are obligatorily marked for number. The morpheme used for this appears to generate associative plurals elsewhere—for example, when it is attached to proper names. We illustrate this in (48) for Mandarin and in (49) for Japanese. (48) a. Mandarin pronouns

1 2 3

(from Chappell 1996:471)

Singular

Plural

wŏ nĭ tā

wŏ-men nĭ-men tā-men

b. XiăoQiáng-men XiaoQiang-pl ‘XiaoQiang’s group’

(from Lü 1980:342, as cited in Iljic 1994:95)

Interactions between Person and Number 99

(49) a. Japanese pronouns

(adapted from Ishii 2000:1)

Singular

Plural

1 2 3masc

watasi anata kare

3fem

kanozyo

3neut

sore

watasi-tati anata-tati kare-ra/ ??kare-tati kanozyo-ra/ kanozyo-tati sore-ra

b. Taroo-tati Taroo-pl ‘Taroo and those in his group’

(from Ishii 2000:2)

The Japanese paradigm given in (49a) has been simplified for presentational purposes. Japanese has a wide variety of pronouns for all three persons and uses four different number markers that are distinguished in terms of their honorificity.15 We will look at some of these number markers in more detail below. On the assumption that the associative effect in plural first and second person pronouns is a property of the person system rather than the number system, it would seem superfluous for plural pronouns to obligatorily carry an associative marker. The obvious alternative analysis would be to say that suffixes like -tati and -men are regular plurals that can, but need not, be combined with a null associative marker. (This is close to the analysis that Ueda and Haraguchi (2008) and Nomoto (2013) propose; these authors argue that there are two distinct suffixes -tati, one a regular plural, the other an associative plural.) If this analysis is correct, the fact that -tati and -men are attached to singular pronouns to form their plural counterpart does not show that this plural is associative. We will now present some arguments that favor this alternative view. At first sight, the conclusion seems unavoidable that -tati in Japanese and -men in Mandarin do indeed have a nonassociative use, simply because they appear in plural DPs that have a uniform (additive plural) interpretation. Japanese and Mandarin examples are given in (50a) (adapted from Nakanishi and Tomioka 2004:113) and (50b) (adapted from Zhang 2008:408), respectively. (50) a. Otokonoko-tati-ga asonde-iru. boy-pl-nom play-prog ‘The boys are playing.’ b. Jiazhang-men zou le. parent-pl leave asp ‘The parents have left.’

100

Chapter 3

In these examples, the subjects refer to groups consisting only of boys and parents, respectively. They do not include reference to nonboy or nonparent associates. Nakanishi and Tomioka (2004) argue that the conclusion that -tati­ can be used as a regular plural marker is not warranted, however. They propose that -tati always derives a set of individuals whose members are represented by the noun to which the affix attaches. This notion of representation has two instances. In a clear associative plural like (49b), the noun represents other members of the set because the other members are associates of the person referred to by the noun. In an apparently uniform plural like (50a), the noun represents the other members of the set because all members of the set share the property of boyhood encoded by the noun. This proposal manages to unify the different usages of -tati in the above examples, but it does so by expanding the possible readings of associative plurals to readings traditionally regarded as nonassociative. Be that as it may, the crucial prediction of the claim that -tati and -men are exclusively associative is that they will always at least allow nonuniform readings. However, this is not correct.16 The most straightforward case is Mandarin -men, which simply cannot have an associative plural reading when attached to common nouns. As noted by Nomoto (2013:65), “[U]nlike forms with -tati, forms with -men only have a uniform reading [when attached to common nouns]. For example, xuesheng-men ‘student-pl’ cannot be used to refer to a group containing some non-students, no matter how (the speaker thinks) these non-student individuals are related to the students. This fact indicates that -men attaching to common nouns is a plural marker, as Li (1999) claims.” Even in Japanese, there are cases where -tati is contained in a DP that does not allow an associative plural reading. Nomoto (2013:67) points out that examples like (51) can only have the uniform reading in (51a). The associative readings in (51b) and (51c) are not available. (51) san nin-no gakusei-tati three clf-lnk student-pl a. ‘three students’ (uniform; no more than three people; no nonstudents) b. *‘three people, including at least one nonstudent, representative of students’ c. *‘a group consisting of three students and their nonstudent associates’ Nomoto suggests that -tati can only be associative when attached to a referential DP.

Interactions between Person and Number 101

As noted, the alternative to assuming that -tati and -men are associative plural markers would be to say that Japanese and Mandarin have null associative morphemes whose affixation triggers subsequent attachment of a regular plural marker.17 In Mandarin, the associative morpheme selects proper names only, while in Japanese it selects any referential human noun. (A contrast along these lines is not unexpected; after all, associative plurals crosslinguistically impose different semantic restrictions on the category they attach to; see Daniel and Moravcsik 2011.) The facts discussed above follow from such an analysis. In Mandarin, a regular plural marker -men can be attached to both nouns and pronouns. The regular plural marker can also be attached to proper names (see (52), from Zhang 2008:408), giving a uniform plural reading as in (52a). Finally, the regular plural marker can be attached to a proper name initially affixed by a null associative morpheme, giving the associative plural reading in (52b) (cf. (48b)).18 (52) XiăoQiáng-men XiaoQiang-pl a. ‘the group of people all named XiaoQiang’ b. ‘XiaoQiang and other people related to him’ In Japanese, the same possibilities exist. Most of these were already illustrated above, except for the uniform plural reading of a proper name affixed by -tati; this reading is indeed possible (Nomoto (2013:64) states that Tanakatati can mean both ‘Tanaka and others’ and ‘more than one person named Tanaka’). In addition, Japanese allows a number of forms not attested in Mandarin, because the null associative morpheme can attach to any human noun. Since the associative morpheme triggers plural affixation, -tati occurs in more associative plurals than -men does in Mandarin. For example, gakuseitati can mean ‘the students and some nonstudents associated with the students’ (see Nomoto 2013:1). If there are two distinct morphemes in Japanese, a regular plural marker -tati and a null associative morpheme with a fairly unrestricted distribution, we predict the possibility of apparent stacking of the plural morpheme. Since a plural of a human nominal is human itself, the null associative morpheme can attach to a nominal affixed by plural -tati. As always, the null associative morpheme then triggers plural marking, resulting in a sequence N-tati-∅-tati or N-∅-tati-∅-tati. This possibility does indeed exist, as the examples in (53) (from Ueda and Haraguchi 2008:237) show. On the readings given, the example in (53a) has the structure gakusei-tati-∅-tati, while (53b) has the structure Taroo-∅-tati-∅-tati.19

102

Chapter 3

(53) a. gakusei-tati-tati student-pl-pl ‘the students and their associates’ b. Taroo-tati-tati Taroo-pl-pl ‘Taroo and his associates and their associates’ As noted above, -tati is just one of four plural markers in Japanese, which differ in honorificity. This means that double plural marking can take a variety of shapes. In fact, the sequence -tati-tati is often dispreferred compared to sequences in which the two plural markers are phonologically distinct. For example, gakusei-tati-ra ‘student-pl-pl’ is preferred over gakusei-tati-tati ‘student-pl-pl’ (Nomoto 2013:69). The examples in (53) have the structures in (54). It is worth noting that the recursion of assoc and nmb required for the analysis of these examples contradicts the idea that there is a strict sequence of functional heads; it is more in line with theories that capture the distribution of functional heads in terms of their syntactic and semantic selectional properties (as assumed here). (54) a.

NMB ASSOC NMB

NMB

ASSOC NMB

N   b.

NMB ASSOC NMB ASSOC N

NMB

ASSOC NMB

ASSOC

Stacking of the number morpheme is not possible in Mandarin, as the examples in (55) (from Ueda and Haraguchi 2008:237) show. (55) a. *xuesheng-men-men student-pl-pl ‘the students and their associates’ b. *XiăoQiáng-men-men XiaoQiang-pl-pl ‘XiaoQiang and his associates and their associates’

Interactions between Person and Number 103

The ungrammaticality of these examples follows from the selectional restriction of the null associative morpheme in Mandarin. It can only attach to proper names. This immediately rules out (55a). We contend that it also rules out (55b), as the plural of a proper name is not itself a proper name. If the analysis of affixes like -tati and -men as regular plural markers is on the right track, then of course their presence in plural pronouns does not support the conclusion that the plural of pronouns is the associative plural. In fact, we can go one step further. Japanese provides a limited, but persuasive, argument that plural marking on pronouns is not necessarily associative. Among the various plural morphemes, there is one that is not compatible with associative readings. This is the highly honorific form -gata. In our terms, -gata selects for nouns not affixed by the associative morpheme.20 This property of -gata is illustrated in (56) (from Iwasaki 2013:58). (56) a. sensee-gata teacher-pl.hon ‘teachers’ b. *Tanaka-sensee-gata Tanaka-teacher-pl.hon ‘Professor Tanaka and his associates’ Nevertheless, -gata can attach to second person pronouns (see Downing 1996:194). (57) anata-gata you-pl.hon ‘you all (respectful)’ At least in this case, then, the conclusion seems inescapable that pronouns permit nonassociative plurals.21 We conclude that -men in Mandarin and -tati in Japanese are regular plural markers, and that their apparent associative use is the consequence of previous attachment of a null associative morpheme.22 If this analysis is on the right track, these languages do not provide evidence for the claim that associative plurals must be used to build the plural of first and second person pronouns. The only observation that might seem to clash with this conclusion is that these plural markers in Japanese and Mandarin do not behave entirely in the same way as the regular plural marker in a language like English. In particular, whereas plural marking on pronouns is obligatory in Japanese and Mandarin, it seems to be optional on regular nouns (but subject to semantic restrictions involving definiteness or specificity; see Li 1999, Ishii 2000, Kurafuji 2004, Nakanishi and Tomioka 2004, Hosoi 2005, Ueda and Haraguchi 2008, and Nomoto 2013 for discussion).23 In such a system, unmarked nouns are often

104

Chapter 3

referred to as having “general number” (see, for instance, Corbett 2000:9–19): they impose truth conditions compatible with situations in which English would use either singular or plural. This contrast between pronouns and regular nouns in Japanese is used by Nakanishi and Ritter (2009) to support the proposal that first and second person pronouns can be pluralized only by using an associative plural marker, while for other nouns this is merely an option. However, it is quite common for languages with a general number system to have subsets of nouns for which number marking is obligatory. For example, general number systems are widespread in Iranian languages, but not for nouns denoting humans, which have a singular-plural system (Corbett 2000:74, referring to Smirnova 1981). It is quite possible, then, that a language with general number may single out pronouns as a domain in which number marking is obligatory. It is, after all, common for pronouns to be marked for features not expressed on nouns in the same language. In English, for example, pronouns are the only nominals marked for case and gender. 3.4 Conclusion

Two main conclusions can be drawn from this chapter with regard to person. The first is that the associative effect in plural pronouns is real and that it is an inherent property of the person system, rather than a consequence of pronouns taking an associative plural. The nature of the input set for the person system, Si+u+o, is such that in Si all nonspeaker elements must be associates of the speaker and in Si+u all nonaddressees must be associates of the addressee. The second conclusion has to do with the structural position of the prs head with respect to the nmb head. In chapter 2, we argued that prs cannot c-command nmb and that in pronouns nmb cannot be present in the absence of person features. In this chapter, we have argued that languages nonetheless vary in the configuration of prs and nmb, in that some have pronouns in which nmb c-commands prs, while others have pronouns in which nmb and prs are conflated, so that person and number information is contained within a single functional head. We showed that this difference has consequences for the interpretation of the universal number feature aug. In languages in which nmb occurs in a functional head distinct from prs, aug indicates a cardinality larger than 1 in all pronouns. In languages in which nmb and prs are conflated, the semantic effect of aug depends on what person features a pronoun carries: it indicates a cardinality larger than the minimal cardinality allowed by the person specification.

4

Impersonal Pronouns

4.1 Introduction

So far, we have only discussed the referential use of pronouns and their use as bound variables. However, some of the same pronouns for which we have proposed a system of person features are also used as impersonal pronouns. When they are used in this way, their interpretation shifts. For instance, the pronoun you in English is used as a generic impersonal pronoun in examples like You should never drive on the wrong side of the road. In the impersonal reading, you does not refer to the addressee; rather, it means something like ‘people in general’. There is an obvious challenge here for any theory of person, as it must be explained how the same pronoun can have such different interpretations. This challenge becomes even more serious when one considers two further observations. First, not all personal pronouns have an additional impersonal use. For instance, third person singular he/she must be referential in an example like He/She should never drive on the wrong side of the road; it does not permit the kind of generic reading just discussed. Second, various languages have dedicated impersonal pronouns, that is, pronouns exclusively used for impersonal readings. A case in point is English one in sentences like One should never drive on the wrong side of the road. Any theory of pronouns must accommodate the emerging pattern of data. In this chapter, we will argue that the feature system developed in chapter 2 for personal pronouns does not need to be adapted to account for the impersonal use of personal pronouns, nor for the existence of pronouns that are exclusively impersonal. However, we do need to make full use of the options available in the feature system, allowing pronouns that are unspecified for person features, as well as pronouns in which person features are applied in parallel rather than sequentially. As we will demonstrate, the maximum stretch that our proposal allows without tampering with the definitions of the person

106

Chapter 4

features is sufficient to generate an adequate typology of impersonal pronouns and their various interpretations. The chapter is organized as follows. In section 4.2, we provide an overview of the data. In particular, we will establish which personal pronouns allow impersonal readings, and what the readings are that are available for dedicated impersonal pronouns. In sections 4.3–4.6, we then provide an analysis. In section 4.3, we discuss impersonal interpretations of plural pronouns. In section 4.4, we focus on a type of dedicated impersonal pronoun that allows both generic readings (‘people in general’) and arbitrary readings (‘someone or other’). In section 4.5, we give an account of the generic use of second person singular pronouns. In section 4.6, we discuss a second type of dedicated impersonal pronoun, which allows only generic readings. In section 4.7, we discuss why a number of impersonal pronouns have the additional restriction that they must refer to human or at least sentient individuals. Section 4.8 concludes. 4.2 Overview 4.2.1  The Distribution of Generic and Arbitrary Readings

Impersonal pronouns can have various different readings (see, for instance, Cabredo Hofherr 2003 for an overview). The terminology used in the literature for these readings varies considerably, but we can make a basic distinction between what we will label the generic interpretation and what we will label the arbitrary interpretation. The generic reading is exemplified by one in One should not give matches to young children. The arbitrary reading is illustrated by they in They called for you, but they refused to leave their name (note that some speakers of British English do not accept this reading). The table in (1) presents an overview of the interpretations that various pronouns permit. (The referential reading is the regular personal interpretation discussed in the previous chapters.) We use the term imp to refer to any impersonal pronoun that has a dedicated form (that is, any impersonal pronoun not identical to any of the personal pronouns in the language). Not all dedicated impersonal pronouns are alike in the readings they allow. Two types must be distinguished; the first is exemplified by English one and the second by German man. We will refer to these types as imp-1 and imp-2, respectively.1 The table indicates what readings we will argue are available for the various types of pronouns, but of course this does not imply that every language has the full inventory of options identified here.

Impersonal Pronouns 107

(1) 1sg 2sg 3sg 1pl 2pl 3pl imp-1 imp-2

Referential

Generic

Arbitrary

yes yes yes yes yes yes no no

no yes no yes yes yes yes yes

no no no no no yes no yes

We illustrate the various (im)possibilities in (2)–(9). The data are all taken from Dutch, except for those in (8), which illustrate the use of imp-1 pronouns (which Dutch does not have). (2) a. *Op een begrafenis in Italie draag ik nette kleren. (*1sg gnr) at a funeral in Italy wear I formal clothes ‘At a funeral in Italy, people (including me) always wear formal clothing.’ b. *Ik heb voor je gebeld, maar het was niet duidelijk (*1sg arb) I have for you called but it was not clear waarover. where.about ‘Someone (of the group to which I belong) has called for you, but it was not clear what it was about.’ (3) a. Op een begrafenis in Italie draag je nette kleren. (2sg gnr) at a funeral in Italy wear you formal clothes ‘At a funeral in Italy, people (including you and me) always wear formal clothing.’ b. *Je hebt voor ’m gebeld, maar het was niet duidelijk (*2sg arb) you have for him called but it was not clear waarover. where.about ‘Someone (of the group to which you belong) has called for him, but it was not clear what it was about.’ (4) a. *Op een begrafenis in Italie draagt-ie/ze nette kleren. (*3sg gnr) at a funeral in Italy wears-he/she formal clothes ‘At a funeral in Italy, people always wear formal clothing.’ b. *Hij/Ze heeft voor je gebeld, maar het was niet (*3sg arb) he/she have for you called but it was not duidelijk waarover. clear where.about ‘Someone has called for you, but it was not clear what it was about.’

108

Chapter 4

(5) a. Op een begrafenis dragen we hier nette kleren. (1pl gnr) at a funeral wear we here formal clothes ‘At a funeral here, people (including me) wear formal clothing.’ b. *We hebben voor je gebeld, maar het was niet duidelijk (*1pl arb) we have for you called but it was not clear waarover. where.about ‘Someone (from the group I belong to) has called for you, but it was not clear what it was about.’ (6) a. Op een begrafenis daar dragen jullie nette kleren. (2pl gnr) at a funeral there wear you.pl formal clothes ‘At a funeral there, people (including you) always wear formal clothing.’ b. *Jullie hebben voor ’m gebeld, maar het was niet (*2pl arb) you.pl have for him called but it was not duidelijk waarover. clear where.about ‘Someone (from the group you belong to) has called for him, but it was not clear what it was about.’ (7) a. Op een begrafenis in Italie dragen ze nette kleren. (3pl gnr) at a funeral in Italy wear they formal clothes ‘At a funeral in Italy, people always wear formal clothing.’ b. Ze hebben voor je gebeld, maar het was niet (3pl arb) they have for you called but it was not duidelijk waarover. clear where.about ‘Someone has called for you, but it was not clear what it was about.’ (8) a. At a funeral in Italy, one wears formal clothing. b. *One called for you, but it wasn’t clear what it was about.

(imp-1 gnr) (*imp-1 arb)

(9) a. Op een begrafenis in Italie draagt men nette kleren. (imp-2 gnr) at a funeral in Italy wears imp formal clothes ‘At a funeral in Italy, one always wears formal clothing.’ b. Men heeft voor je gebeld, maar het was niet (imp-2 arb) imp has for you called but it was not duidelijk waarover. clear where.about ‘Someone has called for you, but it was not clear what it was about.’

Impersonal Pronouns 109

Several generalizations apparent from the table in (1) require an explanation. One is that all plural pronouns allow at least a generic reading, whereas most singular pronouns must be referential. There are two notable exceptions to the latter part of this generalization, however. Second person singular pronouns permit generic readings in a variety of languages (see the Dutch example in (3a)).2 In addition, dedicated impersonal pronouns are syntactically singular; the English and Dutch examples in (8) and (9) are typical in this respect. In turn, second person singular pronouns and dedicated impersonal pronouns are different in two ways. The former permit only a generic reading, while the latter come in two types, one of which (imp-2) allows an additional arbitrary reading. Moreover, dedicated impersonal pronouns systematically trigger third person agreement. Before we turn to our analysis of these generalizations, we discuss a number of observations that appear to complicate the empirical picture as presented above. 4.2.2  Apparent Exceptions

The three phenomena we discuss in this subsection are the apparent generic use of first person singular pronouns, the apparent referential use of impersonal pronouns, and the apparent arbitrary use of first and second person plural pronouns. We will argue that the data in question do not necessitate adjusting the generalizations outlined in the previous subsection. 4.2.2.1  Can First Person Singular Pronouns Be Generic?

Zobel (2012) argues that German has impersonal readings of first person singular pronouns. The kind of example that illustrates this is also attested in Dutch (although we find such examples somewhat marked). (10) is a case in point. (10) Ik kan toch als bruidspaar niet verwachten dat de gasten I can yet as bride.and.groom not expect that the guests het feest betalen. the party pay ‘One cannot expect the guests to pay for the party when one gets married.’ This example has a generic flavor, because it suggests that no bridal couple should expect their guests to pay for the party, as indicated by the English translation. Nevertheless, we believe the ‘I’ used here is the regular referential one, rather than an impersonal pronoun. Our analysis of examples like this hinges on the presence of the ‘as’-phrase. The function of this phrase is to

110

Chapter 4

describe a guise of the associated argument. This is apparent from examples like (11), where burgemeester ‘mayor’ is one of the functions that the subject has. (11) Ze kan als burgemeester toch niet verwachten dat de gasten she can as mayor yet not expect that the guests het feest betalen. the party pay ‘As a mayor, she cannot expect her guests to pay for the party.’ The generic flavor of this example comes about because the predicate holds of the subject in the guise it adopts and not necessarily in other guises. The implication of this is that, according to the speaker, mayors in general should not expect their guests to pay for their parties. The guise introduced by the ‘as’-phrase can (though it need not) be imaginary, in that the associated referent need not actually have the attribute described. If the guise is imaginary, then access to the referent’s mind is required to make a felicitous statement. This explains why a variant of (10) that contains a third person pronoun is impossible (in contrast to (11)). (12) *Ze kan als bruidspaar toch niet verwachten dat de gasten she can as bride.and.groom yet not expect that the guests het feest betalen. the party pay ‘One cannot expect the guests to pay for the party when one gets married.’ If this suggestion is on the right track, the prediction is that, at least in assertions, second person pronouns cannot be associated with an ‘as’-phrase either if the latter expresses an imaginary guise. After all, a speaker has no access to the mind of the addressee. However, the speaker can ask a question whose answer requires access to the addressee’s mind. Therefore, we expect interrogative counterparts of (10) with second person pronouns to be felicitous. There is a complication in testing this. While we assume that the pronoun in (10) is referential, the construction in which it appears also allows for impersonal subjects, and the impersonal second person is no exception.3 (13) a. Men kan toch als bruidspaar niet verwachten dat de gasten one can yet as bride.and.groom not expect that the guests het feest betalen. the party pay ‘One cannot expect that one’s guests will pay for the party when one gets married.’

Impersonal Pronouns 111

b. Je kan toch als bruidspaar niet verwachten dat de gasten you can yet as bride.and.groom not expect that the guests het feest betalen. the party pay ‘One cannot expect that the guests will pay for the party when one gets married.’ However, the prediction can be tested by using the strong form of the second person pronoun, as this resists impersonal use (unless it is contrastive; see note 2). The expected difference in acceptability between assertions and questions is indeed observed.4 (14) a. *Jij zou toch als bruidspaar niet verwachten dat de gasten you would yet as bride.and.groom not expect that the guests het feest betalen. the party pay ‘One wouldn’t expect the guests to pay for the party when one gets married.’ b. Zou jij als bruidspaar verwachten dat de gasten het feest would you as bride.and.groom expect that the guests the party betalen? pay ‘Would one expect the guests to pay for the party when one gets married?’ The proposed analysis of examples like (10) and (14b) is supported by the observation that the apparent generic reading disappears when the ‘as’-phrase is omitted. (15) a. Ik kan toch niet verwachten dat de gasten het feest betalen. I can yet not expect that the guests the party pay ‘I cannot expect the guests to pay for the party.’ *‘One cannot expect the guests to pay for the party.’ b. Zou jij verwachten dat de gasten het feest betalen? would you expect that the guests the party pay ‘Would you (referential) expect the guests to pay for the party?’ *‘Would one expect the guests to pay for the party?’ Zobel (2012:32–33) explicitly claims that the ‘as’-phrase is optional in examples of the relevant type, which would be incompatible with our analysis. However, she notes that for the first person “the impersonal interpretation seems to be less readily available” in the absence of an ‘as’-phrase. For us, the Dutch equivalent of the example she cites does not allow the apparent

112

Chapter 4

generic reading at all. An example parallel to (15a) with a second person pronoun can indeed have a generic interpretation. However, this is not surprising, as there is after all a generic second person pronoun, which can occur whether or not an ‘as’-phrase is present. (16) Je kan toch niet verwachten dat de gasten het feest betalen. you can yet not expect that the guests the party pay ‘You (referential) cannot expect the guests to pay for the party.’ ‘One cannot expect one’s guests to pay for the party.’ As expected, the generic reading in (16) disappears when the strong pronoun jij is used instead.5 Further examples illustrating the contrast between first and strong second person singular pronouns on the one hand and weak second person singular pronouns on the other hand are given in (17) and (18). The sentence in (17) is of the same type as (10). The point of (18) is that even a context that steers the addressee toward a generic interpretation is not enough to license the apparent generic use of the first person pronoun or the strong second person pronoun in the absence of an ‘as’-phrase or equivalent. (17) ?Als boer in de 15e eeuw moest ik de hele winter doorwerken. as farmer in the 15th century must I the whole winter through.work ‘Farmers in the 15th century had to continue working throughout the winter.’ (18) (Context: ‘Do you know anything about the life of farmers in the 15th century?’) a. *Nou, ik/jij moest de hele winter doorwerken. well I/youstrong must the whole winter through.work ‘Well, one had to work throughout the winter.’ b. Nou, je moest de hele winter doorwerken. well youweak must the whole winter through.work ‘Well, one had to work throughout the winter.’ In sum, the phenomenon at hand does not seem to involve an impersonal use of the first person singular pronoun. Rather, what seems to be crucial is the semantics induced by the ‘as’-phrase. 4.2.2.2  Can Impersonal Pronouns Have a Referential First Person Reading?

Besides their generic and arbitrary readings, some authors distinguish a third interpretation of impersonal singular pronouns. One label used for this reading is specific, as it seems that the apparent impersonal pronoun refers to a specific person, typically the speaker. Instances of this are the use of Swedish man in

Impersonal Pronouns 113

(19a) (see Egerland 2003a,b), the use of Dutch je in (19b) (called “football je” because it is supposed to be typical for the speech of football players who are being interviewed) (see Zeijlstra 2015), the use of West Frisian men in (19c) (see Hoekstra 2010), and the “royal” use of English one in (19d). (19) a. Ja, man hade otur igår. yes imp had bad.luck yesterday ‘Yes, I was unlucky yesterday.’ b. Je kreeg de bal van Huntelaar en toen schoot je op het doel. you got the ball from Huntelaar and then shot you on the goal ‘I got the ball from Huntelaar and then I shot on goal.’ c. Men moat altyd de smoarge putsjes dwaan! imp must always the dirty jobs do ‘I always have to do the dirty jobs!’ d. One is not amused. Our take on this is that the “specific” reading is a particular instance of the generic reading of the pronouns in question (see also Tarenskeen 2010). What is striking about the examples in (19a,b,d) is that they are episodic. Normally, a generic DP occurs in a generic sentence, and one would therefore expect that a pronoun whose reference generalizes over people would occur in a sentence that generalizes over situations. We will argue that the examples in (19) in fact also generalize over situations, but in a way that differs from standard generic sentences. Consider a generic sentence like One can see the Eiffel Tower from that hill. This example expresses that in situations where a person stands on that hill, they can see the Eiffel Tower. Note that the reference of one includes both the speaker and the addressee, something we will discuss in detail in sections 4.4–4.6. What this means is that if the speaker or the addressee stands on that hill, the statement applies to them. What it does not mean, of course, is that every situation in which the Eiffel Tower is seen from that hill includes the speaker and/or the addressee. Note, furthermore, that the situations over which a generic sentence generalizes include both actual and counterfactual ones. Thus, the example just given is true even if the speaker and the addressee have not yet been able to climb the hill in question. Indeed, when showing off plans for a new invention, a mad scientist might say, If one steps onto the silver plate, one is instantly transmogrified, where the events generalized over have no actual instantiation at all (at least as yet). What we suggest is special about the examples in (19) is that all instantiations of the generic statement that hold of the actual world hold of the speaker only; all instantiations that do not involve the speaker are counterfactual. The

114

Chapter 4

reference of the pronoun, however, is still ‘people in general (including you and me)’. The consequence is that the use of the generic pronoun in examples like these gives rise to the implication that the statement that holds of the speaker in the actual world would be true of other people if they were to find themselves in the same situation. Note that, on this construal, sentences of the type in (19) can be episodic, as they may describe an event in which the speaker is or has been involved (and in which others would be involved given similar circumstances).6 This appears to be in line with descriptions in the literature of the interpretive effect of the “specific” use of impersonal pronouns, as compared with the use of a first person singular pronoun. Thus, Pettersson (1978), as cited by Coussé and Van der Auwera (2012), argues that the specific use of Swedish man expresses the personal experience of the speaker and at the same time indicates that anyone in the same situation would experience the same thing. Similarly, Zeijlstra (2015:484) states that the use of Dutch “football je” differs from the use of the first person singular in that “in the actual world it was the speaker, but that in all situations where the hearer would be in the same situation, the hearer would do the same thing.” (We would broaden this statement and not limit it to the addressee; see below.) Hoekstra (2010) notes that the specific reading of Frisian men can be paraphrased as ‘people like us’ or ‘people like me’. Thus, an example like (19c) does not simply mean that I always have to do the dirty jobs; it means that other people in a situation similar to mine must do so, too. Finally, although for obvious reasons we do not have intuitions about royal one, this use of the pronoun, too, appears to express that people in the same situation as the speaker (being the monarch and being confronted with something inappropriate) would not be amused, either. This suggestion comes very close to Zeijlstra’s (2015) analysis of “football je” in Dutch. However, Zeijlstra explicitly claims that the pronoun in examples like (19b) is the referential one, rather than the generic one. One of his arguments is that examples of this type can be episodic, something already addressed above. He also argues that not all languages that have a generic second person singular pronoun appear to permit the specific use of this pronoun. This is true, but it is equally true that not all languages that have a referential second person pronoun (that is, all languages) permit it. There is nothing in Zeijlstra’s account either that would explain why specific je appears in Dutch but not, say, in English. Indeed, as far as we know, no one has an account for this distribution. This is not unexpected, since even in Dutch not all speakers have “football je,” even though all speakers have both the referential and the generic second person singular pronoun. An indication that we

Impersonal Pronouns 115

are really dealing with the generic pronoun is that, to the best of our knowledge, the strong form jij does not occur in “football je” contexts. This is expected if “football je” is generic, as jij permits a generic reading only when contrastive (see note 2). We conclude that the specific use of impersonal pronouns is, in fact, a generic use. 4.2.2.3  Can First and Second Person Plural Pronouns Have an Arbitrary Reading?

Regarding plural pronouns, the generalization that emerges from (1) is that these all allow a generic reading, while arbitrary readings are restricted to the third person. The latter part of this generalization may not be immediately obvious, given the existence of examples like (20a–c). In particular, (20a,b) easily allow a construal in which the pronouns refer to a group of people containing speaker and addressee, respectively, but where either speaker or addressee has not personally been involved in the decision to expand in China. For example, speaker and addressee may be part of the WDN Corporation, but not of the board of directors that made the decision in question. This is known as the corporate reading (see Cabredo Hofherr 2003). (20) a. We have decided to expand in China. b. Have you decided to expand in China? c. They have decided to expand in China. We assume that on its corporate reading, a plural pronoun is not, in fact, impersonal. It is referential, but it refers to a collective. In all contexts where a pronoun with a corporate reading can be used, a collective DP can be used as well (with number agreement adjusted). Subsequent resumption by a pronoun is unproblematic, as expected if the pronoun is referential. (21) a. The WDN Corporation has decided to expand in China. They expect to make a big profit there. b. UCL is opening a campus in Dubai. That’s because we expect to make a big profit there. c. Is the University of Edinburgh also planning to recruit abroad, or are you concentrating solely on the local talent pool? For third person plural pronouns, an arbitrary reading is available in contexts that do not easily permit the use of a corporate DP. For example: (22) (Context: The speaker notices a cigarette butt in the hallway of the WDN Corp. headquarters.) a. Hey, they have been smoking here. b. *Hey, the WDN Corporation has been smoking here.

116

Chapter 4

That the arbitrary reading of they is really different from the corporate reading is apparent from the fact that the speakers of British English we consulted find the corporate reading unobjectionable, while the arbitrary reading is judged as either quite informal or even unacceptable. Crucially, even for those speakers who accept the arbitrary reading of they, a comparable reading is not available for first or second person plural pronouns; compare the examples in (23) and (24) with (22a). (23) (Context: The speaker, CEO of WDN Corp., notices a cigarette butt in the hallway.) *Hey, we have been smoking here. ‘Someone from the group I belong to has been smoking here.’ (24) (Context: The speaker, a visitor to the WDN Corp.’s headquarters, notices a cigarette butt in the hallway.) *Hey, you have been smoking here. ‘Someone from the group you belong to has been smoking here.’ Thus, only third person plural pronouns allow a truly arbitrary reading. We will now turn to the analysis of impersonal pronouns in terms of the person system developed in previous chapters. This analysis should provide answers to the following questions: 1.  Why do plural pronouns differ from singular pronouns in always allowing a generic reading? 2.  Why is the second person singular able to support a generic reading, while the first and third person singular cannot do this? 3.  Why are dedicated impersonal pronouns singular? 4. What is the difference between the two types of dedicated impersonal pronoun? 5.  Why do third person plural pronouns permit an arbitrary reading, while first person plural, second person plural, and second person singular pronouns do not? 4.3  Generic and Arbitrary Readings of Plural Pronouns

A desideratum of any theory of pronouns is that the same feature system postulated in the analysis of personal pronouns suffices for the analysis of impersonal pronouns. Our aim in this section is to show that this is possible given the proposal developed in chapters 2 and 3. Our core claim is that the analysis of impersonal pronouns does not require any adjustments to the definitions of the person features prox and dist as we have used them in the analysis of personal pronouns. We first consider plural impersonals.

Impersonal Pronouns 117

We assume that impersonal readings of pronouns arise when a regular pronoun is bound by a generic or arbitrary operator. For concreteness’ sake, we will assume that the relevant operators are directly merged with the pronoun, but it would make no difference to our analysis if their position is, in fact, higher in the clause (as assumed in Rezac and Jouitteau 2016 and Polinsky 2017). We will refer to these operators as gnr and arb; the exact analysis of their semantics is not relevant for our current purposes, but for concreteness’ sake we will assign them universal and existential force, respectively. We should note, however, that generic operators, as opposed to universal operators, permit exceptions (see Lawler 1973, Carlson 1977, Krifka et al. 1995, and Gelman 2003). An illustrative example is (25) (where Piet should be taken to be Dutch). (25) a. #Although all Dutchmen like coffee, Piet actually hates it. b. Although the Dutch like coffee, Piet actually hates it. Thus, if generic readings are to be modeled with a universal quantifier, the restrictor must contain a slot for a contextually given norm. For more discussion, see Moltmann 2010 and Malamud 2012, among others.7 In the case of arbitrary readings, a random selection is made out of the relevant set, much as in the case of existential quantification. There is some discussion about how many types of arbitrary reading there are (see, for example, Cabredo Hofherr 2003). We will not provide an analysis of these fine-grained distinctions here. The third person plural provides the most straightforward illustration of how the impersonal use of a plural pronoun is derived from its personal use. In our system, a third person pronoun is specified as [dist] and, given the definition of this feature, refers to a subset of the input domain that contains only o’s. If gnr is combined with this pronoun, a reading results in which there is quasiuniversal quantification over the set of o’s. If arb is merged instead, a reading results in which there is quasi-existential quantification. b. Arbitrary 3PL

(26) a. Generic 3PL GNR

NMB PRS

NMB AUG

ARB

PRS DIST

Gnr(x), x ∈ Si+u+o – Si+o

NMB PRS

NMB NΠ

AUG

PRS DIST

Arb(x), x ∈ Si+u+o – Si+o



118

Chapter 4

The expectation is that such readings will not be available for singular pronouns, simply because selection from a singleton set is uninformative. When a quantifier applies to a nonsingleton set, different output sets are generated depending on the nature of the quantifier. Therefore, an existential statement over a set of boys does not mean the same thing as a universal statement over the same set. However, in the case of a singleton set, this is no longer true. Any universal or existential statement over a singleton set must necessarily be about the sole member of that set, leading to an interpretation indistinguishable (in terms of truth conditions) from a simple referential reading of a singular pronoun. (For more discussion, see section 2.8.) b. *Arbitrary 3SG

(27) a. *Generic 3SG GNR

NMB

ARB PRS

NMB PRS DIST Gnr(x), x ∈{o} (= {o})

NMB PRS

NMB NΠ

PRS



DIST Arb(x), x ∈{o} (= {o})

Hence, quantification over the set delivered by singular pronouns is uninformative, and therefore we do not expect any generic or arbitrary readings of such pronouns. As discussed above, this is largely correct, but there are two exceptions: the generic reading of the second person singular pronoun, and the arbitrary and generic readings of dedicated impersonal pronouns, which are singular in their syntax. We will analyze these exceptions below. But first, let us turn to the other plural pronouns. As illustrated above, first and second person plural pronouns differ from third person ones in that they do not allow arbitrary readings, although generic readings of such pronouns are available. The generic reading is straightforward: as discussed in detail in the previous chapters, the person and number features that characterize first person plural pronouns deliver a subset of the input domain that contains more than one member (one of which must be the speaker), so that generic quantification over this set is possible. Mutatis mutandis, the same is true for second person plural pronouns.

Impersonal Pronouns 119

b. *Arbitrary 1INCL

(28) a. Generic 1INCL GNR

NMB PRS

NMB AUG

ARB

PRS PROX

NMB PRS

NMB NΠ

AUG

PRS



PROX

Gnr(x), x ∈ Si+u (= S′, {i, u} ⊆ S′) Arb(x), x ∈ Si+u (≠ S′, {i, u} ⊆ S′) The question, then, is why arbitrary readings of these pronouns are excluded. In principle, given that the set delivered by the person/number features has multiple members, existential quantification over this set should be possible. The problem is that existential quantification over Si or Si+u runs afoul of a condition introduced in chapter 2 as an implementation of Preservation of Φ. In particular, we argued that bound variable pronouns are constrained by the condition in (29). (29)  Let Rφ be a set over which a variable ranges (i.e., the set mentioned in its restriction), and S be the set of elements selected from Rφ under variable binding. If inclusion of i in Rφ is encoded, it must be guaranteed that S contains i, and if inclusion of u in Rφ is encoded, it must be guaranteed that S contains u. If this constraint applies to variables in general, and not just to variables introduced by bound pronouns, it also applies to the variable in the restrictor of arb and gnr. Thus, if the inclusion of i or u in this restrictor is encoded by φ-features, it must be guaranteed that these elements are selected under quantification. At least in the case of arb, this cannot be guaranteed, since this quantifier selects one or more random elements from its domain. Therefore, if first or second person features are present in a pronoun, then that pronoun will not permit an arbitrary reading. By contrast, the condition in (29) does not preclude generic uses of first and second person plural pronouns. After all, a universal quantifier selects all elements in its restriction and therefore comes with the guarantee that if the restriction contains i and/or u, these elements will also be selected. There is one complication. As noted, the generic operator contains a contextual norm that has the consequence that the quantification will be quasiuniversal: it permits exceptions (namely, for those elements of the input set that are considered to fall outside the norm). We predict, then, that as a consequence of (29) the norm cannot be chosen in such a way as to exclude i or

120

Chapter 4

u, when gnr applies to a first or second person plural pronoun, respectively. This effect is illustrated by the examples in (30). (30) a. In Italie eten we spaghetti. in Italy eat we spaghetti ‘In Italy we eat spaghetti.’ b. In Italie eten jullie spaghetti. in Italy eat you.pl spaghetti ‘In Italy you (plural) eat spaghetti.’ The example in (30a) expresses a general rule whose scope must include the speaker. Thus, if the speaker is Italian, the rule may be about Italians in general. However, if the speaker is not Italian, the example cannot express a rule about Italians. Rather, it must be about another group—for instance, people visiting Italy—and this group must include the speaker. The same is true, mutatis mutandis, for the example in (30b).8 Our suggestion, then, is that the absence of a guarantee of preservation of i and u, respectively, makes arbitrary readings for first and second person pronouns impossible (as well as generic readings based on a norm that excludes first or second person, respectively).9 A similar restriction does not hold of the third person plural pronoun, since the set delivered by its person feature does not include a particular obligatory member. Hence, this pronoun can have an arbitrary, as well as a generic, reading. Let us now turn to singular impersonal pronouns. 4.4  Dedicated Impersonal Pronouns:

imp-2

As noted, given the reasoning presented so far, we would not expect singular pronouns to have impersonal usages. Although this is partly correct, we have encountered two puzzles: the second person singular allows generic readings, and dedicated impersonal pronouns are grammatically singular. In this section, we will discuss one type of dedicated impersonal pronoun: namely, the one we designated imp-2 above. German man and Dutch men are of this type. Although these pronouns invariably behave syntactically as if they are third person singular, they cannot be specified in the same way as personal third person singular pronouns are. First of all, they are semantically plural. This is demonstrated, for example, by their ability to bind a reciprocal.10 (31) In dit land geeft men elkaar cadeautjes met de kerst. in this country gives imp each.other presents with the Christmas ‘In this country, people give each other presents on Christmas.’

Impersonal Pronouns 121

Moreover, the set delivered by imp-2 pronouns may include speaker and/or addressee. An example like (32a) is easily interpreted as generalizing over all people, including speaker and addressee. Compare this with the generic use of third person ze ‘they’ in (32b), which generalizes over people working in the restaurant, exactly because it excludes speaker, addressee, and their associates (on the assumption that these individuals are not working in a good restaurant). (32) a. Dat doet men niet in een goed restaurant. that does imp not in a good restaurant ‘One doesn’t do that in a good restaurant.’ b. Dat doen ze niet in een goed restaurant. that do they not in a good restaurant ‘They don’t do that in a good restaurant.’ Given that generic operators permit exceptions, it is not impossible to use a dedicated impersonal pronoun in such a way that speaker and/or addressee are excluded. For example, a natural reading of (33a) is that people living in Italy eat spaghetti, but this need not apply to speaker or addressee, as these may well be Dutch. In fact, on this reading men can be replaced by generic ze ‘they’, as in (33b).11 (As expected, there is a second reading of (33a) in which speaker and addressee are included: namely, a reading that can be paraphrased as ‘When in Italy, one eats spaghetti’.) (33) a. In Italie eet men spaghetti. in Italy eats imp spaghetti ‘In Italy they eat spaghetti.’ b. In Italie eten ze spaghetti. in Italy eat they spaghetti ‘In Italy they eat spaghetti.’ In addition to generic readings, imp-2 pronouns allow arbitrary readings, as in (34) for example (see also (9b)). (34) O kijk, men heeft hier bomen geplant. oh look imp has here trees planted ‘Oh look, they’ve planted trees here.’ In sum, imp-2 pronouns are used either to refer to an arbitrary group, or in a generic statement that may or may not include speaker and addressee. In conjunction with their ability to bind reciprocals, this fact shows that, while they are singular third person in their syntactic behavior, they are plural in their interpretation.

122

Chapter 4

Our hypothesis about these pronouns, following the essence of Egerland 2003b, is that they do not carry any person or number features but are in fact an instantiation of bare NΠ.12 (35) a. Generic IMP-2 GNR

b. Arbitrary IMP-2 NΠ

Gen(x), x ∈ Si+u+o

ARB



Arb(x), x ∈ Si+u+o

Recall that the denotation of NΠ is the entire input set Si+u+o (see chapter 2). Therefore, the basis for the interpretation of imp-2 pronouns is a nonsingleton set, consisting of speaker, addressee, associates of these, and others. Applying gnr to this set is unproblematic and gives a reading in which a generalization is made over all relevant people, as in (9a) and (32a). Notice that in this case, the contextual-norm characteristic of generic readings may exclude speaker and addressee without violating the implementation of Preservation of Φ in (29). This is because no person features are present and therefore nothing is encoded about the inclusion of speaker and/or addressee in the restriction of the generic operator. For the same reason, applying arb will not violate (29), which explains the grammaticality of examples like (9b) and (34). Contrary to other pronouns, pronouns specified as in (35) cannot receive a nonimpersonal, referential reading. We assume that this is because any referential pronoun must make a selection from the input set. That is, we assume that it is not possible to refer to a (potentially plural) individual unless other individuals are excluded. A pronoun specified as in (35) cannot make any such selection, however, because it lacks all φ-features. The feature specification in (35) also implies that the pronoun in question will receive a spell-out distinct from that of the personal pronouns (barring accidental homonymy). The spell-out rules for personal pronouns refer to prox or dist or a combination of the two. Hence, no spell-out rule for personal pronouns is applicable; (35) simply does not match the structural description of such rules. Therefore, there must be a distinct spell-out rule for featureless pronouns. Notice that this rule will not be applied to any of the personal pronouns, as such overapplication violates Maximal Encoding. Absence of person features in imp-2 pronouns also implies that such pronouns will trigger default agreement on the accompanying finite verb. We have discussed why default agreement is third person in chapter 2. That fact does not by itself explain why imp-2 pronouns always seem to be syntactically singular across languages, while the set over which gnr and arb quantify obligatorily has a cardinality larger than 1 (as for imp-2 pronouns this

Impersonal Pronouns 123

is Si+u+o, which must contain at least i and u as members). This seems to run counter to Maximal Encoding, which requires plural marking even where superfluous (compare the case of inclusive pronouns in absolute number systems, as discussed in chapter 3). However, there is a basic semantic incompatibility between number marking and the semantics of imp-2 pronouns. According to our analysis, what is definitional of such pronouns is that, prior to application of gnr or arb, no selection is made out of the input domain Si+u+o. Recall (from chapters 2 and 3) that number is specified in such a way that it cannot apply to the input set of the person system; we repeat the relevant definition of nmb in (36). (36)  λP.P ∧ |P| ∈ Snumber, where P ≠ Si+u+o Thus, if the input to the number system happens to be Si+u+o, the derivation crashes. Therefore, no number marking is possible in the case of imp-2 dedicated impersonal pronouns. Of course, the absence of number marking yields a syntactically singular pronoun—as argued in chapter 3, singular is characterized by the absence of number features. 4.5  Generic Second Person Singular Pronouns

Generic use of second person singular pronouns is “cross-linguistically very common” (Creissels 2013:53). Siewierska (2004:212) states that “the second person singular … is a common means of impersonalization throughout Europe. It is used in the Germanic, Romance and Slavonic languages, in Hungarian, Estonian, Komi, Turkish and Abkhaz. … Outside of Europe the impersonal use of the second person singular occurs in, for example, Godie, Gulf Arabic, Hindi, Kashmiri, Koromfe, Koyra Chin, Kurdish, Mandarin, Marathi, Mauwake, Maybrat, Makushi, Modern Hebrew, Mundani, Nkore-Kiga and Tuvaluan.” An example with Dutch je ‘you (singular)’ is given in (37) (see also (3a)). (37) In Italie eet je spaghetti. in Italy eat you spaghetti ‘When in Italy you (impersonal) should eat spaghetti.’ Like imp-2 pronouns, second person singular impersonal pronouns are semantically plural. (38) In dit land geef je elkaar cadeautjes met de kerst. in this country give you each.other presents with the Christmas ‘In this country, one gives one another presents on Christmas.’

124

Chapter 4

Notwithstanding this similarity, second person singular impersonal pronouns differ in their interpretation from imp-2 pronouns. Whereas (33a) allows a construal that excludes speaker and addressee, as discussed above, (37) expresses a generic statement in which speaker and addressee must be included. This is not an idiosyncratic property of the Dutch pronoun; rather, it appears to hold universally for second person impersonal pronouns. As Siewierska (2004:2012) states, “[T]he second-person singular, like the first-person nonsingular, when used impersonally, includes both speaker and addressee amongst the set of potential referents.” Arbitrary readings of second person singular pronouns are excluded, as illustrated by (3b). We should perhaps further clarify at this point what it means for speaker and addressee to be obligatorily included in the reference of a generic pronoun. It cannot mean that the predicate with which the generic subject appears must in the actual world hold of speaker or addressee. For example, the statement in (37) does not imply that the speaker or the addressee has ever eaten spaghetti in Italy, or even been there. This is because, as discussed in notes 7 and 8 generic statements have a conditional aspect to them. What (37) expresses is that, if anyone finds themselves eating in Italy, then spaghetti has to be part of the menu. This rule holds as much for speaker and addressee as it does for other relevant individuals in the discourse. However, (37) does not say anything about who may or may not be eating in Italy. The contrast with (33b) is clear: (33b) does not have any implication for speaker or addressee; even when actually eating in Italy, they can stick to their nonspaghetti diet. The core questions to which the impersonal use of second person singular pronouns gives rise, then, are the following: 1.  Why is the second person singular able to support a generic reading, while the first and third person singular are not? 2.  Why do second person singular pronouns obligatorily include speaker and addressee when used impersonally? Our answers rely on a realizational view of morphology. We propose that the featural contents of second person personal and impersonal pronouns are not distinguishable for the spell-out system and so will result in the same form at PF. The two types of pronouns contain the same features, namely, one instance each of prox and dist. However, clearly these two person features cannot be applied in the same way in the impersonal pronoun as in the referential one, as their distinct readings would then remain unaccounted for. Recall that in the referential pronoun, prox and dist apply sequentially, with dist applying

Impersonal Pronouns 125

after prox. We suggest that in the impersonal pronoun, these features apply in parallel. (39) a. 2SG

b. Generic 2SG

c. *Arbitrary 2SG

GNR

ARB

NMB NMB

PRS

PRS PROX



PRS

PRS



PROX+DIST

PRS

PRS



PROX+DIST

DIST {u}

Gnr(x), x ∈ Si+u+o

Arb(x), x ∈ Si+u+o

Although this proposal involves an extension of the inventory of possible feature structures, it does not require any change in the inventory of person features itself, in that the definitions of the features prox and dist need not be adapted for the impersonal reading and no additional features need to be introduced. We assume that the + operator in (39) is defined over features. Its semantic effect is given in (40), where the functions F1 and F2 correspond to the semantics of the features that the + operator combines. (This definition parallels the way function addition works in mathematics.) (40)  If F1(S) = S1 and F2(S) = S2, then F1+F2(S) = S1∪S2. This means that the output set delivered by (39b) is the unification of the sets selected by prox and dist. Since dist selects the outer layer of the input domain, while prox selects everything except the outer layer, it is easy to see that [prox+dist] applied to Si+u+o returns Si+u+o. When gnr applies to this, a reading results in which the interpretation of the pronoun ranges over the entire person space. This is exactly the reading that examples like (37) have.13 In contrast to imp-2 pronouns, second person singular pronouns do not allow arbitrary readings (see, for example, (3b)). This follows from our earlier condition in (29). The pronoun in (39) contains prox. This encodes that speaker and addressee must be part of arb’s restriction. However, as explained before, it cannot be guaranteed that these elements are selected under existential quantification, which means that (29) is violated. For the same reason, the generic reading of second person singular pronouns, while allowing for exceptions to its universal force, does not permit speaker or addressee to be treated as such

126

Chapter 4

an exception. We already noted that speaker and addressee are indeed obligatorily included in the relevant reading of the pronoun (see (37)). Thus, the core of our proposal is that second person generic pronouns and second person referential pronouns contain the same features, though arranged differently. We assumed before that the spell-out rule for second person pronouns in Dutch has the following form:14 (41) {D prox dist} ⇔ /je/ Given that the structural description of this rule mentions features, but not the structure they appear in, it will apply to both (39a) and (39b), as desired. The above discussion demonstrates that it is possible for the feature structures in (39) to receive the same phonological realization, but it does not force this. To rule out the possibility of separate realizations for these two structures, we propose that spell-out rules in general are subject to the following constraint on possible structural descriptions:15 (42)  The structural description of a spell-out rule is an unordered set of feature tokens. Given that (42) specifies the structural description of a spell-out rule as an unordered set, it is not possible for spell-out rules to specify any structure among the features they mention. This means that no spell-out rule can distinguish between (39a) and (39b). At the same time, (42) does not preclude the existence of spell-out rules that distinguish between [prox] and [prox–prox]. This is because the structural description {prox} requires the presence of one token of the feature prox, while the structural description {prox prox} requires the presence of two such tokens.16 Thus, the condition in (42) gives just enough leeway to distinguish inclusive and exclusive first persons, but does not overgenerate in the domain of generic pronouns. The analysis explains not only why pronouns with the feature specification in (39b) have the same form as referential second person pronouns, but also why they must be singular. Recall the definition of nmb in (36), which states that number cannot be interpreted with regard to the full person space Si+u+o. Given that [prox+dist] delivers exactly this set, number marking is impossible. The introduction of the + operator does not lead to unattested pronouns. In fact, (39b) is the only licit structure that yields a new interpretation. Given that the + operator combines two features, and given that our inventory of person features is prox and dist, there are only two further relevant feature conjunctions we need to look at, namely, [prox+prox] and [dist+dist]. The semantics of the + operator, as defined in (40), is such that conjunction of

Impersonal Pronouns 127

identical features necessarily gives the same output as the unconjoined feature. Suppose F(S) = S′; then F+F(S) is the union of S′ and S′, which is S′. The implication is that, although [prox+prox] and [dist+dist] are in principle grammatical, they do not add anything of substance to the inventory of possible pronouns. The same is obviously true of more complex combinations like [prox+prox+prox], which is equivalent to [prox], or [prox+prox+dist], which is equivalent to [prox+dist].17 4.6  Dedicated Impersonal Pronouns:

imp-1

In section 4.4, we discussed dedicated impersonal pronouns that allow both generic and arbitrary readings. In section 4.5, we looked at second person singular impersonal pronouns, which allow only generic readings. We now turn to the second type of dedicated impersonal pronouns, which are like second person singular impersonals in that they must be generic. This type is instantiated by English one, West Frisian men (Hoekstra 2010), and Icelandic maður (Sigurðsson and Egerland 2009). The contrast between the two types of dedicated impersonal pronouns is apparent from the fact that arbitrary readings of the kind allowed by Dutch men (a case of an imp-2 pronoun) cannot be expressed by English one. (43) a. O kijk, men heeft hier bomen geplant. oh look imp has here trees planted ‘Oh look, they have planted trees here.’ b. *Oh look, one has planted trees here. These pronouns invariably behave syntactically as if they are third person singular. However, they cannot be specified in the same way as personal third person pronouns are, because the set they deliver includes the speaker. Indeed, it has been observed for English one that it is speaker-oriented (Moltmann (2010) speaks in this context of generalizing detached self-reference). An example like (44a) is a statement of a generalization over all people, including speaker and addressee. Compare this with the use of generic they in (44b), which is a statement of a generalization over people in Italy, not including speaker or addressee. Thus, (44a) comes with the implication that the speaker/ addressee, too, will eat spaghetti when in Italy, while no such implication is associated with (44b). (44) a. In Italy, one eats spaghetti. b. In Italy, they eat spaghetti. Our proposal is that imp-1 pronouns have the structure in (45).18

128

Chapter 4

b. *Arbitrary IMP-1

(45) a. Generic IMP-1 GNR

PRS PRS

Gnr(x), x ∈ Si+u+o

ARB NΠ

PRS PRS



Arb(x), x ∈ Si+u+o

The defining property of an imp-1 pronoun is that it does have a prs node, but one that otherwise remains unspecified. This sets it apart from imp-2 pronouns, which have no prs node at all, as well as from other pronouns, which have a prs node that carries a further feature specification. The behavior of imp-1 pronouns follows. Since the prs node only introduces an identity function (see chapter 2), its output is Si+u+o. This has the immediate consequence that number marking is impossible, for the same reason as in the case of imp-2 and second person singular generic pronouns: the function introduced by nmb is not defined if its input set is Si+u+o. Consequently, imp-1 pronouns will trigger third person singular agreement (because this is the default). As with the other impersonal pronouns, gnr can be applied to Si+u+o, giving rise to the generic reading. The presence of an underspecified prs node encodes that the restriction of the quantifier contains i and u, because such a node delivers as its output the entire input set to the person system. This is different from what we have seen with imp-2 pronouns, where there was no prs node, and therefore nothing was encoded. For this reason, imp-1 pronouns differ from imp-2 pronouns in that not having speaker and addressee included in the range of the quantifier will violate (29). As a consequence, i and u cannot fall outside of the norm under which gnr operates. For the same reason, imp-1 pronouns cannot receive an arbitrary reading. In the above discussion, we have distinguished between two types of dedicated impersonal pronoun, mainly on the basis of their possible semantics. However, the proposed analysis differentiates them on the basis of their structure: imp-1 pronouns, but not imp-2 pronouns, have a prs projection dominating NΠ. If this analysis is on the right track, one would expect this structural difference to have syntactic consequences as well. There is at least one phenomenon that seems to indicate that the two types of pronoun indeed behave differently in syntax: imp-2 pronouns have only a nominative form, whereas imp-1 pronouns can bear any case. The basic data have been observed by several researchers (see Egerland 2003b, Cabredo Hofherr 2008, and Hoekstra 2010, among others). We use Dutch men (imp-2) and English one (imp-1) to illustrate this. In contrast to

Impersonal Pronouns 129

English one, Dutch men cannot surface in object position or as the subject of an exceptional-Case-marking (ECM) complement (see also Haeseryn et al. 1997, sec. 5.2.9.2).19 It can, however, be a nominative subject, whether basegenerated or derived (as in a passive). This is illustrated in (46) and (47). (46) a. MenNOM moet geen stereo-installatie kopen zonder hem eerst imp must no stereo-installation buy without him first te testen. to test ‘One shouldn’t buy a stereo without first testing it.’ b. *Ze vragen menACC steeds om bijdragen. they ask imp constantly for donations ‘They constantly ask one for donations.’ c. *Helaas hebben ze menACC de boogaloo zien dansen. unfortunately have they imp the boogaloo see dance ‘They unfortunately saw one dance the boogaloo.’ d. MenNOM wordt hier goed behandeld. imp is here good treated ‘One is treated well here.’ e. *Men’sPOSS gezondheid is belangrijker dan wat dan ook. one’s health is important.cmp than what than also ‘One’s health is more important than anything else.’ (47) a. b. c. d. e.

OneNOM shouldn’t buy a stereo without first testing it. They constantly ask oneACC for donations. They expect oneACC to be able to dance the boogaloo there. OneNOM is treated well here. One’sPOSS health is more important than anything else.

It is important to distinguish the generalization just mentioned from a generalization put forward in Cinque 1988, according to which impersonal pronouns generated in object position cannot have an arbitrary reading. Cinque’s generalization is about the distribution of readings of impersonal pronouns (arbitrary versus generic), while the generalization above is about the distribution of forms (imp-2 pronouns versus imp-1 pronouns).20 The fact that imp-2 pronouns cannot bear accusative case supports the idea that they differ in their syntactic structure from imp-1 pronouns, as we have proposed. Given that pronouns with a very similar semantics, such as generic you, can bear accusative, and given that the contrast is not one between subject and object position but one between nominative and accusative case, it is not likely that we are dealing with a semantic fact.

130

Chapter 4

There are various ways in which the restriction on the distribution of imp-2 pronouns can be implemented in the grammar. We speculate that it is linked to the theory of extended projection (as introduced in Grimshaw 1991). Suppose that all functional projections that introduce φ-features are of level F1 (their respective order being determined by semantics, as well as Preservation of Φ, as discussed before). Suppose, in addition, that Case is of level F2 (or higher). Then attaching Case without first attaching person or number (or gender) results in an ill-formed extended projection in which an F2 head attaches directly to the lexical head. This will rule out accusative instances of imp-2 pronouns, as these lack any φ-features. Nominative instances of such pronouns are allowed, however, on the assumption that nominative is not a Case but in fact the absence of Case (see Jakobson 1936, Andrews 1982, Kiparsky 1989, Falk 1991, Taraldsen 1996, and Neeleman and Weerman 1999).21 4.7  Why Inclusive Impersonal Pronouns Are

human

It is a well-known observation that certain impersonal pronouns may only include humans (or at least sentient beings) in their reference. This holds, in particular, of those impersonal pronouns that include the speaker/addressee in their reference, such as English one and the generic second person singular pronoun. Thus, One/You is/are not allowed on board if one/you weigh(s) too much can only be interpreted as the airline discriminating against overweight people, not as a restriction that also applies to the allowable weight of individual pieces of luggage. This appears to be surprising on our analysis, since generic you, imp-1 pronouns, and imp-2 pronouns all deliver Si+u+o as the output on which the operator works, and hence include o’s (which can, of course, be nonhuman elsewhere). We propose that this observation is due to a requirement that pronouns must have a uniform interpretation, in the sense that the elements in their reference must either all be human or all be nonhuman. Independent evidence for this notion of uniformity comes from the observation that referential pronouns adhere to it. To begin with, the associates contained in first and second person plural pronouns must be human. Thus, if someone is traveling with a particular suitcase, one might expect it to be possible to treat the suitcase as an associate. However, utterances like the following are marginal at best: (48) a. #We were supposed to travel together, but my suitcase ended up in Amsterdam. b. #You were supposed to travel together, but your suitcase ended up in Amsterdam.

Impersonal Pronouns 131

Examples of this type improve if the nonhuman entity can be “humanized”— for example, if a musician is talking about a trusted old instrument. (49)  We always travel together, but my Strad wasn’t accepted as cabin luggage. This also indicates that “human” is not quite the right description of the feature at stake here, but we will keep the term as a shorthand for the effect we are interested in. Relevant discussion on this issue can be found in Holmberg and Phimsawat 2016.22 The effects of the uniformity restriction are even clearer with third person plural pronouns. These can refer to both human and nonhuman entities. Crucially, however, they resist reference to a mixed set of entities. This is illustrated in (50). (50) a. (Where are those late suitcases?) They are all still in the lounge. b. (Where are those late passengers?) They are all still in the lounge. c. (Where are those late passengers and those late suitcases?) #They are all still in the lounge. Given that impersonal pronouns that include speaker and addressee contain some humans in their reference, the uniformity requirement implies that all individuals in their reference must be human. This account would allow third person impersonal pronouns to have two interpretations. They should be able to refer either to sets consisting of humans only or to sets consisting of nonhumans only. For generic third person impersonals, this is correct. (51) a. (Talking about trombone players:) They always play too loudly. b. (Talking about expensive trombones:) They are always underinsured. However, arbitrary third person pronouns appear to be restricted to humans. For example, (52) can mean ‘Some person(s) or other damaged this tree’, but not ‘Some lightning strike(s) or other damaged this tree’. (52)  They have damaged this tree. We therefore need an additional explanation specifically for this case. We do not have a very firm proposal to make, but suggest that the existential operator that accompanies this type of pronoun comes with the restriction that its domain can only contain human individuals. This is somewhat less ad hoc than it seems, since there are other contexts in which existentially bound arguments show the same restriction. In particular, the implicit argument of impersonal passives must be human (Pollmann 1975), as must arbitrary pro (Cabredo Hofherr 2006) and the implicit argument that accompanies arbitrary si in Italian (Chierchia 1995).

132

Chapter 4

4.8 Conclusion

In this chapter and the previous ones, we have developed a feature system that captures the crosslinguistic inventory of persons and their interpretations. This feature system is particularly sparse, in that it is based on only two privative features, prox and dist. These features represent functions that apply to a universal input set introduced by a nominal head NΠ. In the present chapter, we showed that this system needs only minimal elaboration to deal with impersonal pronouns and impersonal readings of regular pronouns. In particular, we must permit pronouns not specified for person, as well as pronouns in which two features apply in parallel. Crucially, the definition of the features themselves need not be altered in any way. This concludes our discussion of the feature system that characterizes person. We now turn to the external syntax of person-marked categories. Person is one of a limited set of features that enters into syntactic agreement relations. In chapters 5 and 6, we will explore the nature of person agreement as well as the structural restrictions it is subject to.

5

The Symmetric Nature of Agreement

5.1  Agreement: Symmetric or Asymmetric?

There is a basic asymmetry between nouns and verbs when it comes to the semantics of φ-features. Person, number, and gender are notions that bear on the interpretation of nouns, but not on the interpretation of verbs. One question that this fact invites is whether the interpretive asymmetry is reflected in syntax. Many theories of agreement assume that this is indeed the case. The presence of verbal φ-features or feature values is supposed to be dependent, in some way or other, on the presence of identical nominal φ-features or feature values. Chomsky (2000, 2001), for example, argues that φ-features of arguments start out valued, while φ-features of verbs are initially unvalued. They acquire a value through the operation of agreement. Frampton and Gutmann (2000) and Pesetsky and Torrego (2007) suggest a subtle adjustment of Chomsky’s proposal, namely, that agreement is an operation of feature sharing. However, they continue to assume that verbal φ-features have no initial value and acquire one through the sharing operation. Like Chomsky’s proposal, this alternative is therefore asymmetric in the sense intended above. Finally, Bobaljik (2008b) treats agreement as the copying of φ-features from an argument to a verbal head, suggesting an even sharper syntactic asymmetry than Chomsky’s: verbal φ-features are not just unvalued initially, but absent. Other theories of agreement (in particular, unification-based theories) are compatible with the view that there is no syntactic asymmetry between nominal and verbal φ-features. In such theories, the presence of φ-features in the verb does not depend syntactically on the presence of matching φ-features in the agreeing argument. Note that this is not the same as saying there is no syntactic asymmetry between argument and verb. Since agreement is a syntactic dependency, and since such dependencies are inherently asymmetric in that one of the categories involved must c-command the other, it goes without saying that

134

Chapter 5

the elements involved in the agreement relation do not have an interchangeable syntactic status (see chapter 6 for more discussion). In this chapter, we will argue in favor of the view that verbal φ-features are generated independently of nominal φ-features, so that the agreement relation is symmetric in this sense. We will base our argument on a phenomenon that we will refer to as subset control. Subset controllers are arguments that, at least at face value, are specified for fewer φ-features than the verb that agrees with them. An example involving number is found in Inari Saami, where dual can be expressed in the verb, but not in nonpronominal subjects (pronominal subjects do express the dual) (example from Toivonen 2007:229). (1) a. Almah kuáˈlástava onne. man.pl.nom fish.3du today ‘The two men are fishing today.’ b. Almah kuáˈlásteh onne. man.pl.nom fish.3pl today ‘The men are fishing today.’ An example involving gender (from Amela Camdzic, personal communication) is given in (2). First and second person pronouns in Serbo-Croatian are not specified for gender. Nevertheless, the gender of speaker and addressee is identified through past participle agreement. (2) a. Ja sam otisla na posao. I am gone.fem.sg to work ‘I (a woman) have gone to work.’ b. Ja sam otisao na posao. I am gone.masc.sg to work ‘I (a man) have gone to work.’ One may wonder whether data of this kind actually involve syntactic agreement in the first place, or whether they could alternatively be analyzed as instances of semantic agreement. While we cannot exclude an analysis along these lines for the examples in (1)–(2), we will discuss a case of subset control below that must involve syntactic agreement. If there are indeed cases of subset control that involve syntactic agreement, then the problem arises that a feature appears to be interpreted in a position different from the one in which it is generated. On the standard assumption that φ-features in the verb are dependent on the presence of identical features in the subject, such examples must involve partial lack of exponence of the subject’s features. However, we will argue that, at least in some cases of subset control, there is no evidence for the syntactic presence of these features in

The Symmetric Nature of Agreement 135

the subject. Therefore, asymmetric theories of agreement cannot account for the data. We also show that symmetric theories of agreement can deal with the existence of subset controllers. They allow features to be generated in the verb independently of the φ-feature specification of the subject, and we will argue that a mechanism of feature spreading ensures that the verbal φ-features are interpreted in the right position at LF. As feature spreading by hypothesis takes place in covert syntax, there is nothing remarkable about the lack of exponence in the subject. What we propose, then, is that subset control should be captured at LF, rather than through (lack of) PF exponence. The rest of this chapter is organized as follows. In section 5.2, we lay out the details of feature spreading and introduce a number of general predictions that follow from our proposal. Sections 5.3 and 5.4 deal with Spanish “unagreement” (Hurtado 1985), which we argue must involve subset control in syntactic agreement. Section 5.3 discusses cases where the controller is an R-expression, whereas section 5.4 deals with cases where it is a quantifier. In section 5.5, we discuss the link between unagreement and pro drop. Section 5.6 concludes. 5.2  Feature Spreading 5.2.1  Feature Spreading in Regular Agreement

We begin by describing aspects of the syntactic derivation insofar as it is relevant to agreement. We take agreement to be a two-step process. The first step consists of the association of the target and the controller. This is a purely syntactic process, subject to locality conditions and the like, which will be discussed in more detail in chapter 6. Once target and controller are associated, several LF processes need to take place in order to yield an interpretable representation. Chief among these are feature spreading and deletion of uninterpretable features. This distinction between the establishment of the syntactic agreement relation and operations on feature content is not new; it goes back at least to Chung 1998 and is reminiscent of the distinction between “matching” and “valuation” in theories of agreement such as those outlined in Chomsky 2000, 2001, Boeckx 2001, and elsewhere. In this chapter, we will take the syntactic establishment of agreement relations as a given, and concentrate on what happens with the agreeing features at LF. The theory of agreement we adopt is symmetric in the sense that φ-features are generated in the verb independently of the φ-features of the argument. The theory of feature checking in Chomsky 1995 is a precursor of this idea. The implementation we adopt is based on the condition in (3), which states

136

Chapter 5

that at LF each occurrence of a φ-feature must be licensed. Licensing is not the same as interpretation, as not every φ-feature has semantic import. For example, French table ‘table’ is a feminine noun, but it is not feminine in its interpretation. The condition must therefore accommodate φ-features inherent in nouns.1 (3)  Φ-feature licensing a. At LF, each φ-feature F must be licensed in each position L with which it is associated. b. F is licensed in L iff (i) F is inherent in L’s lexical specification or (ii) F receives a semantic interpretation in L. Φ-features can be licensed in nominal, but not verbal, locations.2 This implies that a verbal φ-feature poses a potential problem that must be dealt with at LF, before semantic interpretation takes place.3 The solution consists of two operations. As a first step, the verbal φ-feature spreads to the nominal it is associated with under agreement. We can model this process using insights of autosegmental phonology (see Goldsmith 1976 and Williams 1976, among others). In autosegmental phonology, phonemes are not unstructured feature bundles. Rather, features are placed on a tier separate from the tier that contains the position in which they are anchored. Crucially, features can spread from one segment to another (leading to multiple association) or be reassociated with different segments after deletion of their initial segmental host. Let us assume that, in parallel to autosegmental features in phonology, φ-features are represented at LF on a tier separate from the category that hosts them. Thus, categories in the syntactic representation mark positions comparable to the segmental positions in phonology. The φ-features are represented on a separate tier and can be associated with a category on the “segmental” tier. Agreement involves spreading of the verbal φ-feature bundle to the nominal category. In other words, the verbal φ-feature bundle ends up associated with both the verb and the nominal. (4)  Φ-feature spreading DP ... V → DP ... V [F1 ... Fn]

[F1 ... Fn]

Subsequent to spreading, a rule of dissociation must apply that removes the link between the verb and the features it hosts, since these features are in danger of violating (3) when associated with a verb.

The Symmetric Nature of Agreement 137

(5)  Dissociation X →

X

[F1 ... Fn]

[F1 ... Fn]

This simple formulation of the rule suffices. There is no need to add a structural description that limits dissociation to agreeing verbs. If an association line anchored in the verb is deleted while the one linked to the subject is spared, as in (6a), an LF representation is generated that can be interpreted, as we will show. However, if the association line anchored in the subject is deleted, a representation results that violates (3), as shown in (6b). (6) a. DP ... V



[F1 F2] b. *DP ... V  

[F1 F2] →

[F1 F2] c. *DP ... V

[F1 F2]

DP ... V [F1 F2]



[F1 F2] d. *DP ... V

DP ... V

DP ... V [F1 F2]



DP ... V [F1 F2]

If both association lines are removed, as in (6c), we are left with a feature bundle not associated with any syntactic position. We assume that the grammar does not allow such floating features at LF, for reasons discussed below. This ban on floating features also explains why dissociation of features anchored in the verb leads to an ungrammatical result in the absence of agreement. In (6d), the feature bundle associated with the verb is left floating after application of (5). Notice that if any of the output representations in (6b–d) were ruled in, it would be possible to generate sentences containing verbs that host φ-features but fail to agree, contrary to fact (modulo cases of default agreement; see note 3). The ban on floating features at LF is reminiscent of similar conditions in early versions of autosegmental phonology. Goldsmith (1976), for example, assumes that every autosegmental feature must be anchored to a segmental position. In subsequent work in phonology, this condition was relaxed in

138

Chapter 5

some instances, with authors arguing for deletion or tolerance of certain unanchored autosegmental features at the end of the derivation (see, for example, Halle and Vergnaud 1981 and Pulleyblank 1986; for an overview, see Odden 1995). Be this as it may, it is a standard assumption in the syntactic literature that φ-features are subject to Full Interpretation: they must be interpreted in the mapping from LF to semantics (modulo our earlier remarks on inherent features). It is hard to see what interpretation could be assigned to a floating φ-feature, given that φ-features codetermine the semantics of nominal categories. Therefore, floating φ-features violate Full Interpretation and are consequently banned. (Note that it would be highly undesirable to allow LF deletion of floating φ-features. If features could be freely deleted at LF, Full Interpretation would be void of content.)4 To sum up, the rule in (5) is freely applicable. However, if it is applied to a subject or to a verb that does not agree with the subject, its output violates Full Interpretation. After feature spreading, then, we have an intermediate representation at LF as in the output of (6a). The next question is how the features on the DP can satisfy the licensing principle in (3). This is not trivial, since as we have argued at length in chapter 2, φ-features are interpreted not on the DP node itself but in specific positions within the nominal extended projection, namely, nmb and prs. Our proposal is that there is a process of feature identification that applies under domination and is restricted to extended projections. This process identifies features on an XP with matching features within XP. Once a feature on DP is identified with a feature in a position that admits interpretation, the association line between DP and F can be removed as an instance of the general process of dissociation, as in (5), leading to a representation that satisfies (3). (7)  Φ-feature identification [DP [ X ]] → [DP [ F

F

X

]]

F

We should stress that feature spreading, dissociation, and identification operate in the LF branch of the grammar. This means they cannot affect the phonological realization of the categories they apply to. If they did, no language could show overt verbal agreement. The proposal can be illustrated with number agreement in the following Dutch examples:

The Symmetric Nature of Agreement 139

(8) a. Zij lop-en. they walk-pl ‘They walk.’ b. *Hij lop-en. he walk-pl ‘He walks.’ c. *Zij loop-t. they walk-3sg ‘They walk.’ The example in (8a) satisfies the condition on φ-feature licensing after feature spreading (leading to (9a)), dissociation in the verb (leading to (9b)), feature identification (leading to (9c)), and dissociation in DP (leading to (9d)).5 (Recall that plurality is universally marked through the aug feature; see chapter 3.) (9) a.

[AUG DIST] [DP [

NMB/PRS

]] ... V

[AUG DIST] b. [DP [  

NMB/PRS

]] ... V

[AUG DIST] [AUG DIST] c. [DP [

NMB/PRS

]] ... V

[AUG DIST] d. [DP [

NMB/PRS

]] ... V

[AUG DIST] In (8b), the verb is specified as plural, while the subject is singular (in form and interpretation). This means that in the syntax aug is associated with the verb, but not with the subject. After feature spreading (see (10a)), and dissociation in the verb (see (10b)), we can identify the instances of dist on DP and in nmb/prs, followed by dissociation of dist in DP (see (10c) and (10d)). As the DP contains no feature that permits identification with aug, however, it is not possible to associate aug with a position in which it can be interpreted; see (10d). This leads to a violation of (3).

140

Chapter 5

(10) a.

[DIST] [DP [

NMB/PRS

]] ... V

[AUG DIST] b. [DP [  

NMB/PRS

]] ... V

[AUG DIST] [DIST] c. [DP [

NMB/PRS

]] ... V

[AUG] [DIST] d. *[DP [ [AUG]

NMB/PRS

]] ... V

[DIST]

Finally, (8c) shows that agreement is obligatory where it is available. This does not follow from the mechanisms discussed above. The representation of (8c) after feature spreading and dissociation in the verb is given in (11). This representation is identical to the output of dissociation as applied to (8a) (see (9d)) and therefore does not violate any conditions that hold at LF. (11) [DP [  

NMB/PRS

]] ... V

[AUG DIST]

This is not just a quirk of our model. Neither in earlier theories of feature checking nor in the current Minimalist probe-goal system is there a principled syntactic reason why agreeing verbs do not contain fewer φ-features than can be licensed by the subject. However, we have already introduced a principle that captures the obligatoriness of agreement. This is Maximal Encoding, as discussed in chapter 2. Previously, we mentioned applications of this principle in the mapping between semantics and LF and in the mapping between PF and phonology. It stands to reason that it should also restrict the mapping between LF and narrow syntax. In other words, a given LF representation must be associated with a syntax that encodes as many features of this LF representation as possible in as many locations as possible. But (11) is not the richest syntactic representation compatible with an LF representation that contains a plural subject in a language that permits verbal agreement for number. That is one in which the verb has an aug feature in syntax (as in (9a)), licensed through the procedure described above. Therefore, by Maximal Encoding, the representation in (9a) will block

The Symmetric Nature of Agreement 141

a representation in which there is no aug in the verb. (For agreement, the effect of Maximal Encoding is identical to Schütze’s (1997) Accord Maximization Principle.) In sum, asymmetric and symmetric theories of agreement can both capture the “normal” situation: a full match between the φ-features of the verb and the subject. We will argue, however, that symmetric theories are better placed to deal with subset controllers. 5.2.2  Feature Spreading in Unagreement

As mentioned in section 5.1, we use the term subset controller to refer to situations in which the agreeing argument has fewer φ-features than the verb, as in the syntactic representation in (12). (12) [DP [ X ]] … V   [F1] [F1 F2] As noted, asymmetric theories of agreement cannot be reconciled with representations like (12). This is because they are based on the assumption that φ-features in the verb must reflect the presence of identical φ-features in the agreeing argument. In (12), the F2 feature present in the verb would be illicit, as there is no corresponding feature in the argument. So, if there are genuine subset controllers, asymmetric theories of agreement are disqualified. In contrast, a theory of agreement that does not adopt a morphosyntactic asymmetry between the features in controller and target is well placed to deal with subset controllers. There is nothing in the process of feature spreading introduced above that would stop it from taking place when the DP recipient does not have matching features. Following dissociation in the verb, the result would be as in (13). (13) [DP [ X ]] … V   [F1 F2] [F1] In the previous subsection, we argued (for a case like (13) involving “superfluous” number; see (10)) that this situation is usually ruled out because F2 cannot meet the condition in (3): it cannot be identified with a feature lower in DP and it cannot be interpreted on DP either. However, we have not spelled out exactly why interpretation of φ-features on the DP node is impossible. The reason for this is that the features in question rely for their interpretation on a specific input set: Si+u+o for person features and ℕ for number. These sets are introduced in specific heads lower in the DP and are therefore absent in the top node.

142

Chapter 5

Suppose that all languages allow generation of the features Π and #, which encode the relevant input sets, in a finite verb’s inflection, alongside the standard verbal φ-features. If these features spread to DP under agreement, interpretation of a person or number feature on DP is possible even in case it is not identified with a feature lower in DP. We assume that this is what happens in languages showing subset controllers in agreement. Of course, not all languages have this property, and therefore there must be a parameter that regulates its distribution. One possibility is that the parameter in question belongs to a larger set of parameters that specify which particular features are involved in verbal agreement in a language. After all, there is language variation in general in this domain. If languages can differ in whether or not finite verbs agree for person or for number features, then they can also differ in whether or not finite verbs agree for Π or #. So, assume that there is a parameter according to which some languages allow spreading of Π and #, while others do not. In the latter class of languages it is still possible to generate these features in the verb, but usually this will result in an ill-formed LF representation. As explained in the previous subsection, features that are left on V, or are left floating after dissociation with V, are not interpretable. Therefore, the feature # will not occur in verbs in this type of language, while the feature Π will occur in a verb only in cases of default agreement, where it provides the input set for dist (see note 3). A parameter that divides languages into those that allow spreading of Π or # at LF and those that do not may seem ad hoc, but we will argue in section 5.5 that this is, in fact, the pro drop parameter. The remaining issue is how, in those languages that permit subset control, the interpretation of the extra φ-features on the DP node is connected to the interpretation of the DP as a whole. We propose the following rule of interpretation: (14)  Let NMAX be the top node of an extended nominal projection. Let SINT be a set built up compositionally within NMAX. Let SEXT be a set encoded in NMAX but not built up compositionally within NMAX. Then SINT and SEXT must be taken to be identical. This rule applies to a broader range of cases than just subset control in agreement. In particular, it is necessary to account for the interpretation of close apposition structures like we the fans in an example like (15). (15)  This was the time we the fans could actually start believing that a revolution is taking place within our club. We follow Lekakou and Szendrői (2012) in assuming that close apposition involves a multiheaded structure: the top node of the structure is a projection

The Symmetric Nature of Agreement 143

of both DPs simultaneously. This means that the top node inherits the reference of each of these DPs. In terms of (14), the reference of the first DP counts as SEXT on NMAX from the perspective of the second DP, while the reference of the second DP counts as SEXT on NMAX from the perspective of the first DP. The result is that, by (14), the reference of the two DPs must be taken to be identical. In the example we the fans, this means that the set of fans must be equal to Si or Si+u. This gives the correct interpretation: the expression refers to a set of fans exhaustively populated by the speaker and associates (and possibly the addressee and associates, on an inclusive reading of we). For an example of how the system works in cases of subset control, we return to the Inari Saami data in (1), assuming for the sake of the argument that a description in terms of subset controllers is correct. Two questions need to be addressed. First, how can we capture the interpretive effect of the verbal inflection in (1a)? Second, how can the distribution of subset controllers be restricted in the right way? In particular, we need to rule out subset control involving pronominal subjects. The dual form of the verb has a feature min that is dependent on aug according to the number system developed in chapter 3, as well as a dist feature that encodes third person. The presence of min expresses that the plural is minimal. On the view adopted here, these φ-features are represented on a separate tier. If the subject is a nonpronominal DP, its feature specification will not contain min (as there are no dual forms for nonpronominal DPs in Inari Saami). If there is indeed such an asymmetry between verb and subject in (1a), feature spreading results in a surplus of features on DP, compared to what is present within DP. This is schematized in (16b,c). (16) a. [DP [

NMB/PRS

]] ... V

(input)

[#–AUG] [#–AUG–MIN Π–DIST] b.

[#–AUG] [DP [



NMB/PRS

(feature spreading) ]] ... V

[#–AUG–MIN Π–DIST] c.

(dissociation in V)

[#–AUG] [DP [

NMB/PRS

]] ... V

[#–AUG–MIN Π–DIST] The situation in which there are more φ-features on DP than there are within it cannot be dealt with by partial feature identification. Suppose we were to

144

Chapter 5

identify [#–aug] in DP with its counterpart in nmb, followed by dissociation in DP. We would be left with a min feature. This is an uninterpretable structure, since the semantics of min depends on an input set delivered by [#–aug]. Given that feature identification is therefore impossible in this situation, the structure can only be interpreted if the rule in (14) is applied. In the case of (1a), SINT is the set delivered by [#–aug] on the basis of the input set contributed by ‘man’; in other words, it is a set of men containing more than one member. SEXT is the set delivered by [#–aug–min Π–dist]; that is, it is a set of two o’s. The rule in (14) equates these two sets, leading to reference to a set of two men that are o’s. Note that the range of interpretive effects that can result from subset control is quite limited. This is because the rule in (14) implies that the interpretation acquired by a DP must be one independently admitted by its original feature specification. The rule will fail to deliver an interpretable result if the features within DP deliver a set that is inherently different from the set delivered by the features that are spread to DP. There are several situations in which these sets are incompatible. One situation arises when the φ-features inside DP and the φ-features acquired by DP deliver sets with clashing identities, especially when they encode a different part of the person space. Another such situation arises as a result of the workings of Maximal Encoding. As discussed in chapter 2, this principle has the effect that the absence of a feature in a given structure triggers an interpretive effect in case the language also has available structures that include that feature. In particular, in such circumstances the absence of a feature is interpreted as the negation of that feature. Thus, in a language that has aug in its feature inventory, the absence of aug in the nmb node of a DP implies that this DP must be interpreted as singular. But that in turn implies that spreading aug to this DP will not yield an interpretable result. By (14), the reference of the DP would have to be singular and plural simultaneously. Hence, a number feature cannot spread to DPs that do not have the feature in question but could encode it. The same holds for person features. In general, then, our proposal predicts that the interpretive effect of subset control is one of selecting from a range of independently permitted interpretations, rather than one of imposing a reading otherwise not available for the DP. (17)  If an agreement mismatch has semantic effects, these must always be toward a more specific interpretation of the controller. Our proposal thus explains why subset control is not available for pronominal subjects in Inari Saami. As already mentioned, pronouns in this language differ

The Symmetric Nature of Agreement 145

from nonpronominal arguments in that they can express dual. Maximal Encoding therefore blocks the use of a regular plural pronoun if the intended interpretation is dual. The proposal makes two further predictions. Our basic assumption is that the agreeing DP in relations of subset control is genuinely poorer in syntactic feature content than the verbal head it agrees with. We therefore do not expect agreement phenomena internally to the DP that reflect the relevant verbal feature. (18)  If F is generated on the target only, agreement within the controller will not be for F. Recall that feature spreading is an LF process. Therefore, features acquired by this rule will never have effects on surface forms. Finally, we predict that subset control must always have an interpretive effect, even when it is not an instance of semantic agreement. As already mentioned, φ-features need not be interpreted if inherent in a noun (see (3)). However, in the case of subset controllers, the extra feature present in the agreeing head cannot be licensed in this way (otherwise it would have been generated in the subject as well, as demanded by Maximal Encoding). This implies that the feature in question must have semantic import. (19)  If F is generated on the target only, it must be interpreted. This interpretive effect is of course a core observation about examples like (1a). Our argument for a symmetric theory of agreement only holds water if genuine examples of subset controllers can be found. Proponents of the idea that the presence of φ-features in the verb must reflect identical φ-features in the DP in principle have two ways in which they can argue against the existence of subset controllers (assuming that we are, indeed, dealing with syntactic rather than semantic agreement). Taking the Inari Saami example again, the first option would be to hypothesize that nonpronominal subjects can be specified for dual after all, but that this is not expressed morphophonologically, either as a result of impoverishment or because there simply is no overt affix expressing dual that can be attached to nonpronominal nouns. We will call this type of approach a hidden-feature analysis. It is schematized in (20a) for a case of subset control involving the dual. Second, one could hypothesize that the apparent controller of agreement is a dislocated or appositional category, while the real controller is a covert pronoun that fully agrees with the verb. We call this kind of approach, schematized in (20b,c), a hidden-controller analysis (or hidden-subject analysis in case the controller is a subject).

146

Chapter 5

(20) a. Feature hidden by impoverishment [IP [DP #–aug–min] [VP … [V #–aug–min] … ]] → [IP [DP #–aug] [VP … [V #–aug–min] … ]] b. Hidden subject and dislocation of visible DP [DP #–aug]1 [IP [pro #–aug–min]1 [VP … [V #–aug–min] … ]] c. Hidden subject and apposition of visible DP [IP [[pro #–aug–min] [DP #–aug]] [VP … [V #–aug–min] … ]] To be clear, we are not confident that a hidden-feature analysis or a hiddensubject analysis of the Inari Saami data can be excluded, nor are we confident that the Inari Saami data do not involve semantic agreement. We have used the language only as an initial illustration of the issues involved. However, there are examples of subset controllers that do not lend themselves to either alternative analysis, as we will argue in the next section. 5.3  Referential Unagreement 5.3.1 Analysis

A strong case can be made for the existence of subset controllers in Spanish.6 In this language, subjects that are R-expressions may appear with verbs that carry first or second person plural agreement. This phenomenon has been widely discussed and goes by the name unagreement (see Hurtado 1985 and Jaeggli 1986, among others). It comes in two flavors: the subject may be a regular R-expression, as in (21) (from Corbett 2006:132), or a quantifier, as in (22) (compare Moravcsik 1978:351; see also Rivero 2008). We refer to the two subtypes of unagreement as referential and quantificational.7 In this section, we will discuss referential unagreement. Quantificational unagreement is discussed in section 5.4. (21) a. ¡Qué desgraciad-as somos las how unfortunate-fem.pl be.1pl def.fem.pl ‘How unfortunate we women are!’ b. ¡Qué desgraciad-as sois las how unfortunate-fem.pl be.2pl def.fem.pl ‘How unfortunate you women are!’ (22) a. Cada alumno hablamos differente. each student speak.1pl differently ‘Each of us students speaks differently.’ b. Cada alumno habláis differente. each student speak.2pl differently ‘Each of you students speaks differently.’

mujer-es! women-pl mujer-es! women-pl

The Symmetric Nature of Agreement 147

Note that unagreement cannot be an instance of semantic agreement. If it were, the interpretation of an example like (21a) would be such that the φ-features on the copula would function as a restriction on the predicate’s argument variable, yielding the semantics in (23). (23)  the women λx, x ∈ Si, unfortunate(x) This is not the correct semantics for (21a). While it implies, correctly, that all women in the context are part of Si, it permits, incorrectly, that there are elements in Si that are not women. (Compare an example like Every student did her homework, which does involve semantic agreement—namely, for gender. This example implies that all students in the context are women, but not that all women in the context are students.) In other words, while (23) can be paraphrased as ‘the women among us are unfortunate’, the correct paraphrase of the example in (21a) is ‘we women are unfortunate’ (see also note 14).8 If we take the data in (21) at face value, it appears that the verbal head contains a richer constellation of φ-features than the subject. On our theory, R-expressions cannot contain person information, because they do not carry the Π-feature that introduces the input set on which person features operate (see chapter 2). However, the verb is specified as [Π–prox] (or [Π–prox–prox]) in (21a) and as [Π–prox–dist] in (21b). Recall that the parameter that is responsible for a language allowing subset control or not is a parameter that regulates the spreading of Π and # under agreement. Spanish must be of the type that permits this. (This is as expected if this parameter is indeed the pro drop parameter, given that Spanish is a pro drop language; see section 5.5.) The derivation of an example like (21a) then proceeds as follows. The starting point is the configuration in (24a). Feature spreading transforms this configuration into (24b), after which dissociation in V yields (24c). (24) a. [DP [

NMB

]] ... V

(input)

[#–AUG] [#–AUG Π–PROX] b.

[#–AUG] [DP [



NMB

(feature spreading)

]] ... V

[#–AUG Π–PROX] c.

(dissociation in V)

[#–AUG] [DP [

NMB

[#–AUG Π–PROX]

]] ... V

148

Chapter 5

The features on the DP in (24c) cannot be licensed via feature identification and subsequent dissociation. This may be a solution for # and aug, but since R-expressions do not have person features, Π and prox would remain. Instead, the structure must be interpreted as it is, through the rule in (14). In this case, SINT is the set that results from applying [#–aug] to an input set given by the noun ‘woman’ in (21). SEXT is the set generated by applying [#–aug] to an input set delivered by [Π–prox]. The rule in (14) equates SINT and SEXT, with the result that the interpretation of the example in (21a) is such that the DP refers to a set of women consisting of the speaker and the speaker’s associates (plus the addressee on the inclusive reading). Examples involving a second person feature bundle have an analogous derivation. As noted, the rule in (14) is also responsible for the interpretation of close apposition structures of the type in (25), with the consequence that the interpretation of the subject in (21) is identical to that of the DP in (25). (25) nosotros las mujer-es we def.fem.pl women-pl ‘we women’ If the same rule of interpretation applies to structures of close apposition and unagreement alike, it is not surprising that, crosslinguistically, restrictions on close apposition correlate with restrictions on unagreement. This correlation was noted first by Höhn (2016); we will come back to it below. Our analysis makes five predictions. The first is that feature spreading must lead to a narrowing down of the possible readings of the controller, rather than imposing a reading on the subject that it could not have otherwise (see (17)). This is correct for the case at hand. As we discussed in chapter 2, a plural R-expression can refer to a set of individuals that includes speaker or addressee. Thus, while the Dutch in (26a) is most naturally taken to exclude speaker and addressee, the possibility of coreference in (26b) and (26c) implies that the Dutch may have a reference comparable to ‘we the Dutch’ or ‘you the Dutch’. Therefore, unagreement in (21) has the effect that a particular interpretation is selected for the DP out of a larger set of interpretations that are independently available. (26) a. Anyone who knows the Dutch realizes they no longer wear wooden shoes. b. Anyone who knows the Dutch realizes we no longer wear wooden shoes. c. Anyone who knows the Dutch realizes you no longer wear wooden shoes.

The Symmetric Nature of Agreement 149

At the same time, it is not the case that any possible enrichment of the subject’s feature structure through spreading is permitted as long as it is consistent with such narrowing of interpretation. Maximal Encoding implies that, if a given interpretation of a DP can be encoded within the DP itself, it must be encoded (see section 5.2). The consequence is that unagreement cannot narrow the interpretation of a DP to a reading that that DP could itself encode morphologically. Thus, it is not possible for a singular subject in Spanish to acquire plurality through referential unagreement. After all, plural can be expressed on R-expressions just as well as on pronouns. This blocks the examples in (27). (27) a. *¡Qué desgraciad-as somos la how unfortunate-fem.pl be.1pl def.fem ‘How unfortunate we women are!’ b. *¡Qué desgraciad-as sois la how unfortunate-fem.pl be.2pl def.fem ‘How unfortunate you women are!’

mujer! woman mujer! woman

The examples in (21) differ from those in (27) in that the φ-features expressed exclusively in the verb only include person features, which cannot be generated in R-expressions. Therefore, there are no more highly specified forms of the subjects in (21) that could block unagreement. The second prediction our account makes is that unagreement cannot occur if the subject resists inclusion of speaker or addressee in its reference to begin with. In that case, narrowing of interpretation is not possible. A relevant example is that of coordinated singular R-expressions. A singular R-expression cannot normally be used to refer to the speaker or the addressee (possibly for pragmatic reasons; see subsection 2.7.2 for some discussion). This has the consequence that (28b) and (28c) are ungrammatical. Given that neither Jan nor Piet can be understood to be the speaker or the addressee, and given that these individuals make up the plural referent of the coordinate structure, it is impossible to refer back to that plural referent using a first or second person plural pronoun. In line with this, unagreement is impossible when the subject consists of coordinated singular R-expressions (see (29)). (28) a. Anyone who knows Jan and Piet realizes they no longer wear wooden shoes. b. *Anyone who knows Jan and Piet realizes we no longer wear wooden shoes. c. *Anyone who knows Jan and Piet realizes you no longer wear wooden shoes.

150

Chapter 5

(29) a. *¡Qué desgraciad-as somos Juanita y Carmen! how unfortunate-fem.pl be.1pl Juanita and Carmen ‘How unfortunate we, Juanita and Carmen, are!’ b. *¡Qué desgraciad-as sois Juanita y Carmen! how unfortunate-fem.pl be.2pl Juanita y Carmen ‘How unfortunate you, Juanita and Carmen, are!’ If, for pragmatic reasons, singular R-expressions cannot refer to speaker or addressee, we also expect that unagreement will typically be limited to the plural. That is to say, spreading of first or second person feature structures from singular verbs to singular subjects is not expected to occur. By and large, this appears to be correct. In Spanish, for example, unagreement must involve plural verbs (see Höhn 2016:sec. 6.1). Höhn does note a limited set of cases that allow unagreement with singular verbs in Greek. These always seem to involve what Höhn calls “emotively marked” nouns, such as vlakas ‘idiot’. This pattern can be understood if emotively marked nouns can defeat the pragmatic aversion of using singular R-expressions for self-reference and reference to the addressee. We do not know to what extent this is, in fact, the case. We may note, however, that it also seems to be precisely this kind of noun that can occur in close apposition with a singular first or second person pronoun. Examples like You idiot! are fine, whereas something like You student! is acceptable only if student is construed as emotively marked. A third prediction of our account is that unagreement with pronominal subjects should be impossible. This is because pronouns are already specified for person features. Therefore, spreading of a distinct set of person features to the pronoun would lead to a contradictory semantics, given the rule in (14). For example, if the subject is a pronoun specified as [dist] (third person) and the verbal inflection carries [prox] (first person inclusive), spreading would have the effect that the DP must simultaneously refer to only the outer layer and only the inner layers of Si+u+o. Clearly, this is logically impossible. This prediction is correct. For example, if the verb carries first person plural inflection, the only permissible pronominal subject is nosotras ‘we’. It is impossible to insert a differently specified pronoun and use unagreement to force a first person plural interpretation (see (30a)). The same is true, mutatis mutandis, for verbs inflected for second and third person plural (see (30b,c)). (30) a. ¡Qué desgraciad-as somos nosotras/*vosotras/*ellas! how unfortunate-fem.pl be.1pl we/you.pl/they ‘How unfortunate we (women) are!’ b. ¡Qué desgraciad-as sois vosotras/*nosotras/*ellas! how unfortunate-fem.pl be.2pl you.pl/we/they ‘How unfortunate you (women) are!’

The Symmetric Nature of Agreement 151

c. ¡Qué desgraciad-as son ellas/*nosotras/*vosotras! how unfortunate-fem.pl be.3pl they/we/you.pl ‘How unfortunate they (women) are!’ Thus, the data in (29) and (30) confirm the prediction in (17). The fourth prediction is that in referential unagreement there should be no evidence internally to the subject for the presence of prox or dist. Since feature spreading is an LF process, it will not have PF effects and should therefore not affect the morphophonology of the DP. Indeed, although Spanish has DP-internal gender and number agreement, material within the controller is never inflected for person, not even in cases of unagreement. Our account finally predicts that referential unagreement should involve a feature that is not inherent in the noun and that therefore must be licensed by interpretation (see (3) and (19)). This is certainly the case in Spanish, as R-expressions lack person. 5.3.2  A First Alternative: Hidden Features

We now turn to alternative analyses of referential unagreement that maintain the more common view of agreement as syntactically asymmetric (meaning that the syntactic presence of verbal φ-features depends on there being matching features in the subject). We will argue that none of these alternatives is satisfactory. An obvious asymmetric alternative of the Spanish data would be to assume that person features can be generated in R-expressions, so that R-expressions can trigger first or second person agreement. Of course, the analysis would have to assume that these features are never realized on the R-expression itself, possibly as a result of a rule of impoverishment. This is what we have called a hidden-feature analysis in (20). There are several facts that are difficult to reconcile with this type of analysis. First, as noted, there are no R-expressions in Spanish that have inherent person features, while there are R-expressions that have inherent number and/ or gender features. In a hidden-feature analysis, it remains mysterious why there should be such an asymmetry between the features for which—according to that analysis—an R-expression can be lexically specified. The asymmetry follows from our analysis, as R-expressions are not lexically specified for person features to begin with.9 Second, although it might be a coincidence that the features postulated to characterize R-expressions under the hidden-feature analysis are not overtly realized in Spanish, it would require a conspiracy to account for the apparent universal absence of a spell-out of such features on R-expressions. As far as we know, this universal absence is a fact, however. We do not know of any language in which we linguists is expressed using a first person ending attached

152

Chapter 5

to the noun linguists, but omitting the pronoun, or a language in which a noun like linguists is inflected differently depending on whether its reference includes speaker and/or addressee.10 Third, in Mancini et al.’s (2011) event-related potentials (ERP) study, a three-way difference was found between regular agreement, unagreement, and agreement errors. That is to say, the neurophysiological processing correlates of unagreement could not be identified as errors (reflecting the fact that the relevant examples are grammatical), but also could not be unified with the processing correlates of regular agreement. These results have a natural interpretation in our theory, where, compared to regular agreement, unagreement involves additional steps in processing (namely, the discovery that there are surplus features on DP compared to what is present within it, followed by interpretation via the rule in (14)). Indeed, Mancini et al. propose a comparable analysis of unagreement, which they dub “reverse agreement.” Such results seem much harder to reconcile with a hidden-feature analysis, which after all aims to equate unagreement to regular agreement.11,12 These considerations lead us to conclude that a hidden-feature analysis is not the way to approach the Spanish data. 5.3.3  A Second Alternative: Dislocation

The second type of alternative analysis that maintains a syntactically asymmetric theory of agreement is what we have called the hidden-controller analysis. In such an analysis, the apparent subject is assumed not to be the subject at all. Instead, there is a null subject that is fully specified for all φ-features realized on the verb. What seems to be the overt subject is either claimed to be dislocated (or occupy some other type of nonargument position; compare Hurtado 1985) or taken to be adjoined to the null pronominal subject, parallel to close apposition in examples like we the boys (compare Suñer 1988).13 The hidden-subject analysis seems more promising than the hidden-feature analysis for the Spanish data, given that this language allows empty subjects in general. This is not a coincidence: all languages that are reported to have unagreement allow pro drop (see Höhn 2016). Moreover, as argued by Barbosa (1995, 2000), Alexiadou and Anagnostopoulou (1998), and Ordoñez and Treviño (1999), apparent preverbal subjects in pro drop languages may be dislocated, rather than occupying a true argument position. The subject position could then be occupied by a silent pronoun. Thus, the kind of structure required by this particular version of the hidden-subject analysis has been argued to be available independently. Nonetheless, the success of a hidden-subject analysis depends on the extent to which the properties of unagreement are compatible with the syntax of null subjects and the syntax of dislocation or apposition. Our assessment is that it

The Symmetric Nature of Agreement 153

is not. We begin by discussing an analysis of unagreement in which the apparent subject occupies a dislocated position, while clause-internally the subject position is occupied by a null pronoun that agrees with the verb in the standard way (compare Olarrea 1996). An immediate problem for this analysis is that the distribution of unagreeing subjects matches that of regular subjects. For example, Torrego (1996) points out that they need not appear in a peripheral position, but may also occur sandwiched between the verb and an object (see (31)).14 Even theories that take preverbal subjects to be dislocated recognize this position as a regular argument position. (31) Firmamos los lingüistas la carta. signed.1pl the linguists the letter ‘We linguists signed the letter.’ A second difficulty concerns the information-structural status of clitic leftdislocated categories in Spanish. Clitic left-dislocation has been analyzed by Arregi (2003) as an instance of contrastive topicalization. Thus, the dislocated category must be a contrastive topic (or contain one). This is in line with a general typological pattern: fronted topics, as opposed to fronted foci, can be resumed by pronominal elements (see Legate 2002, Frascarelli and Hinterhölzl 2007, and Vermeulen 2013). Hence, if unagreeing subjects are left-dislocated and resumed by a null pronoun inside the clause, we would expect that they must be topics and cannot be foci. However, there is no such requirement. For example, in the context of the question ‘Who signed the letter?’, (32) is an unobjectionable answer (Antonio Fábregas and Luisa Martí, personal communications). This shows that los lingüistas can be a focus. (32) La firmamos los lingüistas. it signed.1pl the linguists ‘We linguists signed it.’ Finally, the dislocation analysis only shifts the problem. In other structures with a dislocated R-expression, the sentence-internal pronoun that resumes the R-expression cannot be first or second person.15 This is of course best illustrated in languages that lack pro drop, such as Dutch. (33) a. De jongens, ze zijn aan elkaar gewaagd. the boys they are to each.other weighed ‘The boys, they are well-matched.’ b. *De jongens, we zijn aan elkaar gewaagd. the boys we are to each.other weighed c. *De jongens, jullie zijn aan elkaar gewaagd. the boys you are to each.other weighed

154

Chapter 5

Assuming that apparent subjects in unagreement structures are dislocated thus begs the question why no feature matching is necessary between the hypothesized null subject and the dislocated category associated with it. One way out would be to assume that Spanish has an additional type of construction in which a null subject pronoun is doubled by a full DP. This hypothetical construction, which we will call low dislocation, must be distinguished from both hanging-topic left-dislocation and clitic left-dislocation in two respects. First, it should not require feature matching. Second, the category associated with the null pronoun should appear clause-internally, rather than in a peripheral position. This may appear ad hoc, but clitic-doubling constructions could provide some independent evidence for low dislocation. Clitic doubling is attested with certain types of objects in Spanish. In these cases, the object behaves as if it is clause-internal. Moreover, mismatches between the φ-features of the clitic and its double can be observed, parallel to those found with subjects and unagreeing verbs (examples from Hurtado 1985:201). (34) a. Nos denunciaron a las mujeres. us denounced.3pl to the women ‘They denounced us women.’ b. Vos denunciaron a las mujeres. you.acc denounced.3pl to the women ‘They denounced you women.’ Thus, clitic doubling and unagreement with subjects could be grouped together as cases of low dislocation: both structures would involve a pronominal argument (silent in the case of subjects), doubled by a DP in an Ā-position internally to the clause (compare the analysis of clitic doubling in Kayne 1975). Although there is a parallel between the two structures, we do not think that an analysis in terms of low dislocation is feasible. Clitic doubling is no longer taken to involve a pronominal argument coindexed with a DP in an Ā-position, partly because the relevant DP behaves like a regular object in certain respects. Instead, it is regarded as a case of agreement between a DP argument and a functional category (the clitic) generated either within the verbal extended projection (Sportiche 1998) or within the nominal extended projection (Uriagereka 1995, Papangeli 2000, and Kayne 2002; our analysis of agreement in chapter 6 is compatible only with this second option). However, if clitic doubling involves agreement rather than low dislocation, all we can conclude from the Spanish data in (34) is that the problem of unagreement extends to object agreement (if the clitic starts out DP-externally) or DP-internal agreement (if the clitic starts out DP-internally). The problem remains one of lack of feature matching between agreeing categories.

The Symmetric Nature of Agreement 155

5.3.4  A Third Alternative: Apposition

There is a second incarnation of the hidden-subject analysis for referential unagreement. The apparent subject could be assumed to be in apposition with a null pronominal subject specified for all features relevant to verbal agreement (Suñer (1988) suggests a related analysis for quantificational unagreement; see below). Thus, (35a) would have a structure much like that of its English translation in (35b), except that the pronoun is not spelled out. (35) a. ¡Qué desgraciad-as somos [DP [DP pro] [DP las how unfortunate-fem.pl be.1pl def.fem.pl mujer-es]]! women-pl b. How unfortunate [DP [DP we] [DP women]] are! Notice that overt apposition in Spanish is allowed in the relevant contexts. Therefore, the examples in (21) could be related to the structures in (36) through pro drop. (36) a. ¡Qué desgraciad-as somos nosotras las mujer-es! how unfortunate-fem.pl be.1pl we def.fem.pl women-pl ‘How unfortunate we women are!’ b. ¡Qué desgraciad-as sois vosotras las mujer-es! how unfortunate-fem.pl be.2pl you def.fem.pl women-pl ‘How unfortunate you women are!’ This version of the hidden-subject analysis has two obvious advantages. First, the semantics of referential unagreement is identical to that of close apposition. Second, apposition does not require feature matching for person between the argument and the category adjoined to it. Examples like we the people, you boys, and so on, are unobjectionable. This means that the analysis may provide a genuine way of reconciling referential unagreement with standard views of agreement. Nonetheless, the analysis faces a serious difficulty, which has to do with the distinction between weak and strong pronouns. Weak pronouns differ from strong pronouns in their syntactic distribution, their prosody, and their interpretation (see Cardinaletti and Starke 1999 for an overview). The distinction is relevant here because weak pronouns, as opposed to strong pronouns, cannot be combined with other material to form a DP. For instance, modification of a weak pronoun is impossible (see Cardinaletti and Starke 1999, among others). Given this generalization, we expect that apposition with weak pronouns should also be ungrammatical. This turns out to be correct, as we will now show.

156

Chapter 5

In Dutch, there is a formal opposition between the strong first person plural pronoun wij and its weak counterpart we (see (37a)). That we is indeed a weak pronoun is shown by the fact that it cannot be coordinated or modified, in contrast to strong wij, as illustrated in (37b,c). As it turns out, only the strong form can be inserted in appositional structures like (37d). Similar data exist in English (with you/ye). The pattern repeats itself for the German first person plural pronoun. The strong form wir ‘we’ has a weak counterpart ma (38a,b) in certain varieties. Only the former can host an appositional DP (see (38c)). (German examples are from Klaus Abels, personal communication.) (37) a. Wij/We gaan dan uit eten. we go then out eat ‘We will go for dinner then.’ b. Wij/*We en de studenten gaan dan uit eten. we and the students go then out eat ‘We and the students will go for dinner then.’ c. Wij/*We die taalkunde studeren gaan dan uit eten. we that linguistics study go then out eat ‘We who study linguistics will go for dinner then.’ d. Wij/*We studenten gaan dan uit eten. we students go then out eat ‘We students go for dinner then.’ (38) a. Ich hoffe daβ wir/ma uns dann amüsieren können I hope that we us then amuse can ‘I hope that we can amuse ourselves then.’ b. Ich hoffe daβ wir/*ma und die Studenten uns dann amüsieren I hope that we and the students us then amuse können. can ‘I hope that we and the students can amuse ourselves then.’ c. Ich hoffe dass wir/*ma die Linguistik studieren uns dann I hope that we that linguistics study us then amüsieren können. amuse can ‘I hope that we who study linguistics can amuse ourselves then.’ d. Ich hoffe daβ wir/*ma Studenten uns dann amüsieren können. I hope that we students us then amuse can ‘I hope that we students can amuse ourselves then.’ Cardinaletti and Starke (1999) argue in some detail that null pronouns must be classified as weak (see also Frascarelli 2007 and Roberts 2010). Indications

The Symmetric Nature of Agreement 157

of this are that, in contrast to strong pronouns, they can have nonhuman referents, cannot be coordinated, and can be used as subjects of weather verbs. We illustrate these properties for Spanish in (39). (39) a. Pro es bonito. pro is pretty ‘He/It is pretty.’ b. *pro y el de Juan son bellos. pro and that of Juan are beautiful ‘This one and that of Juan are beautiful.’ c. pro hace mucho frío. pro makes much cold ‘It is very cold.’ But if silent pronouns must be classified as weak, and if weak pronouns cannot appear in appositional structures, then the Spanish examples in (21) cannot involve hidden apposition. An analysis closely related to the one just discussed is developed in Höhn 2016. The main difference is that Höhn does not treat structures like nosotros las mujeres ‘we the women’ as close apposition; rather, he treats them as involving a functional head hosting person features in the extended projection of the nominal head. This head can be spelled out as a pronoun (nosotros in the example just mentioned). When the pronominal head is left unpronounced, apparent cases of unagreement can arise. We do not see that this analysis circumvents the problem just outlined, however. Whether combinations of a pronoun and a DP involve close apposition or complementation to a functional head, the fact remains that such combinations are impossible if the pronoun is weak. Empty pronouns should therefore be banned from the structure in question. Höhn presents an interesting argument in favor of his analysis, which relates to the crosslinguistic distribution of the unagreement phenomenon. He suggests that unagreement is only found in languages in which apposition with a pronoun can involve a full DP, including a determiner. This is possible in Spanish, as the grammaticality of nosotros las mujeres ‘we the women’ shows. It is not possible in Italian, for example: noi (*gli) studenti ‘we (the) students’. In apparent correlation with this, Italian does not show unagreement, even though it is a pro drop language (so in our terms allows spreading of Π and #). (40) *Gli studenti lavoriamo molto. the students work.1pl much ‘We students work a lot.’

158

Chapter 5

Höhn shows that this generalization extends to a number of other languages. He assumes there is a parameter that divides languages into a group in which the head hosting person features in a DP is separate from the head hosting definiteness and a group in which these features coproject from a single head. In the former type, unagreement is possible along the lines just sketched, and personal pronoun and definite determiner can be spelled out separately. In the latter type, definite determiner and pronoun compete for spell-out of the same node, and, for reasons not relevant here, the pronoun systematically wins. This accounts for the ungrammaticality of (40), as compared to Noi studenti lavoriamo molto ‘We students work a lot’. If Höhn’s generalization turns out to be correct, we would propose an alternative parameter, based on the notion that sets delivered by nominal features can be definite, indefinite, or unmarked for definiteness. Sets delivered by DP are marked for definiteness, because the D head is either definite or indefinite. Lower projections within R-expressions are not marked for definiteness. Finally, all sets derived from Si+u+o count as definite (given that referential pronouns generally behave as definites). Our alternative parameter states that a DP already marked for definiteness can or cannot be the target for definiteness marking. Italian has a negative value for this parameter. This means that it cannot have close apposition of a pronoun with a DP. This follows from our earlier assumption that DPs are marked for definiteness, in conjunction with the fact that close apposition with a pronoun is an instance of definiteness marking as well, given that the pronoun is definite. Italian does allow close apposition with a lower extended projection of N, which we will simply refer to as NP+, as NP+ is not marked for definiteness. This is illustrated in (41). (41) a.

DP [DEF] pronoun [DEF]

DP [DEF]

  b.

DP [DEF] pronoun [DEF]

NP+ [ ]

The same parameter codetermines whether unagreement is possible. We would contend that unagreement with person features is an instance of definiteness marking. This is because person features are inherently definite. In unagreement, they spread from verb to DP and cannot be identified with features inside DP. Unagreement with a DP therefore leads to a situation of double marking for definiteness. Hence, in a language like Italian such unagreement is not possible. If a language tolerates double marking for definiteness,

The Symmetric Nature of Agreement 159

however, both unagreement and close apposition of pronouns with a DP are permitted. This is the situation in Spanish. Note that feature spreading to NP+ is possible in principle, regardless of the setting of the proposed parameter. This may seem to lead to the prediction that, in languages of the Italian type, unagreement is possible with determinerless NPs, contrary to fact. (Examples are from Vieri Samek-Lodovici, personal communication.) (42) a. In questo ufficio lavorano sempre studenti. in this office work.3pl always students ‘Students always work in this office.’ b. *In questo ufficio lavoriamo sempre studenti. in this office work.1pl always students ‘We students always work in this office.’ We assume that (42b) is ungrammatical because of a restriction on possible arguments in languages that have a determiner system.16 Arguably, in such languages any nominal argument must be introduced by a D (or Q) (see Stowell 1989 and Longobardi 1994, 2001). This implies that apparent bare NP arguments are, in fact, DPs with a null determiner. Consequently, despite appearances, (42b) still contains a DP that is the target for definiteness marking, in violation of the negative setting of the parameter introduced above. The overall conclusion of this section is that none of the alternatives to an analysis of referential unagreement in terms of feature spreading appears to be satisfactory. Referential unagreement is a genuine case of subset control, implying that agreement is symmetric in nature. 5.4  Quantificational Unagreement

We now turn to quantificational unagreement, as exemplified in (22). Let us first consider the contribution that φ-features make to the interpretation of a quantifier. In general, the complement of a quantificational element like each or every functions as its restrictor. This is the only place within QP where φ-features can be interpreted. Thus, each of us ranges over the set delivered by the person features of us. With this in mind, consider the Spanish examples in (43) (where (43a) repeats (22)). Our analysis of these examples runs largely parallel to our analysis of referential unagreement. The only difference comes from the way φ-features are interpreted in QP. The φ-features of the verb in (43a,b) cannot narrow down the reference of the subject, simply because the subject is not referential. Instead, φ-features in this case serve to determine the interpretation of the set on which the quantifier operates.

160

Chapter 5

(43) a. Cada alumno hablamos/habláis differente. each student speak.1pl/speak.2pl differently ‘Each of us/you students speaks differently.’ b. La mayoría podemos/podéis conducir con una mano. the majority can.1pl/can.2pl drive with one hand ‘Most of us/you can drive with one hand.’ Quantificational unagreement, like referential unagreement, involves feature spreading from a target to a controller that is internally specified for fewer features. For concreteness’ sake, we illustrate the proposal using first person features encoding an inclusive reading, but the derivations for the first person exclusive and the second person cases are entirely parallel. The starting point is (44a). Feature spreading results in the representation in (44b). After dissociation in the verb takes place, we are left with (44c). (44) a. [QP [ NP+ ]] ... V

(input)

[#–AUG Π–PROX] b. [QP [ NP+ ]] ... V



(feature spreading)

[#–AUG Π–PROX] c.

[QP [ NP+ ]] ... V

(dissociation in V)

[#–AUG Π–PROX]

Feature identification and subsequent dissociation cannot get rid of #–aug or Π–prox on the QP node, since these features are not present within the QP. Hence, the interpretational rule in (14) applies. By this rule, the set encoded by the features on QP must be identified with the set encoded within QP. The result is that the features on QP further narrow down the restrictor of the quantifier, as it is the restrictor to which φ-features within QP pertain (see above). This yields the right interpretation for (43a), shown in (45). (45)  each x, x ∈ SStudent ˄ SStudent = Si+u ˄ |Si+u| > 1, x speaks differently Note that in the examples in (43) unagreement involves not only person features but also a number feature. As discussed, this situation is blocked by Maximal Encoding in the case of referential unagreement, given that referential DPs have a plural counterpart. Quantifiers like cada ‘each’, however, lack a plural counterpart, so that no blocking effects can obtain. For those quantificational expressions that have both a singular and a plural form, however, blocking effects are expected, on a par with those found with referential

The Symmetric Nature of Agreement 161

unagreement. In other words, for such QPs unagreement should be impossible in the singular, but permitted (for person features) in the plural. This appears to be correct, as illustrated for algun ‘some’ and todo ‘every’ in (46). (46) a. *Algún paciente hemos/habéis llamado a la doctora. some patient have.1pl/have.2pl called to the doctor ‘Some of us/you patients have called the doctor.’ b. Algunos pacientes hemos/habéis llamado a la doctora. some patient.pl have.1pl/have.2pl called to the doctor ‘Some of us/you patients have called the doctor.’ c. *Todo niño creemos/creéis en los Reyes Magos. every child believe.1pl/believe.2pl in the Reyes Magos ‘All of us/you children believe in the Magi.’ d. Todos niños creemos/creéis en los Reyes Magos. every child.pl believe.1pl/believe.2pl in the Reyes Magos ‘All of us/you children believe in the Magi.’ Of course, quantifiers that only have a plural form (such as ambos ‘both’ or tres ‘three’) also allow for unagreement in person features. In other words, the generalization that emerges is that quantificational unagreement is allowed with plural quantifiers, and with singular quantifiers as long as they do not have a plural counterpart. If our analysis is correct, we would expect our predictions about the subject in referential unagreement (see subsection 5.3.1) to carry over to the complement of the Q head in quantificational unagreement. We cannot test this for the full range of nominal expressions discussed above, because not all of these can function as restrictors of quantifiers. However, there is a sharp contrast between restrictors that are R-expressions and pronominal restrictors. When QPs have nonpronominal restrictors, unagreement is possible, whether the restrictor is a PP (see (47a)) or an NP (see (43a) and (46b,d)). But just as referential unagreement is blocked because of the semantic incompatibility of sets encoded by different person features in case the subject is a pronoun, so quantificational unagreement is blocked if the quantified subject has an overt pronominal restrictor (see (47b,c); note that pronominal restrictors must be contained in a PP). (47) a. Todos de los que dejaban lo vimos/visteis. all of the that left him saw.1pl/saw.2pl ‘All of us/you that left saw him.’ b. Todos de nosotros/vosotros lo vimos/visteis. all of us/you.pl him saw.1pl/saw.2pl ‘All of us/you saw him.’

162

Chapter 5

c. Todos de ellos lo vieron/*vimos/*visteis. all of them him saw.3pl/saw.1pl/saw.2pl ‘All of them/us/you saw him.’ As expected, these data parallel interpretive restrictions observed in English. (48) a. All of the boys could have done it, because they were all in the schoolyard. b. All of the boys could have done it, because we were all in the schoolyard. c. All of the boys could have done it, because you were all in the schoolyard. (49) a. All of them could have done it, because they were all in the schoolyard. b. *All of them could have done it, because we were all in the schoolyard. c. *All of them could have done it, because you were all in the schoolyard. We now turn to alternative analyses of quantificational unagreement. As with referential unagreement, a hidden-feature analysis is a nonstarter. Person is never expressed on quantifiers, making it unlikely that these elements can carry prox or dist. When a quantifier does not take an overt complement, as in (50), there might be a silent pronominal complement with the relevant person specification (this is essentially the analysis in Suñer 1988). But this analysis does not carry over to cases of unagreement in which the quantifier takes a nonpronominal complement (as in (46b,d)), as such a complement does not carry person features either.17 (50) Todos hemos ido de excursión alguna vez. all have.1pl gone of excursion some time ‘Every one of us has gone on a trip sometime.’ A hidden-subject analysis based on dislocation of the apparent subject is also unlikely to work for quantificational unagreement, as the problems outlined in subsection 5.3.3 carry over. An analysis of unagreement in terms of apposition of the apparent subject to a silent pronoun faces an additional problem in the case of quantificational agreement, simply because quantificational categories cannot function as appositional phrases in structures of close apposition. So, even if we ignore the fact that weak pronominal forms do not allow apposition (see above), the analysis must be rejected in view of examples like (51a,b). These examples show that the universal quantifier phrase todos niños cannot even stand in

The Symmetric Nature of Agreement 163

apposition to a strong pronoun. (We should stress that the same conclusion can be drawn with respect to other quantifiers that allow unagreement, such as negative quantifiers.) (51) a. *[Nosotros [todos niños]] creemos en we all boys believe.1pl in ‘All of us boys believe in the Magi.’ b. *[Vosotros [ todos niños]] creéis en you.pl all boys believe.2pl in ‘All of you boys believe in the Magi.’

los Reyes Magos. the Reyes Magos los Reyes Magos. the Reyes Magos

A more plausible alternative would be to say that it is not the entire QP that stands in apposition to a silent pronoun, but just its complement (on a par with each of us boys, all of you boys, etc.). Of course, this suggestion faces the by now familiar difficulty that weak pronominal forms do not permit apposition to begin with. But irrespective of that, it appears to be insufficiently general. For example, apposition with an overt pronoun is not possible with the NP complement of cada ‘each’ (see (52)). However, we have already pointed out that cada+NP partakes in unagreement (see (22)/(43a)).18 (52) a. *[Cada [nosotros alumno(s)]] hablamos differente. each us student(s) speak.1pl differently ‘Each of us students speaks differently.’ b. *[Cada [vosotros alumno(s)]] habláis differente. each you student(s) speak.2pl differently ‘Each of you students speaks differently.’ We conclude that, as with referential unagreement, an analysis of quantificational unagreement in terms of subset control is best suited to capture the data. 5.5  The Link between Unagreement and Pro Drop

We have described the crosslinguistic distribution of unagreement in terms of two parameters. The first states whether or not Π and # partake in feature spreading. The second states whether or not a category that is marked for (in-) definiteness can be the target of further definiteness marking. It may seem uneconomical to use two parameters to regulate the distribution of a single phenomenon. However, each of these parameters is also linked to other data. We already discussed this for the second parameter, which regulates close apposition with a pronoun. The first parameter, we claim, is rather better known. It is the pro drop parameter.

164

Chapter 5

There are, roughly speaking, two prominent approaches to agreementrelated pro drop.19 The first is based on the idea that agreement is syntactically asymmetric, so that any feature in the verb must have a counterpart in the subject. On this view, there must be an empty subject pronoun specified for the relevant features. This implies that one has to postulate at least six empty pronouns, corresponding to the six different agreement endings in languages like Italian and Spanish. The downside of this analysis is that one might expect there to be languages that make an arbitrary selection out of these six empty pronouns, giving rise to unattested patterns of subject omission.20 It is therefore necessary to add some restriction to the theory that removes this prediction. The second view is that the agreement ending on the verb is the subject argument, and that there is no structural subject present (see the papers in Ackema et al. 2006a for an overview and discussion). This neatly solves the problem of arbitrariness, but it requires a more elaborate theory of subject positions that allows assignment of a θ-role to either inflection or a DP subject. The parameter we have proposed makes a third option available: namely, that there is a single empty element in subject position, specified as nominal but nothing else. Pro drop can then be seen as the most extreme case of subset control: the entire feature content of the subject, apart from its nominal nature, originates in the verb and is transferred to the subject at LF.21 Note, however, that the result of such spreading will only be interpretable if Π and # are included in the information transferred. Otherwise, there is no input set for the person and number features to operate on. If a language does allow spreading of Π and # along with the person and number features, interpretation can proceed on the basis of the rule in (14). In this case, SINT is not subject to any lexical restrictions, as the element in subject position is radically underspecified and so has no inherent semantics. Consequently, identification of SINT and SEXT, which is the set encoded by the φ-features, yields a set that is indistinguishable from SEXT. In other words, the interpretation of the subject is identical to that of a regular pronoun with the same φ-features as the verb. Thus, this parameter, which is relevant to the availability of unagreement, also determines whether a language allows agreement-related pro drop. As an example, we give the derivation of the Italian sentence in (53), which involves pro drop of a first person subject. (53) Mangio la pizza. eat.1sg the pizza ‘I eat the pizza.’ The structure presented to the LF interface is (54a). Feature spreading results in (54b), dissociation in V in (54c).

The Symmetric Nature of Agreement 165

(54) a. N



(input)

V

[# Π–PROX–PROX]



b. N



(feature spreading)

V

[# Π–PROX–PROX] c.

N



V

(dissociation in V)

[# Π–PROX–PROX]

The rule in (14) results in an interpretation for the subject in (54) as first person singular. Given this view of pro drop, any connection between the availability of pro drop and rich agreement morphology on the verb must be indirect. Let us first outline our view of the distinction between rich and poor agreement. Since the rise of realizational theories of morphology, it is widely assumed that the features encoded in morphophonology need not correspond one to one with the features present in morphosyntax. We assume that languages that have any person agreement at all will have full person agreement in syntax. The same holds for number agreement. That is to say, the morphosyntactic parameters governing verbal inflection determine whether person and number can be encoded in the verb, but they do not specify that only a subset of person and number features is present. Maximal Encoding then requires that the verb contain the most specific feature set compatible with the intended interpretation of the structure. This implies that the opposition between rich and poor agreement has to do with morphophonological realization only. Our pro drop parameter, however, does not relate to morphophonological realization. Indeed, it would be conceptually very odd for it to do so, given the wellknown observation that, universally, syntactic processes are not conditioned by word-level phonology (for discussion, see Miller, Pullum, and Zwicky 1997, among others). However, we do expect there to be an indirect relationship between the occurrence of pro drop and the occurrence of rich agreement, for reasons relating to Ariel’s (1990, 1991) Accessibility Theory. Ariel argues that, in general, there is an inverse correlation between the amount of information overtly encoded in an anaphoric element and the accessibility of its antecedent, where accessibility is defined by such factors as closeness and prominence of the antecedent. Ariel models this using two hierarchies. The first is a universal accessibility hierarchy of antecedents. The second is a hierarchy of anaphoric elements. This second hierarchy must be language-specific, since languages

166

Chapter 5

have different inventories of strong, weak, and/or null forms for anaphoric elements. The status of a null anaphoric element in terms of Accessibility Theory depends on two factors. The first is the inventory of other anaphoric expressions that compete with the null form. If, in a given syntactic position, a language allows both strong and weak pronouns, in addition to the null form, then the antecedent for the null form will tend to be more highly accessible than if the position in question does not permit insertion of a weak pronoun (leaving the simple opposition null versus strong). Second, given that null anaphoric elements have no inherent content, the interpretive restrictions they impose in processing depend on the information overtly encoded in the verb. When person/number combinations are not realized unambiguously in morphophonology, a null form will require an antecedent more highly accessible than when this is the case. Hence, while the possibility of pro drop in a language as such does not depend on rich agreement morphology, rich agreement will make use of the null form possible in more contexts. (While this is not identical to the classic pro drop parameter, the idea that a distinction should be made between formal licensing of the null element and recovery of its content can be found already in Rizzi 1982, 1986; see also Deal 2005.) Finally, it is possible that there is some degree of crosslinguistic variation in how the universal accessibility hierarchy of antecedents and the language-specific hierarchy of anaphoric expressions are linked. This linking needs to be consistent in that no anaphoric expression can be used to refer to a discontinuous selection of the accessibility hierarchy, with a different expression referring to the intermediate span. However, this leaves some room for variation in the choice of cutoff points. There are certain advantages to having an indirect link between rich agreement and pro drop. In general, richness of agreement does not completely determine the extent to which pro drop is available. First, languages whose agreement systems are equally rich can nonetheless show variation in the use of pro drop. In certain person/number combinations, Italian permits strong pronouns, weak pronouns, and null forms in subject position (see Cardinaletti and Starke 1999). By contrast, Spanish only ever has an opposition between strong pronouns and null pronouns in subject position. In correlation with this, restrictions on the use of pro drop are stricter in Italian than in Spanish, in terms of the accessibility of the antecedent (see Filiaci 2010 for experimental evidence). This difference exists even though agreement is unambiguous in both languages. Second, Early Modern Dutch and Modern Dutch have agreement paradigms that are comparable: both have three distinct forms to distinguish six person/number combinations (although the distribution of these forms is not

The Symmetric Nature of Agreement 167

identical). Nevertheless, these languages differ with respect to the setting of the pro drop parameter. Modern Dutch does not allow pro drop, but Early Modern Dutch permits subject omission, albeit under extremely limited circumstances. Pro drop is only possible if there is an antecedent available in a nonfinite clause that has been moved to a position preceding the empty subject, as in (55). We have argued in Ackema and Neeleman 2007 that this restriction can be understood if the antecedent must be even more highly accessible than in a language with rich verbal agreement. (If pro drop is this restricted, a language-learning child may not encounter robust evidence for pro drop, leading to a resetting of the parameter and hence the Modern Dutch situation.) (55) Dese reden Joufvrou Wintergroen verstaende, heeft pro gheseght this speech Miss Wintergroen understanding has said daer toe niet te willen verstaen. there to not to want allow ‘When Miss Wintergroen heard this speech, she said she did not want to allow this.’ (anonymous, Wonderlicke Avontuer van Twee Goelieven; 1624) Third, this view on the relation between pro drop and agreement meshes well with the existence of so-called partial pro drop. Partial pro drop, as found for example in Hebrew and Finnish, is usually described as pro drop that is restricted to first and second person (see Vainikka and Levy 1999, Koeneman 2006, and references mentioned there). However, Gutman (2004) argues that the impression of partial pro drop arises as the consequence of a requirement that the antecedent for the empty subject must be highly accessible in the relevant languages. She shows that pro drop of third person subjects is, in fact, possible in the relevant languages, but only if the antecedent is high on Ariel’s accessibility hierarchy. Since speaker and addressee are always highly accessible by their very nature, pro drop in first and second person is possible without further restrictions. By contrast, third person pro drop depends on the antecedent being linguistically expressed in the local context (for related discussion, see Holmberg 2005 and Holmberg, Nayudu, and Sheehan 2009). Crucially, verbal agreement in the relevant structures is as rich in Hebrew and Finnish as it is in classical pro drop languages like Italian and Spanish, suggesting once more that there is no automatic link between pro drop and agreement morphology. 5.6 Conclusion

Agreement relations involve a semantic asymmetry between controller and target, as φ-features can only be interpreted in the controller. In this chapter,

168

Chapter 5

we have investigated to what extent there is also a syntactic asymmetry between controller and target, in that φ-features can only be syntactically present in the target if there are identical φ-features in the controller. According to a number of theories, agreement is indeed morphosyntactically asymmetric in this way. We have argued that the data discussed in this chapter can be analyzed more successfully if the presence of the features in the verb does not depend on the presence of features in the subject. The two φ-feature sets are generated independently. Agreement involves processes of feature spreading, identification, and dissociation that result in the features being in a position in which they can be interpreted at LF. Our case was based on the existence of subset control: structures in which the controller is specified for fewer features than the target. If such structures indeed exist, this is incompatible with the idea that features in the target always depend on identical features in the controller. Proponents of asymmetric theories of agreement have dealt with apparent examples of subset control by arguing for a hidden feature in the controller or a fully specified hidden controller. We have shown that at least in the case of Spanish unagreement such analyses must be abandoned. In the alternative we propose, features are generated freely in target and controller, as long as their distribution at the interface with semantics is such that all φ-features can be licensed. The hypothesis that agreement is symmetric in nature, at least as far as morphosyntax is concerned, should not imply that there are no restrictions on the relative syntactic positions of target and controller in specific cases of agreement. After all, these positions cannot normally be reversed. In the next chapter, we will argue that, while in principle the controller can c-command the target or vice versa, specific instantiations of agreement typically require one of these configurations as a result of locality conditions.

6

Syntactic Restrictions on Agreement

6.1 Introduction

In the previous chapter, we argued that there is no syntactic asymmetry between φ-features on target and controller in an agreement relation. They are generated independently of each other, and whether or not a feature is interpreted does not depend on where it is generated initially. This syntactic symmetry appears to be at odds with the common assumption that agreement relations are subject to a condition of c-command that introduces an asymmetry. Some theories require that the target always c-command the controller (“downward agreement”; see Chomsky 2000, 2001, Preminger 2013, and Preminger and Polinsky 2015, among others). Other theories embody the opposite view: namely, that the controller always c-commands the target (“upward agreement”; see Chomsky 1995, Wurmbrand 2012, Zeijlstra 2012, and Bjorkman and Zeijlstra 2014). Both camps have strong arguments. An argument for downward agreement comes from structures in which the relation crosses clause boundaries. In that case, the target always seems to be located in the higher clause and the controller in the lower clause (compare Preminger and Polinsky 2015). (1) a. [ … V-agr1 … [CP … DP1 … ] … ] b. *[ … DP1 … [CP … V-agr1 … ] … ] An argument for upward agreement comes from the observation, going back to Koster 1987, that all grammatical dependencies share a set of properties known as the configurational matrix. One of these is obligatory c-command of the dependent category by the antecedent. Thus, predicates are c-commanded by their arguments, anaphors are c-commanded by their antecedents, traces are c-commanded by the moved category, and so on. If agreement is valuation, a view adopted by authors on both sides of the debate,

170

Chapter 6

it is a grammatical dependency in which the controller is the antecedent and the target the dependent category. On this basis, one would expect that the controller should c-command the target (see Zeijlstra 2012 and Bjorkman and Zeijlstra 2014). In response to this kind of argument, Preminger and Polinsky (2015) argue against premature unification of different syntactic dependencies. While we agree that it is not necessarily fruitful to analyze all syntactic dependencies as one (in terms of an abstract “Agree” operation that encompasses more than actual agreement), the point of the configurational matrix is that it generalizes over different dependencies. C-command by the antecedent is a hallmark of syntactic relations. Agreement would constitute a unique exception among its peers in this respect if the target must c-command the controller and agreement is valuation of the target. In our perception, what is missing from the debate is a discussion of why agreement should be upward or downward. The way to approach this question is to consider in more detail the general theory of grammatical dependencies and determine why antecedents c-command dependents. The next step is to consider how this interacts with the nature of agreement. The expected directionality of agreement will be different under different conceptions of agreement. If agreement involves feature valuation, and hence an inherent syntactic asymmetry between target and controller, then, as just noted, we would expect agreement to be upward. However, if this is not the case, as in our account in chapter 5, it is in principle possible for the target to c-command the controller or vice versa. We will argue that this is a desirable outcome. Of course, in a large number of concrete cases of agreement the c-command relation between target and controller is fixed. Hence, for us the challenge is to explain observed restrictions on the directionality of agreement in a theory in which these restrictions cannot follow from inherent properties of the agreement relation. A large part of this chapter is devoted to this issue. The picture that emerges is not unlike the theory of agreement developed in Baker 2008. Baker, too, argues that there is no universal directionality to the agreement relation. There are specific instances of agreement that are upward, specific instances that are downward, and instances that allow for variation in direction. Baker argues that the choice between these options is regulated by a macroparameter that generalizes over all functional heads in a language. Agreement between a functional head and a DP always requires c-command, but languages may in addition stipulate that the DP must c-command the functional head (or exceptionally that the functional head must c-command the DP). While we leave open the possibility that such macroparameters exist, our aim here is simply to argue that the syntactic symmetry of

Syntactic Restrictions on Agreement 171

agreement translates into variation of directionality and to explore what factors might restrict the direction of agreement in specific cases. In section 6.2, we address the general theory of syntactic dependencies and the way that agreement fits in. In section 6.3, we discuss the issue of “semantic agreement” and how it can be distinguished from syntactic agreement. This is important, since only the latter is encoded by a syntactic dependency. In section 6.4, we make the case that agreement has no inherent direction: we show that there are instances of both upward and downward agreement. The discussion includes cases whose (un)grammaticality is explained most straightforwardly under the assumption that a single head agrees upward with one argument and simultaneously downward with another. We then discuss, in sections 6.5 and 6.6, how syntactic restrictions governing specific instances of agreement follow from locality conditions. There are two types of such conditions. One type, discussed in section 6.5, has the effect that target and controller must be in the same domain. The other type, discussed in section 6.6, has to do with the notion of minimality: agreement relations cannot cross certain interveners. Both types of locality conditions apply equally to upward and downward agreement, but as we will argue they can nonetheless result in specific cases of agreement being possible only in downward fashion or only in upward fashion. Section 6.7 contains a summary and conclusions.1 6.2  Agreement and the Theory of Syntactic Dependencies

We begin by considering the nature of syntactic dependencies. Our starting point is Koster’s (1987) observation that all syntactic dependencies have a set of constant properties: obligatoriness, c-command by the antecedent, locality, uniqueness of the antecedent, and nonuniqueness of the dependent. It is this set of constant properties that Koster calls the configurational matrix. Neeleman and Van de Koot (2002) argue that the core properties of the configurational matrix can be made to follow from the theory of phrase structure, and in particular from Inclusiveness, the main principle that governs labeling (see Chomsky 1995). We will limit our discussion here to showing how Inclusiveness gives rise to obligatory c-command by the antecedent. We adopt the following definition of Inclusiveness: (2)  Inclusiveness The syntactic properties of a nonterminal node are fully recoverable from its daughters; the syntactic properties of a terminal node are fully recoverable from the lexicon.

172

Chapter 6

There is no dispute that the inheritance of categorial features is governed by the condition in (2).2 On standard assumptions, label inheritance is upward and cannot skip any node in a projection line. Inclusiveness has a reach that extends beyond projection, however: it applies to all syntactic properties of nodes. Among these, we argue, are the properties that characterize an element as the dependent in a syntactic relation. If such properties are indeed subject to Inclusiveness, then it follows that the configurations in which dependencies can be established are subject to certain restrictions. To see why, consider the tree in (3), where δ is a syntactic dependent (say, an anaphor or a predicate), and α is the antecedent with which it is associated. For the sake of explicitness, we assume that the fact that δ must be linked to an antecedent is encoded by a selectional requirement (sr). In (3), the selectional requirement is satisfied by α (we use † to indicate this). … (3)  α … …

δ [SR†]

The relation established between α and δ determines properties of δ, simply because the selectional requirement in δ cannot be satisfied again. Suppose, for example, that α is an argument, δ is a predicate, and the selectional requirement is a θ-role. Once the selectional requirement is satisfied by α, δ no longer qualifies as a θ-role assigner. Therefore, the fact that the selectional requirement is satisfied represents a syntactic property of the node in which it occurs. Given that δ does not dominate α, the fact that the selectional requirement in (3) carries † cannot be recovered from the internal structure of δ. Hence, δ violates Inclusiveness. This problem is not restricted to thematic relations. For example, we may characterize the dependent nature of an anaphor as a selectional requirement satisfied by its antecedent in a configuration like (3). Thus, the standard, chainlike encoding of binding also violates Inclusiveness. Mutatis mutandis, the same is true of other syntactic dependencies. The logic of the problem dictates what alternative encoding of syntactic dependencies we must adopt. Only if the selectional requirement is copied upward recursively to the node that immediately dominates α, as in (4), can it be satisfied without violating Inclusiveness. As long as no node in the path between δ and α is skipped, upward copying satisfies this condition, because each new copy on a node can be recovered from its daughters. Downward satisfaction of the selectional requirement by α also obeys Inclusiveness, because the element that determines the status of the selectional requirement is a daughter of the node that contains it.

Syntactic Restrictions on Agreement 173

(4) 

… [SR†] α

… [SR] δ [SR]



Note that α in (4) does not acquire any syntactic properties in virtue of satisfying the selectional requirement. (If it did, this property on α would violate Inclusiveness.) For example, if the selectional requirement is a θ-role, there is no sense in which δ “assigns” a θ-role to α or α is the “recipient” of a θ-role. Thus, the theory we propose implies that the argumenthood of α is expressed through its relation to δ rather than as a syntactic property of α itself. In contrast to downward satisfaction, as in (4), sideward satisfaction under sisterhood violates Inclusiveness: the fact that the selectional requirement in (5) is satisfied in α’s sister node cannot be recovered from material dominated by this node. … (5)  α … [SR†] δ [SR]



So, Inclusiveness forces a decomposition of syntactic dependencies into two primitive operations: the upward copying and downward satisfaction of a selectional requirement. Two key properties of syntactic dependencies follow from this. The first is that such dependencies may, in principle, span large distances. This is because the copy relation can be established recursively: a copied selectional requirement can itself be copied. As a result, the path along which a selectional requirement travels up the tree can be indefinitely long (as long as applicable locality conditions are adhered to). On the other hand, Inclusiveness requires the relation between the antecedent and the highest node that contains the selectional requirement to be extremely local: the structure in (6), where β intervenes between α and the node containing the selectional requirement, is ruled out, because the status of sr in the top node cannot be determined by looking at β. Access to α is required, but α is not a daughter of the node containing sr†. (6)  *

… [SR†] … [SR]

β α





δ [SR]

174

Chapter 6

The import of the proposed encoding of syntactic dependencies is that, while Inclusiveness is a restriction on phrase structure, it also explains why a dependent must be c-commanded by its antecedent. One clarification is in order. The feature that satisfies a selectional requirement can itself have been copied. Given that this is possible, can it really be guaranteed that the antecedent occupies a c-commanding position? For example, could the feature of α that satisfies the selectional requirement in (6) be copied to β and satisfy it there? This problem is only apparent, however. The grammar does not permit gratuitous percolation of features. Copied features become properties of the node they are copied to. Therefore, the relevant feature can only be copied to β if it is taken to characterize this node, in which case the selectional requirement is in fact satisfied by β. Of course, copied selectional requirements also become properties of the node they are copied to. However, whereas features indicate what a node is, selectional requirements indicate what a node asks for. In many cases, a node cannot be two things at the same time (say, a verb and a preposition). But it can be one thing (say, a verb) and ask for another (say, a prepositional complement). Thus, copying of selectional requirements is less restricted than copying of features. We now consider what consequences this theory of syntactic dependencies has for agreement. The theory developed in chapter 5 does not postulate an inherent syntactic asymmetry between target and controller. At the moment when the agreement relation is established, both the DP and the verb contain φ-features that are interpretable in principle. (The fact that they are interpreted in DP is an LF matter, which cannot bear on the syntactic encoding of the dependency.) In contrast to agreement, other syntactic dependencies involve an element that has a featural deficit. For instance, in a binding relation the anaphor typically lacks φ-features that are present in other pronominal elements (Chomsky 1981, Bouchard 1984, Reinhart and Reuland 1993), while in movement relations the trace can be argued to lack any inherent properties (see the implementation of the copy theory of movement in Neeleman and Van de Koot 2010). It is not surprising, then, that the inherent asymmetry resulting from the deficit of features in one of the elements translates into a syntactic asymmetry, such that the element in question is the dependent in the grammatical relation. If there were a similar asymmetry in agreement relations, then only the target could be the dependent, so that agreement would have to be upward. But if target and controller do not differ in the syntactic status of their features, as argued, then either the target or the controller can function as the dependent in a syntactic agreement relation, at least in principle.

Syntactic Restrictions on Agreement 175

Of course, the selectional requirement that encodes agreement must always be copied upward, so in this sense agreement is upward, like any other syntactic dependency. However, the notions of antecedent and dependent do not coincide with the notions of controller and target. If the target is the dependent, then agreement is upward; if the controller is the dependent, then it is downward. The two situations are depicted in (7), where ctr stands for the selectional requirement that links a target to a controller, and trg for the selectional requirement that links a controller to a target. (Since it is standard to speak of “upward agreement” and “downward agreement,” we will stick to this terminology, but it should be kept in mind that the selectional requirement that encodes the agreement relation is always introduced by the lower category.) (7)  a. Upward agreement

b. Downward agreement … [TRG†]

… [CTR†] DP

… [CTR] …

V [CTR]

V

… [TRG] …

DP [TRG]

The main advantage of this view, it seems to us, is that it reconciles the general observation that in syntactic dependencies the antecedent must c-command the dependent category with the fact that in at least some agreement relations the target c-commands the controller. Notwithstanding the basic symmetry of the agreement system, a number of syntactic asymmetries can be observed in specific instances of agreement. In general, a given agreement relation does not have the choice of being either upward or downward: it has to be one or the other. Moreover, there seem to be asymmetries in locality in upward and downward agreement (compare (1)). In what follows, we will analyze these asymmetries. Our overall account will be based on the hypothesis that such asymmetries are never inherent in the agreement relation itself, but are a consequence of other factors. There are two such factors that we consider. The first relates to Case and the second to locality. Case plays a role because nominatives can act as controllers more easily than DPs with other Cases (we will make this more precise in section 6.4). Locality comes in two types, the “barriers” type and the “relativized minimality” type. Barrier-type restrictions (discussed in section 6.5) state that agreement must take place within a local domain. This can force upward agreement when there is a domain boundary immediately below the agreeing head. Similarly, downward agreement can be forced when there is a domain boundary

176

Chapter 6

immediately above the agreeing head. We schematize this in (8), where B simply stands for barrier at this point, and agreement relations are indicated by coindexation. (8) a. DP1 … X1/*2 … [B … DP2 …] b. DP1 … [B … X*1/2 … DP2 …] Relativized minimality–style conditions (discussed in section 6.6) state that agreement relations span minimal distances. On the path between the elements that introduce and satisfy a selectional requirement encoding agreement, there cannot be other elements that are of the same type (that is, potential dependents or antecedents). This condition, too, can lead to asymmetries between upward and downward agreement. This is because the properties of the selectional requirement introduced by a DP are inherently different from the properties of the selectional requirement introduced by a head. The former looks for a particular type of head to satisfy it, while the latter looks for a particular type of DP. Before we look at restrictions on syntactic agreement, however, it is necessary to demarcate the domain of inquiry. Agreement-like phenomena abound, and not all instances involve syntax. In particular, so-called semantic agreement cannot belong to the same system. This means that we must have criteria to determine whether a specific case of agreement is syntactic or semantic. If we do not, there is a danger that instances of semantic agreement will be incorrectly taken to inform the theory of syntactic agreement. 6.3  Syntactic Agreement versus Restrictor Agreement

Traditionally, a distinction is made between syntactic agreement and semantic agreement (see, for instance, Corbett 2006:155ff.). In this section, we argue that, while there are indeed two types of agreement, these do not map perfectly onto the traditional bifurcation. One type of agreement is indeed syntactic. The nonsyntactic type of agreement, however, has two subtypes, only one of which has semantic import. The section has three parts. First, we discuss agreement between pronouns and their antecedent. Then, we turn to adjectival agreement. Finally, we consider the criteria by which syntactic agreement can be distinguished from nonsyntactic agreement, especially in case the latter makes no semantic contribution. 6.3.1  Nonsyntactic Agreement with Pronouns

The distinction between syntactic agreement and nonsyntactic agreement is often seen as a distinction between agreement on the basis of formal features

Syntactic Restrictions on Agreement 177

of the controller and agreement determined by the controller’s interpretive properties. In the previous section, we discussed how syntactic dependencies are established. Syntactic agreement is an instance of such a dependency. Thus, one of the elements involved introduces a selectional requirement that is satisfied, after having been copied upward, by the other. The question, then, is in what ways nonsyntactic agreement is different. The issue of demarcation is especially important since, as we will now discuss, the crucial split is not one between agreement for formal features and agreement on the basis of semantic properties. If syntactic agreement is indeed constrained by the properties of the configurational matrix, then at the very mimimum a syntactic agreement relation cannot be established across sentence boundaries. One criterion for distinguishing the two types of agreement, therefore, is structural connectedness: if an agreement relation crosses a sentence boundary, it must be nonsyntactic. Such agreement does exist, of course. For example, there is gender and number agreement between a pronoun and its antecedent, whether they are contained in the same sentence or not. (9) a. Mary thought that she/*he would like a new hat. b. Mary walked down the street. She/*He was wearing an old hat. Nevertheless, cross-sentential agreement with a pronoun can be determined by the formal features of the antecedent. We can demonstrate this using Dutch data. Dutch, as opposed to English, has a distinction between two morphological genders, neuter and nonneuter (or “common”) gender. Consider first inanimate nouns, which of course do not refer to something that is biologically male or female. With such nouns, the choice of gender of a coreferential personal pronoun is predominantly determined by the morphological gender of the antecedent. This is illustrated in (10).3 (10) a. Hoe zit het met het fornuis? Heb je ’t/?*’m nu eindelijk how sits it with the.neut stove have you it/him now finally laten repareren? let repair ‘How about the stove? Have you finally had it repaired?’ b. Hoe zit het met de televisie? Heb je ’m/*’t nu eindelijk how sits it with the.comn television have you him/it now finally laten repareren? let repair ‘How about the television? Have you finally had it repaired?’ These facts are particularly clear because there is no possibility of agreement on the basis of semantics (modulo the issues mentioned in note 3). At the same

178

Chapter 6

time, pronominal agreement on the basis of semantics is not impossible. In particular, there are morphologically neuter words that refer to a male or female person, such as diminutives derived from nouns referring to people (all diminutives are grammatically neuter in Dutch). In examples involving such nouns, a coreferential pronoun can be chosen either on the basis of the formal features of the antecedent or on the basis of its semantics, even if the antecedent and the pronoun are not contained in the same sentence. Semantic agreement is often preferred, but both options exist, as illustrated in (11) (for more discussion, see Van Beurden and Nijen Twilhaar 1990, Audring 2009, and Ackema 2014). (11) a. Het jongetje kwam dichterbij. Ik zag dat ’t/ie een pleister op the.neut boy.dim came closer I saw that it/he a plaster on z’n wang had. poss cheek had ‘The little boy came closer. I saw he had a plaster on his cheek.’ b. A: Heb je dat kleine meisje met dat rode jasje onlangs have you that.neut small girl with that red coat recently nog gezien? yet seen ‘Did you see that little girl with the red coat recently?’ B: Ja, ik zag ’t/’r gisteren nog op straat spelen. yes I saw it/her yesterday yet on street play ‘Yes, I saw her playing in the street only yesterday.’ Hence, if an agreement relation is computed using semantic properties of the controller, it must be nonsyntactic (that is, not established through a dedicated syntactic dependency). However, if an agreement relation is computed using formal features of the controller, it can be either syntactic or nonsyntactic. In other words, agreement for formal features is a necessary but not sufficient condition for characterizing an agreement relation as syntactic. As with the coreferential pronouns in (11), the choice of bound variable pronouns can be determined either by the semantics of the antecedent or by its formal features, as discussed in chapter 2. We give two representative examples in (12). (12) a. [Elk meisje]i denkt dat heti/zei de wedstrijd gaat winnen. every girl thinks that it/she the match goes win ‘Every girl thinks that she will win the race.’

Syntactic Restrictions on Agreement 179

b. [Elk raar mannetje]i in dat verhaal denkt dat heti/hiji de every strange man.dim in that story thinks that it/he the enige bewoner van het bos is. only inhabitant of the forest is ‘Every strange little man in that story thinks that he is the only one living in the forest.’ The standard analysis of this kind of data invokes two distinct mechanisms. The φ-features of the pronoun either introduce a presupposition or enter into a syntactic agreement relation with the binder. However, the observations pertaining to coreferential pronouns imply that only a single mechanism might be at play—namely, nonsyntactic agreement—but that this mechanism may or may not bear on the semantics. Indeed, also when agreement is for formal features, the relation between the bound pronoun and its antecedent need not comply with the c-command restrictions on syntactic dependencies. This is shown in (13). (13) De moeder van [elk jongetje]i denkt dat heti/hiji de competitie the mother of every boy.dim thinks that it/he the competition gaat winnen. goes win ‘Every boy’s mother thinks that he will win the competition.’ If nonsyntactic agreement cannot simply be equated with semantic agreement, what is it? Our proposal is that nonsyntactic agreement does not involve a selectional requirement for agreement at all. Rather, the φ-features on the apparent target function as restrictors on a variable that is introduced by this element and subsequently associated with the apparent controller. Henceforth, we will refer to this type of agreement as restrictor agreement. The mechanism involved in restrictor agreement can be schematically represented as in (14), where the cosuperscripting of the variable and the φ-features expresses that these features impose a restriction on the variable. (14) DP1 … [X1k φk] If X is a semantic variable, the φ-features are interpreted as presuppositions about possible values of X. This yields semantic agreement. If X is the syntactic representation of the variable, the φ-features impose a formal restriction on possible antecedents of X. This yields restrictor agreement for formal features. This analysis can be applied to pronominal agreement, as long as we assume that pronouns contain a syntactic representation of a semantic variable. For bound variable pronouns, it is a truism that they introduce a semantic variable, and it is uncontroversial that this semantic variable is represented in syntax.

180

Chapter 6

Following Partee (1978), we assume that the same is true of unbound pronouns. If a pronoun remains unbound, its reference is determined pragmatically as a particular choice for the variable. This analysis is straightforward for semantic and formal restrictor agreement under bound variable binding, and for the semantic subcase under coreference. However, restrictor agreement for formal features under coreference requires more discussion. It is not obvious how a pronoun’s features can restrict the possible formal features of potential antecedents when there is no syntactic relation between pronoun and antecedent. There are two possibilities. First, we could assume that in cases of cross-clausal antecedence there is a syntactic dependency after all, though not with the apparent antecedent, but with an empty category that has the same semantic content in the local left periphery. Alternatively, we could assume that discourse representations are richer than traditionally assumed, in that they contain formal features of the DPs that introduce discourse entities. If so, pronouns that link to a discourse antecedent can be sensitive to these features. A possible argument for the second option comes from cases where a hidden antecedent would not provide a solution, because this antecedent is itself an anaphoric element that needs an antecedent. In the case of referential pronouns, this problem of regress can be circumvented by assuming that the local silent antecedent is a covert R-expression (for example, a silent representation of the discourse topic, per Collins and Postal 2012). With some other types of anaphoric element, however, regress seems unavoidable. Consider crosssentential sluicing, as in (15). (15) A: They managed to assemble this IKEA bed. B: Do you know how __? Whether sluices need a linguistic antecedent is a somewhat controversial issue (they do, according to Merchant 2001 and much subsequent literature, but see Ginzburg 1992 and Culicover and Jackendoff 2012 for potential counterexamples). Whatever the answer might be in the general case, in the specific case of (15) we have not been able to construct situations that would license the reply in the absence of a linguistic antecedent for the elided material. For example, even if someone were looking in desperation at the contents of an IKEA flatpack, it would still be infelicitous for them to ask their friend Do you know how? If elided material in sluices of this kind needs a linguistic antecedent, then any silent IP antecedent for the sluice would itself have to have a local silent IP antecedent, which means that assuming a silent antecedent merely shifts the problem. Quite apart from this, it is hard to see what function the silent IP would have other than to provide a local antecedent for the elided material.

Syntactic Restrictions on Agreement 181

On the other hand, there are data that would seem to fit more naturally into an approach assuming null local antecedents. These involve instances of pronoun resolution in the absence of an overt linguistic antecedent, as in (16) (compare with (10)). (16) a. (Context: The speaker is pointing to the stove.) Heb je ’t/*’m nu eindelijk laten repareren? have you it/him now finally let repair ‘Have you finally had it repaired?’ b. (Context: The speaker is pointing to the television.) Heb je ’m/*’t nu eindelijk laten repareren? have you him/it now finally let repair ‘Have you finally had it repaired?’ As these examples show, in such cases the features of the pronoun can still be determined by the grammatical features of the linguistic expression associated with the pronoun’s referent (as fornuis ‘stove’ is a neuter word in Dutch, while televisie ‘television’ is nonneuter). If there is a possibility of having null antecedents in discourse that are contextually licensed, these data are unproblematic. If one wants to avoid this, it must be assumed that one of the properties of discourse referents consists of the φ-features of their linguistic representation. We will not try to assess here how feasible this position is (see Dowty and Jacobson 1988 for discussion). 6.3.2  Restrictor Agreement with Adjectives

As a further example of how restrictor agreement works, we return to agreement in French between a predicative adjective and its subject, as discussed in chapter 2. Relevant examples are repeated in (17). They contain the second person plural pronoun vous, which can have a singular referent when used as a polite form of address. (17) a. Vous êtes loyal. you.pl be.2pl loyal.masc.sg ‘You (single male addressee, polite) are loyal.’ b. Vous êtes loyaux. you.pl be.2pl loyal.pl ‘You (addressee plus associate(s), potentially polite) are loyal.’ Crucially, polite vous must be syntactically characterized as second person plural, in view of the agreement on the finite verb.4 The alternation in the adjective must therefore be explained on the basis of the interpretation of vous, which refers to a single male addressee in (17a), but to a group including the addressee in (17b). On the other hand, not all instances of adjectival agreement

182

Chapter 6

are interpreted semantically. For example, inanimate nouns in French trigger agreement for gender, even though this feature is not interpreted. Thus, the different forms of the adjectives in (18a,b) depend on the grammatical gender of the subject. (18) a. Le journal est nouveau. the.masc newspaper is new ‘The newspaper is new.’ b. La maisonette est nouvelle. the.fem duplex is new ‘The duplex is new.’ We assume that adjectival agreement in French is another instance of restrictor agreement. The predicative adjective introduces an argument variable that is syntactically represented as a θ-role. Much like the φ-features in pronouns, the φ-features of the adjective can be linked either to the semantic variable or to its syntactic representative. The former case results in traditional semantic agreement, while the latter results in restrictor agreement for formal features. The data in (17) can then be analyzed as follows. If we represent the fact that the adjective takes a subject through λ-abstraction, the φ-features marked on the adjective can be mapped onto a restrictor of the λ-operator, as in (19a). The semantic effect is that the domain of the λ-operator is limited to male singular referents. If the formula in (19a) is subsequently applied to vous under predication, as in (19b), it follows that the referent of vous is presupposed to be masculine and singular.5 (19) a. ⟦loyal⟧: λx | masculine(x) ∧ singular(x) | loyal(x) b. ⟦vous⟧: λx | masculine(x) ∧ singular(x) | loyal(x) A mechanism like this is independently required to deal with structures in which a verb carries a morpheme that expresses that one of the verb’s arguments has a specific set of semantic properties, often referred to as “classifier incorporation”; see Mithun 1984 and Rosen 1989 for discussion. (Classifier incorporation should not be treated as syntactic agreement, as this would lead to an unacceptable proliferation of φ-features. For instance, it seems undesirable to treat the property of being a fish as a φ-feature in order to deal with structures like ‘my father fish-bought eight bullheads’, which are found in Mohawk.) The data in (18) can be understood if the φ-features on the adjective are linked to its external θ-role in syntax, as in (20). If so, they restrict the kind of element that can satisfy this θ-role. This is not a theoretical novelty, given

Syntactic Restrictions on Agreement 183

that c-selectional requirements are typically linked to particular θ-roles. For example, the fact that wonder must take a CP complement, whereas ask can also take a DP complement, is encoded by linking c-selection for a CP to the internal θ-role of wonder, but not to the internal θ-role of ask (for discussion, see Grimshaw 1979). If a θ-role can impose restrictions on the categorial features of its recipient, it is not surprising that the information contained in the ending of an adjective can be combined with the adjective’s external θ-role in a similar way. … [θ†]

(20)  DP

… [θ] A [θk FEMk]

… A [θ]

[FEM]

We should note that, in contrast to the relation between a pronoun and a coreferential DP, there is a syntactic dependency that connects a predicative adjective and its subject, namely, external θ-role assignment. This relation, like other syntactic dependencies, is established via a selectional requirement, in this case introduced by the adjective. As a result, the relation between the adjective and the subject must meet the restrictions imposed by the configurational matrix, which implies, among other things, that it must be local and that it requires c-command. If adjective and subject must stand in such a syntactic relationship in order for predication to be possible, then, of course, agreement will also only occur when the relevant syntactic conditions are met. Nevertheless, these syntactic restrictions are not imposed by a syntactic agreement relation initiated by the φ-features on the adjective; rather, they are inherited from the predication relation. In general, then, agreement relations that appear to be restricted by properties of the configurational matrix are not necessarily instances of syntactic agreement, but can be instances of restrictor agreement. On the above analysis, there should be a thematic relation between an agreeing adjective and the associated controller. One might think that restrictor agreement should therefore not hold between a raising adjective and its syntactic subject. This depends, however, on the analysis of raising. On the general view that syntactic dependencies are encoded through a selectional requirement that originates in the dependent category (see section 6.2), traces of movement must introduce selectional requirements as well. We can make an educated guess about the nature of the selectional requirement introduced by the NP trace (the trace involved in A-movement) if we take as our starting

184

Chapter 6

point the desideratum that it should be possible to generalize over subjects of different classes of verb. How to achieve this is not entirely trivial. Subjects of transitive and unergative verbs receive external θ-roles, whereas subjects of passive and unaccusative verbs do not (their subjects are moved from a lower thematic position). There are two ways in which one might develop a unified view of subjects. The first relies on a positional definition of subjects. For example, one could assume that subjects of unaccusative and passive verbs move to the position in which subjects of transitive and unergative verbs are base-generated (the standard view in early Government-Binding Theory). Alternatively, one could assume that all subjects move from a thematic position to a designated position external to VP (see Kuroda 1988, Koopman and Sportiche 1991). In either case, a DP qualifies as a subject if and only if it occupies a particular structural position. The alternative is to define subjects relationally: the subject of VP is the category that satisfies VP’s external θ-role. Normally, this θ-role is introduced by the verb, but in passive and raising constructions it is introduced by NP trace (see Williams 1987, 1994, Neeleman and Weerman 1999, Neeleman and Van de Koot 2002, 2010). If the relevant θ-role is satisfied VP-externally, subjects of passive and unaccusative verbs are indeed relationally equivalent to subjects of transitive and unergative verbs, as illustrated in (21). In (21b), the trace receives the verb’s internal θ-role and simultaneously introduces a θ-role that is copied into the verbal projection (this θ-role appears in boldface for traceability). This role is satisfied by the VP-external argument, just like the external θ-role in (21a). If the trace corresponds to a variable, equated with Bill through θ-role assignment, then—by transitivity—Bill will be interpreted as an argument of the verb. … [θ†]

(21)  a. DP he

V [θ θ†] V [θ θ] fired

DP Bill

…[q†]

b. DP Bill

… [q] Aux was

V [q θ†] V [θ] fired

tDP [q]

Syntactic Restrictions on Agreement 185

This analysis permits restrictor agreement with raising adjectives: the φ-features of the adjective can be associated with the θ-role copied upward from its complement, as in (22). … [q†]

(22)  DP

… [q] A [qk FEMk]



… [q θ†]

A [FEM] A

[FEM]

tDP [q]

V [θ] …

We cannot test whether this occurs in French, as the language lacks raising adjectives. Indeed, we are not aware of languages that have raising adjectives as well as robust adjectival restrictor agreement, and must therefore leave this matter to future research. Agreement between participles and their objects in French can be modeled in the same way as agreement between adjectives and their subjects. Participial agreement can be seen as an instance of restrictor agreement, because there is a thematic relationship between the verb and its object. Moreover, both instances of agreement are for the same features (number and gender). The biggest obstacle to analyzing participle agreement as nonsyntactic is that the syntactic context determines whether the agreement shows up or not. No participial agreement is observed if the object remains in situ, but agreement does show up when the object moves (see Kayne 1989 and Fuß 2005:87). (23) a. Marie a fait(*-e) la robe. Marie has made(-fem) the dress ‘Marie has made the dress.’ b. La robe a été fait*(-e). the dress has been made-fem ‘The dress has been made.’ c. Marie l’a fait*(-e). Marie it.has made-fem ‘Marie has made it.’ d. Combien de tables Paul a repeint*(-es)? how.many of tables Paul has repainted-fem.pl ‘How many tables has Paul repainted?’

186

Chapter 6

The analysis proposed by Kayne (1989) is that participial agreement requires a specifier-head configuration, and the leftward movement of the object must pass through the specifier of a projection headed by the participle. If so, the agreement relation between participle and object must be syntactic. There is, however, an alternative that does not necessitate this conclusion. This alternative assumes that the φ-features that enter into restrictor agreement are always present in the participle, but are not spelled out if the object remains in situ. This lack of spell-out fits into a general typological pattern related to Greenberg’s (1963:74) Universal 33. Universal 33 states that “when number agreement between the noun and verb is suspended and the rule is based on order, the case is always one in which the verb precedes and the verb is in the singular.” In Ackema and Neeleman 2004, we show that similar effects are found with subject-verb agreement for features other than number. We argue there that, in the languages in question, subject agreement with finite verbs is always present, but the φ-features in the verb are deleted at PF when verb and subject appear in the same prosodic phrase. The rule for prosodic phrase formation is such that in the relevant languages the verb and subject will only be parsed together if the subject follows the verb and there is no maximal projection in between. A detailed discussion of one particular case can be found in chapter 7 below. In parallel to this, one could analyze the French participial agreement data in terms of a rule of impoverishment. The φ-features in the participle are deleted just in case the participle is in the same prosodic phrase as the object, which is the case when the former immediately precedes the latter. Some support for an analysis along these lines comes from the acquisition path taken by children learning Italian. As Borer and Wexler (1992) discuss in detail, there is a stage in the L1 acquisition of Italian in which agreement between participle and object is obligatory regardless of whether the object moves or is in situ. Only later is agreement absent in object-in-situ structures. This is exactly the development predicted by our account. First, upon encountering positive evidence for the presence of object agreement on participles in the language, the child acquires a rule that treats this agreement as restrictor agreement (since regular syntactic agreement with accusative objects is ruled out; see subsection 6.4.1). This means that in production, agreement will be general, leading to overgeneralization in comparison with the target language. Subsequent positive evidence shows the child that agreement is absent if the object remains in situ. The child then acquires a rule of impoverishment that deletes the relevant φ-features in the participle at PF, prior to spell-out, when the participle is in the same prosodic phrase as the object.

Syntactic Restrictions on Agreement 187

6.3.3  How to Distinguish Syntactic Agreement from Restrictor Agreement

In the previous subsections, we have shown that not all instances of agreement are syntactic and that nonsyntactic agreement can be either semantic or formal. The question then arises how to distinguish syntactic agreement from restrictor agreement. We have already suggested that syntactic agreement requires c-command and must be for formal features, while restrictor agreement is only possible if the agreeing element introduces a variable that is bound by the controller. We have used violations of the former criterion to identify unambiguous instances of restrictor agreement. In the same vein, the latter criterion can be used to identify unambiguous instances of syntactic agreement. An example of this is complementizer agreement as found in a variety of Germanic dialects (see De Haan 1997, Zwart 1997, Ackema and Neeleman 2004, and Haegeman and Van Koppen 2012, among many others). The phenomenon is illustrated by the West Flemish data in (24) (see Haegeman 1992, Shlonsky 1994, and De Schutter 1997). (24) a. da-n ik werk-en that-1sg I work-1sg b. da gie werk-t that-2sg you work-2sg c. da ze/Valère werk-t that-3sg she/Valère work-3sg d. da-n wunder werk-en that-1pl we work-1pl e. da gunder werk-t that-2pl you.pl work-2pl f. da-n zunder/Pol en Valère werk-en that-3pl they/Pol and Valère work-3pl Given that there is no semantic relationship between the complementizer and the subject, an analysis as restrictor agreement is not an option. Another example is long-distance agreement. This involves a configuration in which a head in a matrix clause agrees with a controller in an embedded clause, as in the Tsez example in (25) (from Polinsky and Potsdam 2001:584). The absolutive object inside the embedded clause does not entertain a semantic relation with the matrix verb; nevertheless, it agrees with this verb. This agreement relation cannot be an instance of restrictor agreement; rather, it must be syntactic. (25) Enir [užā magalu bāc’rułi] b-iyxo. mother boy bread.III.abs ate III-know ‘The mother knows the boy ate the bread.’

188

Chapter 6

The criteria outlined above still leave examples that permit an analysis in terms of either syntactic agreement or restrictor agreement. In such examples, agreement is for formal features and it holds between c-commanding categories that also entertain a semantic relationship. We propose that in such circumstances the language-learning child is forced to make a choice, and that this choice is guided by a preference for syntactic agreement. The reason for this is that there is simple positive evidence that could trigger a switch to an analysis in terms of restrictor agreement: namely, other examples that show semantic agreement between the same categories in an identical configuration. However, there is no positive evidence that could make the child abandon an analysis in terms of restrictor agreement in favor of syntactic agreement.6 As a result of this preference in acquisition, all rules of agreement that could in principle be of the syntactic type or the restrictor type must in fact be syntactic. One instance of such a rule is the agreement between finite verb and subject in most Germanic and Romance languages. We now turn to restrictions on specific instances of syntactic agreement, considering how these can be accounted for under the hypothesis, advanced in chapter 5, that there is no inherent syntactic asymmetry between the features in controller and target. 6.4  Agreement: Upward and Downward 6.4.1  Encoding Upward and Downward Agreement

Consider syntactic agreement between a verbal head and a DP.7 The symmetric nature of agreement implies that selectional requirements encoding this relation can be introduced either in the verbal head or in heads within the DP. We propose that such selectional requirements are licensed by a lexical rule in categories that have φ-features; see (26). We use the label ctr for the selectional requirement introduced in a verb, given that it will be satisfied by an element that acts as controller in the agreement relation. Similarly, we use the label trg for the selectional requirement introduced by a head in a DP, as it will be satisfied by an element that functions as target. (26) a. [φ] → [φ ctr] b. [φ] → [φ trg] We have assumed that agreement is a two-stage process: in syntax, it is determined which two categories agree, while subsequent LF processes of feature spreading, dissociation, and identification ensure that the φ-features are interpreted in the correct position. The rules in (26) belong to the first stage of this process, as they introduce a selectional requirement used to establish a syntactic agreement relation.

Syntactic Restrictions on Agreement 189

The rule in (26a) applies to verbal inflection positions; we will simply label these positions agr. The rule in (26b) applies to the prs and nmb heads in DP. There is, of course, parametric variation regarding the heads that can initiate agreement. Nevertheless, there are universal restrictions on the introduction of the relevant selectional requirements. These follow, we suggest, from the assumption that there is a basic incompatibility between Case and φ-features, which expresses itself in two ways. The first is that KPs (Case-marked DPs) cannot carry φ-features. The consequence of this is that LF feature spreading from V to KP is blocked. As a result, regular agreement with a KP is impossible. To be sure, KPs can be involved in agreement relations, but only if the φ-feature on V need not be interpreted and so need not undergo spreading. Hence, syntactic agreement with a KP will always be default agreement. It has been argued that there is a basic distinction between nominative and absolutive DPs on the one hand and DPs with any other Case on the other: the former are bare DPs, while the latter have a Case shell, so are KPs (see Falk 1991, Bittner and Hale 1996, and Neeleman and Weerman 1999; see Caha 2013 for related ideas). In other words, nominative/absolutive is, in fact, the absence of Case. If so, it will be nominatives and absolutives that can be involved in LF feature spreading, and hence in regular agreement. The second way in which Case is incompatible with φ-features, we assume, is that KP is a barrier for percolation of trg, the selectional requirement that encodes downward agreement. The result of this is that Case-marked DPs will never initiate an agreement relation. The prs and nmb heads in such a DP could in principle introduce trg, but this selectional requirement could never reach a node external to KP, and hence cannot be satisfied by a verbal antecedent. The result is that downward agreement must be initiated by nominatives/ absolutives. By contrast, upward agreement can in principle be satisfied by either a DP or a KP, with the proviso that upward agreement with a KP will lead to a default form of agr (see above). Notice, however, that there cannot be a rule for agreement with a position that systematically contains KPs rather than nominatives or absolutives. Agreement rules can only be acquired if there is an alternation in the form of the verb that is correlated with the φ-features of the controller. Hence, default agreement must piggyback on the existence of regular agreement with DPs in the same syntactic configuration. In other words, only if a language has agreement with nominative DPs in a particular configuration can it have default agreement with nonnominative DPs in that configuration. If the proposed incompatibility between Case and φ-features is indeed universal, this has implications for the analysis of (nondefault) agreement between

190

Chapter 6

a verbal head and a Case-marked DP. Such agreement cannot be analyzed on a par with regular verbal agreement; instead, it must either be restrictor agreement or involve clitic doubling. Recall from the discussion of Spanish unagreement in chapter 5 that clitic doubling may itself involve syntactic agreement within a “big DP.” This would be compatible with our proposal, as long as this agreement relation is established below the KP level. After the agreement relation has been established, the clitic can move out of DP and attach itself to a verb, giving rise to apparent object agreement. There is one complication with the assumption that upward agreement relations are introduced by the rule in (26a). As it stands, the object in a transitive clause would act as intervener for the relation between the verb and the subject (see section 6.6 on intervention effects). This may provide insight into certain patterns of agreement in ergative-absolutive languages. In such languages, absolutive agreement is agreement with the closest c-commanding argument, as predicted. In nominative-accusative languages, however, the agreement relation with the subject of a transitive clause is established across the object. This means that the rule in (26a) is too coarse-grained. However, it is in the nature of selectional requirements that they impose restrictions on possible antecedents. The selectional requirement introduced by agr can therefore specify that the element satisfying it must be a subject.8 If so, objects cease to be interveners, resulting in a nominative agreement pattern. This does not imply that there can be no agreement with nonsubjects in nominative-accusative languages, as it is possible for such languages to have rule (26b) in their grammar (alongside (26a)). Thus, if there are contexts that permit nominative objects, such objects can agree with the verb. As an initial illustration of this possibility, we consider nominative-dative inversion in Dutch. Dutch is a nominative-accusative language, which, as just noted, implies that verb-initiated agreement is subject agreement. Accusative objects cannot introduce a selectional requirement for agreement, and therefore we do not expect to find object agreement. However, Dutch allows nominative objects to appear—namely, in subjectless sentences. The fact that Dutch has subjectless sentences can be shown on the basis of impersonal passives like (27).9 In impersonal passives, the verb does not initiate an agreement relation, and it therefore shows up in its default form (third person singular, as discussed before). (27) Gisteren werd in Amsterdam weer enorm gefeest. yesterday was in Amsterdam again enormously partied ‘There was again lots of partying going on in Amsterdam yesterday.’

Syntactic Restrictions on Agreement 191

The absence of a nominative subject makes it possible, at least in principle, for there to be a nominative object. Such an object can introduce a selectional requirement for agreement that is satisfied by the verb. Thus, in unaccusative and passive constructions, an underlying direct object can remain in situ and yet trigger agreement with the verb (see, for example, Den Besten 1985 and Broekhuis 1992). The effect is clearest with bivalent verbs. Dutch allows A-scrambling, but in contrast to some other languages (like German or Japanese) it does not allow this process to change the order of arguments. In double object constructions, the only possible order is subject – indirect object – direct object. (28) a. Ik geloof dat de meisjes Jan het boek gaven. I believe that the girls Jan the book gave.pl ‘I believe that the girls gave Jan the book.’ b. *Ik geloof dat de meisjes het boek Jan gaven. I believe that the girls the book Jan gave.pl c. *Ik geloof dat Jan de meisjes het boek gaven. I believe that Jan the girls the book gave.pl In passives and with some unaccusative experiencer verbs, however, two orders are possible. The underlying direct object carries nominative in these structures and determines agreement on the finite verb. This object can either precede the dative argument, as in (29a) and (30a), or follow it, as in (29b) and (30b). The latter order is usually said to display “nominative-dative inversion.” (29) a. Ik geloof dat de boeken Jan gegeven werden. I believe that the books Jan given were ‘I believe that the books were given to Jan.’ b. Ik geloof dat Jan de boeken gegeven werden. I believe that Jan the books given were (30) a. Ik geloof dat de boeken Jan bevielen. I believe that the books Jan please.pst.pl ‘I think that the books were agreeable to Jan.’ b. Ik geloof dat Jan de boeken bevielen. I believe that Jan the books please.pst.pl There is general agreement that in examples of this type the dative argument does not occupy the subject position. Evidence for this includes the observation that attempted control of the indirect object leads to ungrammaticality, while control of the underlying direct object is possible; see (31). Given that PRO must be a subject, it must be the case that datives cannot have this function in Dutch.

192

Chapter 6

(31) a. De boekeni leverden Jan een hoop werk op [zonder PROi hem the books caused Jan a lot work prt without him te bevallen]. to please ‘The books caused a lot of work for Jan without pleasing him.’ b. *Jani gaat vaak naar de bibliotheek [zonder PROi de boeken ooit Jan goes often to the library without the books ever te bevallen]. to please ‘Jan often goes to the library, but the books never please him.’ The order in (29a) and (30a) is not very surprising: it involves raising to subject of the underlying direct object, a process familiar from a number of other languages. Such raising to subject could only be assumed to have taken place in (29b) and (30b), however, if it were possible for the dative argument to subsequently scramble across the subject position. As shown by (28c), however, such scrambling is not possible in Dutch. Hence, we must conclude that the nominative in (29b) and (30b) is still in the underlying object position and agrees from there, in line with the suggestion made above. There are two alternative analyses that avoid the conclusion that there is a nominative DP in object position that agrees with the verb. First, one could argue that the subject position is occupied by a null expletive (see Hoekstra 1984). If so, agreement could be with this expletive. However, this is only possible if some kind of mechanism is introduced that transfers the features of the nominative object to the expletive. At the same time, these features should not be phonologically realized on the expletive pronoun; if they were, the form inserted would presumably be that of a personal pronoun. In the absence of evidence for such a mechanism, an analysis along these lines must remain stipulative, quite apart from the stipulative nature of null expletives themselves (see note 9). The second alternative would be to assume covert movement of the nominative object to subject position (see, for instance, Neeleman and Weerman 1999). The main problem for this analysis is the lack of evidence for such movement. For example, covert A-movement should permit the moving argument to take scope over the argument it moves across. It can be shown, however, that such scopal effects are absent. Double object constructions exhibit a well-known freezing effect: the scope between indirect and direct object is fixed (see Bruening 2001a and Williams 2005). This holds for Dutch as well. Crucially, this same freezing effect is found under nominative-dative inversion. Thus, there is a sharp contrast between (32a) and (32b) in that only the latter allows a wide scope construal of the universal. This would be

Syntactic Restrictions on Agreement 193

unexpected if in (32a) the universal covertly raises to subject position. (For a comparable argument against covert raising in English there-constructions, see Den Dikken 1995.) (32) a. Ik geloof dat een meisje ieder boek gegeven werd. I believe that a girl every book given was ‘I believe that a girl was given every book.’ (∃ > ∀, *∀ > ∃) b. Ik geloof dat een meisje ieder boek gelezen heeft. I believe that a girl every book read has ‘I believe that a girl has read every book.’ (∃ > ∀, ∀ > ∃) Now that we have seen that Dutch allows both upward agreement initiated by a finite verb and downward agreement initiated by a nominative DP, the question presents itself whether a single verb can simultaneously be involved in upward and downward agreement. We will show that this is indeed possible. 6.4.2  Simultaneous Upward and Downward Agreement

We have argued above that agreement relations can be initiated by either the target or the controller. The strongest kind of evidence for this view comes from cases in which there is simultaneous upward and downward agreement. In this subsection, we will discuss two such cases, involving Icelandic nominative objects and Dutch clefts. 6.4.2.1  Default Person versus Default Number

Our starting point is an apparent paradox regarding the defaults in the person and number systems. On the one hand, there is evidence that in the person system the default is third person, while in the number system the default is singular. For example, expletive pronouns and verbs that fail to agree (as in impersonal passives) show up in the third person singular, as the following English and Dutch examples illustrate: (33) a. It seems that a solution is hard to find. b. *I/You/They seem(s) that a solution is hard to find. (34) Nog jaren is/*ben/*bent/*zijn naar een oplossing gezocht. still years be.3sg/be.1sg/be.2sg/be.pl for a solution searched ‘People searched for a solution for many years.’ On the other hand, singular agreement can be “overwritten” by plural agreement in certain contexts, but in those same contexts third person behaves differently. Thus, in (35) the expected singular agreement with the subject pronoun is replaced by plural agreement if the clefted constituent is plural, but

194

Chapter 6

not by first person or second person agreement if the clefted constituent is a first person or second person pronoun. (35) a. Het zijn wij die de whisky gestolen hebben. (pl overwrites sg) it are.pl we who the whisky stolen have ‘It’s us who stole the whisky.’ b. *Het ben ik die de whisky gestolen (1st does not overwrite 3rd) it am I who the whisky stolen heeft. has ‘It’s me who stole the whisky.’ c. *Het ben(t) jij die de whisky (2nd does not overwrite 3rd) it are.sg you.sg who the whisky gestolen heeft. stolen has ‘It’s you who stole the whisky.’ d. Het is hij die de whisky gestolen heeft. (no overwriting) it is he who the whisky stolen has ‘It’s him who stole the whisky.’ In chapter 2, we argued that defaults correspond to feature bundles that do not force an interpretation and that only if a φ-feature bundle may denote an empty set can it fail to be interpreted. In the person system, [dist] is the only feature structure that can deliver an empty set. This is because dist selects the outer layer of Si+u+o, discarding the only obligatory members of this set, speaker and addressee. As any o is optional, it is possible to start with an input set without o’s, so that dist will deliver an empty set. All other person specifications select either i or u or both and can therefore not be dummies. This is true even if both prox and dist are absent, as we argued in our discussion of impersonal pronouns in chapter 4. In the number system, only absence of number features is compatible with empty set reference. This is because [aug] and [aug min], the feature specifications used in chapter 3, impose a positive cardinality on the output of the person system. Absence of number features does not, and is therefore compatible with a cardinality of 0. In sum, there is a fundamental asymmetry between the defaults in the person and number systems. The former default has a feature specification, while the latter does not. This asymmetry of course forms a good basis for an explanation for contrasts like the one observed in (35). As we will now argue, the most straightforward explanation relies on double agreement in syntax, with one of the agreement relations being upward and the other downward.

Syntactic Restrictions on Agreement 195

Consider first what Nevins (2011) calls omnivorous number systems. In such systems (illustrated here for Eastern Abruzzese; examples from D’Alessandro and Roberts 2010:45), a participle shows plural agreement when either subject or object is plural. (36) a. Giuwanne Giuwanne b. Giuwanne Giuwanne c. Giuwanne Giuwanne d. Giuwanne Giuwanne

a pittate nu mure. has painted.sg a wall e Mmarije a pittite and Mmarije have painted.pl a pittit ddu mure. has painted.pl two walls e Mmarije a pittite and Mmarije have painted.pl

nu mure. a wall

ddu mure. two walls

Like Nevins, we assume that data like (36a–d) involve multiple agreement (with the proviso that we analyze agreement with the object as restrictor agreement; see subsection 6.3.2). Multiple agreement implies that there are two distinct feature bundles present in the agreeing participle. What makes the situation in Abruzzese interesting is that these bundles must be realized in a single morphophonological slot. (37) a. DP1 … V-φ1-φ2 … DP2 b. V-φ1-φ2 ⇔ /V/-/affix/ In general, where one form realizes two feature bundles, either unification is necessary or arbitration by rules of resolution. We postpone discussion of the latter type of rule to subsection 6.4.3, focusing here on unification. Unification itself is not a unitary phenomenon. It can involve either unification of sets of syntactic feature structures or unification of phonological forms. (The latter is independently motivated by observations in other domains, such as case realization in free relatives in German; see Groos and Van Riemsdijk 1981, among others.) We first consider the effects of syntactic unification. The instances of syntactic unification relevant to the data in (36) are given in (38). These can all be realized without difficulty, as a singular form in (38a) and a plural form in (38b–d). (38) a. b. c. d.

V-[ ]1-[ ]2 ⟶ V-[ ]1/2 V-[aug]1-[ ]2 ⟶ V-[aug]1/2 V-[ ]1-[aug]2 ⟶ V-[aug]1/2 V-[aug]1-[aug]2 ⟶ V-[aug]1/2

Given that third person is different from singular in that it does have feature content, syntactic unification in parallel cases involving person can result in feature bundles with multiple person specifications.

196

Chapter 6

(39) a. V-[dist]1-[dist]2 ⟶ V-[dist]1/2 b. V-[dist]1-[prox (…)]2 ⟶ V-[dist prox (…)]1/2 While realization of the output in (39a) is unproblematic, the feature specification in (39b) makes spell-out impossible, on the assumption that the process is blocked if a single agreement slot contains multiple feature bundles for the same class of φ-features.10 This means that where the input contains conflicting person specifications, spell-out cannot proceed on the basis of syntactic unification. Instead, phonological unification is necessary. Hence, the derivation converges only if the spell-out rules for [dist] and [prox (…)] deliver the same phonological form. (40) a. {dist} ⇔ /a/ b. {prox (…)} ⇔ /a/ c. V-[dist]1-[prox (…)] 2 ⇔ /V/-/a/ We will now discuss two examples of double-agreement structures that contain only a single morphological agreement slot, with the effect that a low DP must have the same person specification as imposed on the verb by the subject.11 One is the Dutch cleft construction already introduced in (35). The other involves the well-known case of nominative objects in Icelandic. 6.4.2.2  Agreement with Nominative Objects in Icelandic

Agreement with nominative objects occurs when the subject carries quirky case.12 However, such agreement is never found with first or second person objects.13 (Examples are from Sigurðsson and Holmberg 2008:254.) (41) a. *Honum líkum við. him.dat like.1pl we.nom ‘He likes us.’ b. *Honum líkið þið. him.dat like.2pl you.pl.nom ‘He likes you all.’ c. Honum líka þeir. him.dat like.3pl they.nom ‘He likes them.’ We follow a strand in the literature according to which the verb agrees with both the quirky subject and the nominative object (see Burzio 2000, Schütze 2003, and Ussery 2013). Thus, Icelandic agreement is regulated by two rules: (i) agree with the subject; (ii) agree with nominatives. This can be modeled by saying (i) that finite verbs in Icelandic obligatorily introduce a selectional requirement that establishes an agreement relation with the subject and (ii)

Syntactic Restrictions on Agreement 197

that nominative DPs optionally introduce a selectional requirement that establishes an agreement relation with the finite verb. It follows from Maximal Encoding that all nominatives must be part of an agreement relation where possible. Therefore, if a nominative DP is not a subject, the second rule must be applied. The result is the structure in (42). (42) 

IP [CTR†] KP

I′ [CTR] [TRG†] I [CTR]

VP [TRG] V

DP [TRG]

Recall that nonnominative DPs have a Case shell, and that this Case shell blocks φ-feature spreading from the verb. Therefore, a quirky subject will trigger default third person singular, just like other categories that lack φ-features (for example, clauses). This is confirmed by examples with a quirky subject in which the object is not nominative (example from Schütze 2003:295). (43) Mig hefur/*hef/*hafa vantað mýts. me.acc have.3sg/have.1sg/have.3pl lacked mice.acc ‘I have lacked mice.’ Structures like those in (41), which have a quirky subject and a nominative object, can then be represented as follows: (44) KP1 … V-[dist]1-φ2 … DP2 The verb carries the default specification [dist] as a consequence of agreement with the quirky subject; it also carries φ2, the feature specification of the nominative object, because this nominative itself initiates an agreement relation. Whether or not (44) can be realized depends on the content of φ2. Consider the various possibilities listed in (45). (45) a. KP1 … V-[dist]1-[dist]2 … DP2 b. KP1 … V-[dist]1-[dist aug]2 … DP2 c. KP1 … V-[dist]1-[prox (…)]2 … DP2 Syntactic unification of feature bundles applied to these structures yields the following: (46) a. KP1 … V-[dist]1,2 … DP2 b. KP1 … V-[dist aug]1,2 … DP2 c. KP1 … V-[dist prox (…)]1/2 … DP2

198

Chapter 6

The feature bundles in (46a) and (46b) are unproblematic as far as spell-out is concerned. The feature bundle in (46c) is not, however, as it contains contradictory values for person. Hence, spell-out must proceed on the basis of the nonunified structure in (45c). If so, the condition that there be a single affix can only be satisfied through phonological unification, which in turn is possible only if the phonological realization of [dist]1 is identical to the phonological realization of [prox (…)]2. Indeed, Sigurðsson (1996) observes that the person restriction on object agreement is lifted (for many speakers) when the first/second person form of the verb is syncretic with the third person form.14,15 (Examples are from Sigurðsson 1996:33.) (47) a. bored.at-3sg ⇔ /leiddist/ b. *Henni leiddumst við. her.dat bored.at.1pl we.nom c. *Henni leiddust þið. her.dat bored.at.2pl you.pl.nom d. ?Henni leiddist ég. her.dat bored.at.1sg I.nom e. ?Henni leiddist þú. her.dat bored.at.2sg you.sg.nom Agreement with lower nominative DPs occurs not only in monoclausal but also in biclausal structures with a raising verb. In the latter structures, the same person restriction is observed as in the former. (Examples are from Sigurðsson and Holmberg 2008:255.) (48) a. *Honum mundum virðast við (vera) hæfir. him.dat would.1pl seem we.nom (be) competent b. *Honum munduð virðast þið (vera) hæfir. him.dat would.2pl seem you.pl.nom (be) competent c. Honum mundu virðast þeir (vera) hæfir. him.dat would.3pl seem they.nom (be) competent ‘They would seem to him to be competent.’ Interestingly, many speakers allow suspension of agreement with the nominative in the biclausal construction. Crucially, the person restriction is lifted when this happens (see (49), from Sigurðsson and Holmberg 2008:255). This is expected: if there is only agreement with the quirky subject, there cannot be conflicting feature bundles in the verb. (49) a. Honum mundi virðast við (vera) hæfir. him.dat would.3sg seem we.nom (be) competent

Syntactic Restrictions on Agreement 199

b. Honum him.dat c. Honum him.dat

mundi would.3sg mundi would.3sg

virðast þið (vera) hæfir. seem you.pl.nom (be) competent virðast þeir (vera) hæfir. seem they.nom (be) competent

Sigurðsson and Holmberg (2008) observe that there is considerable variation in whether suspension of agreement is allowed, preferred, or required. In one variant (Icelandic C), agreement with low nominatives is dispreferred in general, even in monoclausal constructions. We predict that in that variant there should not be a person restriction on nominative objects at all. This appears to be in line with Sigurðsson and Holmberg’s assessment of the relevant data. 6.4.2.3  Agreement in Dutch Clefts

Dutch clefts show almost the same pattern of core observations as Icelandic quirky subject constructions (see also Den Dikken 2014). They have the following properties. (i) Number agreement with a clefted nominative is obligatory (see (50)). (ii) If there is unambiguous person agreement, first and second person nominatives cannot be clefted (see (51)). (iii) Some speakers allow suspension of person agreement with clefted nominatives. In that case, there is no person restriction (hence the % sign on the variants with third singular is in (51a,b)).16 (iv) Where the verb forms triggered by the pronoun in subject position (het ‘it’) and by the clefted nominative DP are identical, the person restriction is lifted for all speakers. This is the case with some modal verbs and in the past tense (see (52)). (50) Het zijn/*is zij die de whisky gestolen hebben. it are.pl/is they that the whisky stolen have ‘It’s them who stole the whisky.’ (51) a. Het %is/*ben ik die de whisky gestolen heeft. it is/am I that the whisky stolen has ‘It’s me who stole the whisky.’ b. Het %is/*ben(t) jij die de whisky gestolen heeft. it is/are.sg you.sg that the whisky stolen has ‘It’s you who stole the whisky.’ c. Het is hij die de whisky gestolen heeft. it is he that the whisky stolen has ‘It’s him who stole the whisky.’ (52) a. Het zal ik/jij wel geweest zijn die de whisky gestolen heeft. it will I/you.sg indeed been be who the whisky stolen has ‘It is likely that it was me/you who stole the whisky.’

200

Chapter 6

b. Het was ik/jij die de whisky gestolen heeft. it was I/you.sg who the whisky stolen has ‘It was I/you who stole the whisky.’ These data allow an analysis similar to that proposed for Icelandic. Dutch requires agreement with the subject (if there is one; see above), as well as agreement with nominatives. If the clefted constituent is a nominative DP, this yields the representation in (53). (53) 

IP [CTR†] het

I′ [CTR] [TRG†] I [CTR]

VP [TRG] V

… [TRG] DP [TRG]

CP

The subject in (53) is an expletive pronoun specified as [dist], leading to the following situation, where only one morphophonological slot is available for [dist]1 and φ2: (54) het1 … V-[dist]1-φ2 … DP2 [CP (OP2) … t2 …] This structure can be realized without problems if the syntactic unification of [dist]1 and φ2 delivers a feature bundle that does not contain multiple person specifications (that is, when φ2 is [dist (aug)]). Where syntactic unification does not lead to such a feature bundle, the derivation may converge under phonological unification (that is, when /[dist]1/ = /φ2/). If neither type of unification allows spell-out, the derivation crashes. This accounts for the person restriction observed in (51). Some speakers allow agreement with the clefted nominative to be suspended under these circumstances (through what we assume to be postsyntactic deletion of φ2). For those speakers, first and second person singular clefted nominatives may show up with a third person singular copula. (55) het1 … V-[dist]1 … DP2 [CP (OP2) … t2 …] There is an interesting twist in the plural. Here, all speakers require number agreement, but there are no effects of the person restriction. (56) Het zijn/*is wij/jullie die de whisky gestolen hebben. it are.pl/is we/you.pl that the whisky stolen have ‘It’s we/you who stole the whisky.’

Syntactic Restrictions on Agreement 201

These data have no parallel in Icelandic quirky subject constructions and cannot be accounted for through phonological unification, since the third person singular form of the copula is is and the first/second person plural form is zijn. However, in contrast to Icelandic, Dutch shows full neutralization of person distinctions in the plural, as illustrated for the copula in (57). This fact can be accounted for in terms of two rules of impoverishment that delete person features in the context of [aug], as in (58) (see chapter 7 for more discussion). (57) a. Ik ben even weg. I am momentarily away ‘I am out at the moment.’ b. Jij bent even weg. you are momentarily away c. Hij is even weg. he is momentarily away d. Wij/Jullie/Zij zijn even weg. we/you.pl/they are momentarily away (58) a. prox → ∅ / ___ [aug] b. dist → ∅ / ___ [aug] If the rules in (58) apply to the output of syntactic unification of the two feature bundles on the verb, they will remove the conflicting person specifications, leaving only [aug], and therefore the structure will be realized with the plural form of the copula. We give the derivation for a case with a clefted first person plural pronoun in (59). (59) a. het1 … V-[dist]1-[prox aug]2 … DP2 [CP (OP2) … t2 …] (syntactic output) b. het1 … V-[dist prox aug]1/2 … DP2 [CP (OP2) … t2 …] (output of unification) c. het1 … V-[aug]1/2 … DP2 [CP (OP2) … t2 …] (output of impoverishment) Interestingly, no person restriction is observed when the pronoun used as subject is not the weak pronoun het ‘it’ but the strong pronoun dat ‘that’. This is illustrated with a first person pronoun in (60) (but the same holds for second person pronouns). (60) Dat ben ik die op deze oude foto staat. that am I that on this old picture stands ‘That’s me in this old picture.’

202

Chapter 6

This is not a problem for our analysis. There is evidence that suggests that in examples like (60), the personal pronoun ik ‘I’ is the subject, while dat is a topicalized predicate (topicalization is accompanied by subject-verb inversion in Dutch as a consequence of the verb-second requirement). Note first of all that agreement must be with ik rather than dat. (61) *Dat is ik die op deze oude foto staat. that is I that on this old picture stands Moreover, in main clauses in which the initial position is occupied by an adverbial, ik must precede dat, as expected if ik is the subject. (62) a. Volgens mij ben ik dat die op deze according.to me am I that that on this oude foto staat. old picture stands ‘I think that’s me in this old picture.’ b. *Volgens mij ben dat ik die op deze oude according.to me am that I that on this old foto staat. picture stands If so, agreement in (60) is simply with the nominative subject only, as expected. At this point, we should note that the same construction is possible with the weak pronoun het as predicate as well, as shown in (63). (63) Volgens mij ben ik het die op deze oude foto staat. according.to me am I it that on this old picture stands ‘I think it’s me in this old picture.’ One may therefore wonder whether sentences of the type in (51a,b) with ben ‘am/are’ should not be grammatical after all, on an alternative parse with het as fronted predicate, in analogy with (60). However, we can be sure that het cannot be a topicalized predicate in (51). This is because het is a weak pronoun (see Haeseryn et al. 1997) and, like other weak pronouns, it strongly resists topicalization. We can demonstrate this using an object pronoun. (64) a. Ik heb dat/het gezien. I have that/it seen ‘I have seen that/it.’ b. Dat/*Het heb ik gezien. that/it have I seen Hence, in (51) het must be the subject, and the analysis proceeds as explained.

Syntactic Restrictions on Agreement 203

Contrary to what we observed for Dutch, there is full agreement between the copula and the clefted constituent in Icelandic clefts. In contrast to sentences with a quirky subject and a nominative object, there is no evidence for a person clash between the pronoun það and the nominative clefted constituent. This is illustrated in (65).17 (65) Í gær varst það þú sem tókst bókina. yesterday was.2sg it you that took.2sg book.def ‘Yesterday it was you who took the book.’ A first attempt at an analysis might be to equate the Icelandic clefts to the Dutch construction with the strong pronoun dat as predicate, as in (60)–(62). After all, the pronoun used in the Icelandic cleft has the same form as the distal demonstrative, just like Dutch dat, but in contrast to Dutch het. This analysis is compromised, however, by the fact that það not only can, but must, precede the nominative personal pronoun. (66) *Í gær varst þú það sem tókst bókina. yesterday was.2sg you it that took.2sg book.def So it would seem that það really does occupy the subject position. We must conclude, then, that it is possible to delete one of the φ-feature bundles in the verb before spell-out. This is similar to what some Dutch speakers can do, as discussed around (55). The difference is that in Dutch the agreement with the nominative predicate is suppressed, while it is the agreement with the subject that is suppressed in Icelandic.18 Of course, deletion of φ-feature bundles should be regulated, as otherwise any clash in person features in cases of simultaneous agreement can be avoided. We propose that deletion is only possible if the two φ-feature bundles in the verb are coindexed (a restriction suggestive of a local form of recoverability). This will be the case in clefts and specificational copula constructions, as these structures involve predication. The predication relation implies coindexation of subject and predicate (Williams 1994, Neeleman and Weerman 1999). By transitivity, given that each φ-feature bundle is coindexed with its controller, this leads to coindexation of both φ-feature bundles, which will allow deletion of one of them. The main consequence of this condition is that deletion is not a way in which clashes in person agreement can be avoided in quirky case constructions. In copular constructions, deletion is possible in principle, but not every grammar includes deletion rules. Even within a single language, there can be interspeaker variation in this respect, since as we have noted not all speakers of Dutch can suspend agreement with a clefted constituent. This arbitrariness is not surprising, as in general it is the case that the inventory of

204

Chapter 6

impoverishment rules is subject to crosslinguistic and intralinguistic variation. (We will see more examples of this in the next chapter.) 6.4.2.4  Alternatives Based on Unidirectional Agreement

In the above discussion, we have developed an analysis of person clashes in terms of two simultaneous agreement relations, one of which is upward and one of which is downward. The issue at this point is whether it is possible to adopt the hypothesis that these data are accounted for by simultaneous agreement, along the lines suggested above, while maintaining the notion that there is a uniform direction of agreement. We do not think that this is the case, for reasons we will now outline. Our starting point is the well-known observation that antecedents in grammatical dependencies are unique (see, for instance, Koster 1987 for discussion). Thus, a single predicate cannot take multiple subjects, a single anaphor cannot be bound by multiple antecedents, a single trace cannot be linked to multiple moved categories, and so on. Within the framework adopted here, this generalization can be expressed by the following condition:19 (67)  A single node cannot host more than one instance of the same selectional requirement. This rules out the situation schematized in (68), where the node labeled δ may be simplex or complex, with the two instances of sr copied from different daughters. (68)  *

… [SR†]

α1

… [SR] [SR†] α2

δ [SR] [SR] …



As an example, consider a structure in which an anaphor is bound by two antecedents. Such a structure is ruled out: in an example like (69), the anaphor can be bound by the subject or by the indirect object, but not by both simultaneously. Thus, the example does not have a reading where the girls are the agent of an event in which both the girls and the boys are shown to the boys. (69)  [The girls]i showed [the boys]j [each other]i/j/*i+j in the mirror. The nonexistent reading would rely on each other introducing two selectional requirements that encode anaphoric binding, one satisfied by the indirect object and the other by the subject. But (67) rules out a situation in which a

Syntactic Restrictions on Agreement 205

single node contains two selectional requirements encoding the same anaphoric binding relation. The condition in (67) also forces “chain composition” in case a single antecedent is associated with two dependents. Downward-branching grammatical dependencies are attested in structures containing a secondary predicate, a parasitic gap, or multiple anaphors bound by the same antecedent. In such cases, multiple terminal nodes introduce an instance of the same selectional requirement. Following various steps of upward copying, these selectional requirements will end up in the same node. This would violate (67) unless the selectional requirements in question are identified, resulting in a single selectional requirement that has multiple sources. This selectional requirement is subsequently satisfied by a unique antecedent, as in (70) (where δ1 and δ2 may or may not be syntactically complex, with sr being copied from one of their daughters). (70) 

… [SR†] α

… [SR] …

… [SR] δ1 [SR]

δ2 [SR]

This account solves a number of problems in the theory of syntactic dependencies. To give one example, it has been something of a mystery that a single predicate cannot assign more than one θ-role to the same argument while it is possible for a single argument to be assigned multiple θ-roles that originate in distinct predicates, as in (71), where John is the subject argument of both left and drunk. (71)  John left drunk. We can simply maintain the classical θ-Criterion: if the θ-roles of the primary and the secondary predicate are identified, John is assigned a single θ-role. Consider now what consequences (67) has for agreement. First, simultaneous upward agreement with two DPs is ruled out, as it would be an instance of an upward-branching grammatical dependency, which (67) rules out. Indeed, at an empirical level, it is relatively straightforward to show that an analysis of the data in this subsection that involves multiple upward agreement is untenable. This is because person clashes in Icelandic show up in structures in which the nominative object is contained in an infinitival complement to the agreeing verb, as in (48). Hence, at least one of the agreement relations appears to be downward. The only way in which long-distance agreement like

206

Chapter 6

this can be analyzed as upward is by assuming that the controller covertly moves into the matrix clause, to a position above the target. There is no evidence for such movement in Icelandic. (See below for a fuller discussion of long-distance agreement, leading to the conclusion that it is downward in general.) This leaves us with the alternative option that both agreement relations are downward. In our terms, this means that both DPs in the constructions above would introduce a selectional requirement for agreement. The theory of syntactic dependencies would allow this, but only if the selectional requirements introduced by the two DPs are identified, for reasons just discussed. The resulting single selectional requirement can only be satisfied by a unique φ-feature bundle in the verb, given the general uniqueness of syntactic antecedents. The implication is that multiple downward agreement is only possible if all DPs carry the same φ-feature bundle. This is simply not the case in the constructions discussed above. Limiting ourselves to Dutch clefts for the moment, we have discussed examples in which a DP with any type of person or number is clefted, while the pronoun in subject position invariably remains third person singular het. For example, in the plural there is no person restriction on the clefted DP; see (56). Such difficulties do not occur in a theory that assumes the coexistence of upward and downward agreement. This is because the selectional requirements involved are not the same. One selectional requirement is introduced by the verb and looks for φ-features on a DP, while the other is introduced by a nominative DP and looks for φ-features on a verbal node. Such selectional requirements may appear within the same node without violating (67).20 (72) 

IP [CTR†] DP

I′ [CTR] [TRG†] I [CTR]

VP [TRG] V

DP [TRG]

We conclude that agreement indeed does not have a uniform direction. This does raise the question how any observed syntactic asymmetries in specific agreement relations can be accounted for. We think that any such asymmetries result from locality conditions. We will work out the details of this hypothesis in sections 6.5 and 6.6. First, however, we discuss a type of language that differs from Icelandic and Dutch in the way it deals with person clashes.

Syntactic Restrictions on Agreement 207

6.4.3  Omnivorous Person Agreement

The argument for multiple agreement with the same verb developed above relies on the idea that all persons, including the default third person, have a featural specification. This predicts that person clashes are common when a single head agrees with multiple arguments. In other words, omnivorous person effects are predicted to be absent. As it turns out, however, there are languages that show omnivorous person effects. The main difference between such languages and languages like Dutch and Icelandic is that the former allow resolution of a potential clash on the basis of a person hierarchy: the feature structure highest on the hierarchy is realized, while the feature structure lower on the hierarchy is not. A good example is the agreement system in Ojibwe, which is sensitive to a person hierarchy 2 > 1 > 3 (see Valentine 2001 and Lochbihler 2012, among others). The agreement morphology on the Ojibwe verb reflects features of both its subject and its object. That there must be simultaneous subject and object agreement is clearest when considering the so-called theme sign on the verb. This is a suffix that expresses the relative position of subject and object on the person hierarchy. In particular, when the subject is higher on this hierarchy than the object, a “direct” theme sign appears, while an “inverse” form appears when the object is higher on the hierarchy. The form of the theme sign is also determined by whether or not both arguments are local persons (first or second) or only one of them is. Thus, the distribution of theme signs in (73) obtains (adapted from Lochbihler 2008). (73) Ojibwe theme signs

Both subject and object are 1 or 2 Either subject or object is 3

Subject outranks object on 2 > 1 > 3

Object outranks subject on 2 > 1 > 3

-i

-in(i)

-aa

-igw (and allomorphs)

This simultaneous sensitivity to the features of subject and object can only be accounted for under the assumption that both agree with the verb. Only if the features of both arguments are represented in the verb is it possible to have a spell-out system for verbal agreement based on a comparison of their positions on the person hierarchy. In the theme-sign suffixes, then, resolution of person clashes is achieved by spell-out rules that insert a single morpheme as the realization of pairs of feature bundles.

208

Chapter 6

In addition to the theme-sign suffix, the Ojibwe verb carries a prefix that expresses person agreement. Interestingly, this prefix shows omnivorous person effects: it expresses agreement with the argument that is highest on the person hierarchy, regardless of whether this is the subject or the object (grealizes second person, n- first person, w-/∅- third person). Given the discussion above, we know that the person features of both subject and object are represented in the verb. Hence, the behavior of the Ojibwe prefix shows that resolution of a person clash can also consist of nonrealization of the feature structure lower on the person hierarchy. The following examples illustrate the system (from Valentine 2001, cited here from Lochbihler 2008:296–297): (74) a. n-waabm-aa 1-see-dir ‘I see him.’ b. n-waabm-ig 1-see-inv ‘He sees me.’ (75) a. g-waabam-i 2-see-dir(local) ‘You see me.’ b. g-waabm-in 2-see-inv(local) ‘I see you.’ Not all languages that allow resolution of person clashes on the basis of a hierarchy make use of the same hierarchy. There is one crosslinguistic constant, though: third person is outranked by both first and second. The variation lies in the ranking of first and second person, as follows: (76) a. 2 > 1 > 3 b. 1 > 2 > 3 c. 1,2 > 3

(example: Ojibwe; see above) (example: Nocte; see below) (example: Kaqchikel; see below)

We suggest that this crosslinguistic variation comes about through variation in the weight attached to the two conditions in (77). For the purpose of (77b), a feature structure is less uniform if it contains more distinct features. (77) a. prox outranks dist. b. Less uniform feature structures outrank more uniform feature structures. A constraint equivalent to (77a) is present in some form or other in most any theory of person hierarchies, sometimes expressed directly and sometimes expressed in the order of functional projections or in the order of probing

Syntactic Restrictions on Agreement 209

of features (see below). The constraint in (77b) may look unfamiliar, but it is an instantiation of the general idea that feature structures containing more features are marked compared to feature bundles containing fewer. The only innovation is that markedness is assumed not to increase with repetition of the same feature, as in the first person exclusive (characterized by [prox–prox]). If in a given grammar the first condition in (77) is more important than the second, the resulting hierarchy will be 1 > 2 > 3. This is because first person is maximally marked according to (77a), as it only contains instances of prox. By contrast, third person is maximally unmarked, as it contains only dist. Second person is in between, as it contains both prox and dist. If (77b) is more important, second person will be highest in the hierarchy, as this is the only person with a nonuniform feature structure. The relative ranking of first and third person is still determined by the first condition, so that the result is the hierarchy 2 > 1 > 3. Finally, if the two conditions carry equal weight, a hierarchy results in which first and second person are ranked equally, and in which both outrank third person. Nocte is an example of a language that is like Ojibwe, but with first and second person reversed on the hierarchy (that is, it has a 1 > 2 > 3 hierarchy). The following data illustrate this (examples from Das Gupta 1971, cited here from Croft 2003:172; see also DeLancey 1981): (78) a. Nga-ma ate hetho-ang. 1sg-erg 3sg teach-1 ‘I will teach him.’ b. Ate-ma nga-nang hetho-h-ang. 3sg-erg 1sg-acc teach-inv-1 ‘He will teach me.’ c. Nang-ma nga hetho-h-ang. 2sg-erg 1sg teach-inv-1 ‘You will teach me.’ d. Nga-ma nang hetho-e. 1sg-erg 2sg teach-1pl ‘I will teach you.’ As in Ojibwe, an inverse marker appears on the verb in case the object is higher on the person hierarchy than the subject, the only difference being that, since the hierarchy is 1 > 2 > 3 in Nocte, the inverse marker is used when the subject is second person and the object first person. As before, the presence of this kind of morphology can only be understood if there is double agreement, so that the features of both subject and object are represented in the verb. Again as in Ojibwe, there is a second morpheme, in this case a suffix, that agrees in person

210

Chapter 6

with that argument whose feature specification is highest on the hierarchy (the omnivorous person effect). There is an interesting twist when the subject is first person and the object second person, as in (78d). As expected, the person agreement shown by the relevant suffix is with first person. However, the number expressed is an unexpected inclusive plural, rather than the singular. We will not attempt to analyze this observation, but it is another indication that the agreement morphology reflects agreement with both subject and object. The final possibility permitted by (77) is a person hierarchy in which first and second person are equally ranked. This should result in a language that allows resolution of clashes between third person and either first or second person, but not resolution of clashes between first and second person. An example of such a language is Kaqchikel, as discussed in Preminger 2014 (all Kaqchikel data below are taken from this source). In ordinary transitive clauses, the verb agrees with both subject and object, and this configuration of multiple agreement is reflected in two distinct agreement morphemes. (79) a. rat ri achin. x-∅-aw-ax-aj you.sg com-3sg.abs-2sg.erg-hear-act the man ‘You heard the man.’ b. ri achin x-a-r-ax-aj rat. the man com-2sg.abs-3sg.erg-hear-act you.sg ‘The man heard you.’ The interesting twist is that there is a construction in Kaqchikel, known as the agent focus construction, in which the number of agreement slots on the verb is reduced to one. This, of course, creates a situation in which person clashes arise. When one of the arguments of the verb is third person and the other is not, the clash is resolved in favor of the non–third person argument. This is illustrated in (80) for a combination of first person and third person arguments, and in (81) for a combination of second person and third person arguments. (80) a. ja yïn x-in/*∅-ax-an ri achin. foc me com-1sg/3sg.abs-hear-af the man ‘It was me that heard the man.’ b. ja ri achin x-in/*∅-ax-an yïn. foc the man com-1sg/3sg.abs-hear-af me ‘It was the man that heard me.’ (81) a. ja rat ri achin. x-at/*∅-ax-an foc you.sg com-2sg/3sg.abs-hear-af the man ‘It was you that heard the man.’

Syntactic Restrictions on Agreement 211

b. ja ri achin x-at/*∅-ax-an rat. foc the man com-2sg/3sg.abs-hear-af you.sg ‘It was the man that heard you.’ The conclusion must be that, according to the Kaqchikel person hierarchy, first and second person both outrank third person. That first and second person are not ranked with respect to each other is shown by the fact that in the agent focus construction no resolution is possible when both arguments are local. As in Icelandic and elsewhere, unresolved clashes result in ungrammaticality. Thus, the following are impossible, regardless of whether agreement on the verb is first person, second person, or (default) third person: (82) a. *ja at yïn. x-in/at/∅-ax-an foc you.sg com-1sg/2sg/3sg.abs-hear-af me ‘It was you that heard me.’ b. *ja yïn x-in/at/∅-ax-an rat. foc me com-1sg/2sg/3sg.abs-hear-af you.sg ‘It was me that heard you.’ We can refine the analysis just sketched by asking why it should be that in the agent focus construction one of the agreement slots on the verb no longer seems to be available. Our suggestion is that agent focus itself triggers agreement on the verb. That is, there are always two agreement slots, but when the agent focus morpheme (af) appears, it occupies one of them.21 If so, competition for realization does not involve two agreement relations and one slot; rather, it involves three agreement relations and two slots. Given that this competition is regulated by a hierarchy, this hierarchy must also mention af. As we will now show, the data can be accounted for if af is located between the local persons and third person. (83)  Kaqchikel agreement hierarchy 1,2 > af > 3 The procedure for realization of agreement is as follows. First, the two most highly ranked feature bundles are morphologically realized. Second, if the remaining feature bundle can be unified with one of the others, this will happen. Finally, if this is not possible, the remaining feature bundle is left unrealized. Consider what happens in structures in which the subject has the properties that trigger agent focus. If one of the subject or object is a local person and the other is not, then the local person and af will win the competition for agreement realization. The third person argument will not show agreement, as unification is impossible (see also note 22). This is correct, as discussed above.

212

Chapter 6

If both subject and object are third person, then af and one of the third persons will claim an agreement slot. The remaining third person will unify with the first. This, by definition, has no effects on the person morphology, but it should lead to omnivorous number morphology. This is correct: two third person arguments result in third person agreement; if one or both of them are plural, plural agreement results (Preminger 2014:20).22 Finally, if both the subject and the object are local persons, they both claim an agreement slot, leaving no room for af agreement.23 The result is a structure in which the conditions that normally trigger agent focus are present, but in which no af marker shows up, while agreement is with subject and object as in “ordinary” transitive sentences. This is exactly what happens; see Preminger 2014 and Erlewine 2016. Erlewine (2016:461) gives the following example: (84) ja yïn x-at-in-tzët rat. foc me com-2sg.abs-1sg.erg-see you ‘It was me that saw you.’ A prediction of this system is that even if the two local persons are both first or both second person, af should be barred from showing up. This is because in the first step of the procedure that determines the realization of agreement, the two local persons will win the competition for the available slots. Since they are each assigned a slot, these persons will not be unified. Therefore, no room is left for af. It is probably not possible to test this prediction for the first person, since two first person arguments will be interpreted as coreferential, giving rise to reflexivity. However, in recent fieldwork Preminger has found that in examples with two noncoreferential second persons, af is indeed not possible.24 (85) *ja rat x-at-ax-an rat. foc you.sg com-2sg.abs-hear-af you.sg ‘It was you(singular) that heard you(singular).’ Preminger (2014) argues that it is undesirable to appeal to person hierarchies to deal with the Kaqchikel data. He proposes a syntactic account that he argues to be motivated independently, and that derives the effects of the person hierarchy. The account is based on a probe-goal system of syntactic agreement regulated by relativized minimality. In the Kaqchikel agent focus construction, there is one functional head that acts as a probe for person features. This head specifically probes for a participant feature. Given relativized minimality, the highest DP that has a participant feature will act as the goal. However, Preminger assumes, following Béjar and Rezac (2003), that all first or second person features in DPs must be licensed by entering an agreement relation.25

Syntactic Restrictions on Agreement 213

(86)  Person Licensing Condition (Béjar and Rezac 2003:53) An interpretable 1st/2nd person feature must be licensed by entering into an Agree relation with a functional category. The consequence of this is that the lower DP in the agent focus construction cannot be licensed if it, too, is first or second person. In contrast, if the subject is third person, it is skipped in the probe’s search for a participant feature, and agreement will be with the first or second person object. Whether or not an account that appeals to a person hierarchy is more stipulative than this syntactic account can only be evaluated properly when crosslinguistic variation in the effects of person hierarchies is considered. As discussed above, it is not always the case that a clash between first and second person results in ungrammaticality. In some languages, these clashes are resolved, sometimes in favor of first person and sometimes in favor of second person. It seems to us that the only way in which the syntactic account just outlined can deal with such variation is by specifying the features that the probe is searching for. However, the language variation implies that it is not sufficient to specify a fixed feature content for the probe language by language. Probes must be allowed to search for different features, and in addition the features searched for must be ordered such that agreement with some is preferred over agreement with others. Consider a language with a 2 > 1 > 3 hierarchy, for instance. Given that second person defeats first person in a clash, the verbal head must probe for a feature that is unique to second person, say, addressee. Otherwise, it should not be able to skip a first person argument in its search. However, if the probe is specified as addressee also in a context where there is a clash between a first person and a third person argument, the situation would be unresolvable. In order to explain why the third person is ignored in favor of the first person argument, the feature content of the probe must be different. In particular, the probe must now search for a feature that distinguishes first and third person, that is, either speaker or participant. But in the 1 versus 2 situation, the probe cannot be permitted to search for either of these features. The implication is that there is a hierarchy that determines which features are preferably selected as the specification of the probe. Clearly, this is equivalent to a 2 > 1 > 3 person hierarchy. Given the attested language variation, it must be the case that this hierarchy of preferred feature content can vary from language to language. We conclude that there is no difference between the syntactic account and the morphological account proposed here with regard to the necessity of language-particular feature hierarchies.26 The main objection to the syntactic alternative, however, is that it fails to account for those situations in which third person DPs are involved in person

214

Chapter 6

clashes. As noted in the previous sections, the agreement data from Icelandic quirky subject constructions and Dutch clefts can be understood as the result of just such a clash. If the person clash in the Kaqchikel agent focus construction is the result of the Person Licensing Condition in (86), third persons should never lead to a similar problem, contrary to fact. Thus, a unified account of all the data discussed in this section is not possible under a syntactic account based on the constellation of assumptions outlined above. 6.5  The Domain of Feature Spreading 6.5.1 Introduction

Despite our claim that syntactic agreement is a symmetric relation initiated by either a DP argument or a verbal head, there is no denying that specific instances of agreement have a fixed direction. We propose that such asymmetries are an effect of two types of locality condition, the domain type and the relativized minimality type. In this section, we will propose a domain condition for agreement and explore its effects. Minimality effects will be discussed in the next section. 6.5.2  CP as a Domain for Feature Spreading

We propose that CP is a domain for the operation of feature spreading between agreeing elements. This could be stated as an irreducible grammatical condition, analogous to the idea that CP is a bounding node (Chomsky 1981, 1986) or a phase (Chomsky 2001 and subsequent work). However, we will explore the alternative that feature spreading is blocked by an operator attached high in CP. The two options could be teased apart in principle if there are CPs that do not contain the relevant operator, or domains smaller than CP that do contain the relevant operator. In practice, however, the data required to carry out this exercise are hard to come by, and therefore we will simply pursue the second line of analysis here. Recall from chapter 2 that, in order to account for shifts between clauses in the reference of deictic elements, we assume that at least every finite clause contains an operator adjoined to CP that encodes the value of i and u, as well as time and place. Shifts occur if new values for any of these attributes are assigned in an embedded CP. If this analysis is on the right track, establishing an agreement relation across a deictic operator is problematic, at least for person, for the following reason. The system of deictic operators implies that person features must receive an interpretation within the c-command domain of their local deictic operator. If this requirement did not exist, it would be possible to violate the

Syntactic Restrictions on Agreement 215

shift-together condition of Anand and Nevins (2004); see chapter 2 for discussion. For example, two first person pronouns locally c-commanded by the same deictic operator could receive a different reference if one of them could link to a higher operator. In general, then, all person features within the local domain of a deictic operator are bound by that operator. In chapter 5, we analyzed agreement as a relation that triggers feature spreading. This implies that a potential case of person agreement across a deictic operator would lead to a situation in which two instances of the same set of person features are bound by different deictic operators. This is impossible, since the very fact that the feature bundles are bound by different operators implies that they must be interpreted separately. Hence, person agreement across a deictic operator leads to a semantic anomaly. We propose that UG contains a general formal condition that bans this situation. It is formulated in (87). (87)  *[ … φi … [OPdeictic … φi …]] The coindexation of φ-features in this constraint is meant to express that they are instances of the same feature set in different locations. Of course, the constraint does not rule out instances in which two elements that happen to have the same φ-feature specification, without syntactic agreement holding between them, are separated by the deictic operator.27 Notice that the condition does not mention person specifically, but rather mentions φ-features in general, even though the motivation we gave for it is based on complications that arise with nonlocal spreading of person features. The reason for this, we suggest, is that the grammar does not have a specific operation of “person feature spreading,” as opposed to “number feature spreading” or “gender feature spreading.” Rather, there is a general operation of φ-feature spreading triggered by agreement. (Of course, agreement can be just for person or just for number, depending on the φ-feature specification of the elements involved in the agreement relation. This leaves unaffected the fact that the feature-spreading process itself is a general operation of the form “Spread α.”) Since (87) is a condition that ensures proper interpretation of φ-features, it must be an LF condition. We assume that, within the LF branch of grammar, (87) operates as an “anywhere” constraint. That is, after each step in the derivation after overt syntax the condition is checked, and any derivation in which a single feature set is associated with two syntactic positions that are separated by the deictic operator is ruled out. This condition has different consequences for the processes of movement and agreement. We assume that when a movement chain is formed, the

216

Chapter 6

syntactic properties of the trace must be determined. In particular, the φ-features inherent in the moved category may or may not be duplicated in the trace. In other words, both (88a) and (88b) are possible outputs of chain formation. (88)  a. DP



tDP

ϕ b. DP



tDP

ϕ At LF, an operation of chain minimization may take place, which selectively deletes features under identity. Such deletion is presumably necessary to achieve interpretability, under the assumption that features must be interpreted in a single position. Chain minimization cannot operate in (88a), but may reduce (88b) to either (88a) or (89). (89)  DP



tDP ϕ

Movement of a DP across the deictic operator will lead to a problem with the condition in (87) when the φ-features of the moving category must remain associated with both the antecedent and its trace at LF. This problem is avoided if at every point of evaluation of (87) the φ-features are associated only with the head or only with the foot of the chain. We now consider the consequences of the condition in (87) for agreement. Agreement relations involve two elements that each have a set of φ-features, which get identified at LF. In chapter 5, we modeled this using a process of feature spreading that results in a single φ-feature set associated with two positions. Subsequent to this, the features are dissociated from the position in which they cannot be interpreted, and identified with features in the controller. If the element that introduces the selectional requirement for agreement is located below the deictic operator while the potential antecedent is located above it, the condition in (87) requires that the lower of the two elements undergo covert movement to the edge of its clause. Otherwise, feature spreading is blocked by the operator. (There is no alternative way to get the two elements in the same local domain: lowering the higher element into the local domain of the lower element is impossible under the standard assumption that downward movement is impossible.) Whether the relevant LF raising operation can actually take place depends on the element that initiates the agreement

Syntactic Restrictions on Agreement 217

dependency. We will show that asymmetries arise depending on whether the lower element is a verbal head or a DP. In the following two subsections, we will explore these consequences of (87) in two domains: complementizer agreement and long-distance agreement. 6.5.3  Complementizer Agreement

Like other instances of agreement, complementizer agreement can, in principle, be upward or downward. Put differently, the selectional requirement that encodes the agreement relation can be introduced either by C or by a DP. What should not be possible, though, is agreement between C and a DP in case they are separated by a deictic operator. We will now argue that, crosslinguistically, the direction of complementizer agreement may indeed vary, but that there are no instances that are nonlocal in the relevant sense. In this discussion, it should be kept in mind that with complementizer agreement we refer to agreement that is expressed on an element in the C position. This can, of course, be a complementizer, but need not be. It is well-known that a number of languages have a process that moves verbs to C in some contexts. Any agreement relation in a language that involves C will hence extend to such moved verbs. For example, fronted verbs in certain East Netherlandic dialects can show the same agreement morphology as complementizers, morphology that is formally distinct from the regular verbal agreement morphology (see Van Haeringen 1939, Zwart 1997, and Ackema and Neeleman 2004 for discussion). 6.5.3.1  Agreement Initiated by DP

Consider a language that has a nominative-accusative Case system. As we have argued, such languages require that in downward agreement the relevant selectional requirement be introduced by a nominative DP. In upward agreement, if a nominative DP satisfies the relevant selectional requirement, the features encoded in the verb will match those in the controller. An agreement relation between a nonnominative DP and a finite verb can also be established, but the features on the verb will remain uninterpreted, and hence must be the default third person singular. The result is a strong pressure for complementizer agreement to be downward. Consider why. Suppose that we try to establish an upward relation of complementizer agreement—so, with the complementizer itself as the element that introduces the relevant selectional requirement. The deictic operator is adjoined to CP and hence would separate the controller and the target in case the controller is anywhere in the matrix clause. Therefore, in order to prevent a violation of (87), the complementizer would have to move to a position above

218

Chapter 6

the operator at LF. But no such position is available: as the operator is adjoined to CP, there is no higher head position within the verbal extended projection. There is also no possibility for C to move out of CP, since head movement of a functional head into the lexical head of a higher extended projection is impossible (see Baker 1988, 2003:306). There is one XP position above C but below the deictic operator, which therefore is still local enough for C to establish an upward agreement relation with: namely, the specifier of CP. However, Spec,CP is only filled if some category undergoes Ā-movement, and since Ā-movement does not single out nominative DPs, there is no guarantee that there is a controller that can license nondefault agreement on C. Moreover, the elements that move to Spec,CP are wh-expressions and other operator-like elements. Given their nonpronominal nature, these lack person (see chapter 2), which means that even in those circumstances that would allow for agreement, it must be for features other than person. None of these limitations hold when agreement is downward, as the nominative subject DP can initiate an agreement relation whenever C is present, for any φ-features that it contains. As a result, in a nominative-accusative language that shows regular complementizer agreement for person/number, this must be downward agreement, as schematized in (90). The agreement relation is initiated by a selectional restriction on DPy, which percolates upward until it reaches C′, where it immediately dominates the complementizer that will act as antecedent. (90) [DPx … [V V agrx] … [CP OPi,u [CP [C C agry] DPy … ]]] Within the Germanic languages, there are several varieties that have complementizer agreement. (One such variety is West Flemish, as illustrated in (24).) Given that these languages are of the nominative-accusative type, we expect this agreement to be downward, which is indeed the case. There are no examples in these languages of a complementizer agreeing with an argument in the matrix clause. Agreement is always with the local subject below C (see Zwart 1997, Ackema and Neeleman 2004, and Haegeman and Van Koppen 2012, among others). A standard analysis of complementizer agreement since the introduction of Chomsky’s (2000, 2001) probe-goal system is that it is, indeed, downward agreement—in this system, all agreement is downward. However, previously the standard view was that agreement is uniformly upward, being established in a specifier-head configuration. Accordingly, several analyses were proposed in which agreement between complementizer and subject is established in a specifier-head configuration, with the relevant head subsequently moving upward to C past the subject. Of course, our proposal is only supported if

Syntactic Restrictions on Agreement 219

complementizer agreement is really agreement with the complementizer, rather than agreement with a lower head that subsequently undergoes fusion with the complementizer. We argued in Ackema and Neeleman 2004:237ff. that analyses of the latter type face serious difficulties—for example, in accounting for the fact that a complementizer cannot agree when intervening material separates it from the subject position below it. We will not repeat this argumentation here; we merely emphasize that the data suggest that complementizer agreement is genuinely downward.28 6.5.3.2  Agreement Initiated by C

The limitations we mentioned on upward complementizer agreement will not hold when all DPs in a language are Caseless. This is relevant to the analysis of complementizer agreement in some Bantu languages, since at least some of these languages have been argued to lack Case (for discussion, see Diercks 2012 and Van der Wal and Sheehan to appear). In Caseless languages, we expect that C can show systematic agreement with elements in Spec,CP, as the absence of a Case shell makes the features of any element in this specifier accessible. This is in line with Kinyalolo’s (1991) analysis of agreement with wh-phrases in Kilega. In Kilega, wh-phrases can either remain in situ or move to the left periphery of the clause. In the former case, the verb agrees (for noun class) with the subject; in the latter case, it agrees with the fronted phrase. This is illustrated in (91) (from Kinyalolo 1991:21). (91) a. Bábo bíkulu b-á-kás-íl-é mwámí bikí mu-mwílo? 2.that 2.woman 2-a-give-perf-fv 1.chief 8.what 18-3.village ‘What did those women give the chief in the village?’ b. Bikí bi-á-kás-íl-é bábo bíkulu mwámí mu-mwílo? 8.what 8-a-give-perf-fv 2.that 2.woman 1.chief 18-3.village ‘What did those women give the chief in the village?’ Similar observations hold when phrases other than wh-phrases are fronted, as for example with fronted locative phrases or fronted objects: the verb agrees with these fronted elements, rather than with the subject. Our analysis of these data is that the verb in Kilega initiates agreement in the highest position it occupies. In simple declarative clauses, the verb moves to T and therefore agrees with the subject in Spec,TP. When a phrase is fronted to Spec,CP, this is accompanied by verb movement to C. Since the highest position of the verb is now the C position, the verb agrees with the category in Spec,CP.29 Contrary to the view that there is upward agreement from C in Kilega, Carstens (2005) argues that C agrees downward with the category to be fronted, and that it is this agreement relation that triggers the movement to

220

Chapter 6

Spec,CP. Carstens’s evidence for downward agreement comes from data that appear to indicate that the subject stays lower than Spec,TP when the verb agrees with a fronted constituent. This constitutes evidence for downward agreement, Carstens suggests, because a subject in Spec,TP would act as an intervener between the downward-probing head and the base position of the wh-operator. One piece of evidence that the subject is in a low position in the relevant contexts is the lack of subject agreement. However, the complementarity of subject agreement and agreement with fronted categories follows as well if the verb agrees from its highest derived position. Another piece of evidence comes from structures containing an auxiliary and a main verb both displaying agreement with a fronted category. In such structures, the subject follows not just the auxiliary but also the main verb (see (92), from Carstens 2005:238). If subject-verb inversion were the result of verb movement to C crossing a subject in Spec,TP, we would expect only the auxiliary to precede the subject. (92) Ku-Lúgushwá kú-kili ku-á-twag-a nzogu maswá. 17-Lúgushwá 17-be.still 17-a-stampede-fv 10.elephant 6.farm ‘At Lugushwa elephants are still stampeding over the farms.’ We agree that these data suggest that the subject is in a low position when another category is fronted. However, this is not necessarily the result of a potential intervention effect. An alternative account, compatible with our analysis, is that subject movement to Spec,TP is driven by the need for the verb to agree with a local c-commanding category. In the absence of wh-movement or other fronting operations, the subject moves to Spec,TP. However, if there is another element that moves independently to Spec,CP, verbs can agree with this element, voiding the need for subject raising. Note that in (92) not only the auxiliary in C but also the main verb agrees with the element in Spec,CP. This is allowed by the locality condition in (87). This condition merely requires that the deictic operator not separate controller and target, and therefore a verbal head can agree with any c-commanding category in its clause up to and including one in Spec,CP. (For more discussion on multiple agreement structures of this type, see section 6.6.) 6.5.3.3  Agreement Initiated by CP

In the previous subsections, we have argued that it is impossible for two elements to agree across the deictic operator in the left periphery of CP. The consequence was that if C is an agreement target, upward agreement is possible but must be very local: the controller cannot be situated higher than the deictic operator. Interestingly enough, a pattern of agreement that

Syntactic Restrictions on Agreement 221

appears to go against this generalization has been identified in Lubukusu; see Diercks 2013. Lubukusu has an extensive system of agreement for noun class. There are cases where such agreement is realized as a prefix on a complementizer while the controller is the matrix subject, as in (93) (Diercks’s (1)). In (93a), the embedded complementizer is preceded by a class 2 prefix because the matrix subject has this class; in (93b), it has a class 1 prefix in accordance with the class 1 status of the matrix subject. (93) a. Ba-ba-ndu ba-bol-el-a Alfredi ba-li a-kha-khil-e. 2-2-people 2-said-ap-fv 1Alfred 2-that 1-fut-conquer ‘The people told Alfred that he will win.’ b. Alfredi ka-bol-el-a ba-ba-ndu a-li ba-kha-khil-e. 1Alfred 1-said-ap-fv 2-2-people 1-that 2-fut-conquer ‘Alfred told the people that they will win.’ However, the mere fact that the agreement is morphologically realized on C does not necessarily imply that C is the target of the agreement relation. We hypothesize that, in fact, the agreement target is the CP as a whole. In general, nothing rules out that phrases as well as heads can be targets in agreement relations. If a head is the target of agreement (as in the case of Germanic complementizer agreement), it contains a position that hosts the features involved in the agreement relation, as in (94a). (We represent this position simply as agr, but this is just shorthand for whatever features are involved in the agreement relation.) It would require a stipulation to rule out that a phrase can contain a similar position, as in (94b). (94) a. [X X agr] b. [XP XP agr] We suggest that CP complements are a target for noun class agreement in Lubukusu. What this means is that a head that contains noun class features is attached to CP. The controller of the agreement is the matrix subject. Since the deictic operator is adjoined to CP, while the head that encodes the noun class of the CP takes the CP as its complement, agreement must be upward for exactly the same reason that agreement between C and a nominative subject in Germanic must be downward; compare (95) with (90). In both cases, the direction of agreement is forced by the requirement that agreeing elements should not be structurally separated by the deictic operator. (95) [DPx [V agrx V] [CP agrx [CP OPi,u [CP C … DPy … [V agry V] … ]]]] The operator in (95) creates a barrier for agreement by (87). Hence, the agrx position on the embedded CP cannot be associated with a potential controller

222

Chapter 6

below this operator. The agreement morpheme is morphologically realized as a prefix on the first element in CP—so, on C. But since spell-out is a matter of morphophonology, this does not undermine the above argumentation. The hypothesis that the deictic operator acts as a barrier for agreement thus explains why in Lubukusu CP-agreement is with the matrix subject rather than the embedded subject (see Diercks 2013:360–361 and references mentioned there). Diercks (2013:386) shows that agreement with a subject originating in the embedded clause is impossible even when this DP undergoes movement to the subject position in the matrix clause. Lubukusu has a construction known as hyperraising, in which the subject of a finite embedded clause undergoes raising. An example is given in (96a) (Diercks’s (75)). (96) a. Sammy a-lolekhana mbo a-likho a-lwala. 1Sammy 1-appears that 1-prog 1-be.sick ‘Sammy appears to be sick.’ b. *Sammy a-lolekhana a-li a-likho a-lwala. 1Sammy 1-appears 1-that 1-prog 1-be.sick ‘Sammy appears to be sick.’ Two things are remarkable about (96a). First, hyperraising appears to be a genuine instance of A-movement out of a finite clause. The moved DP must occupy the matrix subject position, given that it enters into an agreement relation with the matrix verb. Second, at least a number of speakers allow an overt complementizer in hyperraising. However, Diercks notes that these speakers crucially do not permit this complementizer to partake in agreement (see (96b)). This, in fact, follows from our analysis. The structure after raising in a sentence like (96b) is as shown in (97). (97) [DPx [V agrx V] [CP agrx [CP OPi,u [CP C … tx … [V agrx V] … ]]]] That there can be no direct agreement relation between the agr on CP and the trace will be clear by now (it would be agreement across the deictic operator). What might be more unexpected is that upward agreement between agr on CP and the derived subject is also impossible, while at the same time the matrix verb is perfectly happy to agree with this DP (see (96a)). One analysis that may seem plausible would be that a CP with an agr slot has more structure than a CP without one, and that extraction is blocked somehow by the extra structure. This, in essence, is the proposal advanced by Carstens and Diercks (2013). However, Diercks (2013) notes that this approach cannot account for the fact that raising a DP out of a CP that shows agreement is unobjectionable when the raised DP does not end up in subject position, and so does not act as the controller for CP-agreement.

Syntactic Restrictions on Agreement 223

(98) N-enya Barack Obama n-di a-khil-e. 1sg-want 1Barack Obama 1sg-that 1-win-sbj ‘I want Barack Obama to succeed.’ So, what must be crucial in any account of the infelicitous cases is that it is the agreeing DP that raises out of the agreeing CP. The core of our explanation for this fact is that the constraint in (87), which bans the cooccurrence of identical φ-features above and below a deictic operator, is relevant not only for agreement but also for movement, as suggested above. The consequence is that movement across a finite CP boundary potentially violates (87). As noted, however, such violations only occur if at some point in the LF derivation the φ-features of the moving category are associated with both the head and the foot of the movement chain. If the features are not present in the trace, or if they are dissociated from the antecedent, (87) is satisfied. In the case of Lubukusu hyperraising, the φ-features must be associated with both the antecedent and the trace when the movement chain is created (as in (88b)). This is because agreement must be licensed in both the matrix verb and the lower verb. Since φ-features in a verb are not interpretable, feature spreading from the verb to the controller DP and subsequent dissociation in the verb are required for both the matrix and the embedded verb. To prevent a violation of (87), feature spreading needs to be local, so the matrix verb must spread its features to the antecedent DP while the lower verb must spread its features to the trace, followed by identification with the features in the DP and the trace, respectively. This means that the first step in the LF derivation cannot be chain minimization, as deleting the features in either antecedent or trace would lead to a failure of feature identification and, consequently, the features originating in either the higher or the lower verb would remain uninterpreted. As a result, there is a point in the LF derivation at which the same set of φ-features is associated with both the head and the foot of the movement chain. Unavoidably, then, (87) is violated.30 The only way to rescue the structure is to remove the deictic operator by truncating the top part of the embedded clause. A well-known condition on truncation, going back at least to Rizzi 1994, 1995, is that it is not possible to delete functional material without deleting all dominating functional structure as well. This implies that truncation of the structure that hosts the deictic operator will also remove the agr position, given that this position is located above the deictic operator. As a result, hyperraising of an agreeing subject must lead to suspension of CP-agreement. In contrast to the case of raising to subject, if the raised DP does not land in a position that triggers agreement in the matrix clause, the first step in the

224

Chapter 6

LF derivation may consist of chain minimization, which here dissociates the φ-features from the head of the movement chain and leaves them associated with the trace only. The resulting configuration does not violate the condition in (87), even if the deictic operator is present, and hence truncation is not necessary. This accounts for the fact that CP-agreement is possible in a construction with raising to a nonsubject position, as in (98). Two questions remain. First, why is CP-agreement in Lubukusu obligatory rather than optional when there is no hyperraising? Put differently, why is truncation disallowed in case it is not necessary to save a hyperraising structure? The answer is simply that Maximal Encoding requires agreement where it is possible, as discussed in chapter 5. The second question is more general: why is truncation not universally available as a means of enabling agreement relations? For example, in the Germanic complementizer agreement varieties, why is truncation not used to enable an upward agreement relation with a matrix subject? Our suggestion is that truncation is a language-specific option that must be acquired on the basis of positive evidence. Such positive evidence comes in the form of overt material that is normally present in a particular extended projection, but absent under specific circumstances. In the Lubukusu case, the overt material in question is the agreement morphology on CP. However, it is uncommon for overt material to appear above the deictic operator but within CP, simply because the operator is attached very high. Moreover, since the deictic operator is a silent category, it cannot itself provide positive evidence for truncation. As a consequence, truncation of the structure containing the operator will be rare. An important difference between our analysis and the analyis proposed by Diercks (2013) has to do with the nature of the agreement relation between C(P) and the higher subject. In our proposal, there is a direct agreement relation between these two elements. Diercks considers this possibility as well, but argues instead for an indirect agreement relation, mediated by a null anaphor merged in an Ā-position in the embedded clause. C agrees with this null anaphor, which is itself bound by the matrix subject. Since binding requires matching features, the effect is that the matrix subject appears to agree with the embedded C. Diercks’s analysis is more complex than an analysis involving direct agreement, in that it has to invoke the null anaphoric element. The problem is not so much that the element is null; rather, it is that its function in the embedded clause is unclear. As Diercks (2013:373) notes, the anaphor in question is not assigned any θ-role, hence is nonargumental. The literature contains some examples of nonargumental anaphors (such as English home; see Jackendoff,

Syntactic Restrictions on Agreement 225

Maling, and Zaenen 1993), but we are not aware of anaphors that do not have any grammatical function at all (argument, operator, etc.) in their base position. It appears that the only role of the null element in question is to act as an intermediary for agreement. It would seem, therefore, that an analysis that can do without such an element is preferable. Diercks’s motivation for not adopting a direct agreement account is that there can be intervening nonsubject DPs (or, more precisely, DPs that are not themselves potential agreement controllers) that do not hinder the agreement relation between matrix subject and embedded C(P). For Diercks, this indicates that there is no direct agreement relation, as the intervening nonsubject DPs would induce what is known as a defective intervention effect (see Chomsky 2000, among others). While we do think that such intervention effects exist, there is a difference between the situation in Lubukusu and constructions for which defective intervention has been posited. In the latter, a head (the target) cannot agree with a lower DP (the controller) across an intervening DP. In Lubukusu, the lower element in the agreement relation is the target, rather than the DP. As noted in subsection 6.4.1, the selectional requirement introduced by a verb can specify that it must be satisfied by a subject. If so, objects will not act as interveners. Subjects will, though, which accounts for the fact that CP-agreement in Lubukusu can only be with the subject of the immediately dominating clause (see Diercks 2013:370). For more discussion of intervention effects, see section 6.6. The crucial difference between apparent long-distance upward complementizer agreement in Lubukusu and local upward complementizer agreement in Kilega is the location of the agreement slot. While in both cases the agreement morpheme is realized on an element in C, we assume that it is syntactically attached to CP in Lubukusu, but to C in Kilega (or, more precisely, to V in C). This assumption invites the question, how can we decide whether the position of the agreement morpheme is a sister to C or to CP? There are two considerations that can help in this regard. First, if the agreement morpheme is realized on a verb and shifts along with the verb when the latter moves to T or C, then the relevant morpheme must be attached to the verb, rather than to CP. This is arguably true of Kilega, as discussed above. Similarly, if the agreement morpheme is realized as a suffix on C (as with Germanic agreeing complementizers), it is unlikely (though perhaps not impossible) that its syntactic position is a sister to CP. This would require some form of prosodic inversion in the mapping from morphosyntax to morphophonology, in violation of mapping principles such as those proposed in Sproat 1985, Sadock 1991, and Ackema and Neeleman 2004. Hence, if the agreement morpheme is suffixal, we expect it to express a downward agreement relation or a local upward

226

Chapter 6

agreement relation with the element in Spec,CP. If the agreement morpheme is prefixal, however, its morphosyntactic position could be either on C or on CP. Therefore, for long-distance upward agreement, we predict prefixal morphology (as in Lubukusu). Interestingly, Lubukusu also has a form of local agreement between a complementizer and its specifier, which shows up when an element other than a subject is extracted. It is striking that in exactly this case the agreement morpheme is suffixal (example from Diercks 2009:55). (99) kama-tunda ni-ko ba-ba-andu ba-a-kula likoloba 6-fruit comp-6 2-2-person 2-pst-buy yesterday ‘the fruit that people bought yesterday’ It seems likely that this distinction between prefixal and suffixal agreement on complementizers corresponds with a distinction in syntactic position of the agreement morpheme, namely, as a sister to CP or as a sister to C. 6.5.3.4 Conclusion

In the previous subsections, we discussed three configurations that lead to agreement morphologically expressed on C: downward agreement between C and a nominative DP, upward agreement between C and a category in its specifier, and upward agreement between CP and a subject in the matrix clause. One may wonder whether this exhausts the possibilities that UG allows for C-related agreement. What should be impossible given the discussion above is upward agreement between C and nonlocal DPs, a correct prediction as far as we know. There is one possibility, however, that we have not discussed yet but that, as it stands, our proposal allows. This is a language with Caseless DPs in which C agrees downward with the first DP it c-commands. After all, in such a language any DP can in principle introduce a selectional requirement for agreement. Strikingly, in the languages with Caseless DPs that we have discussed in this subsection, agreement is systematically upward, as they instantiate either the second or the third pattern mentioned above. Of course, this might be just a coincidence; after all, complementizer agreement does not appear to be a very widespread phenomenon. It is possible, however, that there is a deeper connection between the absence of Case in a language and the absence of DP-initiated agreement. In a language without Case, the rule in (26b) would allow any DP to be the lower category in an agreement relation. It seems plausible that such nonselective agreement should be excluded, and that the rule in (26b) will only be part of languages with a Case system (either nominative-accusative or absolutive-ergative).

Syntactic Restrictions on Agreement 227

6.5.4  Long-Distance Agreement

We now turn to long-distance agreement, that is, agreement between a verb and a DP across a finite clausal boundary. Upward and downward agreement are not alike when it comes to long-distance configurations. In fact, it is implied in our account of agreement relations that upward agreement is strictly bounded by finite clauses and should not be possible across longer distances. Suppose a verb in an embedded clause establishes an agreement relation with the subject of the matrix clause. Modulo the possibility of truncation, which for reasons outlined in subsection 6.5.3 is rare, this means that feature spreading in the case of upward long-distance agreement will take place across the deictic operator if embedded verb and matrix subject remain in situ. In principle, this situation could be remedied by moving one of the elements involved across this operator. However, the matrix subject cannot move downward into the embedded clause, as downward movement is impossible in general. The verb cannot move upward to a position above the deictic operator either. Since the deictic operator is adjoined to CP, there is no clause-internal head position available above it that can host a moved verb. Moving a head out of an embedded finite clause into the matrix clause is also impossible (see Baker 1988 and Williams 2003 for discussion).31 In contrast, the system we have proposed does permit downward longdistance agreement. If an agreement relation is established between a DP in an embedded clause and a verb in the matrix clause, the condition in (87) is still potentially violated. However, in this configuration repair is possible. What is required is that the agreeing DP undergoes movement to a position higher than the deictic operator. This movement will bring the DP sufficiently close to the agreeing head to permit feature spreading without violation of (87). Note that in such cases the agreement relation is initiated by the DP, which introduces a selectional requirement that is copied upward and satisfied by the verb. This implies that, as in other cases of DP-initiated agreement, the DP must be Caseless, that is, be nominative/absolutive. Downward long-distance agreement has been observed in a variety of languages; see, for instance, work by Polinsky and Potsdam (2001) on Tsez, Bruening (2001b) on Passamaquoddy, and Branigan and MacKenzie (2002) on Innu-aimûn. The analyses that these authors propose indeed all involve movement of the DP controller to the left edge of the embedded clause. The examples in (100a) (Tsez; Polinsky and Potsdam 2001:584), (100b) (Passamaquoddy; Bruening 2001b:259), and (100c) (Innu-aimûn; Branigan and MacKenzie 2002:388) illustrate the phenomenon in question.

228

Chapter 6

(100) a. Enir [užā magalu bāc’rułi] b-iyxo. mother boy bread.III.abs ate III-know ‘The mother knows the boy ate the bread.’ b. N-wewitaham-a-k [ma=te nomiy-a-w-ik 1-remember-dir-3pl neg=emph see-dir-neg-part3pl mawsuwinuw-ok Kehlis-k]. person-3pl Calais-loc ‘I remember that I didn’t see people in Calais.’ c. Ni-tshissenitamu-ânân-at [mûpishtuât Shûshepa Tshân mâk 1pl-know-1pl-3pl visit Joseph John and Mânî]. Marie ‘We know that John and Marie visited Joseph.’ Note that in these examples the required movement of the controller to the left edge of its clause is covert, given that on the surface the agreeing DP occurs inside the embedded clause. Nonetheless, evidence for movement is available. One of the strongest arguments is that the controller cannot originate inside an island if the target is located outside that island. This is the case in all three languages discussed above. Bruening (2001b:265ff.) argues the point for Passamaquoddy. We repeat two of his examples below: (101a) shows that long-distance agreement is sensitive to the Complex NP Constraint; (101b) shows that it is sensitive to wh-islands. Similar examples are given for Innuaimûn by Branigan and MacKenzie (2002:393). Polinsky and Potsdam (2001:596ff.) argue that for Tsez, too, the relevant movement is detectable through its sensitivity to islands. (101) a. *N-kosiciy-a-k kis-ankuweht-uwon [DP atomupil-ol Piyel naka 1-know.ta-dir-3pl perf-sell-2conj car-inan.pl Piyel and Susehp mil-osk-opon-il]. Susehp give-2.conj.inv-pret-part.inan.pl ‘I know that you sold the cars that Piyel and Susehp gave you.’ b. *Ma=te n-wewitaham-a-wiy-ik Susehp neg=emph 1-remember.ta-dir-neg-3pl Susehp kisi-qecimul-osk [CP keq kisi-htu-htit skitapiy-ik]. perf-ask-2.conj.inv what perf-make-3pl.conj man-3pl ‘I don’t remember if Susehp asked you what the men made.’ The fact that long-distance agreement is subject to island effects of course suggests that the controller must move to make agreement possible. Given the assumptions we made in subsection 6.5.2, it is to be expected that the movement in question can be covert. In fact, it follows from our theory

Syntactic Restrictions on Agreement 229

that it must be covert, at least in the languages under discussion. In these languages, the DP that acts as controller in long-distance agreement also acts as controller for an agreement relation with the verb that heads the embedded clause. This may not be obvious from the glosses in all the examples just given, but judging by the data in the literature cited above it is a consistent property of Tsez, Innu-aimûn, and Passamaquoddy. Hence, the underlying structure of examples involving long-distance agreement can be schematized as follows: (102)  [ … [V V agrx] … [CP [OPi,u [C … DPx … [V V agrx] … ]]]] This implies that the DP that controls agreement on the higher verb must also be involved in φ-feature spreading and identification in the embedded clause. Suppose that the required movement of the DP across the deictic operator were overt, yielding the following structure: (103)  [ … [V V agrx] … [CP DPx [OPi,u [C … tx … [V V agrx] … ]]]] Overt movement leads to a violation of the condition in (87) for the same reason that the (untruncated) Lubukusu hyperraising construction did (compare (97)). The moved DP and its trace in (103) are each in an agreement relation with a verb. This means that when the movement chain is created, the φ-features of the moving category must be associated with both the trace and its antecedent. In addition, the first step in the LF derivation cannot be chain minimization. If it were, the features of at least one of the agreeing verbs would fail to be identified and hence could not be interpreted. But if the φ-features must remain associated with both antecedent and trace, the condition in (87) is violated: antecedent and trace are on opposite sides of the operator. In contrast, if the movement of the controller DP to the left edge of the embedded clause is covert, there will be no problem with the condition in (87). In order to demonstrate this, we will go through the relevant derivation. If the controller DP does not move overtly, the input to LF will be (102), repeated here for convenience: (104)  [ … [V V agrx] … [CP [OPi,u [C … DPx … [V V agrx] … ]]]] This structure does not violate (87), as the φ-features of the higher verb and those of DPx have not been identified yet. (Recall that establishing the agreement relation, which happens in overt syntax, is separate from the processes of feature spreading, identification, and dissociation.) The first steps in the LF derivation consist of spreading of the φ-features of the lower verb to DPx, identification of these features with the features contained in DPx, and dissociation of the φ-features from the verb, as desired. This results in the representation in (105).

230

Chapter 6

(105)  [ … [V V agrx] … [CP [OPi,u [C … DPx … V … ]]]] Next, DPx moves across the deictic operator to the left edge of the embedded CP. In the construction of the movement chain, no φ-features need to be generated in the trace. These are no longer required, as the spreading and subsequent dissociation of the features in the lower verb have already happened. Thus, a movement chain can be created in which only the antecedent contains φ-features, yielding the representation in (106). (106)  [ … [V V agrx] … [CP DPx [OPi,u [C … t … V … ]]]] Since the trace in (106) does not contain φ-features, the condition in (87) is still not violated. Finally, the features in the higher verb spread to DPx, are identified, and are dissociated from this verb. In this process, the operator again does not intervene, since the higher verb and the DP are now both above it. The final result is (107). (107)  [ … V … [CP DPx [OPi,u [C … t … V … ]]]] In short, the crucial difference between the overt and the covert movement cases is the following. In the overt movement case, φ-features must be associated with both the antecedent and the trace in the initial stage of the LF derivation, as the process of feature spreading needs the features to be present in both positions if both the higher and the lower verb agree. The result is that the condition in (87) is unavoidably violated. In the covert movement case, feature spreading from the lower verb is possible prior to movement, so that in the generation of the movement chain it is not necessary to preserve the φ-features in the trace. Hence, at no stage in the derivation are identical φ-features separated by the deictic operator. The hypothesis that the crucial movement of the controller in long-distance agreement must be covert does not exclude the possibility that this controller first undergoes some form of overt movement within the embedded clause. The only condition is that the overt movement not cross the deictic operator; if it does, we will end up with the problem described above. Any overt movement will therefore be followed by covert movement across this operator. The latter operation takes place after feature spreading in the embedded clause and enables feature spreading in the matrix clause. So a structure in which a longdistance-agreeing DP undergoes overt movement in the embedded clause must have the following surface and LF structures: (108) a. Surface [ … [V V agrx] … [CP [OPi,u [ … DPx … tx … [V V agrx] … ]]]] b. LF [ … [V V agrx] … [CP DPx [OPi,u [ … t … t … V … ]]]]

Syntactic Restrictions on Agreement 231

The postulation of short covert movement, as in (108b), is at odds with a common assumption in the literature, namely, that the overt movement that the controller may undergo is the counterpart of the covert movement necessary to license long-distance agreement elsewhere. Our proposal implies that, as opposed to the covert movement, the overt movement must have a landing site lower than the deictic operator. This predicts that the controller can never move overtly into the matrix clause, a prediction that does not follow directly from existing alternative proposals. There is evidence that it is correct, however. Bruening (2001b:275ff.) shows that, in Passamaquoddy, the island effects that reveal movement in the embedded clause (see above) disappear whenever the element controlling matrix agreement is unambiguously located in the main clause. Bruening argues that this means that in this kind of structure, the controller of the matrix agreement is in fact a prothetic DP coreferential with a pronoun that acts as controller of the agreement in the embedded clause. This pronoun is usually null, but can be spelled out in some contexts. We agree with this analysis, but for us the import of the data is that they confirm the predicted ban on overt movement of the controller out of the embedded clause. Our analysis of a Passamaquoddy sentence in which an overtly moved DP controls long-distance agreement must then postulate an additional landing site for covert movement above the deictic operator. This does not require any new stipulations about the structure of the Passamaquoddy clause, as it is already necessary to allow adjunction to CP. When overt movement takes place, the agreeing element can land in a position c-commanding wh-phrases and complementizers, as demonstrated by Bruening (2001b). Our suggestion is simply that overt movement tucks in the controller in a CP-adjoined position below the deictic operator and that covert movement shifts it to a CP-adjoined position above this element. (109) a. Surface [ … [V V agrx] … [CP OPi,u [CP DPx [CP (Wh) C … tx … [V V agrx] … ]]]] b. LF [ … [V V agrx] … [CP DPx [CP OPi,u [CP t [CP (Wh) C … t … V … ]]]]] At first sight, this analysis seems less economical than some of its competitors, since it relies on an extra step of covert movement not required in alternative accounts. However, there is reason to think that this extra step of movement is necessary anyway, when we take into consideration the broader typology of long-distance agreement.

232

Chapter 6

In the languages discussed so far, controllers of long-distance agreement must have a particular discourse function. There is some variation as to what function this is, but by and large the DP must be a topic or a focus, functions that, on many theories, are associated with Ā-positions. However, Fry and Hamilton (2015) show that, in addition to these “classical” long-distance agreement languages, there are two further types of language that allow agreement across finite clause boundaries. In one, the controller must be the subject of the embedded clause. In the other, it must be the DP that has the highest person specification according to a person hierarchy in which local persons (first and second) always outrank third persons (see subsection 6.4.3). The relevant DP can be either the subject, in a clause with direct morphology, or the object, in a clause with inverse morphology. To illustrate, we repeat some of Fry and Hamilton’s examples here. The type of language in which only the subject of the embedded clause can control long-distance agreement is represented by Plains Cree (see Dahlstrom 1991). In neither the direct nor the inverse form can the object act as controller. (110) a. Long-distance agreement with subject, embedded direct clause Ni-kiskeeyim-aaw [George ee-saakih-aa-t o-kosis-a]. 1-know.vta-dir.1>3 George conj-love.vta-dir-3 3.poss-son-obv ‘I know that George loves his sons.’ b. No long-distance agreement with object, embedded direct clause *Ni-kiskeeyim-im-aawa [George ee-saakih-aa-t 1-know.vta-obv-dir.1>4 George conj-love.vta-dir-3 o-kosis-a]. 3.poss-son-obv ‘I know that George loves his sons.’ c. Long-distance agreement with subject, embedded inverse clause Ni-kiskeeyim-im-aawa [George ee-saakih-iko-t 1-know.vta-obv-dir.1>4 George conj-love.vta-inv-3 o-kosis-a]. 3.poss-son-obv ‘I know that his sons love George.’ d. No long-distance agreement with object, embedded inverse clause *Ni-kiskeeyim-aaw [George ee-saakih-iko-t 1-know.vta-dir.1>3 George conj-love.vta-inv-3 o-kosis-a]. 3.poss-son-obv ‘I know that his sons love George.’ The type of language in which the DP with the highest person specification controls long-distance agreement is represented by Mi’gmaq. This is

Syntactic Restrictions on Agreement 233

illustrated by (111a–d) (see Fry and Hamilton 2015:6–7). Note that in the direct form, the subject has a higher person specification than the object, while in the inverse this is the other way around. (111) a. Long-distance agreement with subject, embedded direct clause Geji’-g [Mali ges-al-a-j-i Piel-al aq Je’g-al]. know.vta-3 Mary love-vta-3obj-3-3pl Peter-obv coor Jack-obv ‘I know that Mary loves Peter and Jack.’ b. No long-distance agreement with object, embedded direct clause *Geji’-g-ig [Mali ges-al-a-j-i Piel-al aq know.vta-3-3pl Mary love-vta-3obj-3-3pl Peter-obv coor Je’g-al]. Jack-obv ‘I know that Mary loves Peter and Jack.’ c. No long-distance agreement with subject, embedded inverse clause *Geji’-g-ig [Mali ges-al-gwi’-tit-l Piel-al aq know.vta-3-3pl Mary love-vta-inv-4-obv Peter-obv coor Je’g-al]. Jack-obv ‘I know that Peter and Jack love Mary.’ d. Long-distance agreement with object, embedded inverse clause Geji’-g [Mali ges-al-gwi’-tit-l Piel-al aq Je’g-al]. know.vta-3 Mary love-vta-inv-4-obv Peter-obv coor Jack-obv ‘I know that Peter and Jack love Mary.’ Fry and Hamilton’s analysis is based on the idea that C can contain features that attract the highest matching DP in its complement to the left edge of CP, where this DP can control agreement with the matrix verb. The language variation is accounted for by two factors. First, C can contain either φ-features or discourse-related features. Second, in some languages with inverse morphology the object in an inverse raises to a position above the subject, whereas in other such languages the inverse has no syntactic reflex. If C contains discourserelated features, it will attract either a topic or a focus, regardless of grammatical function or structural position. Other DPs do not intervene, as they do not have the relevant discourse features. If C contains φ-features, the highest DP in its complement is attracted. This will be the subject in languages in which there is no movement in the inverse. In languages in which the inverse features object-over-subject movement, it will be the subject in direct clauses and the object in inverse clauses. On the theory developed in this chapter, the simplest account of the data relies on a parameter that determines whether the selectional requirement for agreement can be introduced only by DPs with a particular discourse function

234

Chapter 6

(such as topic or focus) or by any nominative/absolutive DP. Languages like Tsez and Passamaquoddy are of the former type, while languages like Plains Cree and Mi’gmaq are of the latter type. The facts discussed by Fry and Hamilton can be captured if in the latter type of language only the highest DP in an A-position can initiate an agreement relation. This effect could be attributed to the Case system if only the highest DP in an A-position is Caseless (nominative), while all other DPs carry abstract Case. Since only Caseless DPs can initiate downward agreement (see subsection 6.4.1), it follows that only the highest DP can be involved in long-distance agreement. Regardless of the details of the analysis, however, what is crucial is that it is the syntax of A-positions in Plains Cree and Mi’gmaq that determines which DP can act as controller for long-distance agreement. It is generally accepted, however, that long-distance agreement requires Ā-movement of the controller to the left edge of CP (see the discussion above). This means that at least in these languages there must be an operation that takes the DP in the highest A-position and moves it to an Ā-position in the periphery of the embedded clause. This Ā-movement is not reducible to independently existing Ā-movements triggered by such features as topic or focus, precisely because the agreeing DP does not need to be a topic or a focus in these languages. This is exactly the kind of movement required under our analysis, leaving no argument from parsimony against it. 6.6  Intervention Effects

We now turn to a second type of locality condition on agreement, one reminiscent of relativized minimality. It is a condition on feature spreading, which (like (87)) operates at LF. (112)  In an agreement relation, feature spreading between the antecedent α and the dependent δ is not possible if there is an intervener β such that β c-commands δ and is c-commanded by α. What counts as an intervener depends on what α and δ are in a given agreement relation, since interveners must be of the same type as either α or δ. Upward agreement and downward agreement will therefore differ with respect to the type of element that counts as an intervener. In upward agreement, δ is a verbal head that introduces the relevant selectional requirement. Hence, verbal heads will count as interveners. Categories that can in principle satisfy this selectional requirement are like α, so these will be interveners as well. In the languages under discussion, α must be a subject. Thus, interveners for upward agreement can be defined as in (113).

Syntactic Restrictions on Agreement 235

(113) In upward agreement, β blocks feature spreading iff (a) or (b). a. β is a verbal head. b. β is a subject DP. In downward agreement, δ is a DP argument that introduces the relevant selectional requirement, whereas α is a finite verb. Hence, interveners for feature spreading in downward agreement relations are defined as in (114). (114) In downward agreement, β blocks feature spreading iff (a) or (b). a. β is a DP argument. b. β is a finite verb. Let us consider the consequences of (112)–(114), starting with upward agreement. In this type of agreement, feature spreading cannot take place across a verbal head. Hence, if agreement is upward and there is one agreeing head in a sequence of heads, then that head must be the highest one in the sequence. (115) a. DPi … [aux1-agri [aux2 [V … b. *DPi … [aux1 [aux2-agri [V … c. *DPi … [aux1 [aux2 [V-agri … This captures the pattern familiar from Dutch, English, and many other languages, exemplified here with Dutch. (116) a. dat Jan het boek heeft kunnen lezen that Jan the book have.3sg can.inf read.inf ‘that Jan has been able to read the book’ b. *dat Jan het boek hebben kan lezen that Jan the book have.inf can.3sg read.inf c. *dat Jan het boek hebben kunnen leest that Jan the book have.inf can.inf read.3sg Given that (112) is a condition on feature spreading, it can in principle be satisfied after movement (overt or covert). Note, however, that ungrammatical structures as in (115b,c) cannot be saved in this way. This would require movement of aux-2 or V across aux-1. Such movement violates the Head Movement Constraint.32 Note that an auxiliary will only intervene in an upward agreement relation between a main verb and a DP if the auxiliary itself is structurally sandwiched between the DP and the main verb. If the DP is lower than the auxiliary, the main verb can agree with the DP while the auxiliary does not show agreement. This fits the pattern Aissen (1994) describes for Tzotzil, as her analysis is based exactly on the assumption that the auxiliary c-commands a VP containing the

236

Chapter 6

main verb and its arguments. Such a structure is compatible with Tzotzil word order, which is Aux-V-O-S. Huallaga Quechua shows a similar pattern of agreement (see Baker 2008:239–240) and allows a similar analysis, as its word order is the mirror image of Tzotzil’s, S-O-V-Aux. Therefore, the auxiliary may again occupy a position c-commanding a VP containing V and all its arguments.33 Although it is not possible for a lower verbal head to agree across a nonagreeing higher verbal head, we expect to find cases where both heads agree with the same c-commanding DP controller. This is because grammatical dependencies in general can be downward-branching; that is, a single antecedent can be associated with more than one dependent (see subsection 6.4.2 for discussion). We therefore predict that, next to cases like (115a), it should be possible for a DP antecedent to be associated with an uninterrupted sequence of agreeing verbal heads. (117) a. DPi … [aux1-agri [aux2-agri [V … b. DPi … [aux1-agri [aux2-agri [V-agri … Let us spell out why we do not expect intervention effects in such structures. If all heads that c-command an agreeing head also agree (with the same DP), then they are part of the same downward-branching dependency. However, interveners by their very nature must not be part of the dependency in which they intervene. To make this explicit, we add (118) to (112). (118)  In a downward-branching dependency {α, δ1 … δn}, no dependent δi is an intervener in the relation between α and δj. There are indeed cases where a DP agrees with a sequence of verbal heads. An example from Palestinian Arabic is given in (119). (The example is given by Hoyt (2007:124), who characterizes ‘stood’ and ‘returned’ in (119) as auxiliated verbs.) (119) qaam naam. raǰaʕ stood.3sg.masc returned.3sg.masc slept.3sg.masc ‘Then he went to sleep again.’ We conclude that in upward agreement verbal heads cannot be skipped. The same is not true of DPs. There can be DPs that c-command a verb and are c-commanded by a DP controller without this leading to ungrammaticality. A clear example of this is Lubukusu long-distance agreement, which we analyzed as a direct agreement relation with the matrix subject, but which, as Diercks (2013) explicitly notes, can skip objects in the matrix clause. The possibility of upward agreement across an intervening object has its basis in the fact that selectional requirements can specify the type of antecedent they

Syntactic Restrictions on Agreement 237

are looking for. In particular, if only subjects satisfy the structural description of a selectional requirement encoding agreement, then no internal argument DP will count as an intervener (see (113)). As a result, such DPs can be skipped. In other words, we predict that in upward agreement there is no “defective intervention.” We do expect, however, “nondefective” intervention effects, in that elements of the same type as the controller cannot be skipped. Indeed, agreement in Lubukusu must be with the most local subject. It is not possible for a CP to agree with a subject across another subject. Let us now consider the consequences of the locality condition in (112) for downward agreement. In this case, interveners are either DP arguments or finite verbs (see (114)). Starting with the latter, we expect that feature spreading in downward agreement is blocked by any intervening finite verb, but not by nonfinite verbs. This seems to be a typical pattern, which can be observed, for example, in Icelandic. As already discussed, a finite verb can agree with a lower nominative DP in this language. This is possible in simple verb-object constructions, but also in more complex constructions where the DP is separated from the finite verb by one or more nonfinite heads. A relevant example is (120) (cited here from Boeckx 2009:6). (120) Mér virðast [hafa verið seldir margir hestar]. me.dat seem.3pl have.inf been sold many horses.nom ‘It seems to me that many horses have been sold.’ However, there are no examples in the literature of long-distance agreement of this type across a finite verb. DP arguments are the second type of intervener for feature spreading in downward agreement relations. Hence, any DP argument in the path that connects a target verb and a c-commanded controller DP should act as an intervener. This yields “defective intervention” effects (see Chomsky 2000, 2001). The prime example is again found in Icelandic, where in biclausal structures agreement between a matrix verb and a nominative in the embedded clause is blocked if the embedded clause contains a DP that c-commands the nominative. This is illustrated in (121) (from Sigurðsson 2000:99). Here, the quirky subject in the embedded clause blocks agreement between the nominative object and the matrix verb. (121) a. Okkur virtist/*virtust henni hafa leiðst us.dat seemed.3sg/seemed.3pl her.dat have found.boring þeir. they.nom ‘It seemed to us that she has found them boring.’

238

Chapter 6

b. Okkur sýndist/*sýndust honum hafa hentað us.dat appeared.3sg/appeared.3pl him.dat have suited pennarnir vel. pens.the.nom well ‘It appears to us that he liked the pens.’ Things are somewhat more complicated if the potentially intervening DP is located in the same clause as the agreeing verb. Sigurðsson and Holmberg (2008:258) report that there is speaker variation in whether or not agreement between finite verb and nominative DP is possible in the following construction: (122) Það líkuðu einum málfræðingi þessar hugmyndir. expl liked.3pl one linguist.dat these ideas.nom ‘A linguist liked these ideas.’ We give the example with agreement on the verb, as in the variant of Icelandic that Sigurðsson and Holmberg call Icelandic A. However, Sigurðsson and Holmberg note that in another variant, which they dub Icelandic B, agreement is suspended.34 This suspension of agreement is presumably an instance of the defective intervention effect. One may wonder, then, why there is variation between Icelandic A and Icelandic B with respect to intervention in (122), while the effect in (121) holds for all speakers. We suggest that this is because Icelandic A and B have different derivations for the string in (122). We begin by specifying what Icelandic A and B have in common. First, the relevant constructions are right-branching, so the DP closest to the verb c-commands the second DP. Second, the verb moves first to I and subsequently to C, in both main and embedded clauses. Finally, we assume that the finite verb agrees with low nominative DPs from I. Our hypothesis is that Icelandic A and B differ with respect to the position in which the expletive það is merged in the underlying structure in (123). (123) [CP __ C [IP __ I [vP V DPDAT DPNOM]]] In Icelandic A, the expletive is merged directly in Spec,CP. The EPP is satisfied by moving the dative to Spec,IP, as in (124). This movement has the side effect that the DP intervener between I and the nominative DP is removed. (It is generally assumed in the literature that the trace of a dative DP that undergoes A-movement does not act as an intervener, a hypothesis we adopt.)35 As a consequence, agreement between the nominative DP and V is possible. (124) [CP expl C [IP DPDAT I [vP V tDAT DPNOM]]]

Syntactic Restrictions on Agreement 239

In Icelandic B, the expletive is merged in Spec,IP and subsequently moves to Spec,CP. As a consequence, the dative DP cannot move to Spec,IP, and it therefore remains in its intervening position between I and the nominative DP. Hence, agreement is blocked. (125) [CP expli C [IP ti I [vP V DPDAT DPNOM]]] Note that the intervention effect in the biclausal structure in (121) cannot be circumvented in any variant, given that the dative NP in the embedded clause does not raise past the matrix verb. Interestingly, within Icelandic B, not every dative DP gives rise to a defective intervention effect (see Holmberg and Hróarsdóttir 2003 and Kučerová 2016). Kučerová demonstrates that there is a correlation such that those dative DPs that can undergo object shift do not give rise to intervention effects in long-distance agreement. Her account is that these DPs can move out from between the nominative controller and the trace of the inflected verb. This account chimes well with the above suggestion that the absence of intervention effects is due to the potential intervener moving to a position higher than the agreeing head. This concludes our discussion of locality conditions on feature spreading in agreement. 6.7 Conclusion

Let us summarize the main conclusions of this chapter. We started with the hypothesis (motivated in chapter 5) that agreement is syntactically symmetric. That is, the features in the target are not inherently dependent on the features in the controller (or vice versa, for that matter). Related to this hypothesis, there is no inherent structural asymmetry between target and controller: either one can appear in a c-commanding position, leading to the prediction that agreement can be both upward and downward. One observation confirming this prediction is that there are structures in which a single head entertains two agreement relations simultaneously, one upward and the other downward. The evidence that such structures exist comes from person clashes that arise when a head tries to realize two conflicting φ-feature bundles. This conflict is not syntactic; rather, it must be realizational, in view of the fact that morphological syncretism can resolve it. Therefore, an analysis of these structures that assumes simultaneous upward and downward agreement is attractive, since it is inherent in such an analysis that the two agreement relations do not interact syntactically: they occupy largely separate regions of the tree. By contrast, in theories that assume uniform upward or

240

Chapter 6

uniform downward agreement the two agreement relations take place in largely overlapping regions of the tree. This would likely give rise to syntactic intervention effects, which would block multiple agreement altogether in the structures at hand. Despite agreement in general being symmetric in nature, specific instances are either upward or downward. It is rare at best that the relative positions of target and controller can be swapped without repercussions. We have argued that such asymmetries follow from two types of locality condition that impose restrictions on the operation of feature spreading. First, there is a domain restriction on where feature spreading is possible. The relevant domain boundary is marked by the deictic operator present at the left edge of a CP. One consequence is that complementizer agreement of the type found in the Germanic languages, which involves person and has C as target, must be downward. By contrast, agreement that has an entire CP as target will be with features located in an agreement slot higher than the deictic operator and must therefore be upward. We have argued that an instance of agreement in Lubukusu is of this type. Another consequence of this locality condition is that cross-clausal agreement requires covert movement of the element introducing the relevant selectional requirement across the deictic operator. Such long-distance agreement must be downward, as DPs can move far enough to cross the deictic operator, but verbal heads cannot. The second type of locality condition on feature spreading is a minimality restriction. It has the following consequences. (i) Upward agreement must be with the most local subject, while nonsubjects do not intervene. (ii) Upward agreement must be with the highest verbal head in a functional sequence that is c-commanded by the controller, or with an uninterrupted sequence of verbal heads starting with this head. (iii) Downward agreement must be with the most local finite verb (at least in languages in which nonfinite verbs do not agree). (iv) Downward agreement can in principle be blocked by any intervening DP, giving rise to defective intervention effects.

7

The Morphological Realization of Person

7.1 Introduction

In chapter 1, we put together the following to-do list for this book: •  Explain why the crosslinguistic inventory of persons is the way it is. •  Explain the interactions between person and number. •  Explain the possible readings and forms of impersonal pronouns. •  Find out whether person agreement should be dealt with through a syntactically symmetric or asymmetric theory. •  Integrate the theory of syntactic agreement with the general theory of syntactic dependencies, and use the result to account for syntactic restrictions on agreement relations. • Account for crosslinguistic patterns of syncretism and language-internal variation in the realization of person with one and the same theory of person features and spell-out. The one remaining item on the list is the last one. It will be the topic of this chapter. We develop an account of person morphology that takes as its starting point the theory of person features proposed in chapter 2. This theory assumes two person features, prox and dist, one shared by first and second person and the other shared by second and third person. In this chapter, we will consider the rules that determine how person features are realized at PF; these include impoverishment and spell-out rules. We will argue that such rules are sensitive to the order in which the functions represented by the person features apply. The main results of the proposed analysis are (i) an explanation of the typological observation that syncretism between first and third person is much rarer than syncretism between either first and second or second and third (see Baerman, Brown, and Corbett 2005, Baerman

242

Chapter 7

and Brown 2011), and (ii) a descriptively adequate analysis of person agreement in Dutch, which is used here as a test case because its two person endings arrange themselves in such a way that there is a 2-3 syncretism in the regular case, a 1-2 syncretism under subject-verb inversion, and an optional 1-3 syncretism with a particular lexical class of verbs (modals). 7.2  Patterns of Syncretism

Any theory of the realization of person features faces two general tasks. On the one hand, it must account for known typological patterns in the morphology of person. On the other hand, it must provide a handle on the often quite intricate person morphology of individual languages. There is a tension between these tasks. The second will require positing a fairly complex system, but the theoretical leeway required for this should not undermine the account of the typology of person morphology. It should not lead to the expectation that any system is possible or as common as any other. The aim of this chapter is to develop a system of person realization that is typologically adequate and that nonetheless can capture the fine detail of person agreement in one language, Dutch. Dutch provides a good test case, because it displays several alternations in person agreement, leading to a situation in which the same inflectional endings arrange themselves into different patterns of syncretism in different contexts. We begin with a brief introduction of the data. In the singular, subject-verb agreement in Dutch shows a syncretism between second and third person (a 2-3 syncretism): second and third person singular subjects trigger a verbal -t ending, whereas first person is not marked overtly. In the plural, there is no person marking: the verb systematically ends in -en. We illustrate this using the regular verb lopen ‘to walk’ in (1) (recall that hon stands for honorific).1 (1)

lopen ‘walk’ 1sg 2sg 2sg hon 3sg 1/2/3pl

loop loopt loopt loopt lopen

Deviations from this general paradigm are conditioned by three factors: (i) subject-verb inversion, (ii) verb class, and (iii) politeness. Let us illustrate

The Morphological Realization of Person 243

each in turn. In (2), the verbal affix that encodes second person singular is omitted under inversion. This results in a syncretism between first and second person (a 1-2 syncretism), since the verb form that surfaces is identical to the first person form (while the third person remains unaffected by inversion; see (2c)). (2) a. Jij loop-t dagelijks met een hondje over straat. you walk-2sg daily with a doggy across treet ‘Every day you walk with a doggy in the street.’ b. Dagelijks loop jij met een hondje over straat. daily walk you with a doggy across street ‘Every day you walk with a doggy in the street.’ c. Dagelijks loop-t zij met een hondje over straat. daily walk-3sg she with a doggy across street ‘Every day she walks with a doggy in the street.’ As (3) indicates, the modal verb kunnen ‘can’ has two distinct stems. Both can be used in the second person singular, again giving rise to an agreement alternation because only one stem allows an overt second person agreement ending. The same holds, with some variation, of certain other modals. The result is that in the singular the relevant modals can be said to display an optional syncretism between first and third person (a 1-3 syncretism). (3) ‘can’ 1sg 2sg 2sg hon 3sg 1/2/3pl

kan forms

kun forms

kan kan kan kan *kannen

*kun kunt kunt *kunt kunnen

As (4) indicates, the verb hebben ‘have’ also has two stem allomorphs. One allomorph, heb, shows up in the first person singular, the informal second person singular, and all plural forms. The other, heef, appears in the third person singular. The polite second person singular verb form can be built on either, which suggests that it can alternate between second and third person singular agreement. In other words, although there is the regular 2-3 syncretism in the verbal endings, this verb distinguishes first and second person marking from third person marking through its stem form (a 1-2 syncretism).

244

(4) ‘have’ 1sg 2sg 2sg hon 3sg 1/2/3pl

Chapter 7

heb forms

heef forms

heb hebt hebt *hebt hebben

*heef *heeft heeft heeft *heven

These alternations interact in several ways. To give one example, it is striking that the alternation illustrated in (2a,b) is not found when the polite second person pronoun is used as subject. (5) Dagelijks loop-t/*loop u met een hondje over straat. daily walk-2sg/walk you.hon with a doggy across street ‘Every day you (polite) walk with a doggy in the street.’ It will be clear that the alternations in Dutch between the basic 2-3 syncretism, the 1-2 syncretism found under inversion, and the 1-3 syncretism found with certain modals form an ideal testing ground for a theory of person. However, as noted, such a theory should also be compatible with what is known about the typology of person syncretism in general. One generalization established in the literature is given in (6). (6)  1-2 and 2-3 syncretisms are far more common than 1-3 syncretisms. Evidence for this generalization comes from work by Baerman and colleagues (see Baerman, Brown, and Corbett 2005:59 and Baerman and Brown 2011). Admittedly, where person agreement is concerned, what Baerman and his colleagues considered in determining the frequency of syncretisms are whole verb forms, and not morphemes expressing person, which is what we are interested in here. The reason for this is that morphological decomposition is often a highly contentious issue and therefore impractical in largescale typological research. However, the null hypothesis must be that what is true of whole verb forms is true of inflectional endings, and this is what we will therefore assume throughout this chapter. It would be very surprising if 1-3 syncretisms were common when considering inflectional endings, but undetectable when looking at whole verb forms. This would require (i) that other types of morphological expression showed a typology of person syncretism very different from inflectional affixation, and (ii) that these other kinds of expression were so common that they significantly changed the overall picture. Matthew Baerman (personal communication) reports his impression that the null hypothesis is correct, but work on the issue by the Surrey Morphology Group is expected to take several more years. A cautious

The Morphological Realization of Person 245

application of (6) to inflectional endings thus represents the current state of knowledge in the field. This strategy may find some support in another observation by Baerman, Brown, and Corbett (2005:60), namely, that the pattern in (6) extends to free pronouns. With respect to pronouns, we should also note Harbour’s (2016) finding, already mentioned in chapter 2, that no language has a systematic 1-3 syncretism (that is, a syncretism found in all its paradigms). Devising a theory of person that can account for the Dutch data is a far from trivial task in itself, but making such a theory compatible with (6) at the same time is a challenge. Various existing theories of person features do not seem to fit the bill (compare also Baerman, Brown, and Corbett 2005:60). As an example, consider the well-known system proposed by Harley and Ritter (2002), who analyze first, second, and third person respectively as [participant (speaker)], [participant addressee], and absence of a participant node altogether.2 Baerman, Brown, and Corbett (2005:60–61) note that, although this system disallows 1-3 syncretisms, it also appears to disallow 2-3 syncretisms (given that second and third person have no features in common). The predictions of the system are more subtle when one allows “elsewhere” spellout rules. In that case, the combination of a spell-out rule for speaker and an elsewhere rule will give rise to a 2-3 syncretism, while the combination of a spell-out rule for addressee and an elsewhere rule will give rise to a 1-3 syncretism. It is not clear, however, why there should be an asymmetry between those two options.3 Since the formulation of the spell-out rules depends, of course, on the exact feature structure that characterizes the various persons, we first repeat the inventory of persons developed in chapter 2. The feature structures for pronouns are given in (7)–(9). (We replace the number feature pl as it appeared in chapter 2 by aug; see chapter 3 for discussion.) (7) 

Singular

a. 1st person

b. 2nd person

c. 3rd person

NMB

NMB

NMB

NMB PRS

NMB PRS

NMB PRS

PRS



PRS

PROX

PROX

PROX

DIST



PRS DIST



246

Chapter 7

(8) 

a. 1st person inclusive

b. 1st person exclusive

NMB

NMB PRS

NMB Plural

AUG

PRS

PRS

NMB NΠ

PRS

AUG



PROX

PROX

PROX c. 2nd person

d. 3rd person

NMB

NMB PRS

NMB Plural

AUG

PRS

PRS

NMB NΠ

PRS

AUG

PROX



DIST

DIST (9) 

a. Singular

b. Plural

NMB PRS

NMB Polite

NMB

PRS

PRS

NMB NΠ

AUG

PRS

PROX

PROX

DIST

DIST

HON

HON



The same feature organization holds for verbal agreement, with one proviso. Verbs do not contain the nominal projection line N – prs – nmb. Instead, feature structures are attached to an inflectional node that we have labeled agr. 7.3  Accounting for Person Syncretisms 7.3.1 Introduction

We now turn to the morphophonological realization of agreement endings and to the issue of person syncretism in verbal paradigms.

The Morphological Realization of Person 247

We assume that all languages that show person agreement in the first place have full agreement in syntax, in the sense that the verb is exhaustively specified for person syntactically (see also section 5.5; we will provide evidence for this claim in subsection 7.4.4).4 The fact that not all languages show rich agreement at an observational level is a matter of variation in morphophonological realization.5 We start by considering paradigms in German, which do not show syncretism in the singular (we consider the plural later). The verb kaufen ‘buy’ is representative. (10)

kaufen ‘buy’ 1sg 2sg 3sg

kaufe kaufst kauft

In this case, there is a distinct spell-out rule for each feature combination, as in (11). The interaction between these rules is regulated by Maximal Encoding (see chapter 2). In particular, (11b) must be used for the second person singular, instead of either (11a) or (11c), because this rule mentions both features present in syntax rather than just one. (11) a. {prox} ⇔ /e/ / V-[ __ ] b. {prox dist} ⇔ /st/ / V-[ __ ] c. {dist} ⇔ /t/ / V-[ __ ] With this much in place, we now turn to singular paradigms that show person syncretisms. (We will ignore paradigms that do not show person distinctions at all. Such paradigms can be accounted for in a variety of ways. For example, the language in question may not have person agreement to begin with, or its grammar may include impoverishment rules that delete all person features in verbs.) 7.3.2  2/3 versus 1

In Dutch, regular verbs display a 2-3 syncretism in the singular. (12)

kopen ‘buy’ 1sg 2sg 3sg

(standard)

koop koopt koopt

Syncretism can of course always be analyzed as accidental homophony. That is, Dutch could have three spell-out rules (parallel to the German ones in (11)), which happen to realize both [prox–dist] and [dist] as -t (and [prox] through

248

Chapter 7

a zero ending). Although accidental homophony exists, it would not be a very satisfactory account for this particular syncretism, given its relative crosslinguistic frequency (see (6)). One way to account for the 2-3 syncretism in (12) is to assume that Dutch lacks any spell-out rule mentioning prox as its input, but does have a rule that mentions dist and inserts -t. This rule applies in the third person, and also in the second person, given that there is no more specific rule mentioning both prox and dist (which would otherwise block this, as in German). However, such an account is insufficiently general, as it is not the case that in all languages with a 2-3 syncretism the first person lacks an overt ending. This even holds within Dutch, which has several dialects in which first person agreement takes the form of a schwa or schwa-n ending, but which do show the same 2-3 syncretism as the standard (see Aalberse 2007:134). (13)

kopen ‘buy’ 1sg 2sg 3sg

(dialectal I)

kope(n) koopt koopt

Such dialects must have the spell-out rules in (14a) and (14b) (which realize first and third person). By analogy, we assume that the standard language has the spell-out rules in (14a′) and (14b).6 It does not have a designated spell-out rule that realizes the [prox–dist] specification of the second person, as German does (otherwise, the 2-3 syncretism would have to be analyzed as a case of accidental homophony; see above). The question, then, is why the second person, which after all carries both prox and dist, is realized through application of (14b) (resulting in a 2-3 syncretism), rather than through application of (14a/a′) (which would result in a 1-2 syncretism). (14) a. {prox} ⇔ /e(n)/ / V-[ __ ] a′. {prox} ⇔ ∅ / V-[ __ ] b. {dist} ⇔ /t/ / V-[ __ ] Notice that the answer to this question cannot lie in Maximal Encoding, given that each of the rules in (14a,a′) and (14b) mentions a single feature present in the second person. We propose instead that all rules that operate on φ-feature structures are subject to the condition in (15) (where a host feature is a feature applied first, while a dependent feature is a feature applied subsequently; see chapter 2).

The Morphological Realization of Person 249

(15)  The Russian Doll Principle Given a feature structure with a host and a dependent feature, it is not possible to apply a rule whose target is the host feature and whose structural description does not mention the dependent feature. In other words, in a structure in which F3 applies after F2 and F2 applies after F1, rules that can apply must have a structural description that mentions F3, or F2 and F3, or F1, F2, and F3. Rules that mention only F1, only F2, or only F1 and F2 cannot apply. (This is parallel to Russian dolls in that you can see an outer doll without seeing the dolls it contains, but not vice versa; for a related idea in a very different framework, see Caha 2009.) In the feature structure that characterizes the second person, dist is dependent on prox (see (7b)). Given the Russian Doll Principle, this means that second person agreement can be realized by (14b), but not by (14a/a′). This allows us to account for the common occurrence of 2-3 syncretisms as a simple case of underspecification in spell-out rules: such syncretisms occur whenever there is no designated spell-out rule for [prox–dist], while there are spell-out rules for [prox] and [dist]. In cases of underspecification in the spell-out system, the Russian Doll Principle thus determines which way the syncretism goes. It is important that there is a principle that does so. In its absence, one would predict the possibility of optionality in the choice of which feature in a complex feature structure is spelled out. This flies in the face of the fact that crosslinguistically such optionality never seems to be a property of the general paradigm of a language. We are not aware of any languages in which the second person can freely alternate (in the same syntactic context) between endings shared with first and third person. (The point obviously holds more generally for inflection; optionality in how number or gender is marked in a particular syntactic context is rare.)7 The Russian Doll Principle will turn out to have effects elsewhere—for instance, in blocking certain rules of impoverishment. 7.3.3  1/2 versus 3

Given this account, how would 1-2 syncretisms arise? These are found in various languages (see Baerman, Brown, and Corbett 2005; compare (6)), including a number of dialects of Dutch. Aalberse (2007:132) notes that in a few dialects first and second person both lack an overt ending, while third person is marked by -t; she also notes that in at least one dialect first and second person subjects both trigger a schwa ending.

250

Chapter 7

(16)

kopen ‘buy’ 1sg 2sg 3sg

(dialectal II)

koop/kope koop/kope koopt

These dialects have the same spell-out rules as other versions of Dutch, namely, those in (14). This must be so in view of the way first and third person are realized. Where these dialects differ, we propose, is in having an additional impoverishment rule that deletes dist in the context of prox. (17)  dist → ∅ / V-[ __ prox] As a result of application of this rule, what is delivered to the spell-out system is [prox] for both first and second person agreement. Hence, the rule in (14a/a′) applies in both cases, yielding the paradigm in (16). If the Russian Doll Principle is correct, this account of 1-2 syncretisms must extend beyond Dutch. Whenever such a syncretism is found, it will be the consequence of a rule of impoverishment that deletes dist (since the Russian Doll Principle forbids realizing [prox– dist] using the spell-out rule for prox if dist is still present). 7.3.4  1/3 versus 2

As stated in (6), 1-3 syncretisms are rare, and certainly less frequent than 1-2 and 2-3 syncretisms. The theory we propose appears to be in a good position to explain this asymmetry. First and third person have no features in common, and therefore—in the absence of any impoverishment—there can be no spellout rule that assigns them the same form. However, 1-3 syncretisms do exist, and so there must be a way to generate them. The plural verbal agreement paradigm of standard German (1st -en, 2nd -t, 3rd -en) is an example. Given our assumptions so far, the only way to generate such a syncretism is to reduce the first and third person feature structures to a bare agr node by impoverishment, while leaving the second person feature structure intact.8 This requires two impoverishment rules of the following form:9 (18) a. prox → ∅ / V-[ __ ] b. dist → ∅ / V-[agr– __ ] The rule in (18a) can apply recursively in the first person, stripping it of both its prox features (or its single prox feature in the plural inclusive). It cannot apply in the second person, because there prox has a dependent feature dist that blocks application of (18a) as a consequence of the Russian Doll Principle.10

The Morphological Realization of Person 251

The rule in (18b) is intended to apply in the third person only. This is achieved by mentioning the agr node in its structural description: in order to be a candidate for deletion, the dist feature must be a direct dependent of agr, as is the case in the third person. (Recall that the dash used in (18b) indicates that the element to its right is a direct dependent of the element to its left.) The rule in (18b) cannot apply in the second person, because in the [agr–prox– dist] structure that characterizes second person agreement, dist is not a direct dependent of agr.11 The overall consequence of (18), then, is that the feature structure of the second person remains intact, while first and third person end up with a bare agr node after impoverishment, thus providing the basis for a 1-3 syncretism. (19) Output of (18)

a. 1st person

b. 2nd person

c. 3rd person

agr

agr | prox | dist

agr

Given this output, there are several combinations of spell-out rules that indeed give rise to a 1-3 syncretism. One possibility is that the only relevant spell-out rule in the language is the one in (20). This rule can apply to (19a) and (19c), but not to (19b), given the Russian Doll Principle. This leads to a realization of (19) as /a/–∅–/a/. (20) {agr} ⇔ /a/ / V-[ __ ] A second possibility is that the language has one of the spell-out rules in (21), either as the only relevant spell-out rule or in combination with a rule like (20). In such a system, (19b) will be spelled out by (21a) or (21b), while (19a) and (19c) will either both be realized by (20) or—in the absence of (20)—not be realized at all, giving rise to /a/–/b/–/a/ or ∅–/b/–∅. (21) a. {dist} ⇔ /b/ / V-[ __ ] b. {prox dist} ⇔ /b/ / V-[ __ ] The first of these two spell-out systems (which consists of only the rule in (20)) is unlikely to be attested. In general, there is a tendency for more articulated feature structures to be associated with more articulated phonological realizations. In other words, morphosyntactic markedness correlates with morphophonological markedness. The effect is discussed by many authors,

252

Chapter 7

sometimes under the rubric of “iconicity” (see Zwicky 1978, Aissen 2003, Croft 2003, and Haspelmath 2006). For example, languages that make a distinction between past and present tense can mark either both or only the past tense, but do not mark the present tense only (Bybee and Dahl 1989). Similarly, languages that distinguish singular and plural may mark both or just the plural, but they do not mark only the singular (Greenberg 1963, Universal 35). We can describe this effect by saying that, for any paradigmatic opposition that results from the presence or absence of a feature, there must be a spell-out rule that mentions that feature in its structural description. Although the effect is perhaps not well-understood, it is pervasive, and crucially the spell-out system that realizes (19) as /a/–∅–/a/ goes against it: it is anti-iconic in that the morphosyntactically marked feature structure in (19b) is exactly the one that remains morphophonologically unmarked. Indeed, as Cysouw (2003:57–64) notes, there are no languages known to exhibit a 1-3 syncretism in their inflectional system such that the second person is zero. In sum, compared to 1-2 syncretisms, which require one impoverishment rule (see (17)), 1-3 syncretisms require an additional impoverishment rule (see (18)), whereas 2-3 syncretisms do not require any impoverishment. The realization of 1-3 syncretisms will be as ∅–/a/–∅ or /a/–/b/–/a/. 7.3.5  Why 1-3 Syncretisms Are Relatively Rare

The fact that 1-3 syncretisms require an additional impoverishment rule is not enough to explain why such syncretisms are rare compared to 1-2 and 2-3 syncretisms. There is no reason to assume that the number of grammatical rules required to derive a particular syncretism determines its relative crosslinguistic frequency. Such an assumption is equivalent to what used to be known as the derivational theory of complexity, which we take to mix up levels of description in the sense of Marr 1982 (for discussion, see Neeleman and Van de Koot 2010 and references mentioned there). Moreover, even if we were to accept the notion that grammatical complexity determines typological frequency, there is an empirical issue, because a further prediction would be made about the relative frequency of 1-2 and 2-3 syncretisms: the former should be rarer, as they require a rule of impoverishment while the latter do not. To the best of our knowledge, there is no convincing evidence for such a typological asymmetry. Our account for the relative infrequency of 1-3 syncretisms is rooted in the way spell-out systems for agreement endings are acquired (on the assumption that the child already has access to the featural makeup of verbs in syntax; this assumption is entailed by our hypothesis that if a language has person agreement in the first place, then all relevant φ-features are present on the verb in

The Morphological Realization of Person 253

syntax). The general idea is that, over time, acquisitional difficulty leads to typological infrequency because acquisitional difficulty makes a grammatical subsystem prone to historical change. Therefore, if we can show that 1-3 syncretisms are harder to acquire than other syncretisms, we will have an explanation for the attested typological pattern. Indeed, it can be argued that for all paradigms, except those involving a 1-3 syncretism, there is a learning path that guarantees success and that minimizes the child’s computational burden. What we mean by a successful learning path is one that arrives at the target system without requiring retraction of a spellout rule once it has been postulated. In other words, in a successful learning path there is no backtracking (compare Dresher and Kaye 1990, Nyberg 1992, Bertolo et al. 1997, and Snyder 2007, 2008). Computational burden is defined in terms of the number of syntactic inputs that must be considered at any one point in order to acquire the spell-out rules that account for a given formal distinction. Suppose that the child observes a distinction between two verb forms. The simplest approach to the task of finding out which of the syntactic feature structures that encode person these forms express is to serially test potential mappings between the two observed forms and two inputs. A comparison of potential mappings to three inputs is, we assume, undertaken only when there is no alternative (because such a comparison imposes a heavier memory load; compare Kapur 1994 and Frank and Kapur 1996).12 When a child first notices a formal distinction in agreement morphology, the best strategy is to hypothesize that this reflects the syntactic feature structures [prox] versus [dist] (rather than assuming that the feature structure [prox–dist] is involved). This strategy will not work when the target grammar has a 1-3 syncretism (a point to which we will return), but it is the only strategy that is successful (in the sense outlined above) in the remaining paradigms. All other strategies may lead to the necessity of retracting spell-out rules, contra the ban on backtracking. To see why, consider first the grammars that according to our proposal underlie the various different output paradigms (impoverishment rules are given in compact format). (22) Paradigm a. /a/–/c/–/b/

b. /a/–/b/–/b/

Grammar {prox} ⇔ /a/ {dist} ⇔ /b/ {prox dist} ⇔ /c/ {prox} ⇔ /a/ {dist} ⇔ /b/

254

Chapter 7

Paradigm c. /a/–/a/–/b/

d. /a/–/b/–/a/

Grammar {prox} ⇔ /a/ {dist} ⇔ /b/ [prox–dist] → [prox–∅] {agr} ⇔ /a/ {(prox) dist} ⇔ /b/ prox → ∅ [agr–dist] → [agr–∅]

We assume that the information available to the child when acquiring person paradigms consists of a pair of verb forms and information about the semantics of the structure in which these were used. If the child’s hypothesized spell-out rules are compatible with the semantics of the input structures, the child will incorporate these rules into their grammar. If they are not, the child will simply not posit any spell-out rules on the basis of the input under consideration. Suppose, then, that the child notices a formal distinction in agreement morphology and hypothesizes that this represents a distinction in the realization of the inputs [prox] and [dist]. If this is in line with the semantics, the child will postulate the two spell-out rules {prox} ⇔ /a/ and {dist} ⇔ /b/. Since these are indeed part of all grammars that have a distinction between first and third person, no backtracking will ever be required. If the target grammar is as in (22b), nothing further needs to be acquired. If the target grammar is as in (22a), the dedicated second person form can be acquired upon noticing a further formal distinction between either first and second or second and third person. If the target grammar is as in (22c), an impoverishment rule will be added to the grammar when the child notices that the form /a/ is used in second person contexts as well. Consider an alternative strategy according to which a difference in form is initially taken to reflect a distinction between the inputs [prox] and [prox–dist]. Even if this is in line with the semantics of the input, it is possible for the child to postulate spell-out rules that need to be retracted later. This is because a formal distinction between first and second person is compatible with the spell-out systems in (22a), (22b), and (22d). Crucially, these spell-out systems do not all have the same rule for the second person, or indeed the first person. Therefore, a child could postulate separate spell-out rules for prox on the one hand and for a combination of prox and dist on the other (as in (22a)), while the target grammar is (22b). Converging on the target grammar will then necessarily require retraction of the spell-out rule for [prox–dist]. Alternatively, the child could postulate separate spell-out rules for prox and dist to capture the opposition between first and second person (as in (22b)), while the target

The Morphological Realization of Person 255

grammar is as in (22a). These spell-out rules would then have the following form: {prox} ⇔ /a/ and {dist} ⇔ /c/. Again, this will lead to retraction, because the spell-out rule for [dist] in the target grammar associates it with a different form than the one hypothesized by the child (namely, /b/). Similar problems arise when the child adopts the strategy of initially matching a formal distinction to the inputs [prox–dist] and [dist]. Again, there are multiple spell-out systems that yield different forms for second and third person, namely, those in (22a), (22c), and (22d). If the child happens to make the wrong choice given the target grammar, retraction will be necessary. For example, a child may posit separate spell-out rules for [prox–dist] and [dist] in order to account for an opposition between second and third person (as in (22a)). However, if the target grammar is (22c), the child will later have to retract the rule for [prox–dist]. In summary, we propose that a child adopts a strategy of matching formal contrasts initially to first versus third person inputs, given that alternative strategies are likely to put the child in situations that require backtracking. As a result, a syncretism between first and third person is harder to acquire than other syncretisms, which in turn explains its relative typological infrequency. The fact that the best initial strategy fails for 1-3 syncretisms does not imply that paradigms exhibiting such a syncretism cannot be acquired at all. In fact, it can be done without backtracking. However, it is not possible, following the failure of the initial strategy, for the child to employ a secondary strategy according to which the relevant formal distinction represents either first versus second or second versus third person. This is because, as we have just shown, this strategy may lead the child to postulate spell-out rules confirmed in the initial context that later have to be retracted when more input is taken into account. Thus, the child cannot match two syntactic inputs to two forms. Instead, the child must associate three syntactic inputs to the two forms under consideration. Only then is it possible to set up a system of spell-out rules that is compatible with the input and will not require retraction. As pointed out above, this is a more complex learning task. Such complexity might be an additional factor disfavoring 1-3 syncretisms, but this must remain a speculation given that little is known about the computational capacity of children in domains of this type. The account just outlined makes one clear prediction, namely, that first and third person agreement forms will be acquired before second person agreement forms. This is a remarkable prediction in view of the fact that second person forms have been observed to be highly frequent in child-directed speech. In fact, the subject you is more frequent in English parentese than any other subject (Laakso and Smith 2007:732; compare also Ambridge et al. 2015:245).

256

Chapter 7

There is no reason to think that child-directed speech shows significant differences in the relative frequency of second person subjects across languages. Hence, if anything, one might expect second person agreement to be acquired early. Nevertheless, our prediction that second person forms are acquired last appears to be correct. The following acquisition studies confirm that, at least in the singular indicative present, the child starts using first and third person agreement before second person agreement.13 (23) Language Basque Catalan Croatian Estonian Finnish German Greek Hebrew Italian Lithuanian Northern East Cree Spanish

Turkish

Reference

Order of acquisition

Austin 2012 Grinstead 2000 Katičić 2003 Lipp 1977 Laalo 2003 Clahsen 1988, Poeppel and Wexler 1993 Christophidou and Stephany 2003 Armon-Lotem 2006 Clark 1985 Wójcik 2003 Terry 2009 Grinstead 2000, Félix-Brasdefer 2006, Austin 2012; Aguirre 2003 Ekmekci 1982

3