Features and Processing in Agreement 1527505731, 9781527505735

Citation preview

Features and Processing in Agreement

Features and Processing in Agreement By

Simona Mancini

Features and Processing in Agreement By Simona Mancini This book first published 2018 Cambridge Scholars Publishing Lady Stephenson Library, Newcastle upon Tyne, NE6 2PA, UK British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library Copyright © 2018 by Simona Mancini All rights for this book reserved. No part of this book may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording or otherwise, without the prior permission of the copyright owner. ISBN (10): 1-5275-0573-1 ISBN (13): 978-1-5275-0573-5

“They’ve a temper, some of them – particularly verbs, they’re the proudest – adjectives you can do anything with – but not verbs. However, I can manage the whole lot!” (Lewis Carroll, Through the Looking-Glass)

TABLE OF CONTENTS

Acknowledgments ...................................................................................... ix Introduction ................................................................................................. x Agreement List of Abbreviations ................................................................................ xiv Chapter One ................................................................................................. 1 The Linguistics of Agreement Features in the Minimalist Program ....................................................... 1 Fine-grained decomposition of agreement projections: Cartography .... 5 Features, anchors and interpretation .................................................... 11 Summary .............................................................................................. 14 Chapter Two .............................................................................................. 16 Sentence and Agreement Comprehension The online study of sentence comprehension ...................................... 16 Sentence comprehension models ......................................................... 37 Features and interpretation: anchoring agreement in comprehension .. 42 Summary .............................................................................................. 44 Chapter Three ............................................................................................ 46 Evidence for Feature-Specific Processing Feature-specific processing: Early evidence ........................................ 48 Testing the FIP: Behavioural evidence ................................................ 50 Testing the FIP: Electrophysiological evidence ................................... 54 Testing the FIP: Neuroanatomical correlates ....................................... 60 Summary .............................................................................................. 65

viii

Table of Contents

Chapter Four .............................................................................................. 67 More on Feature-Specific Processing: Person Asymmetry The Person Asymmetry Hypothesis ..................................................... 67 1st/2nd vs. 3rd person: Person underspecification and contextdependence ..................................................................................... 70 Pronoun representation and interpretive anchors: an electrophysiological investigation ............................................. 80 The featural makeup of pronouns ........................................................ 87 Summary .............................................................................................. 89 Chapter Five .............................................................................................. 90 Functional and Temporal Dissociation of Agreement Mechanisms When disagreement is grammatical: Unagreement .............................. 90 Unagreement processing and the role of interpretive anchors ............. 95 Unagreeing, null and overt subjects ................................................... 109 Summary ............................................................................................ 114 Chapter Six .............................................................................................. 116 Anchoring Agreement From feature bundles to feature anchors ............................................ 117 Representations, algorithms and neuroanatomical bases of agreement ................................................................................. 120 Relation to existing sentence comprehension models ........................ 128 Conclusion ......................................................................................... 130 Notes........................................................................................................ 131 Bibliography ............................................................................................ 134 Index ........................................................................................................ 151

ACKNOWLEDGMENTS

Behind these pages there are many people that I have met and worked with during the last ten years and to whom I am deeply grateful. There are three people in particular to whom I feel indebted: Luigi Rizzi, Manolo Carreiras and Nicola Molinaro. Luigi was the first one to introduce me to the realms of linguistics and to the intricacies of agreement. It was a real pleasure and honour to work with him when I was in Siena, and it still is a pleasure now that I am not there anymore. I truly consider myself lucky for this. Manolo opened the doors of his lab in Tenerife to me in 2009. There I met Nicola and after 8 years, the three of us are still talking about agreement (and many other things), in a less exotic but surely exciting environment (the BCBL in the Basque Country). Without them, almost all the work described here would not have been possible. Another person who significantly contributed to all this is Ileana Quiñones, whose skills, hard work and patience pushed this work on agreement farther than initially expected. My gratitude goes also to my BCBL colleagues and friends, especially to Sendy, whose presence, input and help enrich any conversation, be it agreement-related or not! I also wish to thank Alessandro Laudanna, for guiding me in my first steps into psycholinguistics, and Francesco Vespignani, the first person to show me an event-related potentials (ERP) waveform. My thanks go also to former and current CISCL members, Cristiano, Giuliano, Vincenzo and especially Irene and Ida, whose presence and friendship made room 322 in the San Niccolò building a very pleasant place to work in. Finally, I am extremely lucky to have several “anchors” in my life: Jose, my family, Claudia, Iuri, Emma, Laura, Francesca, Hada, Ercole and Ernesto. Their support has mattered and will always matter more than anything else.

INTRODUCTION AGREEMENT

Agreement is a widespread and varied phenomenon. Its pervasiveness in some languages contrasts with its near absence in others, posing a challenge for the linguists and psycholinguists who attempt to explain the mechanics of its representation and processing. This work explores the intricacies involved in agreement computation, with the aim of unveiling theoretical and psycho-/neuro-linguistic aspects of this crucial syntactic relation. The empirical focus will be on subject-verb (s-v henceforth) agreement and on the features engaged in this dependency: person and number. The analysis will be carried out by making reference to the Minimalist Program (Chomsky 1995; 2001; 2005) and Cartography (Belletti 2008; Cinque 1999; Cinque and Rizzi 2008; Rizzi 1997, 2004; Shlonsky 2010, inter alia) recently developed within mainstream generative grammar. Despite the apparently inherent tension that seems to set Minimalism and Cartography in opposition, the combination of these two lines of research may be extremely fruitful: while the former focuses on the generative devices involved in the derivation of syntactic structures, the latter focuses on the “atoms” of the generated structures (Cinque and Rizzi 2008). The two lines of research can therefore be pursued in parallel, and this work represents an attempt at doing so. Agreement manifests itself when grammatical information appears on a word that is not the source of that information. Early derivational grammars defined agreement as a relation holding between two elements – a controller and a target – that share specific features, with the controller (also called trigger) being the element from which grammatical information originates, and the target the element that inherits the information. S-v agreement is an instantiation of the controller-target relation. What characterises this dependency is the systematic covariance (Steele 1978) existing between the feature sets of the former and latter members of the relation: the subject can vary between singular and plural number and among 1st, 2nd and 3rd person, with the form of the verb that changes accordingly, so that an identity of features is realised. Covariance

Features and Processing in Agreement

xi

is an essential notion: it is not sufficient that two elements happen to share properties; the sharing must also be systematic. An important aspect of agreement concerns the way this dependency is morphologically expressed on the verb. In Italian, as in many other richlyinflected languages, person and number values are expressed by an affix that attaches to the verb stem. Crucially, the same morpheme marks both person and number information. In the Italian sentence in (1), the agreement morpheme –e represents both the 3rd person and the singular number value of the subject ‘il gatto’. (1) Il gatto corre The cat3.sg runs3.sg ‘The cat runs’

This leads us to the main research question raised in this work: does the parser distinguish between person and number during agreement processing, or are these features undifferentiated and accessed as a bundle? This question reflects the theoretical divide existing between the single-cluster analysis of agreement elaborated within standard minimalist analyses (Chomsky 1995 and subsequent work) and the distinct-cluster analysis of agreement features put forth within the cartographic framework (Shlonsky 1989, 2000, 2009, 2010; Sigurdsson 2004; Sigurdsson and Holmberg 2008, inter alia), which will be thoroughly illustrated in Chapter 1. Similarly, within recent psycholinguistic research, an unequivocal answer has not been given yet. In light of this, the studies presented in Chapter 3 represent an attempt to clarify whether a dissociation between person and number can be maintained, both in processing and syntactic structure terms. The behavioural and neuro-imaging experiments will provide convergent results to the effect that a functional dissociation between the two features can be maintained. Fundamental to this dissociation is the role of interpretive anchors, i.e. structural positions where morphosyntactic values are linked – or anchored – and that drive the interpretation of person and number. The Feature Interpretation Procedure in Chapter 1 will provide the parser the flexibility necessary to deal with the different information carried by features. A related issue addressed in this monograph concerns a more finegrained aspect of subject-verb agreement, namely the distinction between 1st/2nd and 3rd person pronouns. Morphological and configurational splits among pronouns have been highlighted in a variety of languages and have been attributed to the different featural specifications of pronouns. A longstanding tradition (Benveniste 1966; Forcheimer 1953; Jakobson 1971) distinguishes 1st and 2nd person from 3rd person pronouns on the basis of

xii

Introduction

the capacity that the former, but not the latter pronouns, have to pick up and identify a specific speech participant (speaker and addressee). This intrinsic difference in terms of discourse-(un-)relatedness is thought to affect the featural makeup of pronouns. As real persons in the speech act, 1st and 2nd person are specified for the person feature, while 3rd person pronouns, which represent entities being talked about without any active role in the speech act, are specified for the number feature (Anagnostopoulou 2003; Benveniste 1966; Harley and Ritter 2002; Kayne 2000, inter alia). Identity with (or inclusion of) a speech participant is thus the criterion used to discern the two classes of pronouns, and the discourse-relatedness of a pronoun is therefore determined solely on this basis. Alternative and less radical feature representations have been developed (Bianchi 2006; Sigurdsson 2004) that do not relate the discourse-relatedness of a pronominal form solely on the basis of the presence of an underlying speech participant. It is true that 1st and 2nd person differ from 3rd person in their reference to a speech role, but there are also similarities between the three forms that have not been captured by former analyses, namely the fact that 3rd person refers to a contextually salient entity, which determines a certain degree of discourse-relatedness (Bianchi 2006). The ERP experiment presented in Chapter 4 will show that the parser is able to differentiate between 1st/2nd and 3rd person agreement and between the different degrees of discourse-relatedness that the two classes of pronouns have. The idea of a controller-target dependency highlights a fundamental aspect of agreement, namely its asymmetrical character. Seeing agreement as an asymmetrical relation implies not only that this phenomenon is a matter of “displaced” grammatical information (Corbett 2006) copied from the controller to the target, but also that the two elements involved in the relation do not play the same role. There are two interrelated ways in which the asymmetry between controller and target manifests itself. First, the controller (the subject) has no choice of feature value, while the target (the verb) does. The target can have different morphological forms available to match the person and number features of the noun: in (1), the verb corre, a 3rd person singular, is one of them, which is chosen on the basis of the person and number values of the subject-controller. The controller, on the other hand, does not have the same availability of morphological forms: a lexical DP comes only as a 3rd person, and the only variance that it is allowed is between singular and plural. In this view, it is the verb that agrees with the subject, and not conversely. Second, the contribution to semantic interpretation is related to the controller rather than to the target: if we shift from a singular to a plural subject in (1), the

Features and Processing in Agreement

xiii

verb varies accordingly, but it is not the plural number marking on the verb that will affect the interpretation. Rather, interpretation will rely on the subject argument. In sum, two assumptions seem rather straightforward and unequivocal about agreement: i) the fact that it is based on a systematic covariance of features, which surfaces with feature identity between controller and target; ii) the unidirectionality of the feature valuing process, which is supposed to operate from subject to verb and not vice versa. Nevertheless, across languages, agreement patterns are found in which a featural mismatch between subject and verb is allowed. Unagreement in Spanish is one such phenomenon, in which the presence of a person mismatch between subject and verb nonetheless produces a well-formed sentence, as in (2). What ensures the grammaticality of these sentences is the superimposition of verbal 1st person plural interpretation onto the 3rd person plural value of the subject. (2) Los lingüistas escribimos un artúculo muy interesante The linguists3.pl wrote1.pl. an article very interesting ‘We linguists wrote a very interesting article.

The relevance of unagreement resides in the opportunity that it offers to test the degree of permeability of agreement processing to semanticdiscourse factors and the directionality of the mechanisms. This issue will be profusely addressed in Chapter 5, where instances of unagreement and standard agreement in Spanish will be compared. The theoretical analysis of these constructions will be integrated with data from behavioural, electrophysiological and neuroanatomical investigations that will shed light on the time course, the mechanisms and the neuro-biological bases of agreement processing. By adopting features as units of analysis, this work will decompose agreement into its basic building blocks and algorithms, providing the first framework to understand its dynamics and neuro-biological bases in morphologically rich languages.

LIST OF ABBREVIATIONS

Acc ACC AG Cl CP Dat Dir DO DP EPP Fem FIP Incl Inv IO IP LC LD LIFG LF Masc MTG NID Nom PCC Pl PF s-v sg STG/STS TP VP

Accusative case Anterior cingulate cortex Angular gyrus Clitic Complementiser Phrase Dative case Direct (Voice) direct object Determiner Phrase Extended projection principle feminine gender Feature Interpretation Procedure inclusive Inverse (Voice) Indirect object Inflectional Phrase Logophoric Centre Left dislocation Left Inferior Frontal Gyrus Logical Form masculine gender Middle temporal gyrus Northern Italian Dialect Nominative Case Person Case Constraint plural number Phonological Form subject-verb agreement singular number Superior Temporal Gyrus/Sulcus Tense Phrase Verb Phrase

CHAPTER ONE THE LINGUISTICS OF AGREEMENT

In this chapter, agreement will be presented from the perspective of the Minimalist Program (Chomsky 1995 and subsequent work) and of Cartography (Cinque and Rizzi 2008, Shlonsky 2010, inter alia), the assumptions of which will lay the foundations of the theoretical and experimental study of s-v agreement that will be carried out throughout this book. The focus will be on the central role that basic building blocks of agreement relations – i.e. features – play at the structural and interpretive level.

Features in the Minimalist Program The centrality attributed to features in the derivation of agreement relations has been assumed only lately. Early derivational grammars tended to provide a general account for the phenomenon of agreement, as the concern was mainly on the extent to which agreement processes could be assimilated to general syntactic mechanisms, while the substance of what did the agreeing, i.e. phi-features, was largely ignored. In Syntactic Structures (Chomsky 1957), agreement was treated as a context-sensitive transformation that followed specific rewrite rules by means of which structural changes were applied. Such rules primarily belonged to the part of grammar that specifies how the pronunciation of syntactic structures is affected, as in (1). (1) Number transformation – obligatory Structural analysis: X-C-Y Structural change: C Æ S in the Context of NPsing ___ Ø in other contexts Past in other contexts (from Chomsky 1957, 112)

Number inflection in non-past sentences is rewritten as -s when preceded by a singular NP, and as zero elsewhere. The structural change

2

Chapter One

transformation is in essence a rewrite rule. To surface with the appropriate form, /s/, /z/ or /iz/, the -s morpheme undergoes morphophonological rules. In Aspects of the Theory of Syntax (Chomsky 1965), the approach had become both featural and syntactic. Agreement was described as a phenomenon whose mechanics basically relied on the asymmetry between a controller and a target. Such an asymmetry was captured by means of a process in which the phi-features on the controller were literally copied or moved to the target. In other words, person, number and gender originated in the noun but were eventually expressed somewhere else in the sentence, for example in a verb or in a demonstrative. Phi-features, the substance of agreement, were thus conceived as atomic elements that were manipulated by syntax. In earlier instantiations of the Government and Binding Theory (Chomsky 1981), a fully-fledged theory of phi-features was still missing, but configurations such as specifier-head agreement were outlined that accounted for the licensing of these features and the realisation of agreement. In other words, s-v agreement was explained by making reference to a specifier-head relation in which the two elements share a specific set of agreement features. More recently, in the attempt to reduce derivational complexity, superfluous structural configurations and relations among elements have been eliminated and substituted by a feature-checking mechanism. By introducing the so-called Checking Theory, the Minimalist Program (Chomsky 1995) aimed at simplifying the earlier Case Theory by eliminating case assignment under government and treating Accusative Case, like Nominative Case, as assigned under specifier-head agreement. The minimalist analysis of s-v agreement is essentially based on the interpretive asymmetry existing between agreement features on the subject and those on the verb. More precisely, the mechanisms operating in the realisation of s-v agreement hinge on the opposition between interpretable and non-interpretable features, respectively located on the subject noun and on the verb. This dualism rests on the assumption that, while phifeatures on the subject provide Logical Form (LF) with fundamental cues to interpret this dependency (e.g. the plural value of the number feature indicates the cardinality of individuals, a 3rd person value distinguishes an entity being talked about from a speaker or an addressee, and so on), the features on the verb are mere copies of those on the subject, i.e. morphological expressions of a formal relation, and hence redundant values that do not add any relevant information.

The Linguistics of Agreement

3

Technically, interpretable features enter the derivation endowed with specific person and number values, while uninterpretable ones need to receive a value from the former by means of a formal relation – checking – that permits the licensing of morphosyntactic features in the course of the derivation. For such a relation to be properly established, a local relationship must hold between subject and verb: movement (of the subject DP to SpecTP) is motivated by the need to check off the verb’s uninterpretable features in a specifier-head configuration, which is the only admissible checking relation. For Full Interpretation to be possible, a verb’s uninterpretable feature must be eliminated after checking, as redundancy is not admitted in an optimal design. Chomsky (1995) makes a very clear point about the conditions under which features can be checked: features cannot be checked under feature mismatch, as this circumstance would lead to cancelling of the derivation. Here a conflict seems to arise between this theoretical standpoint and certain agreement configurations that allow feature mismatches. In other terms, if a configuration with mismatching features is not a legitimate syntactic object, how can we motivate the presence of “grammatical mismatches” such as the ones in (2) and (3) below in Spanish? (2) Los lingüistas escribimos un artículo muy interesante The linguists3.pl wrote1.pl an article very interesting ‘We linguists wrote a very interesting article’ (3) Los lingüistas escribís un artículo muy interesante The linguists3.pl wrote2.pl an article very interesting ‘You linguists wrote a very interesting article’

The analysis of these agreement patterns will be dealt with in Chapter 5. An aspect of the minimalist approach to s-v agreement relevant to the analysis that will be developed here concerns the structural site in which uninterpretable features are located within sentence structure. Chomsky’s (1995) assumption is that person and number features form a cluster hosted under the same syntactic head, normally identified with the T node, which is responsible for the expression of tense, as shown in Figure 1-1. It is in T that all uninterpretable phi-features are clustered, without any structural distinction among them. A straightforward consequence of this implementation is that the checking operation accesses the whole feature bundle in a unique computational step, and not in a series of distinct operations, one for each feature to be checked and valued.

4

Chapter One

Figure 1-1. Uninterpretable person and number (uPerson, uNumber) form a cluster under the T head. SpecTP is occupied by the subject, whose interpretable features enter in a checking relation with those on the verb.

In subsequent proposals (Chomsky 2000, 2001, 2005), Chomsky refines the concept of feature checking and introduces the operation Agree. Agree permits feature valuation at a distance, through a c-commanding relation holding between a higher head whose uninterpretable features must be checked (the probe, i.e. the verb), and the element whose interpretable features (the goal, i.e. the subject noun) are checked against. Agree then supersedes the original motivation for movement (feature checking) and replaces it with a system of formal licensing. Uninterpretable features serve to implement operations and as such they render the goal active, i.e. able to implement an operation, which in this case is the deletion of the probe. The phi-set contained in the probe seeks a goal, i.e. matching features, with which it can establish agreement. Locating the goal, the probe erases under Matching: the erasure of uninterpretable features on the goal is the operation called Agree (Chomsky, 2000, 122). Checking therefore reduces to deletion of the uninterpretable features under matching. Importantly, deletion is taken to be a “one fell swoop” operation dealing with the phi-set as a unit. Its features cannot selectively erase: either all delete, or none (Chomsky 2000, 124). Both Checking theory and Agree imply an asymmetrical and unidirectional relation holding between the probe and the goal, due to their different interpretability status: person and number values are copied from the subject to the verb, which amounts to saying that it is the verb that agrees with the subject, and not the converse (Chomsky 2000, 124). Importantly, Chomsky (1995 and subsequent work) stresses the narrowly syntactic nature that s-v agreement mechanisms have. Both the earlier Checking and the later Agree operations introduced to account for the identity of feature values that surfaces with agreement do not span

The Linguistics of Agreement

5

outside the boundaries of Narrow Syntax. It is within the limits of this component that checking and uninterpretable feature deletion take place, since the syntactic object handed over to the covert component, i.e. Logical Form (LF), will have to be deprived of any uninterpretable element to ensure that the derivation will converge.

Fine-grained decomposition of agreement projections: Cartography Besides the minimalist single-cluster approach to s-v agreement, distinct-cluster analyses of this phenomenon can be also found, which stress the structural differentiation of the features involved in s-v agreement. They provide a detailed mapping – i.e. a cartography – of the projections involved in the realisation of this dependency that best captures the cross-linguistic variance of s-v agreement richness. The cartographic approach to sentence structure has its roots in Pollock’s (1989) Split Infl Hypothesis, according to which two distinct functional projections can be identified in the inflectional area of the sentence (Inflectional Phrase, IP): one for the realisation of s-v agreement, i.e. AgrSP, and one for tense marking, i.e. TP, with this functional projection being in a higher position than AgrSP. The core assumption behind this structural differentiation was that rich agreement potentially correlates with height of verb movement: for instance, Romance finite verbs, which show rich agreement, move higher than both English finite verbs and Romance participles, which agree less fully. A slightly different description of the inflectional area and its functional projections was put forth by Belletti (1990), who proposed that AgrSP is higher in the structure than TP, on the basis of the order in which inflectional affixes appear within verbs in Italian (and other languages), as shown in (4): (4) Parl- ava Root- Tense affix-Agr.affix ‘He talked’

In (4), the tense affix -av- appears closer to the verb root (parl-) than the agreement affix (-a). According to the Mirror Principle (Baker 1985), the tense affix, being closer to the root, occupies the lowest position in the syntactic tree, while the agreement affix occupies the highest one, resulting in a configuration like (5):

6

Chapter One (5) [AgrSP Agr [TP T [VP V]]]

Following Pollock’s (1989) and Belletti’s (1990) seminal work, AgrSP has been progressively decomposed to provide a detailed mapping of the inflectional area, in which separate functional projections that host distinct morphosyntactic features have been identified (Linn and Rosen 2003; Poletto 2000; Shlonsky 1989; 2000; 2009; Sigurdsson 2004; 2009 Sigurdsson & Holmberg 2008, inter alia). In other words, AgrSP has been decomposed into different projections that are responsible for person, number and gender agreement singularly. The grammar can then access phi-features separately, and person, number and gender agreement on the verb would result from the establishment of distinct Agree relations, as the verb moves up in the structure. Data from Hebrew, Arabic, Icelandic and Italian show that by separating the bundle of features involved in s-v agreement and analysing each of them as separate projections, it is possible to explain the syntactic phenomenon of agreement in a way that best captures the fact that not all languages show the same richness of agreement. Let us see how. In his analysis of Hebrew and Arabic agreement patterns, Shlonsky (1989) argues that instead of an Agr node per se that includes a bundle of features, a more articulated structure with separate agreement features can be postulated, as in Figure 1-2. The three features of gender, number and person are separately represented in three different projections: GenderP, NumberP and PersonP, with PersonP occupying the highest position in the tree.

Figure 1-2. Decomposition of the Agr node into separate Person, Number, Tense and Gender projections in Hebrew (Shlonsky 1989).

The Linguistics of Agreement

7

In Hebrew and Arabic, verbs need to associate with features one by one, in successive steps. A verb can adjoin to NumberP if it has previously adjoined to GenderP and, similarly, it can manifest person agreement only if it manifests number and gender agreement. Shlonsky reports data showing that agreement can be defective, as in Arabic V-initial clauses and in Benoni verbs in Hebrew. Arabic V-initial verbs agree with the subject only in gender, because the verb has failed to go further up in the tree and adjoin to NumberP and PersonP. Full agreement in Arabic occurs only when the subject is preverbal. In Hebrew, Benoni1 verbs are [-finite] and, as such, they cannot trigger verb raising: V raises to Gender, then to Number, but the non-finiteness of Tense blocks verb raising to PersonP. The result is that verb and subject agree only in number and gender. Table 1-1 shows Hebrew and Arabic agreement patterns. Table 1-1. Hebrew and Arabic agreement patterns Agreement patterns in Hebrew ‫ސ‬ al ha-xacilim Future and Past forms: Inflected for Ata ti – šmor You 2.m.sg guard on the eggplant gender, number and person You will guard on the eggplants ‫ސ‬ Benoni forms: V agrees with the Ata šomer al ha-xacilim subject in gender and number You guardm.sg on the eggplant You guard/are guarding on the eggplants Agreement patterns in Arabic l-ta ‫ݧ‬aam V is clause-initial: Agreement with ‫ސ‬akal – a l- ‫ސ‬awlaad Atem the boym.pl the food the subject only in gender The boys ate the food Preverbal subject: Full agreement Qult – u ‫ސ‬inna l- ‫ސ‬awlaad ‫ސ‬akal –ul ta ‫ݧ‬aam (person, number and gender) Said-1.sg that the boym.pl ate-3.m.pl the food I said that the boys ate the food

These observations on s-v agreement led to the assumption that agreement features are dependent one upon the other: gender can be represented without number and person (as in V-initial clauses in Arabic), and gender and number can be represented without person. There is no verb that is marked for number and not for gender, and no verb that is

Chapter One

8

marked for person and not for number. This implies the hierarchy represented in (6) (Shlonsky 1989)2: (6) Implicational Hierarchy of Agreement Features: a) If a verb is inflected for number, then it is also inflected for gender; b) If a verb is inflected for person, then it is also inflected for number.

According to Shlonsky, the implicational hierarchy above captures the peculiarities of s-v agreement in Hebrew and Arabic, but it can easily be extended to Slavic and Romance languages, as shown in (7) below for French (from Shlonsky 1989), where the past participle repeintes agrees in gender and number with tables. (7) Je sais I know

combien how many

de tables ils ont repeintes tablesf.pl they have re-paintedf.pl

Contrary to what is assumed by standard minimalist assumptions, within a system of functional projections such as the one postulated by Shlonsky (1989), access to the morphosyntactic features involved in s-v agreement cannot be performed by means of a unique operation. Instead, more operations are necessary, namely one for each morphosyntactic feature that is projected in the structure and that needs to be morphologically realised. This assumption also underlies a recent analysis of Icelandic s-v agreement by Sigurdsson (2004), who proposes that complex functional heads like Infl and v need to be decomposed into different functional categories, with each of them representing a feature, as in (8): (8) Infl: Perss, Numbs, M(ood)s, T (ense)s v: Perso, Numbo, Asp(ect)o, v [CP… [IP Perss, Numbs, M, T , Perso, Numbo, Aspo [vP ….v…..]]]

Besides the Inflectional layer of the sentence, another area is important for s-v agreement, namely the Complementiser Phrase (CP) area, where Sigurdsson (2004, 13) locates the so-called logophoric Agent and logophoric Patient (Ȝa and Ȝp) corresponding to the speaker and the hearer of the speech event, i.e. 1st and 2nd person, as shown in (9): (9) [CP Force..ȜA ȜP..Top..ST..SL[IP..PersS..NumS..M..T..[vP…]]]

The Linguistics of Agreement

9

The relationship between logophoric participants and agreement in the CP area of the sentence will be addressed more in depth later in this chapter, as this will turn out to be of fundamental importance for the analysis of s-v agreement that will be outlined here. Support for the person-number distinction put forth by Sigurdsson comes from the analysis of subject clitics (SCLs) in Northern Italian dialects (NIDs) carried out by Poletto (2000). In her thorough analysis of the distribution of SCLs, Poletto (2000) identifies two agreement domains, corresponding to two different structural layers of the sentence: the area preceding the negation (NegP) and the area following it, which respectively correspond to the CP and the IP layers of clausal structure, as outlined in Rizzi (1997). These two domains are further divided into subdomains, on the basis of the type of clitic involved: invariable (inv SCL), deictic (deic SCL), person (hearerP SCL; speakerP SCL) and number clitic (NumbP SCL), as the configuration in (10) illustrates: (10) [LDP invSCLi [CP deicSCL [FP ti [IP[NegP[NumbP SCL [hearerP SCL [SpeakerP Inflv [TP]]]]]]

Apart from invariable SCLs, which do not vary for any of the six persons of the paradigm, all the other SCLs encode some subject features, without the same feature being repeated twice. Importantly, the subject features realised in the pre-negative field are different from those in the post-negative field. On the one side, deictic clitics encode a [± deictic] feature that distinguishes speech participants from non-speech participants. It should not be surprising that these types of clitics are syntactically represented in the CP area of the sentence, exactly where Sigurdsson (2004) encodes speech act-related features. On the other side, number and person SCLs express person, number and gender features: they are merged in the IP area of the sentence, where the morphosyntactic realisation of agreement is dealt with. Interestingly, person information is split into two positions: hearerP SCL and speaker InflV. Neither of these expresses the distinction between singular and plural number: hearerP SCLs express the [±hearer] distinction but do not distinguish between singular and plural, nor does the speakerP SCL. An important fact concerns 1st person singular and plural, which are not realised by unambiguous clitics, but are either expressed by the same deictic clitic that also expresses 2nd person singular and plural, or by a SCL that agglutinates to the verb, which has moved to SpeakerP, as shown in (10). This split strongly suggests that the morphological concept of person does not correspond to any single functional projection where all six persons (or even more in certain languages) are mapped: the syntactic component seems to take into

Chapter One

10

account more basic distinctions, such as the deictic distinction between participants and non-participants in discourse. This will turn out to be important when addressing the issue of the Person Asymmetry Hypothesis in Chapter 4. A last remark concerns gender information, which is parasitically represented in the NumberP SCLs, as postulated also in Di Domenico (1997) and Ritter’s (1993) analysis of grammatical gender (see Chapter 4 for further details). In a more recent proposal, on the basis of Icelandic dative-nominative (DAT-NOM) constructions, Sigurdsson and Holmberg (2008) analyse Person (Pn) and Number (Nr) as distinct probing phenomena, as illustrated in (11): (11) [CP …Top…Fin [TP …Pn…Nr…T…v…DAT…NOM]

Nr and Pn probing are activated by T-raising. T cannot probe for DP number/person unless it has joined the Number (Nr) and Person (Pn) projections separately. Also, Nr and Pn probing must take place immediately after T-raising to Nr and T/Nr raising to Pn, as exemplified in (12) and in (13): (12) þad þótti/þóttu einum malfrædingi þessi rök sterk linguistdat these argumentnom strong Expl thought3.sg/3.pl one (13) Expl Expl Expl Expl

Pn Pn DAT Pn DAT T/Nr/Pn

Nr T [vP DAT V [TP NOM…. Nr T [vP DAT V [TP NOM… T/Nr T [vP DAT V [TP NOM…. DAT T/Nr T [vP DAT V [TP NOM….

The roll-up type of T-movement exemplified in (13) yields the order of tense, number and person markers in Icelandic morphology shown in (14): (14) lœrdum = learn-PAST-PL-1P We learned

The order of affixes in (13) recalls the one in (4) and Baker’s Mirror Principle (Baker 1985): the tense affix is closer to the verb root than the Agr one, suggesting the lower position in the syntactic tree of the former with respect to the latter. Similarly, person may be seen as occupying a higher position than number. The structural distinction between the Person and the Number projections that has just been outlined can be strictly connected to the different interpretive properties associated with the two features, which

The Linguistics of Agreement

11

may have an important role in the operations involved in their licensing and interpretation. Let us now discuss person and number interpretive properties.

Features, anchors and interpretation It is known that the information conveyed by number and person is intrinsically different. While the former feature expresses the mere numerosity of the subject argument, the latter refers to the subject’s role with respect to the participants in the speech act, i.e. speaker and addressee. As Jakobson (1971) observes, “Person characterises the participants of the narrated event with reference to the participants of the speech event.” A deictic component is therefore present in person information that may crucially shape the way this feature is licensed and interpreted. Recent theoretical analyses have indeed emphasised the fact that this feature can be interpreted only in relation to speech act participants (Bianchi 2003, 2006; Sigurdsson 2004; Schlenker 2004). For instance, a 1st person value expresses identity with (or inclusion of) the speaker, while a 2nd person value expresses identity with (or inclusion of) the addressee. Third person indicates exclusion of both speaker and addressee and refers to the entity that is being talked about.3 In this view, speaker and addressee are individuals participating in the speech event, which recalls the Kaplanian representation of context as an index made of several coordinates that directly refer to the actual world of utterance, namely its time, location and participants (Kaplan 1989). The link existing between person specifications and the speech act has been explicitly implemented in recent cartographic analyses of agreement and agreement features that posit a syntactic encoding of speech act and participants in the left periphery of the sentence. Bianchi (2003, 2006) draws a parallel between person agreement and tense marking and identifies the anchoring point for both features in the so-called Logophoric Centre (LC), which constitutes the centre of deixis and hence corresponds to the speech event, with its spatial, temporal and participant coordinates. In structural terms, Bianchi’s LC resides in Fin (following Rizzi’s 1997 approach), the head encoding information concerning the finite or nonfinite nature of a clause. It is the LC that licenses fully-fledged person agreement and absolute tense, by establishing a link – or anchoring – between the IP layer of the sentence (where “morphosyntactic” person is expressed) and the left periphery of the sentence. In a similar fashion, such a link has been captured by Sigurdsson (2004) in terms of a matching relation among features. What characterises

Chapter One

12

person with respect to number is the matching established between clauseinternal positions – the IP system – and the participants in the speech act expressed in the CP system, as in (10), here repeated as (15): (15) [CPForce..ȜA..ȜP..Top..ST..SL[IP..PersS..NumS..M..T..[v…]]]

More precisely, Sigurdsson (2004, 27) subdivides clause structure into three layers, each of them encoding specific features: the lexical layer in the vP shell, with event features; the inflectional layer in the IP area, with grammatical features; and the speech event layer within the CP area, with speech act features (speech participant, speech time and speech location features). For instance, a matching relation ties lexical to grammatical features, and grammatical features to speech act ones. This way, an Agent can be linked to a 1st person pronoun or 1st person verbal morphology, and consequently to a speaker role (a Logophoric Agent, to say it with Sigurdsson), as shown in (16). This would lead to proper interpretation of person. (16) Ĭ = +Person = + ȜA - ȜP Ĭ = +Person = - ȜA + ȜP Ĭ = +Person = - ȜA - ȜP

1st person by computation 2nd person by computation 3rd person by computation

To sum up, for person to be interpreted, matching must necessarily involve speech participants features in the left periphery of the sentence. No such IP-left periphery connection is necessary for number, whose interpretation is independent of the speech role played by the subject argument. A fundamental difference in interpretive requirements therefore lies at the heart of the distinction between person and number. To clarify this point, the notion of “interpretive anchor” will be introduced and used throughout. Let us better define this concept. A tight connection – anchoring – exists between structure and interpretation: to receive a proper interpretation, each morphosyntactic feature entering a derivation activates its “anchor”, a specific feature in the semantic representation of the sentence. The term sigma value will be used throughout to refer to the semantic-discourse value of a feature, i.e. its anchor, as opposed to the phi value, which refers to the morphosyntactic realisation of a feature (Table 1-2, see D’Alessandro 2004 for a similar approach).

The Linguistics of Agreement

13

Table 1-2. Phi and sigma values associated with person and number features Person Phi [+1, -2] [-1, +2] [-1, -2] Number Phi [+sg, -pl] [-sg, +pl]

Sigma [+SPEAKER, -ADDRESSEE] [-SPEAKER, +ADDRESSEE] [-SPEAKER, -ADDRESSEE]

1st person 2nd person 3rd person

Sigma [+ONE, -GROUP/MANY] [-ONE, +GROUP/MANY]

singular plural

The link to the anchor will be activated every time the morphosyntactic feature is involved in operations entailing its licensing and interpretation, as happens when Agree is performed. In the case of number, its interpretive anchor will be represented by the number specification on the subject argument, thus involving no link outside of the specifier-head configuration within which subject-verb agreement occurs. In essence, this amounts to saying that number is interpretable on the subject, in accordance with standard minimalist assumptions . A different interpretive anchor is instead identified for person. In this case, interpretation is made possible by the link activated between clauseinternal positions (the specifier and head positions of IP) and the speech act representation, where speech participant features are encoded. In other words, person’s interpretive anchor resides in the speech act representation. The idea of a link established between the morphosyntactic specification of the person feature in the inflectional area of the sentence and its anchor is not to be taken as an instance of Multiple Agree, as has been postulated to account for phenomena like the Person Case Constraint (Anagnostopoulou 2003, see Chapter 4). Anchoring is here intended as a link established between different structural positions. In light of this, the interpretability status of nominal and verbal person needs to be re-examined: if neither the former nor the latter represent person’s interpretive anchor, they may be seen as two autonomous values, whose interpretation requires separate anchoring to the speech act representation. This obviously marks a fundamental point of departure from standard minimalist analyses of agreement and agreement features, which indicate the subject argument as the locus of person interpretation. The mechanism driving person and number interpretation can be stated as in (17):

Chapter One

14

(17) FEATURE INTERPRETATION PROCEDURE (FIP) Features are structurally differentiated and interpreted in relation to their anchor: a) Person’s anchor resides in the speech act representation; b) Number’s anchor resides in the number specification of the nominal argument.

By introducing the notion of “interpretive anchor”, an approach to person and number agreement has been sketched that seems to accurately account for the intrinsic differences underlying the two features. In the following chapters, it will be shown that the presence of distinct interpretive anchors for person and number can explain behavioural and neuro-physiological correlates of agreement processing.

Summary Linguistic analyses differ in the formal details with which agreement mechanisms are described. Early derivational grammars defined agreement as an asymmetric relation between a controller and a target, with the controller (also called trigger) being the element from which grammatical information originates, and the target the element that inherits the information. Such controller-target asymmetry is central to featurecopying models of agreement, like the one developed within the Minimalist Program (Chomsky, 1995, 2000, 2001, 2005). In essence, minimalist agreement hinges on three basic assumptions: (i)

Feature syncretism: features are expressed as a feature bundle on a single position in the syntactic tree (Tense, or T), and are uniformly dealt with by the syntactic operation of Agree (see Figure 1-1); (ii) Asymmetry: agreement proceeds asymmetrically from the controller to the target. For instance, in s-v agreement, the person and number features expressed on the subject DP are copied onto the verb by the formal operation Agree. Features are valued and interpretable on the nominal argument, hence they are visible to the interpretive system, while they are uninterpretable on the verb, as mere formal copies of the nominal specifications. Agree connects the two positions, and checks and values the features on the verb. (iii) A narrowly syntactic operation: Agree operates within the domain of Narrow Syntax, as uninterpretable features need to be

The Linguistics of Agreement

15

erased from the derivation before these are transferred to the interpretive system. Following the “one morphosyntactic property – one feature – one head” (Cinque and Rizzi 2008) principle, Cartography proposes a distinctcluster analysis of agreement in which minimalist generative devices such as Agree operate on individual features, rather than on bundles. Inspired by Cartography, an approach to s-v agreement computation has been proposed that, while relying on computational devices like the Agree operation, departs significantly from standard minimalist assumptions, in that it posits a distinct-cluster representation of agreement features. The presence of different interpretive requirements for person and number is the key point on which the FIP hinges. In the next chapters, the FIP will be tested in two languages – Spanish and Italian – and in different agreement contexts with the goal of assessing its validity in predicting processing correlates.

CHAPTER TWO SENTENCE AND AGREEMENT COMPREHENSION

Language can be studied from many different perspectives. Chapter 1 has given an example of how linguists attempt to uncover the structure and the computations supporting abstract linguistic knowledge. Understanding how this knowledge is used in real-time comprehension and production, and what the behavioural and neuro-physiological correlates are, belongs to the domain of psycho- and neuro-linguistics. The goal of this chapter is to provide the reader with fundamental tools to test theoretically relevant questions by using some of the experimental paradigms available nowadays for the study of language comprehension. A brief introduction to the experimental techniques that underlie the empirical data presented in the next chapters is provided, spanning from behavioural (online judgements, self-paced reading and eye-tracking) to sophisticated neuroimaging paradigms such as electro-encephalography (EEG) and functional magnetic resonance imaging (fMRI). Because of the relevance of eyetracking, EEG and fMRI paradigms for the understanding of the time course and mechanisms supporting agreement processing, a review of previous agreement-related studies employing these techniques will be offered. This will be followed by a review of mainstream neurocognitive models of sentence processing.

The online study of sentence comprehension The study of language processing can count on the availability of a growing number of sophisticated experimental techniques that have helped us obtain fundamental insights into the mechanisms (how), the spatial (where) and temporal dimension (when) of sentence and agreement comprehension. As will be clear from the following description (and the empirical studies presented in the next chapters), there is no single perfect technique for the study of language comprehension. Rather, it is from

Sentence and Agreement Comprehension

17

investigations involving multiple designs and paradigms that one can get the clearest insights into how, when and where different sources of linguistic information are comprehended.

Behavioural paradigms: online judgements, self-paced reading and eye-tracking Methods measuring participants’ behavioural responses are fundamental in the study of language. Because they represent the “output” of the processing system, response/reading times and accuracy provide the link between a research hypothesis and the more implicit measures derived from the observation of neuro-physiological correlates. Three types of behavioural paradigms are presented below: online and offline judgements, eye-tracking and self-paced reading, each of which offers different insights into online linguistic behaviour.

Online and offline judgements Typical judgement tasks require participants to evaluate the acceptability of a sentence on grammatical and/or semantic/pragmatic bases. They can be administered in the form of offline (timed paper-andpencil or computer-based) questionnaires, thus disregarding the time course of linguistic processing, or online (timed computer-based) tasks, i.e. by imposing a certain amount of time pressure on the participant, who is required to give her/his answer within a handful of seconds. In untimed judgements, participants usually see each sentence as a whole, while in timed versions, stimuli can be presented word-by-word or chunk-bychunk, at a rate set by the experimenter. The material can be also presented auditorily via headphones, in which case no segmentation of speech is carried out. The advantage of online judgements resides in the fact that, besides information concerning the accuracy of the response, the researcher also obtains information concerning the time taken by the participant to express her/his judgement. The type of response available in a judgement task can be binary, i.e. speakers are instructed to indicate whether they judge the stimulus as acceptable or as unacceptable, or can be based on a discrete scale in which speakers indicate e.g. how natural/acceptable they find a sentence from e.g. 1 (unnatural/unacceptable) to 5 (completely natural/acceptable), with the magnitude of the scale being chosen by the researcher. The choice of binary responses or discrete scales depends on the researcher’s objectives. When subtle differences in acceptability between two (or more) types of

18

Chapter Two

stimuli are hypothesised, the wider range of options allowed by a discrete scale may be a better tool to capture them than binary correct/incorrect responses. Nevertheless, participants should always be instructed to use the full range of values provided, and not to concentrate only on the most extreme or on the intermediate points. Whether in an online or offline judgement, the list of materials presented to participants comprises the critical stimuli, i.e. those involving the manipulation of interest, as well as a number of filler items with a different and usually unrelated manipulation, so that participants are kept as naïve as possible with respect to the purpose of the experiment and the critical manipulation (this recommendation applies to psycholinguistic tasks in general). Judgements on the acceptability/grammaticality of sentences have been the first and most used method in linguistic research, as they provide the empirical basis for linguistic theories and for more sophisticated behavioural and neuro-imaging investigations (for an extensive review, see Schütze 2016). In the study of sentence and agreement comprehension, judgements are typically used as an exploratory test to verify the sensitivity of speakers to a specific manipulation (see Chapter 5), before using greater time- and resource-consuming paradigms such as EEG and fMRI. In addition, they are frequently used as ancillary tasks in neuroimaging paradigms, so that the electro- and neuro-physiological responses can be more easily linked to specific behavioural output.

Eye-tracking Readers go through a text with rapid movements – saccades – that bring peripheral visual input onto the central part of the eye retina, the fovea. Between saccades, fixations occur, during which the eyes remain relatively still on a word for about 200–300 milliseconds. The average duration of a saccade during reading is about 30–50 milliseconds, while the time it takes to plan and execute a saccade (latency) is about 150–175 milliseconds (although both latency and duration can change depending on whether reading is silent or oral; see Rayner 1998 and Rayner and Castelhano 2007 for an overview), which suggests that saccade programming could be done in parallel with comprehension processes in reading (Rayner 1998). Eye-tracking during reading has made a substantial contribution to the study of language processing in the last fifty years. Thanks to its exceptional temporal resolution and the possibility that it offers to segment the reading process into distinct components (fixations, re-fixations,

Sentence and Agreement Comprehension

19

saccades), this technique can provide detailed information about momentto-moment cognitive processes with millisecond precision. The recording of eye movements has been successfully applied to the study of many different aspects of language, from orthographic and phonological processing to syllables and morphology, as well as syntactic and morphosyntactic effects (see Clifton et al. 2007 for a review of wordand sentence-level studies). The analysis of a typical reading experiment involves the segmentation of each sentence into different interest areas, the reading times of which are then fractionated into several dependent variables that involve both latency and regressive measures. An overview of eye-tracking measures is provided in Table 2-1 below. Table 2-1. Eye-movement measures typically used in sentence processing studies (see text for further details). Variable First fixation duration First pass/gaze duration Go-past time/regressionpath duration Second pass duration Total reading time Regression probability

Description Duration of first fixation on a word Duration of the first round of fixations through an interest area The time from the first fixation in an interest area until the region is left with a forward saccade to the right (including re-reading of earlier regions) The time spent re-reading a word/area The sum of all fixations on an interest area The probability of making a regression into (to the right) or outside (to the left) of an interest area

Each reading measure is functionally associated with specific processing mechanisms. It is commonly assumed that the processes related to the online integration of a word into its immediate local syntactic context mainly affects the first pass of fixations through a region (also called gaze duration): the more difficult it is to integrate a word, the greater the number of fixations required, hence the longer the time spent on that region. The time taken to regress leftward before moving past the disruption point – the so-called go-past duration – is usually taken to reflect the cost of overcoming processing difficulties, which eventually results in the integration of the critical word in its preceding syntactic context. The time spent re-reading the critical region – the second-pass reading time – is usually associated with reanalysis processes. The overall time spent on a specific region constitutes the total reading time. A typical reading experiment also involves the analysis of the probability and number regressions into and out of an interest area. These measures are usually taken to index reanalysis processes after difficulties in syntactic

20

Chapter Two

integration (see reviews on eye movements measures in Clifton et al. 2007; Clifton and Staub 2011; Rayner and Pollatsek 1989; Reichle et al. 2009; Rayner and Liversedge 2011). A distinctive feature of the eyetracking methodology is the multidimensional view of language processing that it offers. Its dependent variables can be differentiated into early measures (i.e. occurring during the first round of fixations) and late measures (which involve regressions and re-fixations of an interest area). As such, this technique can be particularly useful for the analysis of the time course of sentence processing (see Bornkessel-Schlesewsky et al. 2016 for a discussion on the relationship between eye-tracking and ERP data as far as the time course of sentence processing is concerned). In the following, a brief review of eye-tracking studies on agreement processing is offered, which provides the background for the experiments presented in Chapters 3 and 5.

Eye-tracking and agreement processing Eye movements can provide valuable insights into the mechanisms and strategies adopted to detect and analyse a syntactic dependency. Existing work on agreement carried out with the eye-tracking paradigm report a heterogeneous scenario, with some studies pointing to the parser’s early sensitivity to agreement relations, while others suggest late analysis. However, substantial methodological differences can be found across studies, which partially explains the variety of findings. Early sensitivity to agreement information has been reported in Hebrew by Deutsch (1998) and Deutsch and Bentin (2001), who tracked eye movements during the reading of sentences containing gender mismatches between subject and predicate in Hebrew. The reading of such anomalies produced sizeable consequences in the first pass and in the total time spent in the verb region. Kreiner et al. (2013, Experiment 2) report sensitivity to the presence of morphosyntactic mismatch arising as early as in go-past reading times. In this study, the authors tried to temporally dissociate the grammatical and conceptual processing of agreement by contrasting collective and non-collective nouns. Results showed that collective readings (i.e. “the family wish”) generated go-past effects that were equivalent to those elicited by ungrammatical singular noun-plural verb sentences (“*the widow wish”), while only the effect of unagrammatical constructions persisted in the second pass. According to the authors, this effect evidenced that morphosyntactic factors control the initial computation of agreement, while notional information drives later processing.

Sentence and Agreement Comprehension

21

The findings described above contrast with those by Pearlmutter and colleagues (Pearlmutter et al. 1999), who do not report effects of agreement attraction and violations in early reading variables, but only in the total reading times of the spillover position, suggesting the parser’s sensitivity to agreement manipulations only upon re-reading of a specific region.

Self-paced reading Like eye-tracking, self-paced reading provides a measure of the time taken to read each word or chunk in a sentence, to assess whether a specific manipulation generates an increase in reading times. It represents a convenient method to investigate linguistic processing, as it only requires a computer and presentation software, which is partly why this paradigm has stood the test of time and is still being widely used in psycholinguistic research. In a typical self-paced reading paradigm, participants see sentences one word at a time or chunk-by-chunk on a computer screen. The participant controls the rate of presentation of the consecutive words/chunks by pressing a button, which allows them to read stimuli at their own pace (hence the name “self-paced reading”). Different versions are available that make it possible to vary how words/chunks are displayed on screen. A moving-window version of the paradigm is available in which words/chunks are progressively added from left to right and occupy a position that roughly correspond to that in the sentence displayed as a whole. In the cumulative version, words/chunks are presented from left to right and remain on screen as others are progressively added, while in the non-cumulative version, old displays are removed as new ones are added. A potential problem with the cumulative version of the paradigm is the fact that participants may show the tendency to press the key in rapid succession so that they can read the sentence once it is displayed rather than one word/chunk at a time. This problem can be partially circumvented by using the non-cumulative version of the task. Finally, in the stationary-window version, each word/chunk that is added overprints the previous one in the same location of the screen. In all of these versions of the task, there is some tendency for spillover effects, i.e. increments in processing times that do not appear immediately at the critical word but in the following display. Another problematic feature of this paradigm is represented by the artificial segmentation of the text, which slows down reading and makes the whole process very far from being ecologic.

22

Chapter Two

Agreement-related studies that have employed the self-paced reading technique in some of its versions have provided substantial evidence for readers’ online sensitivity to attraction phenomena and agreement anomalies (see Wagers et al. 2009 and references cited therein), as well as for the different disambiguating power that gender, number and person features have in the processing of anaphoric relations between a null pronoun and its antecedent (Carminati 2005, see Chapter 3). Yet, because of the unidimensional nature of self-paced reading, which only allows for the recording of reading times at the disruption point, the use of this technique cannot contribute substantial information concerning the time course and the mechanisms underlying agreement processing.

Electrophysiology of language: Electroencephalography (EEG) and Event-Related Potentials (ERPs) ERPs are electrophysiological responses induced by specific stimuli that are embedded in the spontaneous electrical activity of the brain, or encephalogram (EEG, a description of the physiological bases of the EEG/ERP signal, is provided in Box 2-1; see also Luck 2014). Thanks to their exquisite temporal resolution, ERPs have proved a reliable and useful technique to study the online comprehension of language and its time course (a comprehensive review is provided by Swaab et al. 2012). Box 2-1. Physiological bases of the EEG/ERP signal Physiological bases of the EEG/ERP signal

EEG relies on the measurement of electric field variations elicited by neural activity. Neural activity can be divided into two categories: action potentials (APs) and post-synaptic potentials (PSPs). APs are bursts of ion flux through which the brain receives, analyses and conveys information. They correspond to rapid, transient, all-or-none nerve impulses that originate at the initial segment of an axon and propagate to the axon terminal, where neurotransmitters are released, at rates of 1–100 m per second (see Hudspeth et al. 2013 for a detailed description). When the neurotransmitter binds to the receptors on the membrane of the post-synaptic cell, a PSP arises, causing instantaneous bursts of ion exchanges that modify the cell membrane potential. The electrical and magnetic activity produced by PSPs derives from the simultaneous activation of pyramidal cells. These are neurons with a pyramid-shaped cell body that are mainly found in layer V of the cerebral cortex, the hippocampus and the amygdala. Because they are aligned and positioned perpendicularly to the surface of the cortex, the electrical and magnetic

Sentence and Agreement Comprehension

23

activity produced by pyramidal cells will sum together, and the resulting voltage can be measured at the scalp. Electricity spreads out through the conductor and tends to propagate laterally to avoid the high resistance of the skull. As a result, electrical potentials generated in one region of the brain can lead to substantial voltages at quite distant parts of the scalp (Luck 2014) and can therefore be picked up by electrodes located all over the head. In its raw form, EEG represents a conglomeration of hundreds of different neural sources of activity, embedded in which are event-related potentials (ERPs). ERPs are electrophysiological responses induced by specific motor, cognitive, affective or sensory stimuli that are administered during an experiment. As ERPs can be very small (e.g. between 2 and 8 μV for language) compared to those for spontaneous activity (10–100 μV), specific averaging procedures, alongside with artefact rejection or correction, must be used to extract relevant stimulusinduced information from background activity (see Handy 2005; Luck 2014 for an overview of these procedures). Four parameters are used to classify ERP activity: latency, polarity, topography and amplitude. Latency refers to the time point, relative to the presentation of a critical stimulus, at which a change in the electrical potential can be observed. It is measured in milliseconds and usually forms part of the name of an ERP component (e.g. N400, where 400 refers to the latency of the component). Polarity indicates whether the voltage change is positive or negative (indicated with “P” and “N” in a component name, respectively). Topography refers to the scalp distribution of a component and is usually defined in longitudinal (e.g. anterior or posterior) and hemispheric (right or left) terms. Finally, amplitude denotes the strength of an effect and reflects a quantitative change in the neural activity of a source neural population. Several ERP components have been identified that relate to the analysis of syntactic, semantic and contextual aspects of the linguistic input as it unfolds over time. The best-studied ERP component is perhaps the N400, a negative shift arising about 400 milliseconds post-stimulus onset that is larger over centro-parietal regions of the scalp, usually with slightly larger amplitude in the right than in the left hemisphere. It was first reported by Kutas and Hillyard (1980) in response to violations of semantic expectancy during the reading of implausible words like “socks” in “He spread the warm bread with socks” compared to a semantically plausible sentence. The N400 forms part of the typical brain electrical activity in response to a large array of stimuli: visual and auditory words,

24

Chapter Two

acronyms, sign language signs, pictures, environmental sounds and gestures (see Kutas and Federmeier, 2009 for an extensive review). Different factors can modulate the amplitude of the N400, among which are lexical frequency (i.e. low-frequency words elicit a larger N400 compared to high-frequency words), repetition (repeated words show a smaller N400 amplitude compared to non-repeated ones) and the presence of a supportive sentential context (i.e. the more supportive the context, the smaller the amplitude of the N400). The modality of stimulus presentation can influence the topographical distribution of the N400. Visually presented words show a widespread scalp distribution with a medial centro-posterior focus, often with a slight right lateralisation (Kutas and Hillyard 1982). The N400 response to auditory words manifests a more central scalp distribution, while N400 responses to pictures and scenes are more anterior than those to visually presented words. Another language-related ERP component is the P600, a long-lasting positive voltage shift arising about 600 milliseconds after critical stimulus presentation (but see Mecklinger et al. 1995 for earlier positive shifts). The P600 usually presents a widespread centro-parietal topography, although more anterior distributions have also been reported (see discussion below). Both auditory and visual stimuli can induce P600 effects. A consensus is still missing as far as the functional interpretation of the P600 is concerned. P600 effects were first observed in response to sentences that presented syntactic violations (Osterhout and Holcomb 1992), as well as to complex and ambiguous sentences (Friederici et al. 2002; Kaan and Swaab 2003). Because of this, the P600 was initially considered an electrophysiological signature of syntactic reanalysis and repair operations. However, this general interpretation has been called into question by studies that report P600 effects in response to violations of thematic constraints, such as in “The hearty meal was devouring the kids” (relative to “The hearty meal was devoured by the kids”, Kim and Osterhout 2005), in which the agentive function assigned to “the hearty meal” by the verb is at odds with its inanimate status (see also Kuperberg et al. 2003). This led many authors to interpret the P600 as a late stage of reanalysis of an anomalous sentence after both the syntactic and semantic analyses have been completed (Friederici and Alter 2004; Kuperberg 2007). In other words, after all the available information has been processed, a reanalysis could be triggered in order to get a coherent interpretation of the material. Interestingly, according to some authors (Barber and Carreiras 2005; Carreiras et al. 2004; Kaan and Swaab 2003; Molinaro et al. 2008), this

Sentence and Agreement Comprehension

25

reanalysis would be pursued in two subsequent stages. The first stage would correspond to an increased positivity with a broad scalp distribution around 600 ms. This stage would represent the initial integration of all the information concerning the target word with the information concerning the previous sentence fragment that is stored in working memory, in order to detect the source of the incongruence (diagnosis stage). After a correct diagnosis of the anomaly, a repair phase follows (Barber and Carreiras 2005; Carreiras et al. 2004; Friederici et al. 2002; Hagoort and Brown 2000; Molinaro et al. 2008) that surfaces with a prevalently posterior distribution of the effect. Findings from the processing of agreement feature violations (gender vs. number violations: Barber and Carreiras 2005; gender vs. phonotactic violations: Molinaro et al. 2008) have suggested that modulations of the amplitude of the late P600 may be revealing of different repair processes going on depending on the feature that is being violated. Modulations of P600 distribution, especially broad/frontal positivities, have long been discussed in the ERP literature on sentence comprehension. The functional interpretation of this response is not yet clear and agreed upon, and a number of independent proposals have been advanced in the ERP literature. According to Hagoort et al. (1999), frontally-distributed positive effects reflect costs associated with overwriting the preferred, most active structural representation of the sentence, as is the case with non-preferred continuations, which necessitate a revision, while posteriorlyfocused P600 effects are associated with a collapse (and hence a repair) of the structural representation, as happens with ungrammatical continuations. An alternative interpretation has been put forth by Friederici et al. (2002), who capitalises on the fact that frontal P600 effects are not limited to revision processes, as in the case of non-preferred continuations, since they can also arise in the presence of increasing syntactic complexity that does not necessarily involve revision of the previous sentence fragment. Finally, a third view is the one put forth by Kaan and Swaab (2003), who reported frontally-distributed P600 effects due to a large number of referents to be integrated in the same discourse representation. To test whether frontal P600 effects resulted from the processing of ambiguous material that necessitated revision operations, or from increasing syntactic complexity, Kaan and Swaab (2003) compared the processing of sentences that contained one or two NPs before a relative clause: the former represented the unambiguous condition, and it was presented in a grammatical or ungrammatical form (e.g. The man in the restaurant does not like the hamburger(s) that are on his plate); the latter represented the ambiguous condition and it was presented with a preferred (e.g. I cut the

26

Chapter Two

cake beside the pizzas that were brought by Jill), non-preferred (I cut the cakes beside the pizza that were brought by Jill) or ungrammatical (I cut the cake beside the pizza that were brought by Jill) continuation. The frontal P600 effect was found to be insensitive to the type of continuation that followed ambiguous sentences, suggesting that revision processes could not be the sole cause of frontal P600 effects. Rather, the frontal component could be related to the difference in complexity preceding the critical verb: the two-NP condition may have been more taxing than the one-NP condition, both from a syntactic and a discourse point of view. The ambiguity of the relative clause between a high or low attachment makes the two-NP condition syntactically more complex than the one-NP condition. At the same time, the presence of two NPs increased the complexity of the sentence at a discourse level, as more referents had to be integrated in the discourse representation compared to the one-NP condition (see also Garrod and Sanford 1994). Kaan and Swaab’s (2003) interpretation is only partially confirmed by a study in Spanish conducted by Carreiras et al. (2004), who found a frontally-distributed P600 effect in Spanish for relative clauses that were ambiguous between a low or a high attachment resolution. In particular, the frontal P600 effect emerged in the presence of a non-preferred continuation, i.e. the low attachment resolution (Juan felicitó a la cocinera del alcalde que fue premiada y laureada en las fiestas / Juan congratulated the cookfem of the mayormasc, who was awarded a prize and honoured at the party), compared to a preferred continuation (Juan felicitó al cocinero de la alcaldesa que fue premiada y laureada en las fiestas / Juan congratulated the cookmasc of the mayorfem, who was awarded a prize and honoured at the party). Given the similar structural complexity of the two conditions, Carreiras and colleagues concluded that a frontal/broad distribution of the P600 may be unlikely to result from syntactic complexity. Rather, it may be the reflection of ambiguity and of the potential problems with such ambiguity. Finally, the discovery of P600 effects in the presence of non-linguistic stimuli such as musical violations (out-of-chord keys, Patel et al. 1998), arithmetic rules (Nuñez-Peña and Honrubia-Serrano 2004), and abstract sequences (Lelekov et al. 2000) suggests that the functional significance of this component extends beyond the scope of language. Emphasising the role of top-down and executive control mechanisms, recent proposals have advanced a functional interpretation of P600 effects in terms of error monitoring that can explain P600 effects across a wide range of domains. From this perspective, these effects reflect reanalysis after a strong conflict has been detected between the top-down expectations and the bottom-up

Sentence and Agreement Comprehension

27

analysis of the perceived input. The control process is not always active, but only becomes so when bottom-up information signals the presence of a conflict (van de Meerendonk et al. 2009; van de Meerendonk et al. 2011). In the following chapters, the perspective that will be adopted regarding the functional interpretation of the P600 is that of a domain-general component that, within a linguistic context, can nevertheless show distributional modulations depending on the type of linguistic information being manipulated. Various types of syntactic anomalies also elicit an early negative shift in the ERP waveform between approximately 300 and 500 milliseconds post-stimulus onset, which is evident in left anterior areas of the scalp (left anterior negativity, or LAN). These negative shifts have been widely reported across languages in the presence of morphosyntactic violations, such as agreement anomalies (Barber and Carreiras 2005; De Vincenzi et al. 2003; Gunter et al. 1997, 2000; Mancini et al. 2011a; Osterhout and Mobley 1995; Rossi et al. 2005 inter alia), although significant crosslinguistic variability has been reported in the emergence of this effect. As noted (Friederici 2011; Friederici and Weissenborn 2007), languages consistently showing this early negative effect are typically richlyinflected language (Spanish, Italian, German), while in poor-morphology languages this effect is less consistent (e.g. English). LAN effects have also been observed in response to fully grammatical sentences that present long-distance dependencies, such as filler-gap constructions (e.g. “Which cake did you choose?”, where cake has been moved from its original object position, see Felser et al. 2003; Kluender and Kutas 1993). In these cases, LAN effects have been associated with increased working memory costs. 1 LAN effects have been reported as early as between 100–300 milliseconds (early LAN or eLAN) after critical word presentation, in response to grammatical category violations (“The goose was *in the fed” relative to “The goose was fed”, Hahne and Friederici 1999, 2002; but see also Friederici et al. 1993, 1996). Whether eLAN and LAN effects reflect two different components indexing functionally distinct language processes, or whether they reflect a single process that varies in latency is a matter of debate (Swaab et al. 2012).

Chapter Two

28

Table 2-2. Overview of ERP components and their functional interpretation. Component eLAN

LAN

N400

P600

Topography and timing Left anterior negativity arising between 150 and 200 ms Left anterior negativity arising between 300 and 500 ms Centro-parietal negativity arising between 300 and 500 ms Centro-parietal positivity peaking around 600 ms

Functional interpretation Word category and phrase structure analysis

Friederici (2002, 2011)

Morphosyntactic processing

Molinaro et al. (2011); Molinaro et al. (2015)

Lexical-semantic processing

Kutas and Hillyard (1980); Kutas and Federmeier (2000); Lau et al. (2008) Osterhout and Holcomb 1992; Hagoort and Brown 2000 Bornkessel and Schlesewsky (2006) Van den Meerendonk et al.(2009) Kaan and Swaab (2003) Hagoort et al. (1999)

Syntactic repair/reanalysis Generalised mapping Error monitoring

Broad/frontal P600

References

Discourse complexity Overwriting of preferred syntactuic interpretation Syntactic complexity

Friederici et al. (2002)

ERPs and agreement Together with constituent order and thematic role assignment, s-v agreement is one of the linguistic aspects that has received more attention in the ERP literature. Agreement has been typically investigated using the so-called violation paradigm, where the processing of number-anomalous sentences is compared to the processing of agreement-correct sentences. As noted above, cross-linguistically, the ERP components typically modulated by the discovery of an agreement violation are the LAN and the

Sentence and Agreement Comprehension

29

P600. A similar biphasic pattern has been reported for agreement features other than number, e.g. person (see Chapter 3), as well as gender and phonological features in within-constituent configurations. In some cases, fine-grained differences between features have emerged in the topography of the P600 effects elicited. In a study in Spanish (Barber and Carreiras 2005), gender and number anomalies were created between determiners and nouns (elmasc.sg/*lafemc.sg/*/losmasc.pl piano estaba viejo y desafinado, the piano was old and off-key) and between nouns and adjectives (el faromasc.sg es altomasc.sg/*altafem.sg/*altosmasc.pl, the lighthouse is high), varying the position of the violations within the sentence (clause-initially or in the middle of the clause). Independent of the position in the sentence, of the agreement configuration (determiner-noun or noun-adjective agreement) and of the type of feature being violated (gender or number), the pattern elicited was a LAN followed by a P600. Yet, the two features differed in the late segment of the P600, where gender violations elicited a larger effect. The authors interpreted this finding as evidence for the deeper reanalysis operations triggered by a gender compared to a number violation (Faussart et al. 1999, see Chapter 3 for a discussion). While repairing mismatching gender information compels the parser to reinspect the lexical representation of a noun, repairing a number violation requires re-accessing a more superficial level of analysis, namely its morphological representation, whereby different processing costs may arise. Analogous results were obtained in a study in Italian (Molinaro et al. 2008), where determiner-noun agreement was manipulated to create anomalies in gender (lomasc/*lafem sciallemasc, the shawl) and in the phonological constraints determining the use of one of the allomorphs (il/lo) available for the masculine singular article (lo/*il scialle, the shawl). In both cases, the anomalies produced a LAN followed by a P600, but the P600 for the phonotactic mismatch was larger than the one elicited by gender violations. In line with Barber and Carreiras (2005), Molinaro and colleagues concluded that the amplitude of the late phase of the P600 depends on the level of analysis to be re-inspected for repair, which in their study corresponded to the pre-lexical (for phonotactics) and lexical (for gender) stages of processing. In the next chapters, ERP experiments on s-v agreement will be presented that, besides investigating feature dissociations in more depth, will also introduce a variation in the design used to study agreement, namely the unagreement design. .

30

Chapter Two

Neuro-biology of language: functional Magnetic Resonance Imaging (fMRI) The last 20 years have seen an increasing interest in the study of the neural basis of language processing (see extensive review in Price 2012). The main contribution of fMRI to the study of language has been the possibility to ground abstract mechanisms of information processing in neural substrates of the brain (Posner and Raichle 1994; Gazzaniga 2000), thanks to the measurement of changes in blood oxygenation leveldependent responses (BOLD signal, see Box 2-2) during the listening or reading of linguistic stimuli. Although the application of fMRI to the field of the cognitive neuroscience of language represents an undisputed advance, the mapping between brain activation profiles and cognitive mechanisms is not without complications and certainly not straightforward (see Box 2-3 for some key concepts on the design and interpretation of the fMRI studies). Box 2-2. Neuro-physiological bases of fMRI

Neuro-physiological bases of fMRI fMRI is a specialised application of magnetic resonance imaging (MRI) for the non-invasive study of in-vivo brain functions. MRI takes advantage of the magnetic properties of tissues to obtain information about the structure of the brain. It captures the signals produced by protons, i.e. nuclei of hydrogen atoms in brain tissues, which respond to applied magnetic fields by emitting radio waves. Each proton rotates around its axis and acts as a small magnet. In an unperturbed tissue, protons spin around their axes and create individual magnetic fields with random directions. When a vertical external magnetic field is applied to a tissue, the protons align with it and generate a small vertical magnetic field. The application of a second magnetic field characterised by a radio frequency pulse causes the protons to start wobbling in phase (or synchronously) around their axes. This creates a magnetic field that changes in time and that generates an electric current that the MRI equipment can measure. When the radio frequency is turned off, protons relax and de-phase, i.e. begin to move out of phase with one another, progressively re-aligning with their original magnetic field. Functional MRI records changes related to tissue function in successive images. The signal obtained with fMRI is based on local changes in blood flow, oxygen levels and glucose consumption that relate to cellular activity in the brain (Raichle 2009). In particular, fMRI studies use BOLD signal as their dependent variable, which measures the

Sentence and Agreement Comprehension

31

changes in magnetic susceptibility of haemoglobin. The haemoglobin molecule has magnetic properties that depend on whether it is bound to oxygen or not. Oxygenated haemoglobin is diamagnetic, i.e. it weakly counteracts the local magnetic field, while deoxygenated haemoglobin is paramagnetic and can disturb the homogeneity of the magnetic field. When neurons are activated as a result of task performance, the supply of blood flow to the active regions increases. The supply of oxygenated haemoglobin to the active region is greater than local oxygen consumption, leading to a greater proportion of oxygenated over deoxygenated haemoglobin, and thus to a signal increase. Temporal resolution in fMRI is limited by the hemodynamic response time. The BOLD lasts about 12-15 seconds and peaks around 5-6 seconds after stimulus onset. This makes fMRI a poor tool to capture the time course of cognitive processes. Yet, its exquisite spatial resolution can give invaluable insights into the brain areas actively involved during a language task, which lesion studies alone could not do. Classical lesion-deficit correlations associated language comprehension mainly with Wernicke’s area in the left posterior superior temporal gyrus (STG), while Broca’s area in the posterior portion of the left inferior frontal gyrus (LIFG) was thought to support language production (Broca 1861; Wernicke 1874; Lichtheim 1885; Geschwind 1965). This stark dissociation is no longer maintained: we now know that language functions are not to be identified in single brain regions. As observed (Fedorenko and Thompson-Schill 2014; Hagoort 2014), the mapping between neurons and cognitive processes relies more on dynamic functional networks than on individual regions, with each single node participating dynamically in other functional circuits as well. The roles of Wernicke’s and Broca’s areas have been expanded, and the arcuate fasciculus is now known to be a bidirectional pathway that connects larger areas of sensory cortex within prefrontal and premotor areas. Besides Wernicke’s and Broca’s areas, additional frontal and temporal regions of the left hemisphere have been found to be involved in language processing. Current models agree that sentence processing relies on a left-lateralised fronto-temporal network (Bornkessel-Schlesewsky and Schlesewsky 2013; Friederici 2011, 2012; Hagoort 2005, 2013; Hickock and Poeppel 2007) that involves posterior/anterior temporal and inferior frontal regions. However, which mechanisms each unit supports is, most of the time, not agreed upon. For example, some authors have argued that the posterior portion of the left superior and middle temporal gyrus supports mechanisms involved in the assignment of thematic roles (Bornkessel-

32

Chapter Two

Schlesewsky and Schlesewsky 2013; Bornkessel et al. 2005; Friederici 2011; Friederici et al. 2003), while other proposals argue for a role of this region in the retrieval of lexical information (Baggio and Hagoort 2011; Hagoort 2005, 2013) and in the linking of words to their corresponding concepts (Dronkers et al. 2004; Lau et al. 2008). A similar debate characterises the functional role of the anterior portion of the superior/middle temporal gyrus. Early studies demonstrated the involvement of this region in sentence-level processing, showing bilateral activation in the presence of syntactically structured material but not for word lists. This led to hypotheses of the essentially combinatorial nature of the mechanisms subserved by this region. Nonetheless, it is under dispute whether the activation of this cortical area is specifically responsive to syntactic combinatorics (Dronkers et al. 2004; Friederici and Kotz 2003; Friederici et al. 2000; Humphries et al. 2006) or whether it is modulated by processes that integrate syntactic and lexico-semantic analyses (Bornkessel-Schlesewsky and Schlesewsky 2013; Pallier et al. 2011; Rogalsky and Hickock 2011; Vandenberghe et al. 2002). Contrasting proposals are also advanced for the role played by the LIFG in sentence comprehension, ranging from domain-specific to domain-general interpretations. Specifically, some authors have argued that this region plays a key role in the language-specific mechanisms involved in form- and meaning-related analysis (Friederici 2011, 2013; Grodzinsky and Friederici 2006; Hagoort 2005, 2013). Friederici (2011, 2013) dissociates LIFG subregions that specifically support syntactic structure building – the pars opercularis (BA44) – from those that subserve semantic analysis at the sentential level, i.e. the pars triangularis (BA45) and orbitalis (BA47). Similarly, Hagoort (2005, 2013, 2014) parcellates the LIFG into syntactic (BA44/45), semantic (BA45/47) and phonologicalunification (BA6) areas. By contrast, other proposals emphasise the domain-general nature of the LIFG, suggesting its involvement in cognitive control and conflict-monitoring mechanisms (BornkesselSchleswesky and Schlesewsky 2013; Novick et al. 2005). In particular, Bornkessel-Schleswesky and Schlesewsky (2013) describe the LIFG as an area that does not support any specific language processing per se, but only task-relevant functions that monitor and regulate behaviour during linguistic processing. Of relevance also is the role played by a region outside of the leftperisylvian network, namely the angular gyrus (AG), whose activity has been shown to correlate with conceptual retrieval and integration in different domains (see Binder and Desai 2011; Binder et al. 2009; Seghier 2013 for recent reviews). Although this area is typically not included in

Sentence and Agreement Comprehension

33

the language network (but see Hagoort 2013, 2014 for a recent theoretical proposal), a number of language-related studies have recently reported increased AG activity in semantic processing at the phrase, sentence and beyond sentence levels. Bemis and Pylkkänen (2012) reported AG involvement in a magnetoencephalography (MEG) study contrasting compositional vs. non-compositional adjective-noun pairs (“red-boat” vs. “xlq-boat”) and concluded that this region plays a major role in conceptual integration processes. At the sentence level, AG involvement has been reported in studies where the context-relatedness of stimuli was manipulated. Kuperberg et al. (2006) reported activity increases in the left and right AGs during the reading of sentences that presented a partial relation with the preceding context, compared to those that were highly related or completely unrelated. Nieuwland et al. (2007) observed significant AG activation when a referentially ambiguous pronoun was found in the sentence, compared to when the pronoun was referentially anomalous. Finally, increased AG activity has also been reported in a series of studies on metaphor processing (Bambini et al. 2011; Mashal et al. 2007; Shibata et al. 2012). The engagement of this heteromodal parietal area in linguistic processing is plausibly mediated by its top-down and bottom-up neuroanatomical connections with fronto-temporal regions (middle, superior and frontal gyri, middle temporal gyrus), as several anatomical connectivity studies have found (Catani et al. 2012; Catani and Mesulam 2008, among others). Importantly, functional coupling between the AG and anterior temporal regions has been recently demonstrated by a study investigating semantic combinatorics in adjective-noun pairs (Molinaro et al. 2015). Language processing also involves regions outside of the perisylvian area that support domain-general processes such as working memory, executive control and attention. This is the case of the dorsolateral prefrontal cortex and of the anterior cingulate cortex (ACC). The dorsolateral prefrontal cortex has been shown to support cognitive functions such as working memory (see Katsuki and Costantinidis 2012 for a review), especially in tasks that require building relationships between items during online processing (Blumenfel and Ranganath 2006; Murray and Ranganath 2007). The ACC, in connecton with the dorsolateral prefrontal cortex (see Hagoort 2005, 2013), is thought to play a major role in processes that prevent behavioural mistakes (Carter and van Veen 2007; Taylor et al. 2007) by monitoring the presence of possible conflicts between expected and perceived input. Several studies on language processing have reported the involvement of the ACC in the processing of anomalous sentences (Folia et al. 2009; Kuperberg et. al.

34

Chapter Two

2003, 2008; Mancini et al. 2017; Nieuweland et al. 2012; Quiñones et al. 2014). Moreover, and importantly, the ACC has been indicated as the neural generator of the P600 effect (Du et al. 2013; Olichney et al. 2010), which further confirms the domain-general functional interpretation of this ERP component. Box 2-3. Basic principles of fMRI data interpretation Interpreting fMRI data

Among the variety of stimuli presentation procedures available for fMRI studies (parametric, blocked, event-related, hybrid designs; see Huettel et al. 2014 for a review), blocked and event-related designs are the most frequently employed. In blocked designs, participants perform tasks (e.g. reading/listening to words/sentences) for a given period of time (e.g. 30 seconds). Brain activity while performing the task is then compared with brain activity while performing another type of task (e.g. a non-linguistic baseline such as fixating a cross). In contrast, in eventrelated designs, the different tasks that compose an experiment are intermixed and can occur at any point during the experiment. In both types of design, contrasts between conditions typically follow the logic of the subtraction method, whereby activations of interest can be isolated by subtracting the activation in a control condition (e.g. correct agreement) from the activation in a critical condition (e.g. incorrect agreement). The activations shown in the contrast maps reported in scientific articles do not reflect absolute differences in neural activity between two conditions/tasks. Rather, they indicate the statistical likelihood with which one can reliably assume that the two conditions/tasks differ from one another. In other words, for researchers to say that one area of the brain is more active than another in a specific task, a statistically significant difference must be found between the BOLD signals emitted by the two regions during the performance of the same task. An excellent review of fMRI methods and designs can be found in Huettel et al. (2014). An important point to be kept in mind when interpreting fMRI data is that they are purely correlational in nature. This means that any activation pattern cannot be interpreted as evidence for the necessary involvement of a specific brain region in a specific cognitive process. A good example of how inaccurate it is to draw a causal relation between activation patterns and cognitive function is the role of the right hemisphere, whose involvement in language processing is reported in several studies, especially in association with prosodic and contextual processing (see Vigneau et al. 2011 for a review). Yet, studies with patients affected by

Sentence and Agreement Comprehension

35

language pathologies have shown that lesions in these areas do not necessarily have disruptive effects in language processing. Therefore, as noted by Bornkessel-Schlesewsky and Schlesewsky (2009), in spite of the enormous contribution and advantages offered by fMRI, the interpretation of the results from studies employing this technique still benefits from the supplementary information provided by studies with patients.

Neuroanatomical correlates of agreement processing Early fMRI studies on language processing were primarily aimed at the identification of the neuroanatomical correlates associated with syntax and semantics processing. Manipulation of morphological information on the verb and its contrast with semantically anomalous sentences was generically used to isolate the neuroanatomical correlates of syntactic processing, with no specific focus on the comprehension of agreement (Kuperberg et al., 2003, 2008; Newman et al., 2003; Ni et al., 2000). In Ni et al. (2000), participants listened to sentences that could be either grammatically correct or could contain a verb finiteness violation (e.g. “*Trees can grew...”) or a semantic violation (e.g. “Trees can eat...”). A dissociation between syntactic and semantic processing was found: while the former violations triggered significantly increased activity in the left inferior frontal gyrus, the latter activated several other regions in both hemispheres including the middle and superior frontal gyrus and the superior temporal and parietal regions. In an attempt to specify the contribution of the inferior frontal cortex during syntactic and semantic processing, Newman et al. (2003) compared sentences with number mismatches between a subject and the second verb of a coordinated structure (e.g. “*The lady praises the sister and meet the artist in the night”) against sentences including an extra verb (e.g. “*The woman thanked the barber and paid the receptionist knew at the desk”). They found increased activation in the pars opercularis of the inferior frontal gyrus for the syntactic violation, whereas the pars triangularis was found to be more sensitive to the extra-verb violation. Additionally, they observed increased activation in the left posterior temporal region for both types of processing. However, as noticed by Quiñones and colleagues (Quiñones et al. 2014), since the nature of the violation involving the extra verb is difficult to determine, the activation observed in this study may not exclusively reflect s-v vs. semantic processing difficulties. Kuperberg et al. (2003) carried out an fMRI study in which they presented the participants with grammatically correct (e.g. “We couldn’t

36

Chapter Two

sleep at night because the baby would cry.”), finiteness-anomalous (e.g. “...because the baby would cries.”) and pragmatic-anomalous sentences (e.g. “...because the baby would remember.”). Relative to the grammatically correct sentences, similar neural regions were recruited by morphosyntactic and pragmatic information, but activation patterns differed. An increased response to finiteness anomalies relative to pragmatic violations was evident in bilateral parietal areas, while the same contrast revealed decreased activation in left temporal and inferior frontal areas. The same design was used in a follow-up study (Kuperberg et al. 2008) where a distinction between two different types of semantically anomalous sentences was added: real-world pragmatic violations (“Every morning at breakfast the boys would plant the flowers.”) and semantic– thematic violations (“Every morning at breakfast the eggs would eat toast and jam.”). The common neuroanatomical network recruited by both finiteness and animacy semantic–thematic violations (relative to grammatically correct sentences) included a widespread bilateral frontoparieto-temporal response. Some of these regions exhibited more activity in response to the finiteness violations than to the animacy semantic– thematic violations, namely the left inferior parietal lobule, bilateral anterior cingulate cortex and medial frontal gyrus. Overall, these controversial results show that the regions implicated in the processing of verbal morphology during the computation of an agreement relation are still uncertain. Different factors can be identified that contributed to these discrepancies, such as the different types of verb morphology manipulations, differences in the experimental designs adopted (see Quiñones et al. 2014 for a discussion) and presentation modality. Moreover, except for Kuperberg (2003, 2008), none of these studies compared anomalous and well-formed sentences in both directions, that is by subtracting the activation for the correct from the activation of the incorrect sentence and vice versa. The fMRI studies that will be presented in Chapters 3 and 5 introduce important theoretical and methodological variations in the study of agreement and morphosyntactic processing. Specifically, the contrast between person and number will be crucial to show the differential sensitivity within the language network to feature-based mechanisms. The contrast between unagreement and standard agreement discussed in Chapter 5 will further corroborate the idea that the computation of an agreement dependency does not rely on a monolithic process. Rather, it consists of two functionally, neuroanatomically and temporally distinct mechanisms – checking and anchoring – that flexibly adjust to the distinct

Sentence and Agreement Comprehension

37

types of information and structural configurations available in the linguistic input.

Sentence comprehension models Understanding how language is comprehended in real time has been one of the main goals of psycholinguistics ever since the cognitive revolution took place in the 1950s. Whether in written, oral or signed modality, understanding language implies mapping the perceived form onto its corresponding meaning (Bornkessel-Schlesewsky and Schlesewsky 2009; Marslen-Wilson 1973). Models of sentence comprehension represent formalisations of how mental representations are generated, transformed and stored during language processing, which involves not only the outcome – what listeners and readers eventually understand from written, oral or signed language – but also the online, moment-to-moment processes that lead to successful comprehension. Existing models of sentence comprehension have typically attempted to explain how syntactic and semantic analysis interact (if they interact) during online processing, and how these two sources of information are represented at the neural level. Depending on the temporal relation between syntactic and semantic analysis, three types of models can be identified, namely serial, interactive and cascaded models. A strictly staged architecture is postulated by serial models such as the one proposed by Frazier and Fodor (1978), Frazier (1987) and Frazier and Clifton (1996). In this model, syntactic processing is modular and interacts with other linguistic information only at the output level (Frazier 1987; Frazier and Clifton 1996; Frazier and Rayner 1982; Friederici, 2002). The parser initially constructs the simplest syntactic structure on the basis of lexical-semantic information, while during a second stage, thematic role assignment is carried out. If syntactic and thematic structure cannot be mapped onto one another, reanalysis takes place (Frazier 1987; Frazier and Rayner 1982). Crucially, this architecture is innate and universal: the syntax-first analysis carried out by the parser and the separation between syntactic and semantic processes hold cross-linguistically. By contrast, interactive models posit mutual influence between the syntactic domain and multiple sources of linguistic constraints throughout the computation of sentence meaning (Bates and Mac Whinney 1987; MacDonald et al. 1994; Mac Whinney and Bates 1989). Syntactic and semantic information undergo parallel processing: both types of information are directly and simultaneously extracted from the input. Because the parser operates in a constraint-based modality, the relative

38

Chapter Two

weight of syntactic and semantic information depends on probabilistic constraints that may vary across languages. In sum, both serial and interactive models assume that syntactic and semantic information are integrated during language perception to achieve understanding. What differentiates them is the time at which integration takes place: early for constraint-based models; late for serial, syntax-first models. A further distinction between the two classes of models concerns the assumptions that they make about processing stages. From an interactive viewpoint, comprehension is a matter of congruency between the different types of information that are simultaneously extracted from the input and language-specific constraints. Crucially, in interactive models, the checking of this congruency occurs throughout sentence processing and not at a specific stage only. On the contrary, a modular approach makes it possible to identify distinct stages of processing: sentence comprehension goes through various independent phases, each of which must be completed before proceeding to the next. A third view incorporates both serial and parallel aspects in a cascaded organisation. Like serial models, cascaded models (see Bornkessel and Schlesewsky 2009; Bornkessel-Schlesewsky et al. 2016 for a review) assume a hierarchy of processing stages in which the analysis of syntactic information dominates other types of analysis. Yet, unlike serial models, cascaded architectures assume a constant flow of information between processing stages. This makes a partial temporal overlap between stages possible, so that Stage 2 can be initiated on the basis of partial (rather than full) input coming from Stage 1. This partial temporal overlap represents a sort of “integrating window” during which modulating influence from the analysis of one information type on another can occur. If, for example, a syntactic violation is detected during this integration window and Stage 1 processing fails, semantic processing can be aborted. However, if the detection of a syntactic violation in Stage 1 occurs after the integrating window, when Stage 2 has already reached a specific threshold, semantic processing cannot be blocked anymore and will proceed.

Neurocognitive models of sentence processing Let us now turn to the presentation of three neurocognitive models of sentence comprehension that combine insights from both ERP and fMRI studies. As will be clear, the three models not only differ in the theoretical framework they adopt, but also in the temporal relationship between

Sentence and Agreement Comprehension

39

syntactic and semantic analysis that they assume, reflecting the main architectures underlying serial, interactive and cascaded models. Friederici’s (2002, 2011) Neurocognitive Model of Sentence Comprehension (NMSC, henceforth) grew out of the aim to provide a neurocognitive implementation of the classical assumptions of two-stage models such as Frazier and Fodor’s (1978). This model claims that comprehension can be subdivided into three processing phases (Table 2-3). Phase 1 encompasses basic and early processes of constituent structuring that draw exclusively on word category information. Initial local structure building would rely on anterior portions of the STG and the pars opercularis. Violations in phrase structure building generate almost immediate processing disruptions that are reflected by eLAN effects. Once local phrase structure has been built, morphosyntactic and lexical-semantic processing takes place in phase 2, with the goal of thematic role assignment. Importantly, in this phase, morphosyntactic and semantic properties are processed in parallel but independently of one another. Anomalies detected at the morphosyntactic level give rise to LAN effects, as opposed to N400 effects that index the detection of semantic and thematic role assignment violations. Table 2-3. The three processing stages and relative neuro-physiological correlates in Friederici’s (2002, 2011) neurocognitive model. Processing stage Phase 1 Phase 2

Phase 3

Processing operation Initial local structure building Independent analysis of morphosyntactic and semantic information

ERP component and timing ELAN (150–200 ms) LAN/N400 (300–500 ms)

Neuroanatomical substrate Anterior STG and BA44 Posterior STS/STG and BA44 Middle/posterior STG/MTG and BA45

Integration and interpretation Repair/reanalysis

P600

Posterior STG and basal ganglia

At the neuroanatomical level, the building of local syntactic structure in phase 1 is subserved by the anterior superior temporal gyrus (STG) in connection with the pars opercularis in the left inferior frontal gyrus (LIFG). Morphosyntactic analysis relies on the tight interaction between the posterior portion of the superior temporal sulcus/gyrus (STS/STS) and Broca’s areas (BA44), while the pars triangularis (BA45) is thought to

Chapter Two

40

play a major role in semantic analysis at the sentence level. Finally, the integration and repair/reanalysis processes occurring approximately 600 milliseconds post-stimulus onset involve the connection between the posterior STG and basal ganglia. A similar hierarchical organisation can be found in the extended Argument Dependency Model (eADM) proposed by Bornkessel and Schlsewsky (2006, 2008, 2009). Three stages are identified for the comprehension of core relations within sentences, but the partial temporal overlap between phases make this model a good example of cascaded architecture (Bornkessel-Schlesewsky and Schlesewsky 2008, 2009). An overview of the model and its architecture is provided in Table 2-4. Word category analysis is a prerequisite for the relational analysis that takes place from phase 2 onwards. All aspects of the form-to-meaning mapping are situated in phase 2, where argument role assignment (Actor, Undergoer) is carried out, on the basis of a restricted set of information (case, animacy, definiteness and person) extracted during nominal analysis, and whose weight vary cross-linguistically. Table 2-4. Overview of processing stages and ERP components and neural substrates in the eADM (Bornkessel and Schlesewsky 2006, 2008, 2009, 2013) Processing stage

Processing operation

Phase 1

Phrase structure building without relational information Compute Prominence Establish Agreement and Compute Linking Generalised Mapping Well-formedness/ repair

Phase 2a Phase 2b Phase 3

ERP component and timing eLAN

Neuroanatomical substrate Pars opercularis and Anterior superior temporal sulcus

N400 LAN/P600 Late positivity

LIFG (pars opercularis) pMTG LIFG and ACC

Agreement is analysed within the establish agreement step of phase 2. The parser checks that the element that agrees with the verb corresponds to the one that has been associated with an Agent during nominal analysis. If a match is found, a link is established. The identification of a morphosyntactic incongruence or the recognition of an alternative assignment of agreement

Sentence and Agreement Comprehension

41

is expected to trigger a LAN effect. Similarly to Friederici (2002, 2011), morphosyntactic processing is thought to be mainly subserved by the pars opercularis in connection with the posterior portion of the MTG. In phase 3 of the eADM, all the modulating, reparatory and evaluative aspects of processing take place, using information from further domains (discourse, prosody and plausibility). Specifically, this stage incorporates two steps: an explicit generalised mapping, which is independent of possible repair operations, and a well-formedness step, which is responsible of repair mechanisms that underlie P600 effects. Neuroanatomically, integration and evaluation mechanisms are supported by LIFG areas. Finally, a lexicalist and interactive approach is adopted in the Memory, Unification and Control (MUC) framework (Hagoort, 2003, 2005). This parsing account is lexicalist in the sense that all syntactic nodes are retrieved from the mental lexicon. In other words, chunks of syntactic structure are stored in memory. Each word form in the lexicon is associated with a structural frame consisting of a three-tiered unordered tree. The top layer of the frame consists of a single phrasal node (e.g. NP), which forms the root node and gets connected to one or more functional nodes (e.g. subject, head, direct object) in the second layer of the frame. The third layer again contains phrasal nodes to which lexical items or other frames can be attached. Importantly, there are no syntactic rules that introduce additional nodes: the only grammatical rule is Unification, which is supposed to bind syntactic frames with identical root and foot nodes and to check agreement features. In the online comprehension process, structural frames associated with the individual word forms incrementally enter the so-called Unification workspace, neuroanatomically identified in the LIFG area. Here, syntactic, semantic-pragmatic and prosodic constraints are all taken into consideration to explore all the possible combinations of the incoming material. Importantly, a parcellation of the Unification workspace is assumed in which BA44/45 is thought to support syntactic unification, while semantic Unification relies on BA45/47. In the MUC model, the analysis of agreement features is not temporally and functionally distinct from semantic-pragmatic and prosodic information analysis: all types of constraints are simultaneously taken into account to successfully bind frames with each other, without any informational/temporal precedence for syntax. According to this model, LAN activity reflects a binding failure as a result of a negative outcome of the agreement check, or failure to find a matching category node. The P600 is instead related to the time it takes to establish Unification links of sufficient strength: its amplitude is

Chapter Two

42

determined by the degree of competition among alternative unifications, by syntactic complexity and by semantic influences. Table 2-5. Overview of Hagoort’s MUC model, and their relative ERP components and neural substrates (Hagoort 2003, 2005, 2014). Processing stage Memory

Unification

Control

Processing operation Representation of phonological, morphological and semantic information, as well as the syntactic templates associated to each word Generation of larger structures from the lexical frames that are retrieved from memory Executive control functions that are necessary for correct language selection, turntaking, attention etc.

ERP component and timing N400: preactivation of a word based on context

Neuroanatomical substrate Temporal cortex and AG

LAN: failure to bind two lexical frames at the morphosyntactic level. N400: difficult context integration P600

BA44/45: syntax BA45/47: semantics BA44/6: phonology

Dorsolateral prefrontal cortex and ACC

Features and interpretation: Anchoring agreement in comprehension Irrespective of their theoretical framework and architecture, the models presented above converge in identifying three basic mechanisms at the base of the form-to-meaning mapping that underlies sentence comprehension: (i)

a mechanism dealing with structural analysis, via morphological decomposition of lexical elements and extraction of features;

Sentence and Agreement Comprehension

43

(ii) a mechanism – checking – that handles the integration of incoming input with previously encountered elements, based on the evaluation of their morphosyntactic and semantic fit, to ensure proper building of agreement relations among sentence parts; (iii) a mechanism for the constant monitoring of the input and output of the other systems, to detect possible conflicts. While agreement is recognised as a core mechanism of online sentence processing, none of the cognitive architectures described in this chapter appear to be equipped with computations and algorithms that can “dissect” an agreement relation to analyse the different information carried by its features. The goal of the experimental work reviewed and discussed in this monograph is to fill this gap by testing the validity of the FIP introduced in Chapter 1, and thus to enrich current processing models. A series of processing hypotheses derive from the structural and interpretive requirements of the FIP. Firstly, the structural dissociation postulated by the FIP between person and number features makes it plausible to hypothesise that the access to morphosyntactic information contained in high nodes of the syntactic tree (i.e. in the PersonP projection, following Shlonsky 1989 and Sigurdsson 2004) may produce different processing costs compared to accessing information contained in lower positions (e.g. NumberP). Secondly, the system that handles the integration of incoming input may be sensitive to the type of information that is being checked. Connections are established between sentence parts with the goal of checking the consistency of features at different levels: morphosyntactic and semantic-discourse. In line with current minimalist models, a mechanism akin to the minimalist Agree operation is assumed to handle the checking of morphosyntactic consistency of both person and number, while the linking of the two features to their corresponding sigma values, i.e. their semantic-discourse correlates, is dealt with by the Anchor operation. Critically, while both Agree and Anchor connect two positions with the goal of feature checking, the sets of features and the position that they target (i.e. the interpretive anchor) are different. Whether a clauseinternal or a clause-external position is reached may have a significant effect on processing. Operations that connect IP-internal positions, such as Agree for person and number, or number anchoring, are expected to drive different processing correlates compared to operations that establish a link with IP-external positions, such as when the compatibility of the discourse participants invoked by subject and verb person is verified. It follows that

44

Chapter Two

the breaking of IP-external compared to IP-internal anchoring may yield different processing disruptions. The crucial role of interpretive anchors should be visible in cases of mismatch, when they should sanction whether the relation can be interpreted or not. If a match can be established between the phi and sigma values of the feature, then the resulting representation is passed on for further semantic analysis. If a mismatch is found, the process takes different shapes depending on the nature of the disagreement. In the presence of mismatching but grammatical agreement patterns, such as unagreement, inspection of the person anchor licences the relation, based on the compatibility of the subject and verb person indices. In contrast, in the presence of an outright person and number mismatch, the parser sanctions a syntax error that signals either the breaking of a crossdimensional anchoring (IP-external anchoring) or a more local conflict (IP-internal anchoring) that can be solved within the inflectional layer of the sentence, for example by re-processing the inflectional information of the controller as in number agreement violations Thirdly, fundamental scaffolding for the incremental implementation of checking operations is provided by domain-general working memory, conflict-monitoring and attention mechanisms. These systems assist the parser in maintaining items available in memory for integration to occur and for the triggering of error-related signals when integration cannot be achieved. In the following chapters, empirical evidence will be presented for a precise neuro-physiological characterisation of the representations and algorithms that support agreement comprehension.

Summary The goal of this chapter was to provide a brief introduction to the experimental techniques and sentence processing models that constitute the background of the studies presented in the next chapters. As should be clear from the preceding sections, there is no unique and perfect design or paradigm for the study of sentence processing. ERPs and eye-tracking allow one to obtain millisecond-precise information concerning the time course of linguistic processing, which unidimensional behavioural techniques such as self-paced reading cannot offer. The response obtained with judgements, although not informative to draw conclusions on the time course of linguistic analysis, can be extremely relevant to get a snapshot of the final representation elaborated by speakers during comprehension. Finally, the analysis of the neural correlates of sentence

Sentence and Agreement Comprehension

45

processing using fMRI does not provide precise temporal information, but neuroanatomical dissociations in the processing of linguistic input can be suggestive of underlying functional dissociations in comprehension mechanisms. Similarly, there is no agreed-upon architecture of language. Three models have been presented that make different claims about the timing of syntactic and semantic analysis and their potential interaction, ranging from those that assume an early and independent analysis of syntactic information (serial models), sometimes allowing for a certain temporal overlap between the two sources of information (cascaded models), to those that predict early interaction between the two sources of information (interactive models). Critically, none of the existing models makes explicit hypotheses about possible fine-grained mechanisms in agreement and features processing: the empirical work presented in the next chapters will be precisely aimed at filling this gap.

CHAPTER THREE EVIDENCE FOR FEATURE-SPECIFIC PROCESSING

In Chapter 1, two theoretical approaches to s-v agreement were illustrated, namely the analysis given to this phenomenon within the Minimalist Program (Chomsky 1995 and subsequent work) and the one outlined by Cartography (Shlonsky 1989, 2010; Sigurdsson 2004; Sigurdsson and Holmberg 2008, inter alia). As pointed out, the two theoretical accounts diverge on two fundamental points. First, the different structural representation attributed to phi-features: clustered and structurally undifferentiated under the same T head in the former analysis, while structurally separated in the latter account. The first straightforward consequence of this divergence is the different checking mechanism underlying the two representations: a unique checking operation accessing the whole feature cluster on one side, as opposed to several separate checking operations – as many as the number of features to be checked – on the other side. Second, it was shown that under a cartographic analysis of agreement, it is possible to capture the intrinsically different interpretive properties associated with person and number, which a single-cluster analysis such as the one put forth by standard Minimalism fails to represent. In light of this, an account of the checking and interpretation of agreement features was proposed that relied on the notion of an interpretive anchor. In this chapter, the results of behavioural, electrophysiological and fMRI studies on s-v agreement processing in Italian and Spanish will be described. The aim of these experiments is to directly contrast the processing of person and number in order to ascertain whether the parser distinguishes between them, despite their synchretical representation in the same inflectional affix, and in in light of the distinct interpretive properties that were outlined in Chapter 1. If person and number are structurally distinct, instead of being part of a set, differences may emerge between person and number agreement anomalies at the critical point, i.e. the verb, as a consequence of the different structural position occupied by the two

Evidence for Feature-Specific Processing

47

features (Baker 2008; Bianchi 2003, 2006; Shlonsky 1989; Sigurdsson 2004; Sigurdsson and Holmberg 2008) and their different interpretive anchors. Recall that the number specification on the subject noun has been identified as the natural interpretive anchor for the number feature, while person’s anchor has been identified in the speech act participants’ representation. Based on this, differences are expected in the effects that person and number anomalies may have on sentence comprehension. Let us illustrate them. Consider a sentence like (1) below (Italian, from Mancini et al. 2014a): (1) *Il giornalista hanno scritto un articolo The journalist3.sg have3.pl written an article

Here the parser will detect an error in the copying of the number specification from subject to verb. The interpretive anchor will be activated and inspected, and a reliable specification to repair the ill-formed dependency may be provided. Therefore, here a local and straightforward repair strategy is supposedly applied that may eventually lead to interpretation. On the other hand, a person mismatch is likely to generate a stronger perturbation: (2) Il giornalista hai scritto un articolo The journalist3.sg have2.sg written an article

The inconsistency detected between subject and verb should activate recourse to the anchor in the speech act participants’ representation. Here, a local diagnosis is impossible, contrary to what is supposed to happen in the presence of a number anomaly. Recall that under the feature-anchoring approach to agreement introduced in Chapter 1, nominal and verbal person represent two autonomous values; hence, both are interpretable and equally anchored to the semantic-discourse representation. The two discordant values presuppose two underlying discordant speech participants: the anchor is therefore unable to provide the parser with a reliable cue to repair the conflict. A person violation is thus likely to generate serious interpretive problems that may culminate in the impossibility to interpret the sentence. As described below, longer latencies, greater ERP effects and different brain activation patterns for person compared to number anomalies could therefore emerge. On the other hand, if person and number are structurally undifferentiated and the parser is not sensitive to the two features’ different interpretive properties, no differences should emerge from a comparison of the two violations.

48

Chapter Three

This chapter is organised as follows: after reviewing early evidence for feature-specific processing and its theoretical and experimental limitations, results of behavioural, electrophysiological and fMRI studies in Italian and Spanish will be illustrated that provide relevant insights into the representation and processing of person and number agreement.

Feature-specific processing: Early evidence One way to investigate the use and representation of agreement features is to study them in a contrastive manner, for example contrasting two features in a similar linguistic task. While a large number of studies have dealt with the processing of gender and number features, person has received less attention, perhaps because the language that has been more prominently tested, i.e. English, is inflectionally rather poor. This makes Romance languages particularly suitable for the study of s-v agreement. Recent psycholinguistic studies have investigated the processing of agreement features in Italian as well as in other languages, with behavioural as well as ERP studies (Barber and Carreiras 2005; Carminati 2005; De Vincenzi 1999; De Vincenzi and Di Domenico 1999; Hagoort et at. 1993; Hinojosa et al. 2003; Nevins et al. 2007; Osterhout and Mobley 1995; Silva-Pereyra and Carreiras 2007; inter alia). The goal of these studies was mainly to investigate at what stage of parsing the processor makes use of the information provided by agreement features, and to ascertain whether these features are differentially processed. However, these studies have not led to convergent results, and so far an unequivocal answer to the question of how features are accessed and processed has not been provided. De Vincenzi and Di Domenico (1999) and De Vincenzi (1999) showed that gender information is not used by the coreference processor to select the appropriate antecedents at the same stage as number information is. In fact, while both gender and number are used sentence-finally to selectively reactivate the matching antecedent, only number information reactivates the antecedent at a sentence-internal position. It thus seems that the parser dissociates gender and number information, which can be ascribed to the different structural representations of the two features. Such a claim is substantiated by a proposal put forth by Ritter (1993) and Di Domenico (1997), who claim that number autonomously projects in the syntax, while gender is projected in the syntax either with number or with the noun in the respective projections, as shown in Figure 3-1.

Evidence for Feature-Specific Processing

49

Figure 3-1. Number and gender features representation in Ritter (1993) and Di Domenico (1997) This would be compatible with a modular theory of language processing (see Chapter 2) that hypothesises the presence of distinct processing subsystems, each associated with its own processing task. One of these subsystems is the syntactic processor, which at an initial stage of processing operates to assign a constituent structure analysis to the input using only the information relevant to that syntactic level of analysis. If gender is not a head, with its correlated functional projection, it will not be visible to the syntactic parser, and its information will not be computed initially. Number, on the contrary, will be available at this initial stage, and the processor will be able to use it to establish a coreference between a noun and a pronoun. The processing of gender and number agreement in comprehension has also been studied using grammatical priming with word pairs, leading to interesting results that support a distinct representation associated with the two features. Faussart et al. (1999) reported longer recognition times when words disagreed in gender than when they disagreed in number, in French as well as in Spanish. To account for this finding, Faussart and colleagues propose that the nature of the feature being violated plays an important role in detecting and repairing a morphosyntactic anomaly. More specifically, they claim that in the presence of a violation, the language processor is compelled to repeat one or more operations involved in the lexical decision process, depending on the nature of the violation. They propose a three-stage model, based on Bradley and Forster’s (1987) lexical retrieval model: in the first stage, lexical identification is carried out by locating the correct lexical entry. Subsequently, in the second stage, called the readout stage, the relevant content of the entry is made available to the specific processors (the parser for purely grammatical features, and the interpreter for semantic properties). Together, these two stages form lexical retrieval (Bradley and Forster 1987). After lexical retrieval, Faussart et al. (1999) add a third stage, called evaluation, where the

50

Chapter Three

appropriateness of the target to the context is evaluated in order to integrate the material. In the presence of a gender violation, the processor may be uncertain about whether or not the correct lexical entry has been located in Stage 1, and it may then be compelled to go back to the lexical identification stage, therefore repeating the evaluation and readout stages. This is compatible with a view of gender as an inherent property of the stem (Ritter 1993; Di Domenico 1997). In contrast, number is not inherently represented in the stem but is attached to it by rule, and in the presence of an anomaly, the processor may only need to re-do the evaluation stage to confirm that the output of the readout stage was adequate. Evidence for a distinct representation and processing of number and gender has also been found in language production data. For example, the rates of number and gender agreement errors are usually greater for the former than for the latter feature, which could be interpreted as evidence for different agreement mechanisms at work (Vigliocco et al. 1996). In addition, Igoa et al. (1999), in an analysis of error production, found that exchange errors mostly affected number suffixes, seldom number and gender simultaneously, and never gender alone. These differences in error rates suggest that gender is retrieved directly from the lexicon and assigned to the phrase structure together with the lemma, while number is derived by rule. In sum, previous studies concerning number and gender processing seem to suggest that there are differences between these two features and that these differences are related to their representation. However, they fail to clarify how and when such differences influence agreement during syntactic processing. In this respect, substantial contribution comes from techniques with finer temporal grain such as ERPs.

Testing the FIP: Behavioural evidence The existing literature on the processing of the person feature is less abundant compared to other features, and the results obtained from studies where it was contrasted with number and gender provide a much less homogeneous picture. A recent study on the processing of person, number and gender in Italian suggests the presence of a distinction among the three features, based on their degree of disambiguating power. Using a series of self-paced reading experiments, Carminati (2005) aimed at testing the validity of the so-called Feature-Strength Hypothesis, according to which a different degree of cognitive strength is to be attributed to person, number and gender, on the basis of Greenberg’s (1963) Feature Hierarchy

Evidence for Feature-Specific Processing

51

Person>Number>Gender. This hierarchy captures the distribution of agreement features across languages: if a language has a feature, then it must have all the ones above it in the hierarchy. If it possesses gender, it also has number. If it has number, then it must have person, but not necessarily gender. This implies that person is more important or basic to language than number and gender. According to Carminati, two key findings support the presence of a Feature Hierarchy. In an experiment where number and gender were contrasted, disambiguation of the null subject pro was facilitated (i.e. was faster) when the identification of the antecedent was made via number information. In another experiment, number disambiguation was compared to number and person disambiguation. The results indicated that the penalty for disambiguating pro was significantly reduced when the disambiguation relied on number and person information together, compared to when only number information was available. Carminati (2005) concludes that these results support a hierarchical organisation of features and the presence of different representations associated with each of them. It should be noted that a comparison of person-only and number-only effects in the disambiguation of pro could have clarified to what extent agreement feature processing reflects the Feature Hierarchy proposed by Greenberg (1963). In the two self-paced readings described below, Mancini and colleagues (Mancini et al. 2014a) used a stationary-window self-paced reading task in which person and number agreement were compared using both pronominal (Experiment 1) and lexical DP subjects (Experiment 2). With respect to Carminati (2005), sentences with overtly expressed subjects were used (instead of null-subject sentences), and person and number were manipulated separately to create agreement anomalies between subject and verb. A factorial design with person agreement and number agreement as the two factors was used, each with two levels (Match and Mismatch), as shown in Table 3-1. If the parser differentiates between the two features and if person anomalies are perceived as more severe than those of number, faster reading times are expected for number compared to person agreement. Moreover, and importantly, the authors hypothesised that an interaction should emerge driven by the significantly longer reading times for the double-violation condition (i.e. where both person and number are violated) compared to number agreement, but not relative to person agreement anomalies. In contrast, if the parser does not differentiate between person and number agreement, no difference is expected between the two types of violations. Similarly, the double-violation condition

Chapter Three

52

should not differ with either the person or the number agreement conditions. Table 3-1. Design and sample of the materials used in the two self-paced reading studies in Italian (Mancini et al. 2014). The critical word is in italics. Experiment 1 Person

Match

Mismatch

Experiment 2 Match

Number Match Io ho letto un libro I1.sg have1.sg read a book *Io hai/ha letto un libro I1.sg have2.sg/3sg read a book Number Match Il giornalista ha scritto un libro The journalist3.sg has3.sg written a book

Person Mismatch

*Il giornalista hai/ho scritto un libro *The journalist3.sg have2.sg/1.sg written a book

Mismatch *Io abbiamo letto un libro I1.sg have1.pl read a book *Io avete/hanno letto un libro I1.sg have2.pl/3.pl read a book Mismatch *Il giornalista hanno scritto un libro The journalist3.sg have3.pl written a book *Il giornalista avete/abbiamo scritto un libro *The journalist3.sg have2.pl/1.pl written a book

Interestingly, the two experiments revealed reading patterns that converge with both hypotheses, as shown in Figure 3-2. In the first selfpaced reading experiment, as expected, the data revealed the presence of strong person and number agreement effects, indicating that readers were aware of the agreement manipulations. However, the failure to find a difference between person-incorrect and number-incorrect sentences, as well as among the three anomalous sentences, seems to suggest that no functional difference can be drawn between the two types of agreement violations. In this respect, these data converge with findings from Silva-

Evidence for Feature-Specific Processing

53

Pereyra and Carreiras (2007) in Spanish, in which neither qualitative nor quantitative ERP differences were found between the two types of agreement. An alternative explanation exists that concerns the type of stimuli utilised, which may have obscured differences between person and number anomalies. It is indeed known that pronominal forms are each characterised by a different underlying representation in terms of type and number of speech participants (Benveniste 1966; Harley and Ritter 2002, see also Chapter 4). While I and Yousg entail unique speech participants with specific roles in discourse – a speaker and an addressee, respectively – the 3rd person indicates exclusion of both speaker and addressee, as it indexes a contextually salient individual that does not have any active role in the unfolding speech event (Bianchi 2006). This distinction emerges even more clearly in plural forms. As a matter of fact, the uniqueness of I and Yousg prevents the I–We and Yousg–Youpl alternations from representing true pluralisations, i.e. shifts from a singular individual with a specific role to a multiplication of identical speakers and addressees. As a matter of fact, We and Youpl refer to composite groups respectively formed by a speaker and associate (i.e. We = I+ Yousg or he/she) and an addressee and an associate (Youpl = Yousg + he/she). It follows that a true pluralisation can be produced only with the 3rd person, which permits shifting from an individual to a multitude of individuals equally deprived of speech role. In light of this, the distinction both in numerosity and speech roles of the I–We and Yousg–Youpl alternations used to create number violations in Experiment 1 may have produced a spurious response, obscuring possible differences between person and number violations. If this is on the right track, the manipulation of person and number agreement between lexical subjects and verbs should represent a more favourable testing ground for the FIP. The results from the second self-paced reading experiment corroborated this hypothesis: the reading of person-anomalous verbs was found to give rise to quantitatively different reading times compared to number-anomalous verbs, but not to person-and-number violated ones. Importantly, this effect emerged both at the critical point and in the following segment, thus indicating that the disruption of a violation involving person agreement spills over from the anomalous point and reaches subsequent positions. Taken together, these results point towards the presence of distinct degrees of complexity for person and number agreement violations, with the former generating greater processing difficulty than the latter. As discussed by Mancini and colleagues (2014), this complexity may be due to the different levels of analysis into which

54

Chapter Three

the interpretation of person agreement taps (i.e. the morphosyntactic and the discourse one), as will be discussed in more depth below.

Figure 3-2. Mean reading times at verb position for the first (left panel) and second (right panel) self-paced reading experiment (adapted from Mancini et al. 2014). Bars represent standard error.

Testing the FIP: Electrophysiological evidence The quantitative dissociation found between person and number in Italian was also tested in an ERP paradigm in Spanish, with the goal of assessing whether qualitative differences could be found between the two types of agreement, as a result of their distinct interpretive properties. Before proceeding to detail the results of the study, a brief review of studies testing feature dissociations with the ERP paradigm is provided. A large body of ERP studies has investigated the online processing of agreement relations, and specific patterns of ERP effects have been attested in the presence of processing difficulties due to agreement anomalies, namely a P600 effect often preceded by early negative effects (see Chapter 2 for a general review). However, only a few studies have investigated whether ERP components are sensitive to fine-grained manipulations that involve the consistency between the features involved in a relation. Barber and Carreiras (2005) investigated the processing of number and gender agreement in determiner-noun and noun-adjective pairs, varying the position of the violation within the sentence (initial or mid-sentence position). Both number and gender violations gave rise to a

Evidence for Feature-Specific Processing

55

LAN-P600 pattern, independent of the position in the sentence and of the agreement dependency involved. A difference between gender and number was found in the late phase of the P600 and was evidenced by a larger P600 effect for the former compared to the latter anomalies. This was taken to suggest the presence of differences in repair strategies associated with the two features, but not in mismatch detection procedures. Similar findings and conclusions were obtained in a study in Italian (Molinaro et al. 2008) where gender and phonotactic features were manipulated within determiner-noun pairs, with the latter features generating a larger P600 in the late phase of this effect. While a considerable number of studies on agreement have dealt with number and gender, person has been given less attention, and fewer studies on this feature can be found in the ERP literature. In a study in German (Rossi et al. 2005), a LAN followed by a P600 effect was found for acoustically-presented sentences containing person agreement violations between subject and verb, which were compared to word category and combined category-agreement violations. Nevins et al. (2007) manipulated subject-verb agreement to create gender, number, gender+number and person+gender violations in Hindi, to assess whether ERPs elicited by agreement violations vary as a function of the nature and/or the number of incorrect agreement features. Compared to correct sentences, a P600 effect was elicited in all grammatically incorrect conditions. No amplitude differences were found among gender, number and combined gender+number violations, while a significantly larger P600 effect was reported for the combined person+gender violation. The authors attributed this finding to the privileged linguistic status of person compared to number and gender. Previous sentence processing findings (Carminati 2005) and typological analyses support this analysis and argue for the presence of an implicational hierarchy among features (Greenberg 1963) – Person>Number>Gender – that seems to govern the distribution of agreement features across languages. However, there are two caveats in the conclusion drawn by Nevins et al. (2007) concerning person’s greater cognitive salience. Firstly, as pointed out by the authors, the larger amplitude of the effect driven by person+gender violations may result from the orthographic markedness that these violations have, due to the varying representation that the syllable containing the person agreement morpheme has in the Devanagari script. Secondly, in their materials, number and gender were manipulated to create both single-feature violations (i.e. gender violations and number violations) and double-feature violations (i.e. gender+number anomalies), whereas person was manipulated only in multiple-feature violations

56

Chapter Three

(person+gender mismatch). This imbalance in the material may have caused person violations to be more salient, hence the larger amplitude of the P600 effect. A different scenario is depicted by the findings reported by SilvaPereyra and Carreiras (2007), who carried out a comparison of person and number anomalies in Spanish and found no differences between the two features, either at early or late stages of processing. Critically, these findings recall those from the first self-paced reading study by Mancini et al. (2014), and the same methodological caveat can be applied. To overcome this caveat, the ERP experiment in Spanish that Mancini and colleagues (Mancini et al. 2011a) designed included anomalies between subjects and verbs whose person specifications gave rise to a participant vs. non-participant opposition (i.e. 3rd person subjects followed by 1st or 2nd person verbs; person violation condition), together with number anomalies between subjects and verbs that displayed a clear cardinality contrast (3rd person singular vs. 3rd person plural), as illustrated in Table 32. The adopted design made it possible to formulate specific hypotheses about the ERP effects that each type of agreement is expected to elicit and thus to test the main claims of the FIP. While predictions for the number mismatch compared to the control are consistent with the LAN-P600 pattern cross-linguistically attested for this anomaly, the clearer-cut person violations included in the design make it possible to hypothesise a qualitative and quantitative difference in the interval associated with early negative effects. Specifically, since a person mismatch affects both the morphosyntactic and the discourse level of analysis, the parser should encounter a persisting difficulty in assigning a coherent interpretation to person, and an N400 effect can therefore be expected. The extra link to the discourse representation of the sentence, where person’s anchor is located based on the FIP, is not necessary for number agreement, the anchoring of which would be circumscribed to the inflectional layer of the sentence. Table 3-2. Design and sample of the experimental material used in Mancini et al. (2011a) Type of Agreement Correct Agreement (CA) Person Mismatch (PM) Number Mismatch (NM)

Los cocineros3.pl cocinaron3.pl un arroz muy rico The cooks3.pl cooked3.pl very tasty rice *El cocinero3.sg cocinaste2.sg un arroz muy rico The cook3.sg cooked2.sg very tasty rice *El cocinero3.sg cocinaron3.pl un arroz muy rico The cook3.sg cooked3.pl very tasty rice

Evidence for Feature-Specific Processing

57

As shown in Figure 3-3, consistently with the hypotheses outlined above, ERPs showed a reliable dissociation between the two types of anomalies. In the 300–500 milliseconds window, the difference between person and number agreement can be ascribed to the different anchors evoked: in the presence of a number anomaly, the inconsistency detected between nominal and verbal specifications is circumscribed to the local s-v agreement configuration, hence within the purely morphosyntactic level of analysis, as the interpretive anchor invoked is the number specification on the subject argument. This emerges in terms of a LAN effect. When a person violation is present, the inconsistency detected between subject and verb invokes an interpretive anchor that is outside of the local s-v configuration, i.e. in the speech act representation. Unlike number’s anchor, that of person cannot supply a correct cue to repair the sentence. Given the autonomous and interpretable status of nominal and verbal person, both specifications and their respective underlying discourse representation may in principle be correct. A persistent difficulty in assigning a coherent interpretation to the structure therefore arises, resulting in the impossibility of interpreting and integrating person information. An N400 effect is therefore elicited. The N400 effect for person violations represents a novel finding within the ERP literature on agreement processing and therefore deserves special attention. Of relevance for the interpretation of these data and for the feature-based approach developed here are findings from Hindi, a headfinal split-ergative language whose case-marking and tense-aspect systems seem to provide important insights into feature licensing mechanisms. In Hindi, a distinction must be operated between ergative-absolutive marking, employed in past perfective sentences, and nominativeaccusative marking, which is employed elsewhere, i.e. in imperfective and non-past sentences. The case marking of Hindi subjects therefore varies on the basis of verbal aspect and is fundamental for the interpretation of the subject’s thematic role: while a nominative argument can be either the agent in a two-argument predicate or the agent/undergoer in one-argument predicates, an ergative-marked argument can only receive an agentive reading. Violating case/aspect association thus entails violating thematic role assignment as well. Choudhary et al. (2009) showed that ill-formed ergative-imperfective and nominative-perfective sentences produced an N400 effect, which they interpret as evidence for the violation of interpretively relevant linguistic rules (as opposed to agreement violations, which, in their view, are supposed to be interpretively-irrelevant and therefore LAN generating). In keeping with the approach developed here, one may plausibly hypothesise that the N400 effect elicited by case

58

Chapter Three

violations in Hindi can be the result of an unsuccessful feature-anchoring operation. Similarly to person and tense, case features are also interpreted by establishing a linking relation with other types of features (or heads) encoded in the syntactic tree, namely voice and aspect (Sigurdsson 2009). Case marking in Hindi is thus anchored to aspect, much in the same way as nominal and verbal person specifications are anchored to the speech participants features. An unsuccessful case-aspect matching would thus produce the same interpretive consequences as a person-speech participant mismatch.1 Hindi ergative-absolutive and nominative-accusative case patterns are therefore syntactic predictors of the verbal tense/aspect marking, which, because of the head-final nature of this language, can only be directly accessed at the end of the sentence. As in many other languages, in Hindi, past tense adverbials can also work as tense/aspect predictors. More precisely, they can be seen as specific semantic cues for a past tense/perfective aspect verbal morphology. Interestingly for the current purposes, Dillon et al. (2012) showed that when verb tense was predicted with a past tense adverbial, an incorrect future tense form elicited an N400-like effect, followed by a broadly-distributed P600 effect, a pattern that closely resembles the one found in the comparison between person and number anomalies in the current experiment. A parallel can therefore be drawn between person and tense, whose licensing is crucially determined by an interpretive anchor located neither on the adverb nor on the verb, but presumably on the Speech Time head, (cf. Sigurdsson 2004). Hence the similar effects elicited by the two violations.2 It should be noted that the same patterns of ERP effects have been found in Italian (De Vincenzi et al. 2003) by comparing number violations with semantic-pragmatic violations (Il nuovo capotreno fischia/*germoglia alla partenza della locomotiva, The new guard whistles/*sprouts at the departure of the locomotive). In particular, number violations were found to elicit a LAN-P600 pattern, as opposed to the N400-P600 one generated by semantic-pragmatic anomalies. The N400 effect found in the presence of person agreement anomalies is at odds with previous findings from Spanish (Hinojosa et al. 2003; Silva-Pereyra and Carreiras 2007) and German (Rossi et al. 2005), in which a LAN effect was reported. The use of different procedures and materials across experiments may explain the divergent pattern of results. As noted by Mancini et al. (2011), an important aspect that cannot be neglected for a proper interpretation of these findings is the fact that the negativity elicited by person anomalies may be the result of a superimposition of LAN and N400 effects. This analysis is supported by

Evidence for Feature-Specific Processing

59

two facts. Firstly, the two conditions did not differ in frontal regions (see Mancini et al. 2011a for detailed statistical analyses). Secondly, the topographically distinct negativities elicited by person and number do not differ either in latency or onset, but emerge in the same temporal window. In processing terms, this would imply that the two features rely on computationally equivalent feature-checking mechanisms, as hypothesised in the FIP outlined in Chapter 1. The two violations differ also in the early (500–800 ms) and late (800– 1000 ms) phase of the P600. In this respect, the results from Mancini et al. (2011a) can inform the debate centred on the functional interpretation of this effect. In the 500–800 ms interval, both person and number agreement give rise to a P600 effect. Interestingly, while in posterior areas the two conditions do not differ, in frontal sites an increase in the P600 effect can be found for person agreement. These data seem to be naturally compatible with functional interpretations of the P600 that posit a dissociation of diagnosis processes. In particular, purely formal and discourse-related processes appear to surface with topographically distinct positivities. The posteriorly-distributed P600 effect found for both anomalies naturally aligns with previous findings on s-v agreement processing in Spanish and other languages (Barber and Carreiras 2005; Hagoort et al 1993; Hinojosa et al. 2003; Nevins et al. 2007; Osterhout and Mobley 1995; Silva-Pereyra and Carreiras 2007), where this effect can be explained in terms of diagnosis (Barber and Carreiras 2005; Carreiras et al 2004; Molinaro et al 2008) and syntactic integration difficulty in general (Kaan et al. 2000; Friederici et al. 2002). Both anomalies cause similar morphosyntactic integration problems, as suggested by the absence of differences between the two conditions in the posterior areas. However, due to incompatibility of the speech participant representation underlying nominal and verbal morphology, a person violation may also cause discourse integration problems, which would be reflected by an increase in the frontal portion of the P600 effect. Such an explanation would fit with the functional interpretation given to this effect by Kaan and Swaab (2003), who indicate discourse-related complexity as a possible explanation for frontal P600 effects. In other words, when diagnosis is performed, the parser would find that the two violations affect distinct levels of analysis. In the 800–1000 ms window, both anomalies produce a P600 effect that has its maximum in posterior sites. Differences between number and person emerge in terms of the amplitude of this effect, which appears to be larger for the person agreement. Assuming that in this time window a repair process is operative (Barber and Carreiras 2005; Bornkessel and

60

Chapter Three

Schlsewsky 2006; Carreiras et al. 2004; Friederici 2002; Hagoort et al. 1999), the larger P600 effect for person agreement may evidence a costlier repair , reminiscent of the one reported for gender vs. number agreement in Spanish (Barber and Carreiras 2005) and phonotactic vs. gender violations in Italian (Molinaro et al. 2008). The amplitude of the P600 seems therefore to be sensitive to the type of information to be repaired. The repair operation triggered by a person mismatch may be more complex, as a correct person specification cannot be reliably recovered to interpret the sentence either in the local subject-verb configuration or in the speech act representation. In contrast, the smaller P600 effect for the number mismatch would evidence a more straightforward repair of the mismatch based on the number specification of the subject argument. In sum, the ERP data just described add another piece of evidence in favour of feature dissociations. The processing system distinguishes between person and number agreement, and the difference between these two types of agreement is not only quantitative (i.e. distinct processing cost) but also qualitative (i.e. different mechanisms involved). More specifically, what differentiates person and number is not the operation underlying their morphosyntactic checking, i.e. Agree, which has the same formal character for both features, but the different position targeted by the Anchor operation.

Testing the FIP: Neuroanatomical correlates There is yet another aspect of the investigation of the Person-Number Dissociation Hypothesis that deserves to be discussed, namely the dissociation found between these two features at the neuroanatomical level. The relevance of this investigation resides in the opportunity that it offers to correlate the performance of Agree and Anchor with neuroanatomical bases and to assess whether any feature-level differentiation can be found. Feature-specific mechanisms are not considered in current sentence processing models, despite their relevance for an in-depth functional characterisation of the language network. Mancini and colleagues (Mancini et al. 2017) sought to fill this gap by investigating the neuroanatomical substrates involved in the construction of agreement dependencies and their interpretations, capitalising on the distinction between person and number. To this end, agreement violations of the type indicated in Table 3-2 were used in an event-related fMRI paradigm. One important advantage of the fMRI paradigm used in Mancini and colleagues (2017) is the opportunity it gives to isolate neural correlates associated with the processing of both correct and incorrect agreement,

Evidence for Feature-Specific Processing

61

thanks to the bidirectional contrasts between conditions that were carried out. One can therefore delineate detailed predictions concerning brain activation patterns for (i) correct agreement vs. agreement violations (i.e. when person and number violations are collapsed); (ii) for agreement violations vs. correct agreement; (iii) for each violation type vs. correct agreement and, finally, (iv) for person vs. number violations. Under the assumption that the processing system attempts to assign an interpretation to both a congruent and an incongruent sentence, it will extract morphosyntactic information for checking and anchoring purposes in both cases, although only in one case (correct agreement) are these operations smoothly carried out. One can therefore expect agreement violations to show partially common activation patterns with correct agreement. Existing work (Kuperberg et al. 2003, 2008; Quiñones et al. 2014) point to the LIFG and the MTG as the cortical areas primarily involved in the comprehension of correct sentences. Within this network, the finding of differential responses to person and number mismatch sentences would evidence that feature-specific mechanisms are at work in the building and interpretation of a sentential relation, in line with theoretical analyses (Baker 2008; Bianchi 2006; Mancini et al. 2013; Sigurdsson 2004, 2008). If this is on the right track, one can hypothesise two possible ways in which the person-number dissociation can arise. Firstly, quantitative differences can emerge between person and number anomalies in brain areas that are thought to support the extraction of features to build syntactic structure, as a result of the different position the two features occupy in a syntactic tree (Baker 2008; Bianchi 2006; Shlonsky 1989; Sigurdsson 2004, 2008). In this respect, the posterior portion of the MTG/STG represents a potential candidate. Secondly, qualitative differences are expected to emerge in cortical areas that support the mapping between morphosyntactic and semantic-discourse information, given the critical interpretive consequences that determining the discourse role of an argument has. In this case, one can expect selective involvement of the anterior portion of the MTG/STG (Bornkessel-Schlesewsky and Schlesewsky 2013; Lau et al. 2013) and the pars opercularis/orbitalis of the IFG (Friederici 2011) for person but not for number agreement. The contrast between the two violations and the correct sentences permits the identification of the neural substrates involved in the checking of feature consistency between subject and verb. In this respect, a potential candidate is represented by the LIFG. If person and number agreement share similar feature-checking mechanisms, the superimposition of the contrast between person mismatch and correct agreement, and between

62

Chapter Three

number mismatch and correct agreement should show overlapping patterns of activation in middle frontal areas. The consequences of processing a morphosyntactic violation cannot be solely linguistic, since its mere detection implies the involvement of extralinguistic, domain-general components, such as the conflict-monitoring and attention systems at a minimum. While conflict-monitoring mechanisms are arguably involved in the processing of both correct and incorrect agreement dependencies, they are however expected to yield a stronger and more sizeable effect in incorrect than in correct sentences. Recruitment of the dorso-lateral prefrontal cortex and the ACC bilaterally is therefore expected to emerge from the comparison between the two mismatch conditions and correct agreement. Existing work on the processing of morphosyntactic dependencies has consistently demonstrated the involvement of this network as a result of a (task-driven) mismatch detection process (see Van den Meerendonk et al. 2009 for a review; Kuperberg et al. 2008; Nieuwland et al. 2012; Quiñones et al. 2014 and Chapter 2). The results point to the comprehension of an agreement dependency as a composite algorithm made of feature-sensitive operations of structure projection and mapping of information. As shown in Figure 3-4, on the one hand, the comprehension of correct agreement sentences (as revealed by the correct agreement > agreement violations contrast) recruited a network that included the anterior portion of the MTG, the posterior MTG and the LIFG. On the other hand, sensitivity to incorrect agreement (as revealed by the incorrect agreement > correct agreement contrast) emerged in a widespread bilateral fronto-parietal network. In line with the hypotheses laid out above, the direct contrast between person and number violations revealed both qualitative and quantitative differences. In particular, a clear quantitative dissociation between person and number agreement emerged in the posterior MTG, while activation patterns in the anterior MTG, as well as in the pars orbitalis and triangularis of the LIFG, suggests a qualitative difference between the two features. Closer inspection reveals that the regions where the person-number dissociation is found belong to the fronto-temporal network recruited by the comprehension of correct agreement. Crucially, this finding suggests that these areas, arguably involved in the building and interpretation of sentential relations, operate in a feature-specific manner. In contrast, the failure to find a dissociation between person and number agreement in areas associated with the processing of incorrect agreement (middle frontal and parietal regions) is indicative of the fact that feature consistency checking and conflict-monitoring mechanisms operate in a feature-insensitive manner. Let us discuss these findings in relation to the FIP.

Evidence for Feature-Specific Processing

63

Consistently with studies and models that assume predominant involvement of the posterior MTG in the extraction of information from linguistic input to build syntactic structure (Hagoort 2005, 2008, 2014; Hickok and Poeppel 2007; Molinaro et al. 2015; Pallier et al. 2011), the quantitative dissociation that emerged between person and number in this area is compatible with the view that the two features occupy different structural positions in the syntactic tree. More precisely, the differential activation for person and number anomalies in this area could result from the different positions in the syntactic structure from where information has been retrieved. Extraction of person information from a higher node compared to number may yield greater processing costs and hence, greater activation of areas supporting this processing mechanism. In contrast, the asymmetry between person and number in the anterior MTG is revealing of important interpretive differences between the two features. Activity in this region has been associated with the building of syntactic structure (Friederici 2011; Friederici and Gierhan 2013), as well as to the analysis of sentence propositional meaning (Bornkessel-Schlesewsky and Schlesewsky 2013), in line with a larger literature that attributes this region a crucial role in semantic memory and conceptual combination (Baron et al. 2010; Patterson et al. 2007). The finding of differential activation profiles for correct and incorrect agreement sentences suggests that this region does not support local syntactic structure building. Indeed, if this were the case, similar profiles would have emerged for the two types of sentences, which critically share the same underlying syntactic structure (regardless of morphosyntactic consistency). In the context of the FIP, the qualitatively different pattern of activation elicited by person compared to number anomalies can be linked to the asymmetry that characterises these two features at the interpretive level. While identifying and assigning a discourse role to the subject is crucial for the derivation of the overarching meaning of the sentence, the identification of whether this argument refers to a single entity or to a multitude of entities is not. Moreover, this finding and its interpretation accord with the N400 effect found for person violations in the ERP study and the hypothesised role of the anterior portion of the left temporal lobe in the generation of the N400 component (Lau et al. 2013, 2008), although further evidence is necessary to validate this hypothesis. Finally, the patterns of activation emerging in frontal areas cast important light on the mechanisms supporting feature consistency checking across sentential constituents. In particular, the analysis of correct agreement (minus agreement violations) and agreement violations (minus correct agreement) revealed a clear-cut dissociation between regions

64

Chapter Three

selectively involved in the analysis of correct dependencies – pars orbitalis and triangularis of the LIFG – and regions that more predominantly responded to agreement inconsistencies, i.e. middle frontal areas. Critically, the functional significance of this dissociation can be better understood by considering the activation profile for the person > number contrasts in the same regions, from which a dynamic interaction of featuresensitive and feature-insensitive checking mechanisms emerges. The pars orbitalis and triangularis of the LIFG appear to be significantly responsive to the presence of correct agreement dependencies, but also to person violations when directly contrasting the two violations. Activity in these two regions has been associated with the analysis of meaning at the sentence level (Friederici 2011; Friederici and Gierhan 2013; Vigneau et al. 2006). In this context, activity in the pars opercularis and triangularis could reflect a constant and incremental evaluation of the semanticdiscourse fit of the elements being processed, i.e. the matching of subjects and verbs in terms of sigma features, to assess whether the two elements can be integrated in a meaningful conceptual and discourse representation. Greater activation for person compared to number anomalies could be due to the greater relevance that the former feature has for the analysis of a sentence’s overall meaning. This interpretation is corroborated by the mirroring activation profile found in the anterior MTG, which is neuroanatomically linked to the pars triangularis-orbitalis by a ventral pathway (Friederici 2011). As mentioned, a different scenario emerged in middle frontal areas, which showed predominant activation for incorrect agreement dependencies, regardless of the type of features manipulated. Several studies manipulating morphosyntactic information have reported involvement of this region in the analysis of inconsistencies (Folia et al. 2009; Kuperberg et al. 2003, 2008; Nieuwland et al. 2012, see Chapter 2), but also in domain-general working memory mechanisms (see reviews in Katsuki and Constantinidis 2012; Rogalski and Hickok 2011). In the context of agreement anomalies processing, this region could be plausibly supporting memory mechanisms involved in the morphosyntactic consistency checking mechanisms that are performed during the processing of agreement. This interpretation is in accord with the common left anterior negative effect elicited by both person and number agreement violations, and confirms the hypothesis about the presence of common phi-featurechecking operations for the two types of agreement. The “division of labour” that emerges between the inferior and the middle frontal gyrus can inform the ongoing debate concerning the domain-generality/specificity of the functions subserved by frontal areas.

Evidence for Feature-Specific Processing

65

The dissociation between correct and incorrect agreement evidenced in these area during agreement processing can plausibly result from the interplay between language-specific mechanisms that incrementally integrate linguistic information coming from temporal areas, and domaingeneral processes that ensure the detection of possible mismatches between perceived and expected input (see Fedorenko et al. 2012 for a relevant discussion on LIFG subregions and their functions). Overall, what these data suggest is that the comprehension of agreement dependencies relies on a left-lateralised fronto-temporal network, where a tight interplay between domain-general and linguistic proper functions permits incremental integration of morphosyntactic and semanticdiscourse information.

Summary The current chapter set out to test two theoretical approaches to s-v agreement: the analysis given to this phenomenon within the Minimalist Program (Chomsky and subsequent work) and the one outlined by cartographic studies (Shlonsky 1989, 2010; Sigurdsson 2004; Sigurdsson and Holmberg 2008, inter alia). The two approaches diverge on two fundamental points: (i) the different structural representation attributed to agreement features, i.e. clustered and structurally undifferentiated under the same T head in the former account, while structurally separated in the latter; (ii) the checking mechanism to be adopted in each account, namely a single operation that checks all of the features in the T head, as opposed to distinct checking operations that deal with each feature separately, which makes it possible to capture the intrinsically different interpretive properties – or anchors – associated with each feature. Three studies were presented that were aimed at testing the so-called Person-Number Dissociation Hypothesis and that relied on similar experimental designs, which were implemented with the use of three different techniques. The self-paced reading paradigm made it possible to assess whether person and number agreement, when violated, gives rise to quantitatively different processing costs. Using ERPs, it was possible to ascertain whether qualitative differences arise in the processing of person and number agreement violations, as a result of the distinct interpretive properties that characterise these features. Finally, the use of fMRI allowed us to identify the neuroanatomical bases of agreement processing in general and person and number agreement specifically. The results from the three studies are concomitant in that a dissociation between the two features is possible not only at a theoretical but also at a

66

Chapter Three

processing level. In this respect, it appears reasonable to link quantitative differences between person and number agreement processing, as revealed by self-paced reading data and by brain activation patterns in the posterior MTG, to the distinct position in sentence structure that the two features occupy. Evidence for formally similar morphosyntactic checking mechanism (Agree) comes from the similar left anterior negative effects that both person and number agreement violation elicit, as well as from the common activation pattern found in middle frontal areas. Finally, the qualitative differences expected for the performance of anchoring mechanisms were reflected in the clearly distinct electrophysiological (N400 vs. LAN) and neuroanatomical (different activation profiles in pars opercularis/triangularis and anterior MTG) correlates elicited by person compared to number violations. All this contributes to conceptualising agreement comprehension as a feature-based set of representations and algorithms that selectively involve a left-lateralised fronto-temporal network.

Figure 3-1. Grand-averageed ERPs (time-locked to thee verb presentaation) and n and perrson agreement condition (relaative to the topographicall maps for the number control condiition). Negativee voltage is plotted upwards. A Adapted from Mancini M et al. (2011a).

Upper panel: siignificant activaation clusters reesulting from th he contrast Figure 3-2. U person mismaatch + number mismatch > co orrect agreemennt (in yellow) and a correct agreement> pperson mismatcch + number miismatch. Middlle panel: superiimposition of the significcant activation clusters resultin ng from the conntrasts person violation v > correct agreeement (in red), number violation > correcct agreement (in ( green). Yellow voxeels represent siignificant activ vated voxels inn both contrasts. Lower panel: signifiicant activationn clusters resultiing from the coontrast person violation v > number violattion. Adapted from f Mancini ett al. (2017).

Figure 4-3. E ERP for 1st/2nd peerson and 3rd peerson number vioolations (in red)) relative to their correct coounterparts (in black). b Negative voltage is plotteed upwards.

Figure 5-1. E ERPs and topoggraphical maps associated a withh unagreement relative r to standard agrreement (left panel) and person mismatcch relative to o standard agreement (rright panel). Negative N voltag ge is plotted uupwards (Adap pted from Mancini et al. 2011).

Figure 5-2. Significant acttivation clusterrs resulting froom the contrasst between standard agreeement and persson mismatch (upper ( panels), unagreement and a person mismatch (miiddle panels) annd superimposition of the signnificant activatio on clusters resulting from m the contrastss between perso on mismatch/unnagreement and d standard agreement. A Adapted from Quuiñones et al. (2 2014).

CHAPTER FOUR MORE ON FEATURE-SPECIFIC PROCESSING: PERSON ASYMMETRY

Across languages, an asymmetry among pronominal forms has been outlined that differentiates 1st and 2nd person from 3rd person, in virtue of the different featural makeup that characterises the former compared to the latter pronouns. Due to their intrinsic reference to the participants in the speech act (speaker and addressee), 1st and 2nd person are typically regarded as context-related forms specified for the person features. In contrast, because they refer to individuals that are being talked about and that therefore do not bear active speech roles, 3rd person pronouns (as well as lexical DPs) are said to be specified only for the number feature. The goal of this chapter is to test the so-called Person Asymmetry Hypothesis to verify whether, during online processing, the interaction between person and number interpretive anchors shapes the comprehension mechanisms associated with 1st/2nd and 3rd person agreement. Before detailing the design and the results of the experimental investigation, a review of linguistic data and analyses that support the Person Asymmetry Hypothesis is provided.

The Person Asymmetry Hypothesis It has long been recognised that a fundamental difference exists between 1st and 2nd person on the one hand and 3rd person pronouns on the other, with this distinction mainly residing in the different individuals and roles picked up by these pronominal forms. This distinction was first noticed by Benveniste (1966) and Forcheimer (1953), who provided distinct but convergent evidence in favour of it. Forcheimer (1953) identified morphological generalisations in support of a split between 1st/2nd and 3rd person, such as the fact that 3rd person agreement is often zero while 1st and 2nd person is normally overt, or that 3rd person is more subject to objective subdivisions such as class, gender and location than 1st and 2nd person are. Adopting a more pragmatic perspective, Benveniste

68

Chapter Four

(1966) and Jakobson (1971) attributed the split to the distinction between participants and non-participants in the speech act. At the heart of this distinction lies the intuition that in the former case, identity with (or inclusion of) a speech participant is expressed, while in the latter case, the individual identified is neither a speaker nor an addressee, but only the entity being talked about. In other words, the split is meant to differentiate elements indicating coincidence (1st and 2nd person) from elements indicating non-coincidence (3rd person) with the participants in the speech event. In a similar fashion, Silverstein (1985) proposed a hierarchy of person values such that within pronouns, 1st and 2nd person dominate over 3rd person (1st/2nd>3rd), on the basis of the different feature composition characterising the two classes of pronouns. On the one hand, 3rd person pronouns bear the same features as lexical DPs do (e.g. animacy, gender, countability, etc.), and thus to different degrees preserve the lexical properties of the underlying nominal expression. On the other hand, the same cannot be said for 1st/2nd person pronouns, which do not stand for lexical DPs, but rather denote and index participants in the speech act. Evidence for the existence of this split in the pronominal system of natural languages has been found, and a number of proposals have been outlined that capitalise on the fact that deixis plays a role in morphology. An influential analysis from this perspective is represented by Harley and Ritter’s (2002) Feature Geometry. Thanks to the analysis of a large data set from various languages, Harley and Ritter implemented a feature geometric analysis of pronominal forms that accounts for the intrinsic differences postulated to exist between 1st/2nd and 3rd person, as illustrated in Figure 4-1. In this geometry, all nominal features are dependent upon a root node called Referring Expression. The PARTICIPANT node and its dependents are used to represent person, specifically 1st and 2nd person. The INDIVIDUATION node and its dependents – Group (plural) and Minimal (marked singular, or dual in combination with group), as well as Augmented (trial/paucal) – are used to represent number systems. Finally, the CLASS node encodes gender and other class information. The primary division of the Referring Expression node is into person and number features, corresponding to Participant and Individuation. First and 2nd person, as participants in the speech act, have a significantly different status compared to 3rd person, which, according to Harley and Ritter (2002), is simply the absence of person. Thus, 1st person is represented by a bare, underspecified Participant node, which receives the default interpretation of speaker. Second person is represented by a Participant node with an addressee dependent.

More on Feature-Specific Processing: Person Asymmetry

69

Figure 4-1. Feature geometry (Adapted from Harley and Ritter 2002)

Inclusiveness is represented when both the speaker and addressee nodes appear as dependents on participant. Number, specified by the individuation node, is encoded as singular by a bare individuation node. Plural is instead expressed by an individuation node with a group dependent. The individuation node thus represents those features of a DP (pronominal and non-pronominal) which are independent of discourse, i.e. number and gender information. The neat distinction between person and number information and between 1st/2nd and 3rd person that underlies Harley and Ritter’s geometry follows a specific criterion, namely discourse-dependence. Importantly, discourse-dependence is here understood as identification with a speech role: this makes it possible to distinguish 1st and 2nd from 3rd person, with the former manifesting a tight relation to specific speech roles that the latter lacks. As will be shown later, such a criterion may be too restrictive, leading to a failure in accounting for certain syntactic and distributional facts concerning 3rd person. Harley and Ritter’s (2002) geometry captures Benveniste’s intuition that 3rd person is in fact not a true personal form: person belongs to I and You, while it is lacking in He (Benveniste 1966). What makes I and You different from He/She/It is thus the specification for person present in the former but absent in the latter forms, which are specified for number (Anagnostopoulou 2003; Harley and Ritter 2002; Kayne 2000, inter alia). This becomes particularly clear when examining the relationship between singular and plural pronouns. The pronoun I, which identifies the speaker, refers to an entity that is unique within discourse, and the same holds for You, the addressee. I and Yousing’s uniqueness makes it impossible to consider I and Youplur as augmentative forms, as mere multiplications of their singular counterparts. Rather, they denote groups that result from the combination (or association) of different individuals bearing different speech act roles (Table 4-1). In other words, We and Youplur have an associative, rather than plural, meaning (Cysouw

Chapter Four

70

2003): the prototypical meaning of We can be indicated as ‘I and my associate(s)’, in which the associate can be either the addressee or another entity (ex: We = I + you/I + he ), while the meaning ‘You and your associate(s)’ can be applied to Youplur (ex: Youplur = yousg + he/they).1 It follows that the word ‘plural’, when part of the terms 1st and 2nd person plural, is nothing but a misnomer (Benveniste 1963; Cysouw 2003; Harley and Ritter 2002; Wechsler 2004). On the contrary, in shifting from He/She/It to They, no modification of the speech participants’ makeup occurs, as only individuals with no speech roles are involved. Within a speech event, only one speaker and one addressee can be present, but there can be more than one entity being talked about: hence the availability of a true plural form for 3rd person and its specification for the number feature. Table 4-1. Speech Participant composition of 1st, 2nd and 3rd person pronouns. Person 1st person 2nd person 3rd person

Singular Speaker Addressee

Plural Speaker + Addressee Speaker + Other(s) Addressee + Other(s)

Other, entity being talked about (non-participant)

Others, entities being talked about (non-participants)

1st/2nd vs. 3rd person: Person underspecification and context-dependence A number of typologically unrelated languages present morphological and configurational person splits that are often accounted for in terms of person underspecification of the element involved. Manifestations of this phenomenon are the so-called Person Case Constraint (or PCC, Bonet 1991) and the agreement patterns of some American Indian languages, but also the morphological realisation of pronouns and their distribution. What follows is a review of these phenomena.

Morphological evidence for subject splits In his analysis of French and Italian person morphemes, Kayne (2000) shows that in these two languages, a distinction in terms of feature specification can be drawn between 1st/2nd person clitics and reflexive sclitics on the one side, and l-clitics on the other side. Kayne observes that while l-clitics (French: lemasc, lafem ; Italian: lomasc, lafem) contain a word marker for gender, 1st and 2nd person morphemes (French: me1st , te2nd;

More on Feature-Specific Processing: Person Asymmetry

71

Italian: mi1st , ti2nd) are pure person morphemes with an epenthetic vowel (Table 4-2). Moreover, l-clitics regularly inflect for plural (les, li), while 1st and 2nd person clitics do not (see nous/vous; ci/vi). Based on this alternation, Kayne (2000) concludes that l-clitics are not specified for person, while 1st and 2nd person morphemes are. Table 4-2. Person and number composition of 1st/2nd person clitics and l-clitics in Italian and French based on Kayne (2000) French l-e l-a m-e t-e s-e l-es l-es nous (*m-es) vous (*t-es) *s-es

Italian l-o l-a m-i t-i s-i l-i l-i ci(e) vi(e) -

Person f.sg 1 2 3 1 2 -

Number Sg Sg Sg Sg Sg Pl Pl Pl Pl -

Gender Masc Masc Masc Masc -

A split is also observed in Hebrew pronouns by Ritter (1995), who distinguishes 1st/2nd person pronouns from 3rd person ones on the basis of their underlying feature specification and their distribution. Given 1st/2nd person pronouns’ complementary distribution with the definite article ‘ha’, they are to be considered DPs with only the functional head D and a full feature set. In contrast, 3rd person pronouns are not DPs, because they can combine with the definite article ‘ha’ to form remote demonstratives. Therefore, they are to be regarded as number elements under the Num head, as in Figure 4-2. The opposition is thus one between full-fledged pronominal forms like 1st and 2nd ones and pronominal forms that only have a number specification. The French and Hebrew data thus seem to suggest that there is no such thing as a 3rd person feature: 3rd person inflections are simply unspecified for person.

Figure 4-2. A. Person split in 1st and 2nd person pronouns in Hebrew. B. Person split in 3rd person pronouns and DPs (Ritter 1995).

Chapter Four

72

Agreement patterns showing a configurational person split A closer inspection of the Hebrew agreement system evidences a dissociation between 1st/2nd and 3rd person agreement. Null subjects in Hebrew are allowed only with past and future tense, and with the negation eyn. In this latter case, a person asymmetry is found. The verbal form of eyn sentences is a participle that, because of its defective nature, cannot licence agreement in 1st and 2nd person, which are thus expressed on the negative head eyn. Crucially, all combinations of number and person agreement on eyn yield grammatical null-subject sentences, with the only exceptions being 3rd person singular and plural, as (1) and (2) below illustrate. (1) *eyn-(ԥn)-o mԥdaber rusit speakmasc Russian Neg3.sg ‘He does not speak Russian’ (2) eyn-(ԥn)-i mԥdaberet rusit Neg1.sg speakfem Russian ‘I do not speak Russian’ (from Shlonsky 2000)

The same behaviour is manifested by overt pronouns: in the presence of eyn, 1st and 2nd person pronouns can occur only above eyn, while 3rd person pronouns enjoy greater freedom and are allowed both above and below negation. The reduced acceptability of (4) compared to (5) and (3) below suggests that 1st and 2nd person pronouns occupy a position above negation, while 3rd person pronouns occupy a post-negative position. (3) eyn hi gveret Levi Not she Mrs Levi ‘She is not Mrs Levi’ (4) ?eyn ?ani/ ?at gveret Levi Not I / you Mrs Levi (5)

?

ani/?at eyn-ni/ex gveret Levi I/ you not-1.sg/2.fem.sg Mrs Levi I/you am/are not Mrs Levi’ (from Shlonsky 2000)

The person split in modern Hebrew can be analysed as an instantiation of the opposition between participants and non-participant-related

More on Feature-Specific Processing: Person Asymmetry

73

arguments (Shlonsky 2009). Shlonsky (2009) proposes an account of the split that treats 1st and 2nd person inflectional endings as overt subject clitics that represent the morphological correlates of a syntactic head, namely the Speech Act Participant head, or Sap (see also the discussion on Bianchi 2006 below), which is located in the left periphery of the sentence and to which the verb in T would adjoin. These subject clitics can be doubled by an overt pronominal located in the specifier position of SapP, which would explain the grammaticality of (5). In contrast, 3rd person endings have a different status: they are referentially dependent, and as such they must have an antecedent in the sentence that binds or controls them and can thus assign a person feature,2 as shown in (6) and (7). No such constraint is operative for 1st person null subjects, as (8) illustrates. (6) *eyn-o lomed Albanit Neg-3.masc.sg studypres.masc.sg Albanian ‘He is not studying Albanian’ (from Shlonsky 2000) (7) Talila ‘amra le-itamari Se proi hicliax Talila said to-Itamar that succededmasc.sg ‘Talila told Itamar that she succeded’ (from Ritter 1995, 434) (8) eyn-eni lomed Albanit Neg-1.sg studypres.masc.sg Albanian ‘I am not studying Albanian’ (from Shlonsky 2000)

A split between Sap and non-Sap persons has been reported in other languages. NIDs, in particular several Veneto dialects, show a positional dissociation between 1st, 2nd and 3rd person, respectively represented by the Hearer Subject Clitic Clause (SCL), Speaker SCL and Numb SCL, as in (9) below (from Poletto 2000): (9) [LDP invSCLi [CP deicSCL [FP ti [IP[NegP[NumbP SCL [hearerP SCL [SpeakerP Inflv [TP]]]]]]

In Poletto’s analysis, the pre-negative field corresponds to the CP system (following Rizzi 1997), while the post-negative field is identified with the inflectional layer of clause structure, the IP system. Apart from invariable SCLs (invSCLs), which are the same for all persons and thus do not imply any participant vs. non-participant distinction (e.g. ‘a’, as in A vegni mi / I come, or A ta vegnat ti / You come, in the Lugano dialect), all

74

Chapter Four

other SCLs encode some subject feature. A clear distinction between 2nd and 3rd person is found in the post-negative field, as represented by Hearer SCLs and Number SCL, respectively occupying the head of the Hearer Phrase and of the Number Phrase. As for 1st person, a gap is found since no specific SCL encodes the [±Speaker] distinction, as a result of the SCL agglutination to the verb, which has moved to the Speaker Phrase projection the of the SCL. Crucially, the subject features encoded in the pre-negative field are not the same as those realised in the post-negative one: one finds fine-grained distinctions of person, number and gender lower than NegP, whereas above NegP, 1st and 2nd person are distinguished from 3rd person only by means of a [+deictic] feature for which the former but not the latter clitic forms would be specified. Configurational person splits have also been observed in the verbal agreement morphology of American Indian languages. In Algonquian languages, the position of agreement markers switches according to the relative animacy of the arguments, in accordance with the Animacy Hierarchy to which these languages are sensitive. When two arguments occur, the argument higher in animacy appears as a prefix on the verb, while the argument lower in animacy appears as a suffix (Linn and Rosen 2003). In Athapaskan languages, subject splits are evident in the position occupied by 1st, 2nd and 3rd person agreement markers on the verb. More specifically, 3rd person agreement markers appear in a different position on the verb form than that of the 1st/2nd person marker: the latter appears next to the verb stem, while the former appears next to the object marker. This is suggestive of the presence of two subject positions: one for 1st and 2nd person, and one for 3rd person (Rice 2000). Similarly, two subject positions are identified by Linn and Rosen (2003) in their analysis of agreement in Euchee, a language originally spoken in the southeast of North America and now circumscribed to Oklahoma. Like other American Indian languages, Euchee is largely polysynthetic and with a rich verbal agreement. Here, 1st and 2nd person pronouns agree with the verb in person and number, while 3rd person ones agree only in gender and number. This strongly suggests a different feature specification for the two types of pronouns: person and number specification in the former case, number and gender in the latter. Also, 1st/2nd person agreement markers appear on the verb in a different position compared to 3rd person ones, as happens in Athapaskan languages. Linn and Rosen (2003) account for this split by adopting a feature-checking approach and proposing that the person and gender features of Euchee subject pronouns are checked in different positions: gender would be

More on Feature-Specific Processing: Person Asymmetry

75

checked in a dedicated projection, namely GenP, located above vP, while person would be checked in AgrSP, in a higher position. In a recent analysis, Baker (2008) posits the presence of a condition – the so-called SCOPA (Structural Condition on Person Agreement) – that rules the occurrence of 1st/2nd and 3rd person agreement in a variety of typologically unrelated languages. Baker’s SCOPA can successfully account for the configurational person splits found not only in American Indian languages but also in Romance and Germanic ones. Baker observes that 1st/2nd person agreement differs from 3rd person agreement in terms of locality: in the former case, for agreement to be properly displayed, subject and verb must be in a strictly local configuration within the same functional projection (TP, or FvP, as Baker calls it), with the 1st/2nd person argument occupying the specifier position. It is only within this configuration that 1st and 2nd person agreement is allowed (a pattern that is reminiscent of the configuration in which Hebrew 1st and 2nd person agreement is licensed in the presence of the negation eyn). In contrast, as shown by (10) below, 3rd person agreement can be realised both locally and at a distance: the verb in T can agree with a 3rd person argument that occupies a position other than SpecTP, for example a relative pronoun in CP: (10) The people whopl Clark thinkpl are in the garden3 (from Kimball and Aissen 1971, quoted in Baker 2008, 86)

Another relevant case discussed by Baker (2008) concerns verb agreement with a quirky subject in Icelandic, as illustrated in (11)–(13) below. Icelandic presents various dyadic predicates in which the thematically higher argument (normally the experiencer or a psych predicate) has a lexically specified case, often dative. Since the dative argument is the highest, it is raised to SpecTP to satisfy the EPP feature of T. However, T cannot agree with its quirky-case marked subject, as shown in (12), and it therefore agrees (only in number) with the lower nominative argument. For 1st/2nd person agreement to be expressed, a local spec-head configuration is necessary, while it is not required for 3rd person agreement. (11) Henni leiddust þeir Herdat was.bored.by3.pl theynom ‘She was bored with them (12) *Henni leiddumst við Herdat was.bored.by1.pl wenom ‘She was bored with us’

76

Chapter Four (13) Mér lík-ar (*lík-a) bók-in Idat like3.sg/*1.sg book-thenom ‘I like the book’ (from Taraldsen 1995; quoted in Baker 2008, 87)

The Person Case Constraint Evidence for a configurational split between 1st/2nd and 3rd person also comes from the analysis of the PCC (Bonet 1991). The PCC is a ban on the co-occurrence of 1st/2nd person dative and accusative elements such as clitics, agreement markers and weak pronouns. It is operative in a number of languages, including Italian, Greek, Spanish and Catalan. In these languages, the combination of a 1st/2nd person dative clitic with an accusative clitic is allowed only if the latter is 3rd person, as shown by the following Greek examples in (14)–(17) and those in Italian in (20)–(25) below. To account for the constraints showed in (14)–(17), Anagnostopoulou (2003) develops a feature-checking account that operates on the basis of the different featural specifications underlying 1st and 2nd person clitics with respect to 3rd person ones, in keeping with those claims that attribute to 1st and 2nd person the status of real persons, participants in discourse, as opposed to the non-participant status of 3rd person (Benveniste 1966; Harley and Ritter 2002; Jakobson 1971). Under this assumption, 1st and 2nd person and reflexive clitics are specified for person, while direct object (DO) 3rd person clitics are not, and are instead specified for number. (14) 1st IO – 3rdDO Tha mu to stilune Fut clgen.1.sg clacc.3.sg.neut send3.pl ‘They will send it to me’ (15) 2nd IO – 3rd DO Tha mu to stilune Fut clgen.1.sg cl acc.3.sg send.3.pl ‘They will send it to me’ (16) *2nd IO – 1st DO *Tha su me sistune Fut clgen.2.sg cl acc.1.sg introduce3.pl ‘They will introduce me to you’ (17) *3rd IO – 2nd DO *Tha tu se stilune Fut clgen.3.sg.masc cl acc.2.sg send3.pl ‘They will send you to him’

More on Feature-Specific Processing: Person Asymmetry

77

According to Anagnostopulou (2005), dative 3rd person clitics follow the same pattern as 1st and 2nd person ones in that they are specified for person, a characteristic following from their referring to human/animate entities that denote a point of view holder. Based on this, Anagnostopoulou analyses the PCC as the result of a competition between two arguments to be licensed by a single head. Because of a person restriction, agreement features cannot be checked simultaneously, and person is thus checked separately from number. This is technically implemented by postulating the presence of two vPs: a lower one that introduces the indirect object (vP-Intr), and a higher one that introduces the direct object (vP-Tr) and whose head possesses a full set of features. Being closer to v-Tr, it is the indirect object (IO) that moves first and checks its person feature in v-Tr. The DO clitic moves next, but at this point no person feature is left to be checked in v-Tr and it can only check number. The derivation converges if the DO clitic is 3rd person, because it does not have to check person, but if it is a 1st or 2nd person clitic, its person feature would be left unchecked and the derivation would therefore crash. Similarly to Anagnostopoulu (2005), Adger and Harbour (2007) explain the PCC as a syntactic phenomenon deriving from the way agreement features are distributed as checked/valued on functional heads. Specifically, they assume that 3rd person dative clitics are specified for person (and thus follow the pattern of 1st and 2nd person ones), while 3rd person accusative ones for number. More precisely, their analysis of the PCC in Kiowa (a Kiowa-Tanoan language of Oklahoma) hinges on the presence of a [participant: ] feature entailing semantic animacy that characterises 3rd person IOs,4 but not 3rd person DOs. In this latter case, only a [number: ] feature is present. A different approach is developed by Bianchi (2006) in her analysis of the PCC in Italian. The crucial assumption on which Bianchi’s implementation of the PCC rests is the presence of a connection between the PCC in languages like Italian and the Animacy Hierarchy of Inverse Systems (contra Anagnostopulou 2003). In Inverse Systems, argument features are given a relative ranking, as shown in (18), and the object must not outrank the subject in person, so that grammatical function prominence and Person/Animacy prominence can align. This way, the direct voice can be used, as in (19a). In the direct voice, the co-occurrence of a 1st/2nd person object and a 3rd person subject, or of an animate object and a nonanimate subject, is banned. A subject and a DO are allowed to co-occur as long as they are equally ranked in the Person/Animacy Hierarchy. When the logical object outranks the logical subject in the Person/animacy scale,

78

Chapter Four

the inverse voice (or passive) must be employed, as shown by the suffix ku in (19b): (18) Animacy Hierarchy: 1st person > 2nd person > 3rd person animate > 3rd person inanimate (19) a. K-ciksotuw-a-nnu-k 2- listen to- dir-1p- 3p ‘Weincl listen to them’ b. K-ciksota-ku-nnu-k 2- listen to- inv-1p- 3p ‘They listen to usincl’

According to Bianchi (2006), Italian has Animacy Hierarchy effects in the interaction of DO and IO, as evidenced in (20)–(25). On a closer look, one can see that: i) a 1st /2nd person DO cannot co-occur with a 3rd person IO, and that ii) a 1st /2nd person DO cannot co-occur with a 1st/2nd IO. These facts can be reformulated as an animacy-based constraint (Bianchi 2006): combinations in which the person specification of the DO outranks that of the IO are ruled out. The DO can instead co-occur with an equally ranked IO (as in ‘glielo’), or with a higher ranked IO (as in ‘me lo’ or ‘te lo’). In sum, in Italian we find effects of an Animacy Hierarchy that rank speech participants, i.e. 1st and 2nd person, above non-participants, i.e. 3rd person. (20) *1st DO – 3rd IO *Mi gli ha affidato cl1.sg cl 3.sg has entrusted ‘He has entrusted me to him’ (21) *2nd DO – 3rd IO *Ti gli ha affidato cl 2.sg cl 3.sg has entrusted ‘He has entrusted you to him’ (22) *1st DO – 2nd IO *Mi ti ha affidato cl 2.sg cl 3.sg has entrusted ‘He has entrusted me to you’

More on Feature-Specific Processing: Person Asymmetry

79

(23) 2nd IO – 3rd DO Te lo ha affidato cl 2.sg cl 3.sg has entrusted ‘He has entrusted him to you’ (24) 1st IO – 3rd DO Me lo ha affidato cl 1.sg cl 3.sg has entrusted ‘He has entrusted him to me’ (25) 3rd IO – 3rd DO Glielo ha affidato cl 3.sg cl 3.sg has entrusted ‘He has entrusted him to him

One could relate the effects of the Animacy Hierarchy in Italian to the person underspecification allegedly characterising 3rd person forms, as assumed in Harley and Ritter’s (2002) Feature Geometry. As Bianchi (2006) observes, there is more to subject splits than a simple (and radical) person underspecification. Specifically, she argues that by reducing subject splits to only a matter of person specification or undespecification on the basis of the presence or absence of a connection to a speech role, one ends up hiding important aspects of 3rd person forms, such as their definiteness. It is true that 3rd person lacks a speech role, but this should not obscure the fact that this pronominal form has a connection with the context of utterance, as evidenced by its definiteness. Third person is thus not completely discourse-unrelated. As a matter of fact, 3rd person makes reference to a contextually salient non-participant individual, either by deixis or anaphora. Based on this, Bianchi (2006) outlines the following feature representation for 1st, 2nd and 3rd person: (26) 1st person: [context-determined], [speaker] 2nd person: [context-determined], [hearer] 3rd person: [context-determined]

The novelty of this approach resides in the fact that all the three person specifications share a feature, namely the [context-determined] feature. The opposition between 1st/2nd and 3rd person is carried out on the basis of the presence/absence of a speech role associated with a given form5, and not simply by postulating person underspecification for 3rd person forms. It should be noticed that in Bianchi’s (2006) approach, the presence of a [context-determined] feature clearly presupposes an anchoring between the morphological realisation of pronominal forms and the speech

80

Chapter Four

act representation. The syntactic implementation given to this anchoring by Bianchi (2006) involves the presence of a Fin head that licenses the projection of a field of deictic heads that includes 1st, 2nd and 3rd person projections, as shown in (27): (27) [FP Force° [FinP Fin°…[SapP 1/2P [3pP 3p…[vP v…]]]]]

The Fin head represents the Logophoric Centre of the utterance (Bianchi 2003), while the field of deictic heads is expressed by the SapP. This makes it possible for person to be interpreted, so that the pronominal argument receives a value. The hypothesis of a syntactic representation of context has been advanced also in a proposal by Sigurdsson (2004)6 who, in line with Bianchi (2006), postulates the presence of a tight relationship – an anchoring – between person in the IP layer of the sentence and the speech act participants’ representation in CP. Sigurdsson captures the anchoring by means of a matching relation (Sigurdsson 2004, 27): (28) Ĭ = +Person = + ȜA - ȜP Ĭ = +Person = - ȜA + ȜP Ĭ = +Person = - ȜA - ȜP

1st person by computation 2nd person by computation 3rd person by computation

In essence, a matching must be established between event features in the lexical layer of the sentence and grammatical features in the inflectional area. These, in turn, must match logophoric features in the speech act representation, in relation to which they are interpreted. What (28) shows is that absence of speech role, as happens for 3rd person, does not imply absence of a connection to the speech act participants’ representation. Third person is anchored to the left periphery of the sentence as are 1st and 2nd person, the only difference being in the negative matching (- ȜA, - ȜP) characterising the former with respect to the latter forms. In other words, absence of a speech role is not absence of matching, it is just negative matching.

Pronoun representation and interpretive anchors: an electrophysiological investigation In Chapter 3, it was proposed that the difference in anchoring points between person and number predicted by the FIP should become crucial in cases of agreement mismatches between subject and verb, and the results obtained from experiments in Italian and Spanish are in line with this hypothesis. The role of the interpretive anchors will be further investigated

More on Feature-Specific Processing: Person Asymmetry

81

here in the processing of number agreement anomalies between pronominal subjects and verbs, with the goal of ascertaining whether the different featural makeup associated with 1st, 2nd and 3rd person pronouns influences the way these anomalies are processed. More specifically, the goal of the following ERP experiment was to find out whether different processing mechanisms are employed by the parser when faced with number violations involving participant and non-participant subjects. To test the hypotheses above, an ERP experiment was designed where responses to number violations with 3rd person and 1st/2nd person pronominal subjects could be compared to 1st/2nd and 3rd person correct sentences. Table 4-3 illustrates the experimental material used. Number agreement violations have been consistently studied using the ERP technique (see review in Chapter 2). A study conducted in Italian by De Vincenzi et al. (2003) reported the finding of a biphasic pattern in the presence of number agreement anomalies between subject and verb (Il cameriere anziano serve/*servono con espressione distratta, The old waiter serves/*serve with vacant look), namely a LAN followed by a posteriorly-distributed P600 effect. Such a biphasic pattern appears to be cross-linguistically quite stable, and similar patterns have been found in a variety of languages (Dutch: Hagoort et al. 1993; Hagoort and Brown 2000; English: Osterhout and Mobley 1995; German: Roehm et al. 2005; Italian: De Vincenzi et al. 2003; Spanish: Barber and Carreiras 2005; Silva-Pereyra and Carreiras 2007). To date, none of the existing ERP studies has explored the issue of the person asymmetry between pronominal forms, and 3rd person subjects have been predominantly used (mainly lexical DPs). The only ERP study that has used pronominal subjects to test agreement feature processing is the study by Silva-Pereyra and Carreiras (2007), where 1st and 2nd person pronouns were used to create number (and person) violations in Spanish. In line with the existing literature on the processing of number agreement, the results showed that number anomalies with 1st and 2nd person pronouns produced an anterior negativity followed by a posteriorly-distributed P600. However, a closer inspection of the experimental material used by Silva-Pereyra and Carreiras (2007) reveals the presence of two possible confounds that may have corrupted the ERP response (see Chapter 3 for further discussion). First, Silva-Pereyra and Carreiras (2007) studied the processing of number agreement using pronominal forms that are allegedly not specified for this feature, i.e. 1st and 2nd person singular pronouns. These forms were lumped together with others whose featural makeup is unclear between a person and a number specification, i.e. 1st and 2nd person plural pronouns. Second, part of the experimental material

82

Chapter Four

used by Silva-Pereyra and Carreiras consisted of polite forms of the type ‘Ustedes abren la puerta’ (You.2.pl open 3.pl the door), which comprise a 2nd person plural pronoun followed by a 3rd person plural verb – an agreement pattern that can be included within the unagreement phenomena (Rivero 2007; see Chapter 5), which will be shown to give rise to a different ERP pattern compared to full agreement and full disagreement patterns. Taken together, these two factors may have produced a spurious electrophysiological response that does not truly permit identification of the effects of a number violation with 1st/2nd person subjects. Support for the parser’s sensitivity to a person split comes from an Italian self-paced reading study by Carminati (2005). In this study, the identification of a null subject’s antecedent was found to be faster with 1st and 2nd person than 3rd person verbal morphology, supporting Silverstein’s (1985) Person Sub-Hierarchy. Let us now dwell on the strategies that the parser may adopt when presented with number violations of different persons. Suppose that the parser is faced with a number agreement anomaly like the one in (29): (29) *Lui scrivono una lettera a casa ogni sera He3.sg write3.pl a letter to home every night ‘He writes a letter home every night’

The performance of Agree would quickly lead to the detection of a mismatch in the phi values of the subject and verb number features. Similarly, a mismatch would be detected by Anchor when linking phi and sigma values. In line with previous studies on number agreement processing, a LAN effect is expected to arise. Diagnosis processes should be then triggered in the early window of the P600, followed by repair mechanisms, which would entail re-writing the correct number specification on the verb, an operation that involves constituents in a local configuration and that could be straightforwardly accomplished. In both its early and late phase, the P600 is expected to show a posterior distribution. A different situation may arise in the presence of a number anomaly such as the one in (30): (30) *Io scriviamo una lettera a casa ogni sera I1.sg write1.pl a letter home every night ‘I write a letter home every night’

Recall that 1st person plural (scrivono, we write) is not to be regarded as a multiplication of identical entities with identical speech roles, but as a group composed by the speaker and its associates. On this view, a number

More on Feature-Specific Processing: Person Asymmetry

83

anomaly like the one in (30) entails shifting from an individual with a specific speech role – io (I), the speaker – to an association of individuals – scriviamo, (we write) – that may comprise the speaker and the addressee, or the speaker and a non-participant (3rd person, see Table 4-1). In other words, the number anomaly in (30) determines contrasting cardinality representations between subject and verb, but also contrasting discourse representations, as these two elements imply different types of speech participants. The parser could therefore identify a problem for number as well as for person interpretation. There are two hypotheses concerning the processing correlates of this twofold conflict. If 1st/2nd person pronouns are specified for person but not for number (Adger and Harbour 2007; Anagnostopoulou 2005; Benveniste 1966; Harley and Ritter 2002; inter alia), one could expect the early ERP correlates of these number violations to differ qualitatively from the one elicited from 3rd person number anomalies. Specifically, the early negativity could show a different topography and resemble the one found for person anomalies in Spanish, i.e. an N400. Alternatively, assuming the two types of pronouns have a similar morphosyntactic representation (i.e. they are both specified for person and for number) but differ in the type of participants they are anchored to, the parser may become aware of discordant speech participant representations at later stages of processing, i.e. in the early phase of the P600, when diagnosis processes are performed. In this case, a LAN effect would arise as for number anomalies with 3rd person pronouns, but it would be followed by a positive effect with a frontal/broad distribution in its early interval, in line with previous studies (Kaan and Swaab 2003) that found anteriorly distributed P600 effects in the presence of increased discourse complexity (see Chapter 2). Once all the sources of the conflict have been correctly diagnosed, repair mechanisms can be triggered. In this case, the distribution of the P600 effect is not expected to differ between 1st/2nd and 3rd person number anomalies.

Chapter Four

3rd person number violation

3rd person correct

1st/2nd person number violation

1st/2nd person correct

Qualcuno ha detto che Someone has said that Qualcuno ha detto che Someone has said that Qualcuno ha detto che Someone has said that Qualcuno ha detto che Someone has said that *Qualcuno ha detto che Someone has said that *Qualcuno ha detto che Someone has said that *Qualcuno ha detto che Someone has said that *Qualcuno ha detto che Someone has said that Qualcuno ha detto che Someone has said that Qualcuno ha detto che Someone has said that *Qualcuno ha detto che Someone has said that *Qualcuno ha detto che Someone has said that

io scrivo una lettera a casa ogni sera I1.sg write1.sg a letter home every night tu scrivi una lettera a casa ogni sera I2.sg write2.sg a letter home every night noi scriviamo una lettera a casa ogni sera we1.pl write1.pl a letter home every night voi scrivete una lettera a casa ogni sera you2.pl write2.pl a letter home every night io scriviamo una lettera a casa ogni sera I1.sg write1.pl a letter home every night tu scrivete una lettera a casa ogni sera I1.sg write2.pl a letter home every night noi scrivo una lettera a casa ogni sera I1.pl write1.sg a letter home every night voi scrivi una lettera a casa ogni sera you2.pl write2.sg a letter home every night lui scrive una lettera a casa ogni sera he3.sg writes3.sg a letter home every night loro scrivono una lettera a casa ogni sera they3.sg write3.pl a letter home every night lui scrivono una lettera a casa ogni sera he3.sg write3.pl a letter home every night loro scrive una lettera a casa ogni sera they3.pl writes3.sg a letter home every night

Table 4-3. Experimental material sample. Underlined segments indicate subject and verb manipulations.

84

More on Feature-Specific Processing: Person Asymmetry

85

Consistent with the second hypothesis, both agreement violations showed a similar left-lateralised early negative effect between 300 and 500 milliseconds (Figure 4-3). Differences between number violations with 1st/2nd and 3rd person emerged between 500 and 700 milliseconds, when the former type of anomaly (relative to the corresponding correct sentence) gave rise to a more broadly/anteriorly distributed P600 effect compared to 3rd person number mismatches (relative to the correct condition). Subsequently, in the 700–900 millisecond window, both types of number anomalies yielded a posteriorly-distributed P600 effect. The different featural composition of 1st/2nd and 3rd person seems to play a role at later processing stages, namely when information of various kinds is handled and integrated by the parser to diagnose the source of an anomaly. The functional interpretation of broad/frontal P600 effects is not yet clear, and various proposals have been advanced in the ERP literature (see Chapter 2). One proposal maintains that frontal P600 effects are a consequence of an increasing complexity found at a discourse level of analysis (Kaan and Swaab 2003), for instance when the parser is confronted with a great number of referents to be integrated within the same discourse representation. Discourse representation contains a record of the previously processed referents and their roles. When new referents are to be included, appropriate slots must be found, so that interpretation can take place (Garrod and Sanford 1994). On this view, the greater the number of slots to be created and filled to accommodate DPs referents, the greater the processing cost, which would be reflected in the topography of the early P600 effect. Despite the different syntactic context that characterises the occurrence of a frontal P600 in the current experiment with respect to Kaan and Swaab’s (2003) one, one may similarly interpret this effect for 1st/2nd person number violations as a reflection of integration difficulties at a discourse level. As stated, a possible side effect of number violations with 1st/2nd person is the involvement of the speech act participants’ representation, since such an anomaly does not only imply changing the number of participants, it also entails changing the type of participants involved. The discourse representation underlying the verbal form is thus different from the one underlying the subject (see Figure 4-4). This follows from the fact that the meaning of 1st and 2nd person plural is not merely augmentative, but associative (see Table 5-1). The shift from a unique individual with a specific speech role (speaker or addressee) to a group of individuals with different statuses (or from a group of individuals to only one specific participant) may have caused greater integration

86

Chapter Four

problems compared to number anomalies with 3rd person. In this latter case, a shift in number does not entail a shift in the type of participants involved, and integration can be thus more easily accomplished. Further experimental findings seem to corroborate the interpretation of frontal positive effects as a result of discourse integration problems. In particular, Filik et al. (2008) report on a frontal P600 effect in the presence of pronouns that have no antecedent in the previous sentence fragment (e.g. The in-flight meal I got was more impressive than usual. In fact she courteously presented the food as well). Drawing on Kaan and Swaab (2003), they interpret this effect in terms of processing difficulty due to the necessity of adding and integrating new entities into the discourse model. Also of relevance is the parallelism between the ERP data presented in Chapter 3 on the processing of person anomalies in Spanish (Table 4-4). In both cases, broadly-distributed positivities have been seen in the early P600 window. Moreover, the same ERP pattern is found for number anomalies with 3rd person pronouns in Italian and number anomalies with lexical subjects in Spanish. This parallelism suggests that number anomalies with 1st/2nd person pronouns generate an intermediate pattern between ‘pure’ person and ‘pure’ number anomalies with 3rd person subjects. This strongly suggests that the dissociation operated by the parser between the two types of number anomalies is due to its sensitivity to the qualitatively different anchoring relations established between the morphosyntactic and discourse layer of sentence structure, as illustrated in Figure 4-4. Table 4-4. Overview of ERP components across studies and languages. (*Mancini et al. 2011, see also Chapter 3) 3rd person number violations LAN Posterior P600 Posterior P600

Italian 1st/2nd person number violations LAN Broad/anterior P600 Posterior P600

Spanish* Person violations

Number violations

N400 Broad/Anterior P600 Posterior P600

LAN Posterior P600 Posterior P600

The comparison between number anomalies with 1st/2nd and 3rd person pronouns has made it possible to highlight an important fact on agreement processing, namely that the online analysis of an agreement relation is not a monolithic process. Rather, it is a composite procedure that is sensitive

More on Feature-Specific Processing: Person Asymmetry

87

not only to the feature being manipulated, but also to the interaction between them (Figure 4-4).

Figure 4-4. Different speech participant feature specifications underlying 1st and 3rd person agreement. In both cases, upon checking feature consistency between subject and verb (Agree, solid line), a mismatch is detected. A. The person value of the subject and of the verb activate different speech act participants representations, respectively one that underlies the presence of a Speaker (Ȝa, striped line) and one that underlies the presence of a group including the Speaker and its associate (dotted line), here indicated as an Addressee (Ȝp). B. Despite the number inconsistency, the subject and verb’s person values activate the same speech act participant representation.

The featural makeup of pronouns Let us now discuss the theoretical relevance of these findings. Recall that in the introduction, a Person Asymmetry Hypothesis was illustrated that distinguished the former pronominal forms from the latter in terms of person underspecification. Such a hypothesis is typically called upon to explain the configurational and morphological person splits found across many languages (Anagnostopoulou 2003; Harley and Ritter 2002; Kayne 2000; Linn and Rosen 2003; Ritter 1995, inter alia). Besides this hypothesis, two less radical accounts were outlined to the effect that a distinction among 1st/2nd and 3rd person was still drawn in terms of discourse-relatedness vs. non-discourse-relatedness, but on different grounds. Bianchi’s (2006) and Sigurdsson’s (2004) accounts postulate that 1st/2nd and 3rd person are both context-dependent, but differ in the type of

88

Chapter Four

entity they refer to: a discourse participant in the case of 1st and 2nd person, a non-participant in the case of 3rd person, as indicated in (26) and (28) above. Overall, the ERP data discussed above seem to suggest that the difference between 1st/2nd and 3rd person agreement resides in their anchoring to the speech act representation, and not in their morphosyntactic specification. This may be concluded from the fact that the processing stage at which the dissociation arises is one when multiple sources of information are supposedly handled, to correctly diagnose the source of the conflict. The analysis of discourse-related information sanctions a mismatch for 1st/2nd person anomalies, but not for 3rd person violations, as in this case when subject and verb share the activation of non-participant roles. This interpretation clearly does not seem to argue in favour of analyses that explain the asymmetry between 1st/2nd and 3rd person in morphosyntactic terms, i.e. of presence/absence of a morphosyntactic feature. In contrast, proposals that identify the watershed between the two types of pronouns in the positive/negative anchoring to discourse (Bianchi 2006; Sigurdsson 2004) may better fit the set of data just discussed. Nevertheless, for this conclusion to be confirmed, experimental evidence should be gathered where person agreement with 1st/2nd and 3rd pronouns is manipulated. The results of this experiment also have significant implications for the structural representation of agreement features. The distinction during processing that is operated between 1st/2nd and 3rd person leaves room for the hypothesis that distinct projections in the syntactic tree may be responsible for 1st/2nd and 3rd person agreement, in the spirit of recent cartographic approaches to sentence structure (see Bianchi 2006; Shlonsky 2000, 2009; Sigurdsson 2004). This leads to a further decomposition of the Inflectional layer of the sentence (IP) into independent projections that are responsible for 1st/2nd and 3rd person agreement (Figure 4-5).

Figure 4-5. Cartography of person features.

More on Feature-Specific Processing: Person Asymmetry

89

Summary Person and number agreement was further tested in this chapter by focusing on the dissociation between 1st/2nd and 3rd person, the so-called Person Asymmetry Hypothesis. This asymmetry manifests itself in a number of typologically unrelated languages in the form of constraints in the co-occurrence of direct and indirect object clitics (e.g. the PCC), the morphological realisation of pronoun, and also agreement patterns in some American Indian languages. The relevance of the Person Asymmetry Hypothesis for the study of the FIP introduced in Chapter 1 resides first and foremost in the possibility to test the interaction between person’s and number’s anchor during the online processing of s-v agreement. To this end, an ERP study was designed that contrasted 3rd person number with 1st/2nd person number violations, with the goal of ascertaining whether the response to such anomalies changed as a function of the participantrelatedness of 1st and 2nd person, as opposed to the non-participantrelatedness of 3rd person pronouns. The results showed that whether the subject is a participant or not in the speech act (a speaker or addressee) significantly affects the way a number violation is dealt with, although at a later stage, when diagnosis operations are performed. This adds yet another piece of evidence in favour of a processing and structural dissociation between the two features, and also evidences the crucial role that the assignment of discourse roles has for the derivation of the overall meaning of a sentence.

CHAPTER FIVE FUNCTIONAL AND TEMPORAL DISSOCIATION OF AGREEMENT MECHANISMS

The analysis of s-v agreement and of its online processing will continue here with the analysis of unagreement patterns in Spanish. Unagreement is found in some Romance and non-Romance null-subject languages and is characterised by a mismatch in person between subject and verb that, nevertheless, gives rise to a well-formed sentence. The analysis of this special pattern and its comparison with standard agreement and truly ill-formed sentences offers the possibility to further test the validity of the feature-anchoring approach to agreement introduced in Chapter 1. A brief introduction to unagreement, its cross-linguistic distribution and theoretical analyses is provided, followed by the presentation of behavioural, electrophysiological and neuroanatomical findings on the processing of this agreement pattern. Finally, an account that integrates the theoretical and processing perspective is proposed.

When disagreement is grammatical: Unagreement The minimalist analysis of s-v agreement (Chapter 1) is essentially based on the assumption that an interpretive asymmetry exists between the agreement features on the subject and those on the verb. Technically, at the beginning of a syntactic derivation, person and number are valued on the nominal but unvalued on the verb. Agree connects the two positions and permits the checking and valuing of verbal morphology. This operation essentially consists of copying the subject person and number values onto the unvalued verb features: it thus proceeds in a specific direction, making the verb agree with the subject and not conversely (Chomsky 2000). Nominal features are therefore regarded as interpretable, and are assumed to be taken into account by the interpretive system. On the contrary, verbal features, being mere formal copies of the nominal specifications, are regarded as non-interpretable.

Functional and Temporal Dissociation of Agreement Mechanisms

91

These assumptions are, however, controversial. Evidence has been provided for a dissociation between person and number from a structural and an interpretive point of view (see Chapter 1), and the experimental findings illustrated in Chapter 3 confirm this. The fact that the person interpretive anchor is localised outside of the IP area, i.e. in the speech act representation, led to the hypothesis that, unlike nominal and verbal number, nominal and verbal person are in fact two autonomous values (see also Ackema and Neeleman 2013). This hypothesis appears to be corroborated by the fact that, across languages, agreement patterns are found that question the validity of the unidirectionality of Agree and verbal features non-interpretable status. This is the case with Spanish unagreement patterns (Hurtado 1984), sentences where s-v agreement can manifest itself with a person mismatch, as in (1) and (2): (1) Los lingüistas escribimos un artículo muy interesante. The linguists3.pl wrote1.pl an article very interesting ‘We linguists write a very interesting article’ (2) Los lingüistas escribís un artículo muy interesante. The linguists3.pl write2.pl an article very interesting ‘You linguists write a very interesting article’

Unagreement patterns of the kind illustrated above arise in the presence of plural lexical subjects and are characterised by a mismatch between the person feature in the verbal inflection and the one in the subject. The scope of unagreement in Spanish is however larger than (1) and (2) indicate, as this phenomenon occurs also in the presence of postverbal subjects (3) and non-nominative subjects (4): (3) Ayer llegamos los españoles Yesterday, arrived1.pl the Spaniards3.pl ‘Yesterday the Spaniards arrived’ (4) A Ana le gustamos los españoles To Ana cl like1.pl the Spaniards ‘Anna likes us Spaniards’ (from Rivero 2007)

In addition, unagreement patterns are possible with collective nouns, as shown in (5), while outright ungrammaticality is produced when a singular subject is used as in (6):

92

Chapter Five (5) La gente mayor no hablamos muchos idiomas The people3.pl old not speak.1.pl many languages ‘We old people don’t speak many languages (6) *El lingüista hablamos muchos idiomas The linguist3.pl speak1.pl muchos idiomas

In (5), the interpretation of ‘la gente’ is shifted from ‘the old people’ to ‘we old people’, based on the notional plural number of the subject argument. This pattern suggests the presence of an asymmetry between person and number: unagreement is a special pattern that targets only the person feature of the lexical items involved, and not the number one. Number unagreement is indeed not an option of Spanish grammar, as no such sentences (such as (7) and (8) below) are allowed in Spanish. (7) *Los lingüistas habla muchos idiomas The linguists3.pl speak3.sg many languages (8) *El lingüista hablan muchos idiomas The linguist3.sg speak3.pl many languages

Unagreement patterns are also found in Asturian, a Romance language of the West Iberian group, as shown in (9) and (10) below (from VillaGarcia 2010; see also Höhn 2016 and Corbett 2006 for further crosslinguistic descriptions of unagreement): (9) Los llingüistas saben que lo que facemos ye raro The linguists3.pl know that what do1.pl is weird ‘We linguists know that what (we) do is weird’ (10) Los llingüistas saben que lo que facéis ye raro The linguists3.pl know that what do2.pl is weird ‘You linguists know that what (we) do is weird’

Interestingly, this pattern is also attested in a non-Romance null-subject language, i.e. Greek (from Villa-Garcia 2010): (11) Oi foitites kseroume ti theloume The students know1.pl what we want1.pl (12) Oi foitites kserete ti thelete The students know2.pl what we want1.pl ‘We/You students know what we want’

Functional and Temporal Dissociation of Agreement Mechanisms

93

It should be noted that discordant subjects of the type shown in Spanish, Asturian and Greek are not allowed in other genuine Romance null-subject languages like Italian. In Italian, lexical subjects need to be associated with an overt pronominal, as shown in (13) and (14) below: (13) Noi linguisti sappiamo quello che vogliamo (we) want We linguists know1.pl what (14) Voi linguisti sapete quello che volete You linguists know2.pl what (you) want ‘We/you linguists know what we want’

The cross-linguistic distribution of unagreement suggests that a necessary (but not sufficient) condition under which this phenomenon occurs is the activation of the null-subject parameter (Rizzi 1982). Unagreement may thus represent an extension of null subject’s interpretation to cases of overt nominal subjects, which is possible in some pro-drop languages (e.g. Spanish, Greek, Asturian), but not in others (Italian). In other words, the interpretive process taking place in unagreement could resemble the one occurring in the presence of a null subject, in which feature interpretation is forced on the verb. As will be clear, one of the goals of this chapter is to compare unagreement and null-subject configurations from an experimental perspective, to assess the degree of interpretive overlap between them.

Unagreement: Theoretical analyses The pro-drop property of unagreement languages is usually taken as the point of departure to explain the derivation of these patterns. Early accounts proposed that in null-subject languages like Spanish, the real subject of the sentence is the agreement morpheme on the verb (Alexiadou and Anagnostopoulou 1998; Ordoñez and Treviño 1999), in accordance with the claim that Spanish lacks the EPP and has no SpecIP position projected. On this view, the preverbal lexical subject occupies a Topic or left-dislocated position. However, this analysis clashes with two important facts. Firstly, topicalised phrases are typically associated with a specific prosodic contour that unagreement preverbal DPs do not in fact have. Secondly, the availability of post-verbal subjects forcefully suggests that los lingüistas is not a left-dislocated constituent. A more recent analysis by Torrego and Laka (2015) capitalises on the internal structure of the Spanish (and Basque) DP, which, in their view, differs from the internal structure of DPs in non-unagreement languages like Italian. In particular, they propose that in Spanish, a null pronoun occupies the head of an autonomous DP node which is endowed with its

94

Chapter Five

own Number projection. The lexical DP los lingüistas stands in an appositive relation with the null pronoun. Subsequently, in the course of the derivation, the 1st/2nd person plural verb establishes an Agree relation with the null pronoun. Similarly to Torrego and Laka (2015), Höhn (2016) also claims that the presence of unagreement patterns is a direct consequence of the different DP structure that characterises unagreement languages compared to non-unagreement ones. Adopting a distributed morphology framework, Höhn (2016) proposes that what makes unagreement possible is the presence of null person specifications in the Person Phrase of the preverbal DP, as opposed to the overt person specification to be found in standard agreement. Finally, Villa-Garcia (2010) shifts the focus from structure to generative devices and assumes that agreement in Spanish may not be an essentially syntactic phenomenon. More specifically, in configurations where optimal agreement does not obtain, such as unagreement, the operation Agree can leave one feature unchecked/unvalued – e.g. person – which is later deleted for convergence by alternative mechanisms possibly relying on semantic-pragmatic information. Importantly, a similar split-phi probe mechanism (see also Bejar 2003, 2008) would also be operative for the checking/valuing of gender in null-subject configurations. Assuming the T head is endowed with a full array of features (person, number, gender and case) when it enters the derivation, the null subject of a sentence such as “Ha llegado” (prohas arrived) cannot be supplied with any gender value, as pro does not carry one. It follows that, similarly to person in unagreement configurations, gender in null-subject sentences must be handled by a mechanism other than Agree, which possibly connects verbal inflection to higher-level discourse information where the referent of the null subject can be identified and its gender retrieved. Critically, none of the syntactic analyses of unagreement described above accounts for why unagreement is possible with plural but not with singular subjects, which implies shifting focus onto the semantics of person and number features. Person features can be seen as indexicals that denote functions defined with reference to an utterance context, which determines participant roles such as speaker and addressee. In contrast, number features denote functions that identify atomic individuals and pluralities. First and 2nd person singular presuppose the presence of an atomic individual that includes the speaker or the addressee, whereas 3rd person singular presupposes an atomic entity that excludes both speaker and addressee. The ungrammaticality of (15) may thus derive from the impossibility of assigning an atomic individual two different and

Functional and Temporal Dissociation of Agreement Mechanisms

95

incompatible participant roles. In other words, as a singular entity, the linguist cannot be both a speaker and a non-participant in the same speech act (nor could they be both speaker and addressee, if the combination consisted of 1st and 2nd person). (15) *El lingüista escribiste un artículo muy interesante *The linguist3.sg wrote2.sg a very interesting article

By contrast, when pluralities are involved, the groups identified by the 3rd person plural subject and the 1st person plural verb as in (1) above do not mutually exclude. Because a plurality can accommodate different types of entities, the group of non-participants referred to by the linguists can include the speaker participant that the 1st person plural verb invokes (see also Ackema and Neeleman 2013 for a similar analysis but with a symmetric approach to agreement). In this case, therefore, no conflicting assignment of discourse roles is involved. The syntactic and semantic analyses described above suggest that the derivation of unagreement arguably relies on a close interaction between structure, formal operations (Agree) and the interpretive properties associated with the features involved, which makes unagreement an ideal candidate to test the feature-based anchoring approach described in Chapter 1.

Unagreement processing and the role of interpretive anchors In Chapter 3, it was shown how the difference in terms of interpretive anchors can account for the qualitatively different effects produced by number and person agreement anomalies. In this chapter, the focus will be on the role of the person interpretive anchor in dissociating grammatical from ungrammatical agreement mismatches. The studies that will be presented hinge on two crucial comparisons: one between standard agreement and unagreement, to highlight when and how, during comprehension, the two types of agreement start to diverge; and one between unagreement and person anomalies, to disentangle mechanisms associated with pure mismatch detection from those related to morphosyntactic checking. From a psycholinguistic perspective, the unagreement paradigm offers a clear advantage over the violation paradigms traditionally used in psycholinguistics, as it provides the opportunity to open a window onto the extra-syntactic components involved in agreement processing without overwhelmingly taxing the

96

Chapter Five

parser as outright anomalies do, or without “silencing” semantic-discourse factors, as happens in standard agreement sentences. In the presence of a person anomaly, the parser should find itself faced with an agreement inconsistency that can hardly be repaired. Inspection of the interpretive anchor would probably not provide any cue on how to resolve the conflict, due to the incompatible indices that 1st and 3rd person singular have. Person interpretation would be thus blocked, an error sanctioned and repair operations triggered. The N400-P600 effect reported in Chapter 3 for person violations corroborates this hypothesis. One may expect unagreement to be dealt with differently. Similarly to person anomalies, in unagreement, a feature inconsistency is likely to be detected quickly, and recourse to the anchor will be activated. However, instead of sanctioning an agreement mismatch and activating repair strategies, the inspection of the interpretive anchor may give the parser a cue on how to solve the conflict. The parser could interpret person on the verb, as happens in null-subject contexts, and overwrite the 1st person plural value onto the preverbal DP, with a mechanism that will be dubbed Reverse Agree. No repair strategy would therefore be activated, and the apparent conflict should be rapidly solved. In the following, a series of experiments will be presented that were designed to assess (i) whether unagreement processing differs from the processing of standard agreement and true person anomalies and, if so, how and when it differs; (ii) the degree of overlap between unagreement and null-subject sentence processing. Overall, it will be shown that the response of the parser to this agreement pattern changes over time, arguably guided by the information provided by the interpretive anchor, which points to agreement processing as consisting of two functionally and temporally distinct mechanisms.

Judging unagreement grammaticality The first step into unagreement processing involved the combination of offline and online acceptability judgements, in the attempt to assess whether speakers’ correct evaluation of the sentence can be misled by the presence of a feature mismatch between subject and verb. The combination of the two paradigms was motivated by the hypothesis that speakers, in their evaluation of unagreement grammaticality, could be erroneously misled by the featural mismatch, and that this could happen in the early stages of processing. As a matter of fact, while both online and offline judgements permit the assessment of speakers’ evaluation of unagreement, the types of responses that the two paradigms elicit may be

Functional and Temporal Dissociation of Agreement Mechanisms

97

the product of distinct processing stages (Lewis and Phillips 2015; Phillips and Lewis 2013). The response produced in the offline version of the task results from mechanisms that have not been subject to time restrictions, and it therefore captures the final output of late and “slow” interpretive processes. In contrast, the response elicited in an online time-pressured paradigm is likely to be the product of rapid intermediate processing stages that may not have explored all the available options of a grammar. This means that misalignment between the offline and the online output can occur. If the grammatical status of unagreement is clear from the beginning, its person mismatch should not affect its evaluation, and readers should be equally accurate in judging unagreement and standard agreement grammaticality, as well as in rejecting true person mismatches. Moreover, no differences should emerge between online and offline responses. In contrast, if the mismatch temporarily misleads processing, a different degree of accuracy should characterise the evaluation of unagreement and standard agreement, with responses to unagreement being less accurate than those to standard agreement. Misalignment between online and offline responses is therefore expected. Table 5-1 illustrates the design of the experiment (Mancini et al. 2014, Experiment 2). Table 5-1. Unagreement, standard agreement and person violation stimuli included in the grammaticality judgement, ERP, fMRI and one eye-tracking study described in this Chapter. Agreement type Unagreement Standard agreement Person Mismatch

Los lingüistas escribimos un artículo muy interesante The linguists3.pl wrote1.pl a very interesting article Los lingüistas escribieron un artículo muy interesante The linguists3.pl wrote3.pl a very interesting article El lingüista escribiste un artículo muy interesante The linguist3.sg wrote2.sg a very interesting article

The results were in line with the second hypothesis: although unagreement was accurately rated as correct in 94% of the cases in the online task, this percentage was smaller compared to the accuracy with which standard agreement and true person anomalies were evaluated (97% for both). In contrast, no difference in the evaluation of the acceptability of standard agreement and unagreement was found offline (96% and 95.5% respectively, 99% for person mismatch), yielding a clear misalignment between the two tasks.

98

Chapter Five

If online and offline responses reflect the output of distinct processing stages, it is therefore possible that the analysis of unagreement undergoes two phases: one in which feature consistency between subject and verb is verified, and which can produce erroneous parses because of the mismatch, and one in which the overarching meaning of the sentence is obtained (i.e. when the 1st person plural interpretation is derived). To test this hypothesis, an ERP paradigm was adopted that made it possible to assess whether, over time, the processing of unagreement qualitatively differed from the analysis of standard agreement and true person mismatches.

Electrophysiology of unagreement processing While a large body of studies exists on the processing of s-v agreement anomalies (see review in Chapter 2 and 3), the study designed by Mancini and colleagues (Mancini et al. 2011b) was the first one to deal with unagreement. The experimental design adopted is illustrated in Table 5-1. If unagreement triggers processing mechanisms that qualitatively differ from those employed for the analysis of standard agreement patterns, differences are expected to arise in the LAN/N400 window. Here, separate access to morphosyntactic and semantic information has been postulated (Bornkessel & Schlesewsky 2006; Friederici 2002, 2011), which can be assumed to correspond in part to the Agree operation. Due to the apparent mismatch between subject and verb, the parser may initially detect incongruence, which could generate an early negativity. However, before sanctioning the error, the parser may become aware of the compatibility of the subject’s and verb’s person indices. Under the assumption that in unagreement, person interpretation is forced on the verb, one may hypothesise that a Reverse Agree operation is carried out to overwrite 1st person plural values onto the DP person values (whether null or overt). Being a more marked checking option, the performance of Reverse Agree may result in processing complexity. The negative effect that emerged in the LAN/N400 window could therefore extend to the early phase of the P600. In the late phase of the P600, no differences are expected between the two conditions, since the grammatical status of unagreement does not require repair strategies to be performed. As for the person-mismatch condition, an N400-P600 pattern is expected, in line with the results from the experiment in Chapter 3. Alternatively, if unagreement and standard agreement undergo similar processing, no differences between these two conditions are expected either in early negative or in late positive effects.

Functional and Temporal Dissociation of Agreement Mechanisms

99

The results of this experiment are illustrated in Figure 5-1. In line with the former hypothesis, the ERP data showed qualitatively distinct patterns in early as well as late temporal intervals. The posteriorly-distributed negativity elicited by the unagreement condition is revealing of qualitatively different mechanisms at work compared to standard agreement. A closer look at the waveforms and at the topographical maps in Figure 5-1 shows that the maximum of the effect for unagreement is in the left posterior area, exactly as for the person mismatch. Therefore, a straightforward interpretation of this pattern is that these early negative effects reflect the parser’s sanctioning of a morphosyntactic mismatch. In the 500–800 millisecond window, the unagreement condition gives rise to a less positive effect in the centro-posterior sites compared to the standard agreement condition. This reduced positivity may be the final, and hence reduced, segment of the negativity found in the previous interval (300–500 milliseconds), which may evidence the processing complexity engendered by the application of Reverse Agree. A second explanation is, however, possible. Assuming that the first phase of the P600 effect reflects diagnosis processes aimed at identifying the source of a mismatch (Barber and Carreiras 2005; Carreiras et al. 2004), the smaller positivity found here may represent the parser’s attempt at inhibiting the syntax error signal that would trigger a diagnosis operation, as no violation is present here. In the late phase of the P600, no difference is found with respect to the standard agreement condition, suggesting that no integration difficulties are being dealt with by the parser. Unlike person mismatch, no repair operations are therefore triggered in the presence of unagreement. Overall, the set of data just described confirm the biphasic N400-P600 effect for the processing of person mismatch reported in Chapter 3. But more importantly, they suggest that the analysis of unagreement undergoes different processing stages that possibly reflect the online/offline divide evidenced by grammaticality judgement tasks. More in-depth investigation is, however, needed to clarify the nature of the early negative effects and the possible involvement of Reverse Agree operations in the overall interpretation of the pattern. The fMRI and eye-tracking studies that follow precisely attempt to answer these questions.

Neural correlates of unagreement processing Quiñones and colleagues (Quiñones et al. 2014) used the same design as the one used in the ERP study by Mancini et al. (2011b) to investigate the neural correlates of unagreement processing. The goal of this fMRI

100

Chapter Five

study was to isolate the neural substrates involved in agreement computation, with a special focus on both the evaluation of morphosyntactic feature consistency and morphosyntax-discourse anchoring operations. To this end, the authors carried out three types of contrasts among conditions. The first type of contrast was meant to dissociate the neural regions involved in the computation of correct agreement dependencies from those supporting the detection of morphosyntactic incongruities (standard agreement > person mismatch), while the contrast between unagreement and person mismatch (unagreement > person mismatch) was aimed at isolating the neural substrates engaged in the analysis of grammatical mismatch from those involved in the processing of outright anomalies. The second group of comparisons involved reversing the direction of the contrasts used in the previous comparisons, so that the neural correlates associated with morphosyntactic mismatch detection and conflict-monitoring could be dissociated from those related to the analysis of meaningful relations (person mismatch > standard agreement; person mismatch > unagreement). Finally, the third comparison comprised the contrast between unagreement and standard agreement (unagreement > standard agreement), to identify the brain regions involved in the person shift triggered by the interpretation of unagreement. A specific set of hypotheses follows from these contrasts. Based on the data presented in Chapter 3, grammatical sentences (standard agreement and unagreement) in comparison to person anomalies should determine increased activity within an extended left frontotemporal network, including the anterior and posterior temporal cortex and the inferior and middle frontal gyri (Kuperberg et al. 2003, 2008). As for person anomalies relative to grammatical sentences, in line with the findings presented in Chapter 3, activation of left middle frontal and bilateral parietal areas is expected, as a result of the outright mismatch sanctioned by conflict-monitoring operations. Assuming that unagreement and person mismatch share a mismatching morphosyntactic representation, the contrast between these two conditions should evidence overlapping patterns of activation in areas associated with the analysis of morphosyntactic consistency, such as middle frontal areas (see Chapter 3). However, the fact that no P600 effect is elicited by the reading of unagreeing verbs predicts no significant involvement of conflict-monitoring mechanisms, and hence no significant activation of bilateral parietal areas for unagreement compared to person mismatches. Finally, because unagreement and standard agreement share grammaticality, the processing of the two conditions should engender overlapping fronto-temporal patterns of activation. However, the shift in

Functional and Temporal Dissociation of Agreement Mechanisms

101

person interpretation required for unagreement interpretation compared to standard agreement may lead to increased responses in areas associated with conceptual analysis and integration. Two plausible candidates are the angular gyrus and the anterior temporal cortex, for their respective role in mechanisms supporting conceptual retrieval and integration and processing at the interface between the syntactic and semantic level of analysis (Binder et al. 2009; Lau et al. 2008, 2013; Price 2012; Rogalski and Hickok 2009; Seghier 2013; Vigneau et al. 2006; see Chapter 2 and 3). Figure 5-2 illustrates the main findings reported by Quiñones et al. (2014). The contrast between standard agreement and unagreement evidenced the involvement of a left-lateralised fronto-temporal network, including the pars triangularis and orbitalis of the IFG and the anterior and posterior MTG. A similar fronto-temporal network also emerged from the contrast between unagreement and person mismatch, suggesting that the processing of correct sentences, whether fully agreeing or unagreeing, involves akin neural substrates. Nevertheless, close inspection of the activation patterns shows that unagreement and standard agreement differ in the pars orbitalis and in the most anterior part of the middle temporal cortex, as well as in the AG, which appear to be more activated by unagreement. The superimposition of the unagreement > standard agreement contrast on the person mismatch > standard agreement one made it possible to understand the functional significance of middle frontal and AG activations. As Figure 5-2 (lower panel) shows, the AG was specifically activated only for unagreement sentences, while middle frontal activation is shared by both unagreement and person mismatch. Let us discuss these findings in more detail. The similarity in activation patterns between unagreement and standard agreement in the anterior temporal cortex is consistent with the ERP findings described above, and suggests that the two conditions share similar processing mechanisms. The role and function played by this region in the context of sentence processing are still not agreed upon in the literature (see Chapter 2). Some authors argue that this region plays a special role in storing and activating semantic associative, categorical and contextual information (Kuperberg et al. 2008), while others claim that it plays a crucial role in the storage/retrieval of lexico-semantic information or combinatorial semantic processes (Rogalsky and Hickok 2009). A third view assumes that the mechanisms supported by the anterior portion of the temporal cortex are essentially combinatorial in nature, although it is still unclear whether syntactic operations (Dronkers et al 2004; Friederici and Kotz 2003; Friederici et al. 2000; Humphries et al. 2006) or operations at the interface between syntax and lexico-semantic analysis (Bornkessel-

102

Chapter Five

Schlesewsky and Schlesewsky 2013; Pallier et al. 2011; Rogalsky and Hickok 2011; Vanderberghe et al. 2002) are involved. In the context of the study by Quiñones and colleagues, and consistent with the fMRI study illustrated in Chapter 3, it seems plausible to maintain that the anterior portion of the temporal cortex is engaged in mechanisms that lie at the intersection between syntax and higher levels of analysis such as discourse, which are necessary for the assignment of interpretively relevant roles and hence for the derivation of the overarching meaning of the sentence (“who does what”). The similarity in the activation pattern for person mismatch and unagreement sentences in middle frontal areas is consistent with the early negative effect that both conditions elicit compared to standard agreement in the ERP experiment (Mancini et al., 2011b). This finding is also consistent with the fMRI results reported in Chapter 3 and in other studies testing different languages and dependencies (pronoun gender mismatches in Dutch (Nieuwland et al. 2007), article–noun gender violations in Dutch (Folia et al. 2009); finiteness violations in English (Kuperberg et al. 2003, 2008; Newman et al. 2003); verb–object number violations in Basque (Nieuwland et al. 2012)). There are two possible explanations for this common activation in the posterior part of the middle frontal cortex. First, from a more domaingeneral perspective, this common activation may reflect the engagement of working memory mechanisms associated with the evaluation of the morphosyntactic consistency of subject and verb person features (see Katsuki and Constantinidis 2012 for a review about the role of dorsolateral frontal regions in working memory processes). An alternative, more language-specific explanation assumes that this sub-region within the middle frontal gyrus is crucially engaged in checking the morphosyntactic match between two sentence constituents, irrespective of their grammaticality. Nonetheless, the fact that unagreement and person mismatch only share morphosyntactic subject-verb incongruency lends support to the idea that this common brain activation reflects processes involved in the evaluation of morphosyntactic consistency. Of relevance also is the significant activation for unagreement (both relative to standard agreement and person mismatch) in the AG. During unagreement processing, the AG could be actively involved in the performance of the mechanisms that drive the discourse integration of the apparently discordant person values displayed by subject and verb. Support for this interpretation comes from studies that found AG activation with experimental manipulations involving the contextrelatedness of sentences (Kuperberg et al. 2003, 2008) and the referential

Functional and Temporal Dissociation of Agreement Mechanisms

103

ambiguity between pronouns and their referents (Nieuweland et al. 2008), but also from studies on metaphor processing (Bambini et al. 2011; Mashal et al. 2007; Shibata et al. 2012). The comprehension of metaphors requires accessing and elaborating a meaning that extends beyond the literal meaning expressed by the linguistic input, which may therefore trigger the recourse to additional semantic resources. In this respect, the shift in subject interpretation from 3rd to 1st person plural that unagreement triggers may represent the extra-semantic analysis that the parser carries out to overcome the apparent incongruence that it detects. Finally, the involvement of the conflict-monitoring system in the processing of the person-mismatch condition is consistent with the fMRI and ERP studies reported in Chapter 3, as well as with the ERP data described in this chapter, in which a clear distinction in the responses associated with ungrammatical and grammatical sentences emerges, as only person mismatches generated a P600 effect. In sum, the experimental design used by Quiñones and colleagues evidenced that unagreement processing shares computational and representational properties of both standard agreement and person anomalies. With the former, unagreement shares grammaticality, but not the extra computation that is necessary to derive the overall first person plural interpretation. With the latter, it shares the morphosyntactic mismatch, but not the repair mechanisms that are meant to fix the mismatch. This made it possible to identify the neural correlates associated with two functionally distinct processing mechanisms: the morphosyntactic checking operated by Agree – which appears to be supported by the posterior portion of the left middle frontal gyrus – and the linking of morphosyntactic information onto higher-level semantic-discourse information that Anchor performs, in which the anterior part of the temporal cortex and the angular gyrus appear to play a prominent role. Having tested whether (un-)agreement processing relies on two functionally distinct processing mechanisms, the next step involves investigating the time course of morphosyntactic checking and discourse anchoring operations, which was accomplished through the use of the eyetracking paradigm.

Eye-tracking unagreement processing The recording of eye movements can provide relevant insights into the mechanisms and strategies adopted by the parser to detect and integrate constituents within a syntactic dependency. The potentially ecological presentation of the experimental material and the possibility to fractionate

104

Chapter Five

the reading process into several distinct latency and regression variables make it possible to identify not only the disruption point, i.e. the point in the text where the reading process is perturbed, but also the reading and comprehension strategies adopted to overcome such difficulty (Clifton and Staub 2011). The eye-tracking paradigm was profitably applied by Mancini and colleagues (Mancini et al. 2014b) to the study of (un-)agreement processing with two specific goals. Firstly, the authors aimed to assess when and how, during reading, unagreement comprehension differs from standard agreement and unagreement, to find out whether morphosyntactic feature checking and discourse anchoring could be functionally and temporally dissociated. Secondly, by comparing unagreement with other agreement dependencies that involve different types of subjects – null and overt pronouns – the authors meant to verify the potential computational and interpretive similarities between unagreement and null-subject sentences, as well as to the test the validity of the FIP to a wider range of s-v agreement dependencies. For the comparison of unagreement with standard agreement and person anomalies, the same design as in Table 5-1 was adopted (Mancini et al. 2014, Experiment 3). In this case, the analysis of eye-tracking data revealed that, across the different early and late reading variables analysed, the magnitude of latencies and regression probability for unagreement was intermediate between those for standard agreement and person mismatch (Figure 5-3). In other words, the reading of an unagreeing verb determined morphosyntactic integration difficulties that were stronger than those elicited by the processing of a standard agreement dependency, but not as severe as those generated by truly mismatching verbs. There are two possible interpretations for these gradient effects. On the one side, early reading disruptions could reflect the performance of checking operations and the subsequent detection of a mismatch for both unagreement and person violations, while later effects could index the parser’s attempts to repair the mismatch and anchor the mismatching person values to a congruent discourse representation. This operation generates a cost, which is greater for person mismatch than unagreement as a result of the impossibility of deriving an overall interpretation for the former condition. The inspection of the person interpretive anchor in unagreement patterns sanctions the possibility of building a coherent discourse representation, resulting in a remarkably less severe disruption.

Functional and Temporal Dissociation of Agreement Mechanisms

105

Figure 5-3. Average first-pass, go-past and total reading times and probability of regression into the verb area. Latency measures are reported in milliseconds and probability of regression in percentages. Legend: SA= Standard Agreement; UN=Unagreement; PM=Person Mismatch. Adapted from Mancini et al. (2014, Experiment 3).

On the other hand, the intermediate reading disruption of unagreement could indicate that the parser had already become aware of the grammaticality of this agreement pattern during the first pass of fixations

106

Chapter Five

through the verb. This would imply that the performance of discourse anchoring operations was carried out at early and not at later stages. To assess when, precisely, discourse anchoring takes place during agreement processing, an eye-tracking experiment was designed that allowed for a clear-cut dissociation between unagreement sentences that could be smoothly anchored, such as those used so far, and unagreement sentences in which discourse anchoring yields discourse implausibility. This was achieved by manipulating the plausibility of the subject as the speaker of the sentence. More specifically, the subject could be represented either by human beings – e.g. a group of linguists – or by animals – e.g. a group of birds – which could combine with a standard agreement or unagreement verb (Table 5-2). Table 5-2. Sample of the experimental material used by Mancini et al. (2014, Experiment 4). Legend: HU= Human; AN=Animal; UN= Unagreement; SA= Standard Agreement. Being HU

AN

Agreement Los lingüistas escribimos un artículo muy interesante UN The linguists3.pl wrote1.pl a very interesting article Los lingüistas escribieron un artículo muy interesante SA The linguists3.pl wrote3.pl a very interesting article Los pájaros volamos por el cielo UN The birds3.pl flew1.pl in the sky Los pájaros volaron por el cielo SA The birds3.pl flew3.pl in the sky

While in human sentences, the speaker invoked by the 1st person verbal morphology can be plausibly integrated into the group referred to by the subject argument, with animals this should not be possible, as a bird is unlikely to be the speaker of a sentence. Greater processing complexity can therefore be predicted for animal compared to human unagreement. If morphosyntactic checking is what drives early effects, first-pass and go-past measures should be sensitive to the morphosyntactic mismatch that characterises unagreement, regardless of the humanity of the subject argument. The presence of later effects should therefore be modulated by the type of being the subject refers to, leading to greater total reading times for the animal unagreement than for the human unagreement sentences. Alternatively, if morphosyntactic and discourse information interact from early on, the agreement factor should be modulated by the being factor as early as during the first pass through the verb region.

Functional and Temporal Dissociation of Agreement Mechanisms

107

The analysis of latency and regressive measures revealed a scenario compatible with the former hypothesis. The comparison between human and animal unagreement showed a clear dissociation between first-pass and total reading time effects (Figure 5-4).

Figure 5-4. Average first-pass, go-past and total reading at verb and post-verbal position. Bars indicate standard error. Legend: SA= Standard Agreement; UN=Unagreement. Adapted from Mancini et al. (2014, Experiment 4)

While the magnitude of the first-pass and go-past effects did not depend on whether the subject was an animal or a human being, the

108

Chapter Five

magnitude of total reading times did, with unagreement verbs in animal sentences generating a sizeable increase in reading times compared to their standard agreement counterpart. In contrast, no such increase was found for unagreement verbs in human-subject sentences. This asymmetry suggests that a mismatch between subject and verb (even if grammatical) yields initial morphosyntactic integration difficulties that are not modulated by the conceptual features of the subject. These latter features act as triggers for re-reading behaviours to assess how plausible an entity is as the speaker of an utterance.

Agreement and unagreement: Time course and mechanisms The data from the set of eye-tracking experiments reported so far all converge in indicating that agreement analysis is carried out from early stages of sentence processing (Deutsch and Bentin, 2001; Kreiner et al. 2012), and that the reading processes adopted to interpret an agreement relation are primarily influenced by the degree of feature consistency between subject and verb. Across experiments and paradigms, data converged in the fact that, in spite of its grammaticality, the reading of an unagreeing verb produces sizeable and sustained disruptions in the comprehension process, generating early negative shifts in the brain’s electrical activity, increased activation in left middle frontal areas similarly to person anomalies, and a greater number of fixations on the verb. Unagreement can also mislead the parser to the extent that it is judged as ungrammatical in a significant number of cases when a speedy evaluation is required. But as the parser’s analysis of the morphosyntactic and discourse information extracted from subject and verb goes on, the grammaticality of unagreement is unequivocally acknowledged. The interpretive anchor of person in the discourse representation plays a crucial role in shaping interpretive processes towards the sanctioning of irreparable errors or discourse implausibility, or towards the licensing of apparently mismatching patterns. Importantly, what these data reveal is that checking whether a subject and a verb share identical phi values is only one part of the process (although a fundamental one) of leading to agreement interpretation. In order to build a coherent interpretation, both the morphosyntactic and discourse congruence of a relation – the phi and sigma level in the account developed here – must be verified. The function of sigma values – or anchors – has proved to be especially critical when feature mismatches are involved: their inspection guides processing and permits disentangling of

Functional and Temporal Dissociation of Agreement Mechanisms

109

mere feature mismatch from syntactic error, by evaluating whether the indices underlying disagreeing features are compatible or not. Of relevance also is the time course of Agree and Anchor operations. The eye-tracking data just presented point to a sequential implementation of the two operations, with Anchor being performed at a later stage of processing regardless of Agree’s output. Before ending this review of unagreement processing, there is yet another aspect that deserves to be explored, namely whether unagreement shares representational and interpretive similarities with null-subject agreement processing. Such is the goal of the experiment that will be presented below.

Unagreeing, null and overt subjects The degree of interpretive overlap between unagreement and nullsubject configurations was investigated by Mancini et al. (2014b) in a further eye-tracking experiment. In this experiment, the verb could be preceded by a lexical DP (los lingüistas), a null pronoun, and an overt pronoun (nosotros) with features that matched the unagreeing verb, as illustrated in Table 5-3. Table 5-3. Sample of the experimental material used in Mancini et al. (2014b, Experiment 1). Legend: UN=Unagreement; SA= Standard Agreement; NS= Null subject; OP= Overt Pronoun. Agreement configuration Hace mucho tiempo los lingüistas escribimos un artículo muy UN interesante A long time ago the linguists3.pl wrote1.pl a very interesting article Hace mucho tiempo los lingüistas escribieron un artículo muy SA interesante A long time ago the linguists3.pl wrote3.pl a very interesting article Hace tiempo para los lingüistas (pro)escribimos un artículo muy NS interesante Some time ago for the linguists (pro)wrote1.pl a very interesting article Hace tiempo nosotros los lingüistas escribimos un artículo muy OP interesante Some time ago we1.pl the linguists wrote1.pl a very interesting article

110

Chapter Five

The inclusion of the overt-pronoun condition was meant to assess whether the presence of a pronoun in an unagreement context could mitigate, or even cancel, the effect of a morphosyntactic mismatch between the preverbal DP los lingüistas and the verb. Based on the results from the previous eye-tracking experiments, unagreement should generate immediate and sustained reading disruptions (relative to standard agreement and overt pronoun), as a result of the feature mismatch. This disruption should be mitigated, or perhaps even cancelled, by the presence of a 1st person plural pronoun. In this case, an Agree relation would be unequivocally established between the verb and the pronoun’s features, with no processing penalty arising, exactly as in standard agreement. In the absence of a phonologically realised agreement controller, no feature-checking mechanism should be triggered, ruling out possible mismatch effects in early measures such as first-pass and go-past reading times. This means that during this time, the null-subject should not differ from the standard agreement and overt-pronoun conditions, while sizeable differences should emerge in comparison with unagreement. Similar hypotheses can be formulated for later reading measures such as total reading times and regression probability. If the null-subject verb has already been smoothly integrated in the syntactic context, no sizeable differences in reading times and in the probability of regression (both into the verb and into the preverbal position) should arise from the comparison between null-subject and standard agreement/overt-pronoun conditions. In contrast, increased total reading times on the verb and a greater probability of regressions into the preverbal area may emerge for unagreement compared to standard agreement and overt pronoun conditions, as a result of the more complex anchoring operation that takes place in the former compared to the latter conditions. This scenario would suggest that the interpretive mechanisms that characterise unagreement and null-subject configurations are qualitatively distinct. An alternative hypothesis is also possible, namely that in null-subject contexts, re-reading of the preverbal area is consistently triggered to assign the null pronoun the person and number features extracted from the verb. In this case, one may assume that unagreement and null-subject processing undergo similar anchoring mechanisms. Figure 5-5 illustrates reading times and regression probability for the four conditions. As expected, a sizeable disruption was produced by the reading of unagreeing verbs (relative to standard agreement and overt pronoun conditions) during the first pass and go-past pass through the region. Increased latencies on the verb were also accompanied by

Functional and Temporal Dissociation of Agreement Mechanisms

111

regressive saccades into the preverbal area. Critically, such effects vanished when an overt pronominal was present, suggesting that an argumental dependency was computed between the overt pronoun and the verb.

Figure 5-5. Average first-pass, go-past and total reading times (verb position) and probability of regression (preverbal position). Bars indicate standard error. Legend: NS=Null Subject; OP=Overt Pronoun; SA=Standard Agreement; UN= Unagreement. Adapted from Mancini et al. (Experiment 1).

112

Chapter Five

The comparison of null-subject with standard agreement and overtpronoun conditions did not reveal large differences either in first-pass or in go-past reading times. Differences emerged in total reading times, both in the verbal and preverbal area, which indicates that this effect may be caused by re-reading of the two regions. In other words, the absence of an overt controller does not prevent a verb from being smoothly integrated in the syntactic context during the first pass of fixations, but it does make the overall interpretation of the agreement dependency costlier, triggering the re-inspection of the preverbal and verbal area. Let us discuss this point. Overall, a clear asymmetry emerged between the early stages of unagreement and of overt-pronoun/null-subject agreement processing. While the reading of an unagreeing verb gave rise to a rapid and sizeable disruption, the reading of a null-subject verb did not, exactly as happened when an overt pronoun was available. In pro-drop languages like Spanish, the parser is aware that the goal of the Agree operation can be found preverbally, post-verbally or can be phonologically unrealised. In unagreement (and standard agreement) contexts, the nearest and only available matching element with phi-features is the preverbal DP los lingüistas. Agree immediately targets this position: phi-checking is carried out and a mismatch is sanctioned, as the match is only partial (i.e. in number). No such disruption occurs when an overt pronoun is available. In the presence of nosotros, the parser overrides the lexical DP (arguably because of its appositive reading) and establishes an Agree relation with the pronoun. When the pronoun is null, the parser cannot establish Agree with any element in the preverbal or post-verbal area. However, as this option is available in the grammar of Spanish, no mismatch effect arises. A parallelism emerges between unagreement and null subject during later processing stages, when both conditions trigger re-reading of the verb and re-inspection of the preverbal area. For unagreement, this could be aimed at assigning the preverbal DP the 1st person plural interpretation, and thus deriving the overall interpretation of the relation. For null-subject configurations, it may be aimed at assigning the null pronoun the feature values extracted from the verb. In both conditions, the re-reading of the preverbal area could therefore be a behavioural correlate of the Reverse Agree operation hypothesised above.2 These results appear to have implications for structural analyses of unagreement, as those proposed by Höhn (2016) and Torrego and Laka (2015). The fact that the two conditions yield different processing correlates at early stages of processing does not provide full support for the presence of null features (Höhn 2016) or null pronouns in the internal structure of the DP (Torrego and Laka 2015). If the presence of silent

Functional and Temporal Dissociation of Agreement Mechanisms

113

features/pronouns is a feature of the internal structure of Spanish DPs, one would expect the parser to be aware of them as soon as the analysis of the preverbal DP is initiated. The consistent early mismatch effect that arises upon reading an unagreeing verb seems to refute this hypothesis. It has been argued (Höhn 2016) that such a mismatch effect may simply be a performance issue having nothing to do with the abstract knowledge of language (competence). While the nature of the relationship between grammar and parsing is beyond the scope of this discussion, the fact that the processing penalty under discussion arises in an ecologic paradigm such as eye-tracking, in which speakers are under no time pressure and no limitation is imposed on memory and other cognitive functions, certainly represents a strong counterargument. An alternative explanation, which however requires further investigation, is that the parser triggers checking operations as soon as verbal phi values are extracted, and that these operations erroneously target the noun phrase lingüistas because the projection and analysis of the internal structure of the subject DP has not yet been completed. Once the head hosting the null pronoun has been projected, regressions are performed to assign the null pronoun the 1st person plural interpretation. This explanation, which assumes a cascaded architecture where processing of nominal and verbal information partially overlap in time, implies that resolving unagreement is a matter of time: the more time the parser is given to complete structure projection, the less likely it is that an incorrect evaluation will arise, because checking mechanisms can target the correct position in the syntactic structure. In this respect, evaluating the impact of unagreeing verbs in non-adjacent s-v relations would represent a good testing ground, which, at the time of this writing, is under investigation. The analysis proposed by Villa-Garcia seems to provide a better framework for the interpretation of these findings. In both unagreement and null-subject sentences, re-reading behaviour could instantiate postsyntactic agreement handled by a mechanism other than Agree, which could be identified in the Anchor mechanism adopted here. Some caveats are however in order. Implicit in the analysis proposed by Villa-Garcia (2010) are time course assumptions concerning when checking/valuing of features occurs in unagreement contexts. Specifically, person checking/valuing is claimed to be dealt with entirely post-syntactically, and hence after checking/valuing of e.g. number has taken place. Although no comparison of unagreement with number agreement processing has been carried out, several findings from the studies discussed so far seem to clash with this interpretation. Firstly, the early mismatch effect that arises for unagreement compared to

114

Chapter Five

standard, null-subject and overt-pronoun agreement unequivocally shows that person analysis is dealt with from early on. The temporal and functional independence of this effect from conceptual factors (Mancini et al. 2014, Experiment 4) further suggests that it is driven by morphosyntactic analysis, and that it can be reliably identified as a comprehension correlate of the Agree operation. Finally, the fact that early negative effects related to person-mismatch processing, whether in an unagreement or nonunagreement context, emerge in the same temporal window as numberrelated mismatch effects (between 300–500 milliseconds after verb reading) also point to person being initially handled by the same computational operation as number, even in unagreement configurations. One can therefore conclude that Agree is fully operative in unagreement contexts: a connection is established for morphosyntactic checking purposes between the verb and the nearest available nominal element bearing phi values. Agree between the two sets of phi-features is arguably not the only formal operation that the parser performs. A connection is also established between the subject’s and verb’s features and their interpretive anchors, to check their compatibility at the semantic-discourse level. This operation is computationally similar to Agree, in that it connects two positions, but differs from Agree in the set of values that it targets. While Agree between the two sets of phi values sanctions a mismatch, sigma value checking does not, and this rescues the derivation from crashing. If no implausibility arises, the unagreeing s-v relation is eventually interpreted and discourse roles are assigned, deriving the overall 1st person plural interpretation. Yet, if a mismatch is sanctioned at the phi and sigma level, as in outright person mismatches, no discourse role can be assigned and no interpretation can be derived.

Summary This chapter set out with the ambitious goal of enriching current theoretical and psycho-/neuro-linguistic descriptions of agreement by separating the morphosyntactic from the semantic-discourse components of its representation, using unagreement as a testing ground. Importantly, the grammatical mismatch that characterises unagreement allowed disentangling of the cognitive and neuro-physiological processes involved in feature matching from those triggered by mere syntactic errors and by the analysis of extra-syntactic information, commonly confounded or obscured in agreement-violation paradigms. Precise processing stages have been identified, along with the computational operations that characterise them, showing a division of

Functional and Temporal Dissociation of Agreement Mechanisms

115

labour between early operations, aimed at checking the phi-features involved in an agreement dependency, and later processes, which target the interface between morpho-syntax and discourse and that are able to significantly shape processing mechanisms. Interestingly, functional dissociations were not limited to the behavioural and electrophysiological domain, they also emerged at the neuroanatomical level. Within the left fronto-temporal network involved in language processing, a dissociation could be envisaged between areas supporting the performance of checking mechanisms at the morphosyntactic level (left dorsolateral middle frontal regions) and areas involved in integrating representations at higher levels of analysis (anterior MTG and AG).

CHAPTER SIX ANCHORING AGREEMENT

The goal of this work was to carry out a comprehensive investigation of s-v agreement that could shed light on the theoretical, processing and neuro-biological underpinnings of this linguistic phenomenon. The starting point of this endeavour was the observation that the comprehension of agreement dependencies implies the interpretation of distinct types of information – who is the speaker of an utterance, how many entities are involved in an event – which, in languages like Spanish and Italian, are syncretically expressed in the same inflectional morpheme. This many-to-one relationship reflects the theoretical divide between standard minimalist analyses (Chomsky 1995, 2001, 2005) and more recent cartographic approaches (Cinque and Rizzi 2008; Rizzi 2004; Shlonsky 1989, 2000, 2009; Sigurdsson 2004, 2009; Sigurdsson and Holmberg 2008). As Cinque and Rizzi (2008) observe, a fundamental opposition seems to divide the two approaches. The complexity of cartographic representations appears to forcefully clash with the simplicity of the generative devices upon which minimalist syntax relies. However, what appears to be an inherent tension can be regarded as ‘a fruitful division of labour’ (Cinque and Rizzi 2008, 49). On the one side, Minimalism focuses on the generative devices to be employed for the derivation of syntactic structures, which are reduced to extremely simple combinatorial operations (internal and external Merge) plus a search operation (Agree) which the system exploits for the building of structural dependencies. On the other side, cartographic projects make use of the generative devices developed by Minimalism in order to derive fine-grained syntactic structures that identify the ‘atoms’ of syntactic computations (Cinque and Rizzi 2008, 50). In this work, the core tenets of both lines of research have been exploited to derive a unified account of agreement feature representation and interpretation. The FIP serves exactly this purpose and epitomises the fruitful combination of minimalist devices (Agree, in this case) and the cartographic fine-grained mapping of syntactic structures. On the one hand, in keeping with standard minimalist assumptions, the FIP in (1)

Anchoring Agreement

117

below presupposes the same Agree operation at work for the licensing and interpretation of person and number. On the other hand, following a cartographic approach, the FIP predicts a distinction between person and number based on the different interpretive properties associated with the two features (their ‘anchors’), which correlate with positional differences: (1) FEATURE INTERPRETATION PROCEDURE (FIP): Features are structurally differentiated and interpreted in relation to their anchors: a) Person’s anchor resides in the speech act participants representation; b) Number’s anchor resides in the number specification of the nominal argument.

This hypothesis was tested through a series of experiments aimed at unveiling the mechanisms underlying person and number agreement processing within different environments (standard agreement and unagreement) and with different types of subjects. Evidence has been provided that supports a fine-grained decomposition of agreement projections, agreement processing mechanisms and neuroanatomical substrates. In the following, the implications of these findings are discussed from theoretical and neurocognitive perspectives.

From feature bundles to feature anchors The starting point of this investigation on agreement was the comparison between standard minimalist and cartographic approaches to feature representation, from which a substantial opposition emerged. An aspect of the minimalist approach to s-v agreement that was highlighted concerned the structural site where uninterpretable features are located. Chomsky’s (1995) assumption is that person and number form a cluster (together with gender, case and tense) hosted under the same syntactic head – the T node – which is responsible for the expression of tense. It is in T that all uninterpretable features are clustered, without any structural distinction among them (see Figure 1-1). A straightforward consequence of this implementation is that Agree accesses the whole feature bundle in a unique computational step, and not in a series of distinct operations, one for each feature to be valued. The one morphosyntactic property-one feature-one head approach postulated by Cartography (Cinque and Rizzi 2008) predicts that person and number occupy distinct structural positions, in light of their different interpretive properties. The analysis of linguistic data from a variety of

118

Chapter Six

languages has led to the identification of a projection specifically devoted to the checking of person – PersonP – in a higher position within the inflectional layer of the sentence compared to NumberP (see Figure 6-1). Postulating two distinct positions for the two features has an important theory-internal consequence, namely that the two features undergo separate checking (Sigurdsson and Holmberg 2008) and that Agree has a richer search domain. Moreover, the interpretation of the two features should be subject to different constraints: the discourse-relatedness of person should involve the analysis of extra-syntactic information that is encoded in the left periphery of sentence structure. In contrast, the interpretation of the information conveyed by number should rely on the cardinality information available on the subject argument, hosted clauseinternally in the specifier of IP. The FIP was precisely meant to capture the interpretive differences between the two features. The results of the experiments illustrated in Chapters 3 and 4 forcefully point to the presence of a distinct-cluster representation of agreement features that mainly hinges on their distinct interpretive properties. A deictic component is present in person information that may crucially shape the way this feature is licensed and interpreted. As observed by Bianchi (2006) and Sigurdsson (2004), person can only be interpreted in relation to the participants in the speech act. An anchoring is therefore established that connects the purely ‘morphosyntactic’ person head in the inflectional layer of the sentence with the logophoric participants in the left periphery of the sentence (Sigurdsson 2004) or the Logophoric Centre of the utterance (Bianchi 2003). No such IP-CP anchoring is necessary for number interpretation, which is locally fulfilled by anchoring the verbal specification to the nominal one. The processing system has proven to be highly flexible in the way it deals with the different information conveyed by person and number features. Evidence for a person-number dissociation comes from behavioural, electrophysiological and neuroanatomical data and points to qualitative as well as quantitative differences. Evidence in favour of the claim that person and number occupy distinct positions in the syntactic structure comes primarily from the analysis of neuroanatomical correlates of person and number agreement violations. The quantitative difference between person and number agreement that emerged in the posterior portion of the MTG can be interpreted as the consequence of accessing morphosyntactic information at different positions in the syntactic tree. While the extraction of person requires accessing information contained in high nodes of sentence structure, the extraction of number does not,

Anchoring Agreement

119

leading to differential processing costs and hence to differential patterns of activation in this brain region. As previously mentioned, accessing the feature information contained in different syntactic nodes implies the performance of separate, but formally similar, checking mechanisms. The set of ERP data reported in Chapter 3 points to person and number checking mechanisms as taking place concurrently, i.e. within overlapping processing stages, and with formally similar operations. Between 300–500 milliseconds (after verb presentation), the violation of the two features triggers a similar effect in left anterior electrode clusters, but only the effect of person anomalies extended to posterior scalp areas. Mancini and colleagues (Mancini et al. 2011a) interpreted person’s broad negativity as a superimposition of a LAN and an N400-like effect, which suggests that the two features arguably undergo formally similar morphosyntactic (or phi-) checking mechanisms, but different Anchor operations. That person and number undergo similar morphosyntactic checking mechanisms is confirmed also by the two features’ shared activation in the middle frontal gyrus, an area that has consistently shown sensitivity to morphosyntactic anomalies of different types (Folia et al. 2009; Kuperberg et al. 2003, 2009; Nieuweland et al. 2008). Yet, the activation and inspection of distinct anchors by the two features for proper interpretation produce qualitatively different neuroanatomical correlates. A remarkable asymmetry between person and number was revealed in the anterior part of the MTG, where the former produced a significant increase of activation but the latter did not. Activity in this region has been associated with the integration of different types of information to derive the propositional meaning of a sentence (Bornkessel-Schlesewsky and Schlesewsky 2013; Pallier et al. 2011). This asymmetry could be driven by the assignment of interpretively relevant roles (discourse roles) that characterises person but not number interpretation. In other words, while identifying and assigning a discourse role to the subject argument is crucial for the derivation of the propositional meaning of the sentence, the identification of whether this argument refers to a single entity or to a multitude of entities is not, hence the qualitatively different response for the two violations in the anterior part of the MTG. That blocking the assignment of a discourse role has more severe interpretive consequences than the inability to derive cardinality information is further confirmed by reading times for person and number violation obtained with a self-paced reading paradigm (Mancini et al. 2014a). A further distinction can be operated within the person feature, namely between 1st, 2nd and 3rd person, on the basis of the ERP experiment

120

Chapter Six

presented in Chapter 4 (Figure 6-1). Importantly, in the presence of number anomalies, 1st/2nd and 3rd person pronouns have shown to yield qualitatively different electrophysiological responses. This dissociation can be ascribed to the qualitatively different anchoring with speech act participants that 1st/2nd and 3rd person pronouns show – a positive, explicit anchoring in the former case; a negative one in the latter (Bianchi 2006; Sigurdsson 2004) – which naturally follows from of the interpretive divide between person and number. Again, interpretive differences can correlate with positional differences, which, in cartographic terms, can be implemented by locating 1st and 2nd person at a hierarchically higher position than 3rd person, along the lines of Bianchi’s (2006), Shlonsky’s (2009) and Sigurdsson’s (2004) proposals.

Figure 6-1. Cartography of agreement features. First, 2nd and 3rd person are structurally separated, with 1st/2ndP hierarchically higher than 3rdP. These are in turn located in a higher position than NumberP.

The projection responsible for person agreement can thus be decomposed into two independent projections: the 1st/2ndperson phrase and the 3rd person phrase. In the spirit of Sigurdsson (2004, 2009), a matching relation is supposed to anchor the two independent projections to the Speech Participant Phrase in the left periphery of the sentence. Having outlined the theoretical relevance of the empirical findings described in the preceding chapters, let us now discuss how they can be framed in a neurocognitive account of agreement processing.

Representations, algorithms and neuroanatomical bases of agreement Regardless of the architecture on which the online parser operates (serial, interactive or cascaded), three basic mechanisms can be identified

Anchoring Agreement

121

that constitute the backbone of the online comprehension of linguistic input: structure analysis, integration and monitoring processes. In Chapter 2 specific hypotheses were delineated concerning how the FIP could shape and guide the online parser during the performance of these basic mechanisms. The data discussed so far have shown the validity of the predictions made by the FIP and have suggested that the comprehension of an agreement dependency hinges on composite feature-based representations and computations that shape processing and its neuroanatomical correlates. Nominal and verbal features represent basic representations (or atoms) that are used by the parser during the online analysis of agreement relations. The parser takes the features extracted from nominal and verbal elements as input and checks whether they are consistent, so that a relation can be established between them. The output is a feature that has been assigned the appropriate discourse role (person) and cardinality information (number). Two operations drive the conversion of input into output features, namely Agree and Anchor. Agree operates on the phi values of nominal and verbal features (i.e. the morphosyntactic value of person and number) and checks their consistency. Anchor operates on sigma values and verifies the compatibility of features’ indices, to ensure that the presupposition underlying the interpretation of a feature is not violated (i.e. presence or absence of a speaker). The algorithm guiding the comprehension of an agreement relation comprises a series of discrete steps in which Agree and Anchor operate in a serial fashion. The initial stage of agreement comprehension involves structural analysis and the decomposition of the linguistic input to extract features, a mechanism that is sensitive to the distinct position in the syntactic tree that each feature occupies. The extraction of morphosyntactic features triggers a checking mechanism, with the goal of integrating incoming input into a local syntactic structure. Agree establishes a connection between the verb and an element that bears person and number features, and checks the consistency of their phi values. Importantly, given features’ structural dissociation, Agree targets each of them separately and, for each of them, sanctions whether a match or mismatch is found. Regardless of Agree’s match or mismatch output, Anchor is subsequently triggered to connect the phi values of each feature to their respective interpretive anchor (or sigma value). Like Agree, Anchor serves the purpose of connecting two positions and operates on person and number information separately, but the sets of features that it targets are different, and so are its reflexes on processing. Anchoring between the phi and sigma values drives interpretation: in the case of person, interpretation

122

Chapter Six

occurs when a discourse role has been assigned to the subject of the agreement dependency. As for number, interpretation amounts to identifying how many entities (an individual or a plurality) are involved in the speech event, regardless of their discourse role in it. If compatibility is found, the output of Anchor is shipped to further semantic analysis. The critical role of interpretive anchors in this architecture is evident in case of a mismatch, as they can steer processing to the effect that grammaticality is eventually acknowledged even if a mismatch has been sanctioned during morphosyntactic analysis. Transforming input features into their output representations involves the nodes of an extended left fronto-temporal network of brain regions, where feature-based sensitivity is found. Access to different syntactic positions for the extraction of feature information appears to be supported by the posterior portion of the left MTG. Although the data discussed here and in preceding chapters do not provide an exact timing of this operation with respect to the integration phase, existing MEG investigations on Spanish gender and number agreement processing (Molinaro et al. 2013) have evidenced that morphological decomposition of verbs to extract feature information occurs as early as around 170 milliseconds, and actively involves the posterior MTG, suggesting the temporal precedence of this step with respect to integration mechanisms. The information extracted from the linguistic input in the posterior part of the MTG could be made available for morphosyntactic and semantic-discourse integration to further regions in the network, via ventral and dorsal connections to the anterior temporal cortex and left frontal areas (Bornkessel-Schlesewsky and Schlesewsky 2013; Dronkers et al. 2004; Hagoort 2013; Molinaro et al. 2015; Papoutsi et al. 2011; Saur et al. 2008; Wilson et al. 2012). During the integration stage, evaluation of s-v morphosyntactic and semantic-discourse fit is supported by frontal areas, with a division of labour between the middle frontal gyrus and more anterior regions, such as the pars opercularis and triangularis of the LIFG. The former regions appear to be actively engaged in the evaluation of the morphosyntactic fit between subject and verb operated by Agree. This interpretation is supported by previous studies reporting the involvement of this region in the analysis of morphosyntactic mismatches (Folia et al. 2009; Kuperberg et al. 2003, 2008; Nieuweland et al. 2012), but also by the LAN effect that both person and number violations elicited. Studies investigating the processing of mismatches in domains other than language syntax, such as music syntax, have localised the source of this early negative effect in the middle frontal gyrus (Maess et al. 2001), suggesting that this region plays

Anchoring Agreement

123

a crucial role in the checking of consistency between stimuli across domains (see Koelsch et al. 2005 for a discussion). In contrast, the pars triangularis and orbitalis would be engaged in the incremental evaluation of the fit between the phi and sigma values of features carried out by Anchor, in line with views that propose the involvement of anterior regions of the LIFG in the analysis of meaning at the sentence level (Friederici 2011; Friederici and Gierhan 2013; Hagoort 2005; Vigneau et al. 2006). The performance of Anchor also relies on the anterior portion of the MTG, which could play a more specific role in the assignment of the discourse roles that person interpretation implies. This interpretation fits well with the N400 effect found for person violations relative to correct agreement. Indeed, it is possible that the anterior portion of the left temporal lobe is involved in the generation of this negative component (Lau et al. 2008, 2013), although further investigation is needed to confirm this hypothesis. Moreover, the functional connection between anterior LIFG areas and the anterior MTG for semantic-discourse analysis is corroborated neuroanatomically by the presence of a ventral pathway that connects the two brain areas (see Friederici 2011). Incremental integration of incoming input into the morphosyntactic and semantic-discourse representation of a sentence also involves domaingeneral working memory and conflict-monitoring mechanisms that are supported by regions in the dorsolateral prefrontal cortex and the ACC. This network provides the scaffolding necessary for the parser to maintain elements of the linguistic input available in memory, so that relations can be established and error signals can be sent, as happens in the presence of a clear conflict between what is expected (e.g. a correctly inflected verb) and what is encountered (e.g. a verb incorrectly inflected in person or number). Importantly, several studies have linked the anterior cingulate cortex to the generation of late positive ERP effects (Du et al. 2013; Olichney et al. 2010) such as those elicited by both person and number anomalies, which further confirms the tight interaction between these two regions in conflict-monitoring operations. Worthy of note is the sensitivity of the conflict-monitoring system to the different interpretive consequences that person and number agreement violations have. Across agreement violations (person and number) and languages tested (Spanish and Italian), modulations were found in the early stage of the P600 effects, which were associated with a featuresensitive diagnosis stage aimed at detecting the source of the anomaly. Specifically, broad/frontal P600 effects were elicited by person violations and by number violations with 1st/2nd person subjects, while the distribution of P600 effects for number violations (both with lexical and

124

Chapter Six

3rd person pronoun subjects) was prevalently posterior. The errormonitoring system could therefore dissociate between different sources of conflicts in the linguistic domain: conflicts in the anchoring mechanism that links person inflection to the discourse representation, and conflicts generated by a problem in the anchoring of number inflection with the semantic representation of the subject argument.

Towards a comprehensive feature-based account of agreement processing The feature-based approach to agreement comprehension proposed above generates several questions concerning the applicability of its representations, algorithms and neuroanatomical implementation to further features and the relations they are involved in, such as case, tense and gender agreement. Understanding the propositional meaning of a sentence is not limited to determining who the speaker and addressee are and how many entities are involved in the speech event. It also encompasses understanding “who does what to whom”, for which the analysis of case information and its mapping to thematic roles is required. Because of the interpretive relevance that thematic role assignment has, one can plausibly hypothesise the performance of anchoring mechanisms similar to those characterising person interpretation. The processing of case has been mostly investigated using ERP paradigms, with designs that comprised two nominative- or two accusative-marked arguments (German: Frisch and Schlesewsky 2001, 2005; Japanese: Müller et al. 2005, 2007) or two nominative-/two ergative-case marked arguments (Hindi: Choudhary et al. 2009), a manipulation that is meant to block the mapping of each argument onto a specific thematic role. Across languages, a biphasic N400-P600 pattern has been found which recalls the ERP pattern found for person agreement anomalies in Spanish. Further corroboration to this hypothesis comes from the processing of case and person agreement anomalies in Basque, where the two violations were found to elicit a biphasic N400-P600 pattern (Zawiszewski and Friederici 2009; Zawiszewski et al. 2011).1 A legitimate question concerns why violations of case marking in languages like English engender a LAN and not an N400 effect. Coulson et al. (1998) tested case violations in English by replacing accusative pronouns with nominative pronouns in sentences like “The plane took *we/us to paradise”, and found a LAN effect. One explanation for this asymmetry has to do with the type of manipulation used. As suggested by Friederici and Weissenborn (2007), different responses may be found

Anchoring Agreement

125

depending on whether the mismatch is between the number of available arguments and those required by the verb, or between the case marking of a given argument and that required by the verb. In the former case, which corresponds to the double-nominative/accusative manipulations described above, an N400 can be expected, as a result of the impossibility to link arguments to their precise thematic role. In the second case, which corresponds to the manipulation used by Coulson et al. (1998), a LAN is predicted. Indeed, regardless of the wrong case marking on the pronoun, positional information still allows the parser to identify who is the agent and who is the patient, and thus to assign an interpretation to the sentence. In other words, the effects of such violations would be limited within the morphosyntactic level of analysis. This explanation highlights an important parallelism between the interpretive consequences and processing correlates of case violations in languages that do not overtly mark case, and the processing correlates of number agreement violations. In contrast, when case is morphologically marked, as in German or Japanese, and the position of the argument does not provide any strong cue, violating case marking can severely compromise assessing who is the agent and who is the patient, as happens when a person mismatch blocks the assignment of a discourse role. Fundamental to sentence meaning is also the analysis of the tense information expressed on the verb and its congruence with an optional adverb. As illustrated in Chapter 1, a correlation has been hypothesised between person and tense agreement, in light of their common anchoring to discourse (Bianchi 2006; Sigurdsson 2004, 2010). The analysis of tense and person agreement violations should therefore evidence similar anchoring correlates. In line with this hypothesis, Biondo (2017) reports findings from an eye-tracking study in which the processing of person and tense agreement violations showed similar effects in late reading stages, which suggests that similar anchoring mechanisms are at work for the two features. Further investigation is, however, needed to confirm this initial result.2 Crucially, an important testing ground for the tense-person correlation is represented by the so-called tense agreement dissociation evidenced in agrammatism. Previous studies (Friedmann and Grodzinsky 1997; Friedmann 2001) have revealed a dissociation in this acquired language disorder between the better preservation of agreement processing compared to tense, which has been ascribed to agrammatic patients’ inaccessibility to the information contained in the left periphery of sentence structure. Critically, the comparison has been limited to number and tense agreement, while no study has so far assessed whether the asymmetry extends to person and number agreement. The dissociation

126

Chapter Six

found here between person and number therefore calls for a deeper investigation of agreement and tense processing in agrammatism. Finally, the feature-based approach developed here can be extended to the processing of gender agreement. More specifically, predictions can be made concerning an interpretive asymmetry between biological and grammatical gender, in virtue of their distinct interpretive properties. While the grammatical gender assigned to a Spanish inanimate noun like mesa (table) is completely arbitrary, because tables have no inherent property that makes them feminine entities, the feminine gender of a noun like abuela (grandmother) serves to classify the referent entity as female or male, which has a clear interpretive reflex. Differences between the two types of gender are therefore plausibly expected to emerge in anchoring mechanisms. This hypothesis was confirmed by a series of fMRI studies in which Quiñones (2015) investigated the neuroanatomical correlates of grammatical and biological gender agreement using a violation paradigm. The processing of both types of gender agreement elicited the activation of the posterior part of the MTG and of middle frontal areas, while correct gender agreement, irrespective of whether grammatical or biological, generated increased activation in the anterior part of the MTG. Crucially, only biological gender anomalies involved the activation of the AG, an area that has been claimed to support conceptual analysis across multiple domains (see Chapter 2). Thanks to its coupling with the anterior temporal cortex, the AG could plausibly mediate the anchoring of the grammatical and conceptual information conveyed by biological gender. Crucially, AG and anterior MTG involvement is not unique to biological gender processing: the processing of person agreement in unagreement contexts also triggers activation of these two brain regions, which suggests shared interpretive mechanisms between the two features. Further systematic investigation (possibly with multiple paradigms) is certainly necessary to extend the feature-based model proposed here to other languages and relations. Nonetheless, these preliminary parallelisms suggest that features represent an appropriate unit of analysis to explore the processing of a variety of linguistic phenomena and their neurobiological implementation

Relation to existing sentence comprehension models In this section, the relation between the feature-based model proposed here and existing models of sentence comprehension is discussed. The discussion will be centred on three models that share a similar theoretical and neurocognitive orientation, namely the NMSC by Friederici (2002,

Anchoring Agreement

127

2011), the MUC model by Hagoort (2003, 2005, 2014), and BornkesselSchlesewsky and Schlesewsky’s (2006, 2009) eADM. Overall, remarkable differences can be found among these four models as far as the localisation of linguistic functions in the brain is concerned. This is especially evident in the functional role attributed to the anterior and the posterior temporal cortices. Friederici (2011) considers the anterior temporal cortex as the locus of local syntactic structure building, as opposed to the critical role that this region serves for interface processing in the eADM and in the feature-based model proposed here. With respect to the posterior portion of the temporal cortex, Friederici’s (2011) and Bornkessel-Schlesewsky and Schlesewsky (2006) agree in linking this area to mechanisms that map syntactic structure onto thematic roles, while Hagoort (2005, 2014) correlates activity in this region with the extraction of lexical frames from semantic memory. This mechanism appears to be functionally similar to the structure analysis operation proposed by the feature-based model, although the lexicalist framework adopted by the MUC does not predict any difference between person and number agreement processing. The four models also differ in the functional role attributed to the LIFG. In line with Friederici (2002, 2011) and Hagoort (2003, 2005, 2014), the model proposed here assigns the LIFG a critical role in the language-specific mechanisms dealing with form- and meaning-related analysis, as opposed to Bornkessel-Schlesewsky and Schlesewsky’s (2013) domain-general account. Important differences also concern the architecture underlying agreement processing. While agreement’s core function is undisputed in the four models under discussion, the computational, algorithmic and neuroanatomical bases that support the processing of these relations appear to be mostly underspecified in Hagoort’s, Friederici’s and Bornkessel-Schlesewsky and Schlesewsky’s formalizations. Clearly, none of these proposals predicts dissociations in processing mechanisms and neuro-physiological correlates between person and number, or among other features, and their architectures cannot easily make room for them. While Friederici’s (2011) NMSC is the model that could most easily accommodate the findings from Italian and Spanish presented here, due to the serial time course and the theoretical framework on which it rests, it differs from our model in one fundamental aspect. Specifically, phase 2 of the NMSC describes morphosyntactic and verb-argument relation processing as two parallel streams, with this distinction relying on three precise assumptions, namely that the two types of relations have different goals (thematic vs. grammar role assignment), generate different types of

128

Chapter Six

ERP responses (N400 vs. LAN) and involve different neuroanatomical substrates (posterior MTG vs. LIFG). Critically, the tenability of this distinction gets weaker and less justified once a feature-based analysis is adopted, thanks to the fine-grained similarities and differences between features and their processing that are captured. As observed above, person and case processing share the goal of assigning an interpretively relevant goal. The blocking of this mechanism inevitably prevents the parser from determining “who does what” and “who is the speaker” of the utterance, and the biphasic N400-P600 ERP patterns for person and case anomalies seem to underscore this interpretive similarity. Moreover, the processing of agreement information, be it person-, number- or gender-related, actively engages the posterior portion of the MTG/STS, which further weakens the claim of a stark functional distinction between these two linguistic aspects. In the MUC model, the clear-cut division between what belongs to the syntax and what belongs to the semantics domain, also reflected neuroanatomically in the segregation between the syntactic and semantic unification workspaces, does not allow for an articulate account of agreement processing. In this model, agreement pertains to the syntactic domain and is accomplished by unifying two lexical frames with consistent morphosyntactic and semantic information. This coarse-grained distinction between syntactic and semantic processing critically obscures fundamental operations at the interface between the two levels of analysis. As a matter of fact, the electrophysiological and neuroanatomical correlates of Spanish and Italian agreement processing do not fit in either a strictly syntactic or strictly semantic functional characterisation, but are suggestive of composite interface mechanisms. Contrary to the MUC model, emphasis on interface processing is the main feature of the eADM by Bornkessel-Schlesewsky and Schlesewsky (2006, 2009). In this model, argument interpretation and role identification is accomplished at the syntax-semantics interface, on the basis of a crosslinguistically defined set of prominence scales and their language-specific relevance. The authors describe an articulate set of processes that operate on a variety of linguistic configurations, with the goal of mapping arguments onto thematic roles using cues that vary across languages. As noted by the authors themselves (Bornkessel-Schlesewsky and Schlesewsky 2006), the crucial role of at least certain aspects of linguistic structure cannot be easily explained in general cognitive terms and extended to domains other than language. This marks a notable difference between the eADM and the proposed feature-based model. The basic operations carried out by Agree (and Anchor to some extent) can be straightforwardly linked

Anchoring Agreement

129

to more general mechanisms of integration that are required for the building of increasingly larger assemblies of units. Critically, integrative mechanisms of this type are not unique to language, which allows for a parallelism between processing mechanisms in the linguistic and nonlinguistic domain. Indeed, the idea that components of the grammar are language-specific has been long abandoned in the minimalist framework (see Hauser et al. 2002). Non-linguistic counterparts of agreement can be found in music, action control and visuo-spatial processing (see review in Tettamanti 2003), where features function as grouping criteria to organise simple units into higher-order assemblies. Consistent with this, it has been suggested that the establishment of relations in linguistic and nonlinguistic domains may involve common basic neural and computational mechanisms (Tettamanti et al. 2009), as evidenced by shared activation of BA44 in the LIFG during the processing of syntactic relations across words and across symbols. In the context of the feature-based model developed here, this overlap may be indicative of the fact that the processing of relations in the two domains share common feature-checking mechanisms. Needless to say, more research is necessary to assess to what extent the linguistic and the visuo-spatial domain share neural, computational and procedural properties. Nevertheless, features certainly represent an appropriate unit of analysis for a cross-domain comparison.

Linking agreement comprehension and production Also worthy of discussion is the relationship between agreement comprehension and agreement production. Crucially, the dissociation between the phi and sigma level of feature analysis that has been advanced recalls the Marking vs. Morphing distinction elaborated within the agreement production model proposed by Eberhard et al. (2005), the main difference being the direction of encoding: from morpho-syntax to semantics/discourse in comprehension, from semantics/discourse to morpho-syntax in production. In modelling number agreement production, Eberhard and colleagues (2005) describe a process – Marking – in which number information is imported from semantics into syntax. Importantly, this is the level at which agreement production can be influenced by conceptual factors. Anchor and its capacity to see into features’ semantics and discourse values can be regarded as the comprehension counterpart of the Marking operation described by Eberhard and colleagues. Similarly, a parallel can be envisaged between Morphing and Agree checking operations. In Eberhard et al.’s (2005) account, Morphing is the operation responsible for

130

Chapter Six

copying the morphosyntactic information extracted from the subject argument onto the verb. In comprehension, a similar function is performed by Agree in the feature-based model proposed here. This parallelism encourages more systematic research on the relationship between the comprehension and production of agreement. In this respect, one aspect that is worthy of investigation is certainly the comparison between person and number agreement, to assess whether the dissociation found in comprehension extends to the production domain.

Conclusion A neurocognitive model has been proposed that represents the first comprehensive framework for understanding real-time analysis of basic building blocks of language: features and the relations they are engaged in. The primary motivation for this investigation lay in the observation that the understanding of a relationship such as the one between a noun and a verb implies the interpretation of a “bundle” of different types of information expressed in verbal morphology. A precise hypothesis has stemmed from this observation, namely that the processing system and the neuroanatomical substrates that support it could be sensitive to the information conveyed by the single features. To test this hypothesis, feature bundles have been decomposed and the interpretive properties of features have been tested experimentally with a variety of paradigms. Across studies, data converge on the conclusion that there is more to agreement than a formal feature-checking procedure that targets a whole set of feature. Features are accessed separately, and their interpretation is guided by anchors, whose inspection provides the processing system with fundamental coordinates that assign arguments to interpretive roles even in non-optimal conditions (i.e. when features partially mismatch). While the present proposal will need more detailed neuroanatomical, computational and cross-domain characterisation in the future, it certainly provides a promising initial step forward in the endeavour to model the neurobiological bases of language architecture.

NOTES

Chapter One 1

Benoni verbs appear in a wide range of environments: they can express present tense, but they can also be gerundive complements to perception verbs and participles. What distinguishes the last two verbal forms is the absence of any kind of tense specification in them. 2 See also Greenberg’s (1963) Feature Hierarchy: Person>Number> Gender. Person, number and gender stand in an implicational relationship: if a language has gender agreement, then it also has number agreement. If a language has both gender and number agreement, it must also have person agreement. 3 The syntactic implementation of the speech act and speech participants as developed in recent cartographic approaches (Bianchi 2006; Sigurdsson 2004 and subsequent work) may be seen as a revised and updated version of the classic Performative Hypothesis (Ross 1970), according to which every sentence is associated with an explicit illocutionary act, i.e. it is derived from a deep structure containing a performative verb. For example, sentence (i) would be derived from sentence (ii): (i) I’ll write you next week (ii) I claim/I promise that I’ll write you next week.

Chapter Two 1

It should be noticed that the validity of the LAN as an electrophysiological correlate of sentence processing has recently been questioned. Specifically, recent studies have advanced the claim that LAN effects result from artefacts in the grand-average process that obscure possible individual variability in the ERP response (see Molinaro et al. 2011, 2015; Tanner et al. 2014, 2015 for a discussion). 2 The experimental paradigms that have traditionally tested the eLAN effect have been criticised on several methodological aspects. See Steinhauer and Drury (2012) for a detailed explanation.

Chapter Three 1

An extended discussion of the relationship between person and case processing is provided in Chapter 6. 2 For a further discussion on the tense-person correlation, see Chapter 6.

132

Notes

Chapter Four 1

A true 1st person plural may be claimed to exist as a conceptual category. If we interpret 1st person singular as making direct reference to the speaker, we can conceive 1st person plural as making reference to a mass of speakers, a group of people speaking together in unison. Mass speaking, as happens in church services, concerts or sport matches, is one such circumstance under which one can talk about a true 1st person plural. Along the same lines, one can identify a true 2nd person plural in the use of You to address an audience, as happens in class when addressing all of the students present at the time of utterance. In both cases, one is faced with conceptual categories that seem not to be grammaticalised in any language: so far, no language has been found that distinguishes the mass speaking meaning from the associative meaning of We, or the audience address meaning from the associative meaning of You (see Cysouw 2003 and references cited therein). 2 Third person null subjects in Hebrew are allowed only in contexts of nonstandard binding and/or control (Shlonsky 2009, and references cited therein). 3 Such sentences are regarded as acceptable in some non-standard dialects of English. See discussion in Kimball and Aissen (1971) and Kayne (2000). 4 In Kiowa, as in other languages, IOs are always interpreted as semantically animate, i.e. capable of experience, a restriction that is familiar from IndoEuropean languages. 5 A similar analysis has been put forth by Nevins (2007), who proposes a distinction between 1st/2nd and 3rd person based on a binary feature system (and not a privative one as Harley and Ritter (2002) use): 1st person: [+ author, +participant] 2nd person: [-author, + participant] 3rd person:[-author, - participant]. 6 The hypothesis that the speech event is syntactically encoded has been advanced in a number of independent proposals besides those of Bianchi (2006) and Sigurdsson (2004), which have been particularly emphasised here. Among these is the one by Speas and Tenny (2004), who propose to associate the speaker and the hearer of an utterance with the pragmatic roles of the agent and the goal of the utterance, while the utterance content is identified with the theme. This conception of speaker and hearer entails a direct relationship with the projection expressing the illocutionary force of an act, namely the Speech Act Phrase, corresponding to Rizzi’s (1997) ForceP. The representation of the context (and of participants in particular) that they give is quite distinct from the one adopted in the current analysis. In Speas and Tenny’s approach, speaker and hearer are arguments of a speech act denoting head, while in Bianchi’s (2006) and Sigurdsson’s (2004) analysis, they are participants of a speech event, hence purely Kaplanian parameters (Kaplan 1989). Such a difference has important consequences for a cartographic implementation of context, and the choice of either of the two approaches is thus not immaterial. For the current purposes, Bianchi’s (2006) and Sigurdsson’s (2004) analysis will be adopted.

Functional and Temporal Dissociation of Agreement Mechanisms

133

Chapter Five 1

As Rivero (2007) points out, the Spanish polite system can be regarded as another instance of unagreement: i. Ustedes hablan You.2.pl speak.3.pl ‘Youpl speak’ Moreover, unagreement patterns can also involve quantifiers: ii. Ninguno hablamos muchos idiomas No one speak.1.pl many languages ‘No one of us speak(s) many languages’ 2 An alternative explanation is available for the re-reading of the preverbal area in null-subject sentences. As suggested by Alonso-Ovalle et al. (2002) and Carminati (2005), the interpretation of a null pronoun requires the identification of a salient antecedent in the sentence. Re-reading of the preverbal area in null-subject conditions could therefore be due to difficulties in identifying a referent, and not to the assignment of verbal person and number features to the null pronoun. Further investigation is, however, necessary to disentangle this interpretation from the one advanced in the main text.

Chapter Six 1

Díaz et al. 2011 tested for double-ergative anomalies in Basque but failed to replicate the N400 effect found in Hindi, another language that uses ergative-case marking. As noted by the authors, a possible explanation for this difference could be the fact that Hindi is a split-ergative language, while Basque is not. 2 Experimental evidence on the processing of adverb-verb tense violations is rather sparse and mainly comes from ERP studies (Steinhauer and Ullman 2002; Baggio 2008; Qiu and Zhou 2012) in which heterogeneous experimental material was tested (i.e. adverb-verb violations were tested in different languages and at different linear and structural distance), leading to inconsistent results. A review of these studies can be found in Biondo (2017).

BIBLIOGRAPHY

Ackema, Peter; Neeleman, Ad. “Subset controllers in agreement relations.” Morphology, 23, no. 2 (2013): 291-323. Adger, David, and Daniel Harbour. “Syntax and Syncretisms of the Person Case Constraint.” Syntax 10, no. 1 (2007): 2-37. Alonso-Ovalle, Luis, Susana Fernández-Solera, Lyn Frazier, and Charles Clifton. “Null vs. Overt Pronouns and the Topic-Focus Articulation in Spanish.” Italian Journal of Linguistics 14 (2002): 151-70. Anagnostopoulou, Elena. “Strong and Weak Person Restrictions: A Feature Checking Analysis.” Clitic and affix combinations: Theoretical perspectives (2005): 199-235. Baggio, Giosue. “Processing Temporal Constraints: An ERP Study.” Language Learning 58, no. s1 (2008): 35-55. Baggio, Giosuè, and Peter Hagoort. “The Balance Between Memory and Unification in Semantics: A Dynamic Account of the N400.” Language and Cognitive Processes 26, no. 9 (2011): 1338-67. Baker, Mark C. The Syntax of Agreement and Concord. Vol. 115: Cambridge University Press, 2008. Bambini, V., C. Gentili, E. Ricciardi, P. M. Bertinetto, and P. Pietrini. “Decomposing Metaphor Processing at the Cognitive and Neural Level through Functional Magnetic Resonance Imaging.” Brain Research Bulletin 86, no. 3-4 (Oct 10 2011): 203-16. Barber, Horacio, and Manuel Carreiras. “Grammatical Gender and Number Agreement in Spanish: An ERP Comparison.” Journal of Cognitive Neuroscience 17, no. 1 (2005): 137-53. Baron, S. G., S. L. Thompson-Schill, M. Weber, and D. Osherson. “An Early Stage of Conceptual Combination: Superimposition of Constituent Concepts in Left Anterolateral Temporal Lobe.” Journal of Cognitive Neuroscience 1, no. 1 (Mar 2010): 44-51. Bates, Elizabeth, and Brian MacWhinney. “Competition, Variation, and Language Learning.” Mechanisms of Language Acquisition (1987): 157-93. Belletti, Adriana. “Generalized Verb Movement”. Rosenberg & Sellier, 1990.

Functional and Temporal Dissociation of Agreement Mechanisms

135

Belletti, Adriana. “Answering Strategies: New Information Subjects and the Nature of Clefts.” A. Belletti, Structures and Strategies, New York: Routledge (2008): 242-65. Bemis, D. K., and L. Pylkkanen. “Basic Linguistic Composition Recruits the Left Anterior Temporal Lobe and Left Angular Gyrus During Both Listening and Reading.” Cerebral Cortex (Jul 4 2012). Benveniste, Emile. “Problèmes De Linguistique Générale, Tome 1.” Paris, Gallimard (1966): 132-39. Bianchi, Valentina. “On Finiteness As Logophoric Binding.” Temps et Point de Vue/Tense and Point of view, Université de Paris X (2003) : 213-246. Bianchi, Valentina. “On the Syntax of Personal Arguments.” Lingua 116, no. 12 (2006): 2023-67. Binder, J. R., and R. H. Desai. “The Neurobiology of Semantic Memory.” Trends in Cognitive Science 15, no. 11 (Nov 2011): 527-36. Binder, J. R., R. H. Desai, W. W. Graves, and L. L. Conant. “Where Is the Semantic System? A Critical Review and Meta-Analysis of 120 Functional Neuroimaging Studies.” Cerebral Cortex 19, no. 12 (Dec 2009): 2767-96. Biondo, Nicoletta. “The syntactic side of time: processing adverb-verb temporal agreement”.Ph.D Dissertation, University of Trento, 2017. Blumenfeld, Robert S, and Charan Ranganath. “Dorsolateral Prefrontal Cortex Promotes Long-Term Memory Formation Through Its Role in Working Memory Organization.” Journal of Neuroscience 26, no. 3 (2006): 916-25. Bornkessel-Schlesewsky, I., and M. Schlesewsky. “Reconciling Time, Space and Function: A New Dorsal-Ventral Stream Model of Sentence Comprehension.” Brain and Language 125, no. 1 (Apr 2013): 60-76. Bornkessel-Schlesewsky, Ina, and Matthias Schlesewsky. “An Alternative Perspective on ‘Semantic P600’ Effects in Language Comprehension.” Brain Research Reviews 59, no. 1 (2008): 55-73. —. Processing Syntax and Morphology: A Neurocognitive Perspective. Vol. 6: Oxford University Press, 2009. Bornkessel, I., and M. Schlesewsky. “The Extended Argument Dependency Model: A Neurocognitive Approach to Sentence Comprehension across Languages.” Psychological Review 113, no. 4 (Oct 2006): 787821. Bornkessel, I., S. Zysset, A. D. Friederici, D. Y. von Cramon, and M. Schlesewsky. “Who Did What to Whom? The Neural Basis of Argument Hierarchies During Language Comprehension.” Neuroimage 26, no. 1 (May 15 2005): 221-33.

136

Bibliography

Bradley, Dianne C. and Kenneth I. Forster. “A Reader’s View of Listening.” Cognition 25.1 (1987): 103-13. Broca, Paul. “Remarques Sur Le Siège De La Faculté Du Langage Articulé, Suivies D’une Observation D’Aphémie (Perte De La Parole).” Bulletin et Memoires de la Societe Anatomique de Paris 6 (1861): 330-57. Carminati, Maria Nella. “Processing Reflexes of the Feature Hierarchy (Person > Number > Gender) and Implications for Linguistic Theory.” Lingua 115, no. 3 (2005): 259-85. Carreiras, M., L. Carr, H. A. Barber, and A. Hernandez. “Where Syntax Meets Math: Right Intraparietal Sulcus Activation in Response to Grammatical Number Agreement Violations.” Neuroimage 49, no. 2 (Jan 15 2010): 1741-9. Carreiras, Manuel, Elena Salillas, and Horacio Barber. “Event-Related Potentials Elicited During Parsing of Ambiguous Relative Clauses in Spanish.” Cognitive Brain Research 20, no. 1 (2004): 98-105. Carter, Cameron S, and Vincent Van Veen. “Anterior Cingulate Cortex and Conflict Detection: An Update of Theory and Data.” Cognitive, Affective, & Behavioral Neuroscience 7, no. 4 (2007): 367-79. Catani, Marco, Flavio Dell’Acqua, Francesco Vergani, Farah Malik, Harry Hodge, Prasun Roy, Romain Valabregue, and Michel Thiebaut de Schotten. “Short Frontal Lobe Connections of the Human Brain.” Cortex 48, no. 2 (2012): 273-91. Catani, Marco, and Marsel Mesulam. “The Arcuate Fasciculus and the Disconnection Theme in Language and Aphasia: History and Current State.” Cortex 44, no. 8 (2008): 953-61. Cinque, Guglielmo. Adverbs and Functional Heads: A Cross-Linguistic Perspective. Oxford University Press, 1999. Cinque, Guglielmo, and Luigi Rizzi. “The Cartography of Syntactic Structures.” Studies in Linguistics 2 (2008): 42-58. Clifton, Charles, Adrian Staub, and Keith Rayner. “Eye Movements in Reading Words and Sentences.” Eye Movements: a Window on Mind and Brain 27 Ed. By Roger Van Gompel, Martin Fisher, Wayne Murray and Robin Hill. Amsterdam: Elsevier, 2007, 341-72. Clifton Jr, C, and Adrian Staub. “Syntactic Influences on Eye Movements in Reading.” The Oxford Handbook of Eye Movements, Ed. By Simon Liversedge, Iain Gilchrist, Stefan Everling, Oxford: Oxford University Press, 2011, 895-909. Corbett, Greville G. Agreement. Cambridge: Cambridge University Press, 2006.

Functional and Temporal Dissociation of Agreement Mechanisms

137

Cysouw, Michael. The Paradigmatic Structure of Person Marking. Oxford: Oxford University Press, 2003. Chomsky, Noam. “Aspects of Theory of Syntax.” Cambridge: MTT Press (1965). —. “Minimalist Inquiries: The Framework. Step by Step: Essays on Minimalist Syntax in Honor of Howard Lasnik.”, Ed. By Roger Martin, David Michaels, and Juan Uriagereka, Cambridge, MA: MIT Press, 2000, 89–155. —. The Minimalist Program. Vol. 1765: Cambridge Univ Press, 1995. —. “Three Factors in Language Design.” Linguistic Inquiry 36, no. 1 (2005): 1-22. Choudhary, K. K., M. Schlesewsky, D. Roehm, and I. BornkesselSchlesewsky. “The N400 as a Correlate of Interpretively Relevant Linguistic Rules: Evidence from Hindi.” Neuropsychologia 47, no. 13 (Nov 2009): 3012-22. De Vincenzi, Marica. “Differences between the Morphology of Gender and Number: Evidence from Establishing Coreferences.” Journal of Psycholinguistic Research 28, no. 5 (1999): 537-53. De Vincenzi, Marica, and Elisa Di Domenico. “A Distinction among PhiFeatures: The Role of Gender and Number in the Retrieval of Pronoun Antecedents.” Rivista di Linguistica 11, no. 1 (1999): 41-74. Deutsch, Avital. “Subject-Predicate Agreement in Hebrew: Interrelations with Semantic Processes.” Language and Cognitive Processes 13, no. 5 (1998): 575-97. Deutsch, Avital, and Shlomo Bentin. “Syntactic and Semantic Factors in Processing Gender Agreement in Hebrew: Evidence from ERPs and Eye Movements.” Journal of Memory and Language 45, no. 2 (2001): 200-24. Di Domenico, Elisa. Per una Teoria del Genere Grammaticle. Padova: Unipress, 1997. Dillon, Brian, Andrew Nevins, Alison C Austin, and Colin Phillips. “Syntactic and Semantic Predictors of Tense in Hindi: An ERP Investigation.” Language and Cognitive Processes 27, no. 3 (2012): 313-44. Dronkers, N. F., D. P. Wilkins, R. D. Van Valin, Jr., B. B. Redfern, and J. J. Jaeger. “Lesion Analysis of the Brain Areas Involved in Language Comprehension.” Cognition 92, no. 1-2 (May-Jun 2004): 145-77. Du, Xue, Yigui Qin, Shen Tu, Huazhan Yin, Ting Wang, Caiyun Yu, and Jiang Qiu. “Differentiation of Stages in Joke Comprehension: Evidence from an ERP Study.” International Journal of Psychology 48, no. 2 (2013): 149-57.

138

Bibliography

Eberhard, K. M., J. C. Cutting, and K. Bock. “Making Syntax of Sense: Number Agreement in Sentence Production.” Psychological Review 112, no. 3 (Jul 2005): 531-59. Faussart, Christel, Celia Jakubowicz, and Marianne Costes. “Gender and Number Processing in Spoken French and Spanish.” Rivista di Linguistica 11, no. 1 (1999): 75-101. Fedorenko, E., J. Duncan, and N. Kanwisher. “Language-Selective and Domain-General Regions Lie Side by Side within Broca’s Area.” Current Biology 22, no. 21 (Nov 6 2012): 2059-62. Fedorenko, E., and S. L. Thompson-Schill. “Reworking the Language Network.” Trends in Cognitive Science 18, no. 3 (Mar 2014): 120-6. Felser, Claudia, Harald Clahsen, and Thomas F Münte. “Storage and Integration in the Processing of Filler-Gap Dependencies: An ERP Study of Topicalization and Wh-Movement in German.” Brain and Language 87, no. 3 (2003): 345-54. Filik, Ruth, Anthony J Sanford, and Hartmut Leuthold. “Processing Pronouns without Antecedents: Evidence from Event-Related Brain Potentials.” Journal of Cognitive Neuroscience 20, no. 7 (2008): 131526. Folia, Vasiliki, Christian Forkstam, Peter Hagoort, and Karl Magnus Petersson. “Language Comprehension: The Interplay between Form and Content.” Paper presented at the Cognitive Science Society, 2009. Forcheimer, P. The Category of Person. Berlin, 1953. Frazier, Lyn. “Sentence Processing: A Tutorial Review.” Attention and Perception XII: The Psychology of Reading. Ed. by Michael Coltheart, Hillsdale, NJ: Lawrence Erlbaum, 1987, 559-586. Frazier, Lyn, and Charles Clifton. Construal. MIT Press, 1996. Frazier, Lyn, and Keith Rayner. “Making and Correcting Errors During Sentence Comprehension: Eye Movements in the Analysis of Structurally Ambiguous Sentences.” Cognitive Psychology 14, no. 2 (1982): 178-210. Friederici, A. D. “The Cortical Language Circuit: From Auditory Perception to Sentence Comprehension.” Trends in Cognitive Sciences 16, no. 5 (May 2012): 262-8. —. “Neural Basis of Auditory Sentence Processing.” Trends in Cognitive Science (2002). Friederici, A. D., and S. M. Gierhan. “The Language Network.” Current Opinion in Neurobiology 23, no. 2 (Apr 2013): 250-4. Friederici, A. D., and S. Kotz. “Brain Basis of Syntactic Processes.” NeuroImage 20 (2003): S8-S17.

Functional and Temporal Dissociation of Agreement Mechanisms

139

Friederici, A. D., M. Meyer, and D. Y. Von Cramon. “Auditory Language Comprehension: An Event-Related fMRI Study on the Processing of Syntactic and Lexical Information.” Brain and Language 74, no. 289300 (2000). Friederici, A. D., Rueschemeyer, A. Hahne, and C. Fiebach. “T Role of Left Inferior Frontal and Superior Temporal Cortex in Sentence Comprehension: Localizing Syntactic and Semantic Processes.” Cerebral Cortex 13 (2003): 170-77. Friederici, A. D., and J. Weissenborn. “Mapping Sentence Form onto Meaning: The Syntax-Semantic Interface.” Brain Research 1146 (May 18 2007): 50-8. Friederici, Angela D. “The Brain Basis of Language Processing: From Structure to Function.” Physiological Reviews 91, no. 4 (2011): 135792. Friederici, Angela D, and Kai Alter. “Lateralization of Auditory Language Functions: A Dynamic Dual Pathway Model.” Brain and Language 89, no. 2 (2004): 267-76. Friederici, Angela D, Anja Hahne, and Axel Mecklinger. “Temporal Structure of Syntactic Parsing: Early and Late Event-Related Brain Potential Effects.” Journal of Experimental Psychology-Learning Memory and Cognition 22, no. 5 (1996): 1219-48. Friederici, Angela D, Anja Hahne, and Douglas Saddy. “Distinct Neurophysiological Patterns Reflecting Aspects of Syntactic Complexity and Syntactic Repair.” Journal of Psycholinguistic Research 31, no. 1 (2002): 45-63. Friederici, Angela D, Erdmut Pfeifer, and Anja Hahne. “Event-Related Brain Potentials During Natural Speech Processing: Effects of Semantic, Morphological and Syntactic Violations.” Cognitive Brain Research 1, no. 3 (1993): 183-92. Friederici, Angela D., and Stefan Frisch. “Verb Argument Structure Processing: The Role of Verb-Specific and Argument-Specific Information.” Journal of Memory and Language 43, no. 3 (2000): 476507. Friedmann, Na'ama. “Agrammatism and the Psychological Reality of the Syntactic Tree.” Journal of Psycholinguistic Research 30, no. 1 (2001): 71-90. Friedmann, Na'ama, and Yosef Grodzinsky. “Tense and Agreement in Agrammatic Production: Pruning the Syntactic Tree.” Brain and Language 56, no. 3 (1997): 397-425. Garrod, Simon C, and Anthony J Sanford. “Resolving Sentences in a Discourse Context: How Discourse Representation Affects Language

140

Bibliography

Understanding.” Handbook of Psycholinguistics, Academic Press (1994): 675-698. Gazzaniga, Michael S. Cognitive Neuroscience: A Reader. Blackwell Publishing, 2000. Geschwind, Norman. “Disconnexion Syndromes in Animals and Man.” Brain 88, no. 3 (1965): 585-85. Greenberg, Joseph H. “Some Universals of Grammar with Particular Reference to the Order of Meaningful Elements.” Universals of Language 2 (1963): 73-113. Grodzinsky, Y., and A. D. Friederici. “Neuroimaging of Syntax and Syntactic Processing.” Current Opinion in Neurobiology 16, no. 2 (Apr 2006): 240-6. Gunter, Thomas C, Angela D Friederici, and Herbert Schriefers. “Syntactic Gender and Semantic Expectancy: ERPs Reveal Early Autonomy and Late Interaction.” Journal of Cognitive Neuroscience 12, no. 4 (2000): 556-68. Gunter, Thomas C, Laurie A Stowe, and Gusbertus Mulder. “When Syntax Meets Semantics.” Psychophysiology 34, no. 6 (1997): 660-76. Hagoort, P. “Muc (Memory, Unification, Control) and Beyond.” Frontiers in Psychology 4 (2013): 416. —. “Nodes and Networks in the Neural Architecture for Language: Broca’s Region and Beyond.” Current Opinion in Neurobiology 28 (Oct 2014): 136-41. —. “On Broca, Brain, and Binding: A New Framework.” Trends in Cognitive Sciences 9, no. 9 (Sep 2005): 416-23. Hagoort, Peter, Colin M. Brown and John Grothusen. “The syntactic positive shift (SPS) as an ERP measure of syntactic processing.” Language and Cognitive Processes 8 (1993): 439-483. Hagoort, Peter. “How the Brain Solves the Binding Problem for Language: A Neurocomputational Model of Syntactic Processing.” Neuroimage 20 (2003): S18-S29. Hagoort, Peter, and Colin M Brown. “ERP Effects of Listening to Speech Compared to Reading: The P600/SPS to Syntactic Violations in Spoken Sentences and Rapid Serial Visual Presentation.” Neuropsychologia 38, no. 11 (2000): 1531-49. Hagoort, Peter, Colin M Brown, and Lee Osterhout. “The Neurocognition of Syntactic Processing.” In The Neurocognition of Language, 273317: Oxford University Press, 1999. Hahne, Anja, and Angela D Friederici. “Electrophysiological Evidence for Two Steps in Syntactic Analysis: Early Automatic and Late Controlled

Functional and Temporal Dissociation of Agreement Mechanisms

141

Processes.” Journal of Cognitive Neuroscience 11, no. 2 (1999): 194205. Handy, Todd C. “Basic principles of ERP quantification.” Event-related Potentials: A Methods Handbook. MIT Press (2005): 33-55. Harley, Heidi, and Elizabeth Ritter. “Structuring the Bundle: A Universal Morphosyntactic Feature Geometry.” Pronouns: Features and representation. Ed. By Horst Simon and Heike Weise, Amsterdam/ Philadelphia: John Benjamins, 2002, 23-39. Hickok, Gregory, and David Poeppel. “The Cortical Organization of Speech Processing.” Nature Reviews Neuroscience 8, no. 5 (2007): 393-402. Hinojosa, José, Manuel Martín-Loeches, Pilar Casado, Francisco Muñoz, and Francisco Rubia. “Similarities and Differences between Phrase Structure and Morphosyntactic Violations in Spanish: An EventRelated Potentials Study.” Language and Cognitive Processes 18, no. 2 (2003): 113-42. Höhn, Georg FK. “Unagreement is an Illusion.” Natural Language & Linguistic Theory 34, no. 2 (2016): 543-92. Kandel, Eric, Schwartz James, Jessel, Thomas, Siegelbaum, Steven, Hudspeth, A. Principles of Neural Science. New York: McGraw-Hill, Health Professions Division, 2013. Huettel, SA, AW Song, and G McCarthy. Functional Magnetic Resonance Imaging. Sunderland, MA: Sinauer Associates, 2014. Humphries, Colin, Jeffrey R Binder, David Medler, and Einat Liebenthal. “Syntactic and Semantic Modulation of Neural Activity During Auditory Sentence Comprehension.” Journal of Cognitive Neuroscience 18, no. 4 (2006): 665-79. Hurtado, Alfredo. “The Unagreement Hypothesis.” Selected paper presented at the thirteenth linguistic symposium on romance languages, 1985. Igoa, José Manuel, José Eugenio García-Albea, and Rosa Sánchez-Casas. “Gender-Number Dissociations in Sentence Production in Spanish.” Rivista di Linguistica 11, no. 1 (1999): 163-96. Jakobson, Roman. “Language in Relation to Other Communication Systems.” Selected Writings 2, The Hague, Netherlands: Mouton de Gruyter, 1971, 570-79. Kaan, Edith, Anthony Harris, Edward Gibson, and Phillip Holcomb. “The P600 as an Index of Syntactic Integration Difficulty.” Language and Cognitive Processes 15, no. 2 (2000): 159-201. Kaan, Edith, and Tamara Y Swaab. “Repair, Revision, and Complexity in Syntactic Analysis: An Electrophysiological Differentiation.” Journal of Cognitive Neuroscience 15, no. 1 (2003): 98-110.

142

Bibliography

Katsuki, Fumi, and Christos Constantinidis. “Early Involvement of Prefrontal Cortex in Visual Bottom-up Attention.” Nature Neuroscience 15, no. 8 (2012): 1160-66. Kayne, Richard S. Parameters and Universals. Oxford: Oxford University Press, 2000. Kim, Albert, and Lee Osterhout. “The Independence of Combinatory Semantic Processing: Evidence from Event-Related Potentials.” Journal of Memory and Language 52, no. 2 (2005): 205-25. Kimball, John, and Judith Aissen. “I Think, You Think, He Think.” Linguistic Inquiry 2, no. 2 (1971): 241-46. Kluender, Robert, and Marta Kutas. “Bridging the Gap: Evidence from ERPs on the Processing of Unbounded Dependencies.” Journal of Cognitive Neuroscience 5, no. 2 (1993): 196-214. Koelsch, Stefan, Thomas C Gunter, Matthias Wittfoth, and Daniela Sammler. “Interaction between Syntax Processing in Language and in Music: An ERP Study.” Journal of Cognitive Neuroscience 17, no. 10 (2005): 1565-77. Kreiner, Hamutal, Simon Garrod, and Patrick Sturt. “Number Agreement in Sentence Comprehension: The Relationship between Grammatical and Conceptual Factors.” Language and Cognitive Processes 28, no. 6 (2013): 829-74. Kuperberg, G. R. “Neural Mechanisms of Language Comprehension: Challenges to Syntax.” Brain Research 1146 (May 18 2007): 23-49. Kuperberg, G. R., P.J. Holcomb, T. Sitnikova, D. Greve, A.M. Dale, and D. Caplan. “Distinct Patterns of Neural Modulations During the Processing of Conceptual and Syntactic Anomalies.” Journal of Cognitive Neuroscience 15:2, no. 272-293 (2003). Kuperberg, G. R., T. Sitnikova, and B. M. Lakshmanan. “Neuroanatomical Distinctions within the Semantic System During Sentence Comprehension: Evidence from Functional Magnetic Resonance Imaging.” Neuroimage 40, no. 1 (Mar 1 2008): 367-88. Kuperberg, Gina R, Balaji M Lakshmanan, David N Caplan, and Phillip J Holcomb. “Making Sense of Discourse: An fMRI Study of Causal Inferencing across Sentences.” Neuroimage 33, no. 1 (2006): 343-61. Kutas, Marta, and Kara D Federmeier. “Electrophysiology Reveals Semantic Memory Use in Language Comprehension.” Trends in Cognitive Sciences 4, no. 12 (2000): 463-70. —. “N400.” Scholarpedia 4, no. 10 (2009): 7790. Kutas, Marta, and Steven A Hillyard. “The Lateral Distribution of EventRelated Potentials During Sentence Processing.” Neuropsychologia 20, no. 5 (1982): 579-90.

Functional and Temporal Dissociation of Agreement Mechanisms

143

—. “Reading Senseless Sentences: Brain Potentials Reflect Semantic Incongruity.” Science 207, no. 4427 (1980): 203-05. Lau, E. F., A. Gramfort, M. S. Hamalainen, and G. R. Kuperberg. “Automatic Semantic Facilitation in Anterior Temporal Cortex Revealed through Multimodal Neuroimaging.” The Journal of Neuroscience 33, no. 43 (Oct 23 2013): 17174-81. Lau, E. F., C. Phillips, and D. Poeppel. “A Cortical Network for Semantics: (De)Constructing the N400.” Nature Reviews Neuroscience 9, no. 12 (Dec 2008): 920-33. Lelekov, Taïssia, Peter F Dominey, and Luis Garcia-Larrea. “Dissociable ERP Profiles for Processing Rules vs. Instances in a Cognitive Sequencing Task.” NeuroReport 11, no. 5 (2000): 1129-32. Lewis, Shevaun, and Colin Phillips. “Aligning Grammatical Theories and Language Processing Models.” Journal of Psycholinguistic Research 44, no. 1 (2015): 27-46. Lichtheim, Ludwig. “On Aphasia.” Brain 7 (1885): 433-84. Linn, Mary S, and Sara Thomas Rosen. “The Functional Projections of Subject Splits.” Paper presented at The role of agreement in natural language (Texas Linguistic Forum 53), 2003. Luck, Steven J. An Introduction to the Event-Related Potential Technique. MIT press, 2014. MacDonald, Maryellen C, Neal J Pearlmutter, and Mark S Seidenberg. “The Lexical Nature of Syntactic Ambiguity Resolution.” Psychological review 101, no. 4 (1994): 676. MacWhinney, Brian, and Elizabeth Bates. “The Crosslinguistic Study of Sentence Processing.”, Cambridge: Cambridge University Press, 1989. Maess, Burkhard, Stefan Koelsch, Thomas C Gunter, and Angela D Friederici. “Musical Syntax Is Processed in Broca’s Area: An MEG Study.” Nature Neuroscience 4, no. 5 (2001): 540-45. Mancini, S., N. Molinaro, L. Rizzi, and M. Carreiras. “A Person Is Not a Number: Discourse Involvement in Subject-Verb Agreement Computation.” Brain Research 1410 (2011a): 64-76. Mancini, S., F. Postiglione, A. Laudanna, and L. Rizzi. “On the PersonNumber Distinction: Subject-Verb Agreement Processing in Italian.” Lingua 146 (2014): 28-38. Mancini, Simona, Nicola Molinaro, and Manuel Carreiras. “Anchoring Agreement in Comprehension.” Language and Linguistics Compass 7, no. 1 (2013): 1-21. Mancini, Simona, Nicola Molinaro, Doug J. Davidson, Alberto Avilés, and Manuel Carreiras. “Person and the Syntax–Discourse Interface: An

144

Bibliography

Eye-Tracking Study of Agreement.” Journal of Memory and Language 76 (2014): 141-57. Mancini, Simona, Nicola Molinaro, Luigi Rizzi, and Manuel Carreiras. “When Persons Disagree: An ERP Study of Unagreement in Spanish.” Psychophysiology 48, no. 10 (2011b): 1361-71. Mancini, Simona, Ileana Quiñones, Nicola Molinaro, Juan A HernandezCabrera, and Manuel Carreiras. “Disentangling Meaning in the Brain: Left-Temporal Involvement in Agreement Processing.” Cortex 86 (2017): 140-55. Marslen-Wilson, William. “Linguistic Structure and Speech Shadowing at Very Short Latencies.” Nature (1973). Mashal, Nira, Miriam Faust, Talma Hendler, and Mark Jung-Beeman. “An fMRI Investigation of the Neural Correlates Underlying the Processing of Novel Metaphoric Expressions.” Brain and Language 100, no. 2 (2007): 115-26. Mazoyer, Bernard M, Nathalie Tzourio, Victor Frak, Andre Syrota, N Murayama, Olivier Levrier, Georges Salamon, et al. “The Cortical Representation of Speech.” Journal of Cognitive Neuroscience 5, no. 4 (1993): 467-79. Mecklinger, Axel, Herbert Schriefers, Karsten Steinhauer, and Angela D Friederici. “Processing Relative Clauses Varying on Syntactic and Semantic Dimensions: An Analysis with Event-Related Potentials.” Memory & Cognition 23, no. 4 (1995): 477-94. Molinaro, N., H. A. Barber, and M. Carreiras. “Grammatical Agreement Processing in Reading: ERP Findings and Future Directions.” Cortex 47, no. 8 (Sep 2011): 908-30. Molinaro, N., H. A. Barber, A. Perez, L. Parkkonen, and M. Carreiras. “Left Fronto-Temporal Dynamics During Agreement Processing: Evidence for Feature-Specific Computations.” Neuroimage 78 (Sep 2013): 339-52. Molinaro, N., P.M. Paz Alonso, J.A. Duñabeitia, and M. Carreiras. “Combinatorial Semantics Strengthens Angular-Anterior Temporal Coupling.” Cortex 65 (2015): 113-27. Molinaro, N., F. Vespignani, and R. Job. “A Deeper Reanalysis of a Superficial Feature: An ERP Study on Agreement Violations.” Brain Research 1228 (Sep 4 2008): 161-76. Molinaro, Nicola, Horacio A Barber, Sendy Caffarra, and Manuel Carreiras. “On the Left Anterior Negativity (LAN): The Case of Morphosyntactic Agreement: A Reply to Tanner et al.” Cortex 66 (2015): 156-59.

Functional and Temporal Dissociation of Agreement Mechanisms

145

Murray, Linda J, and Charan Ranganath. “The Dorsolateral Prefrontal Cortex Contributes to Successful Relational Memory Encoding.” Journal of Neuroscience 27, no. 20 (2007): 5515-22. Ni, Weijia, et al. “An Event-Related Neuroimaging Study Distinguishing Form and Content in Sentence Processing.” Journal of Cognitive Neuroscience 12.1 (2000): 120-133 Nevins, A., B. Dillon, S. Malhotra, and C. Phillips. “The Role of FeatureNumber and Feature-Type in Processing Hindi Verb Agreement Violations.” Brain Research 1164 (Aug 20 2007): 81-94. Newman, Sharlene D., Marcel Adam Just, Timothy A. Keller, Jennifer Roth, and Patricia A. Carpenter. “Differential Effects of Syntactic and Semantic Processing on the Subregions of Broca’s Area.” Cognitive Brain Research 16, no. 2 (2003): 297-307. Nieuwland, Mante S, Andrea E Martin, and Manuel Carreiras. “Brain Regions That Process Case: Evidence from Basque.” Human Brain Mapping 33, no. 11 (2012): 2509-20. Nieuwland, Mante S, Karl Magnus Petersson, and Jos JA Van Berkum. “On Sense and Reference: Examining the Functional Neuroanatomy of Referential Processing.” NeuroImage 37, no. 3 (2007): 993-1004. Novick, Jared M, John C Trueswell, and Sharon L Thompson-Schill. “Cognitive Control and Parsing: Reexamining the Role of Broca’s Area in Sentence Comprehension.” Cognitive, Affective, & Behavioral Neuroscience 5, no. 3 (2005): 263-81. Núñez-Peña, M Isabel, and M Luisa Honrubia-Serrano. “P600 Related to Rule Violation in an Arithmetic Task.” Cognitive Brain Research 18, no. 2 (2004): 130-41. Olichney, John M, Jason R Taylor, Shiaohui Chan, Jin-Chen Yang, Andrew Stringfellow, Dieter G Hillert, Amanda L Simmons, et al. “fMRI Responses to Words Repeated in a Congruous Semantic Context Are Abnormal in Mild Alzheimer's Disease.” Neuropsychologia 48, no. 9 (2010): 2476-87. Osterhout, Lee, and Phillip J Holcomb. “Event-Related Brain Potentials Elicited by Syntactic Anomaly.” Journal of Memory and Language 31, no. 6 (1992): 785-806. Osterhout, Lee, and Linda A Mobley. “Event-Related Brain Potentials Elicited by Failure to Agree.” Journal of Memory and Language 34, no. 6 (1995): 739-73. Pallier, C., A.D. Devauchelle, and S. Dehaene. “Cortical Representation of the Constituent Structure of Sentences.” Proceedings of the National Academy of Sciences of the United States of America 108, no. 6 (2011): 2522-27.

146

Bibliography

Papoutsi, M., E. A. Stamatakis, J. Griffiths, W. D. Marslen-Wilson, and L. K. Tyler. “Is Left Fronto-Temporal Connectivity Essential for Syntax? Effective Connectivity, Tractography and Performance in LeftHemisphere Damaged Patients.” Neuroimage 58, no. 2 (Sep 15 2011): 656-64. Patel, Aniruddh D, Edward Gibson, Jennifer Ratner, Mireille Besson, and Phillip J Holcomb. “Processing Syntactic Relations in Language and Music: An Event-Related Potential Study.” Journal of Cognitive Neuroscience 10, no. 6 (1998): 717-33. Patterson, K., P. J. Nestor, and T. TRogers. “Where Do You Know What You Know? The Representation of Semantic Knowledge in the Human Brain.” Nature Reviews Neuroscience 8, no. 12 (Dec 2007): 976-87. Pearlmutter, Neal J, Susan M Garnsey, and Kathryn Bock. “Agreement Processes in Sentence Comprehension.” Journal of Memory and Language 41, no. 3 (1999): 427-56. Phillips, Colin, and Shevaun Lewis. “Derivational Order in Syntax: Evidence and Architectural Consequences.” Studies in Linguistics 6 (2013): 11-47. Posner, Michael I, and Marcus E Raichle. Images of Mind. Scientific American Library/Scientific American Books, 1994. Price, C. J. “A Review and Synthesis of the First 20 Years of PET and fMRI Studies of Heard Speech, Spoken Language and Reading.” Neuroimage 62, no. 2 (Aug 15 2012): 816-47. Qiu, Yinchen, and Xiaolin Zhou. “Processing Temporal Agreement in a Tenseless Language: An ERP Study of Mandarin Chinese.” Brain Research 1446 (2012): 91-108. Quinones, I., N. Molinaro, S. Mancini, J. A. Hernandez-Cabrera, and M. Carreiras. “Where Agreement Merges with Disagreement: fMRI Evidence of Subject-Verb Integration.” Neuroimage 88C (2014): 188201. Quiñones, Ileana. “From minimal dependencies to sentence contexts: neural correlates of agreement processing”. PhD thesis, University of the Basque Country, 2015 Rayner, Keith. “Eye Movements in Reading and Information Processing: 20 Years of Research.” Psychological Bulletin 124, no. 3 (1998): 372. Rayner, Keith, and Monica Castelhano. “Eye Movements.” Scholarpedia 2, no. 10 (2007): 3649. Rayner, Keith, and Simon P Liversedge. “Linguistic and Cognitive Influences on Eye Movements During Reading”, The Oxford Handbook of Eye Movements, Ed. By Simon Liversedge, Iain

Functional and Temporal Dissociation of Agreement Mechanisms

147

Gilchrist, Stefan Everling, Oxford: Oxford University Press, 2011, 772-787. Raichle, Marcus E. “A Brief History of Human Brain Mapping”. Trends in Cognitive Science (2008): 118-126. Reichle, E. D., T. Warren, and K. McConnell. “Using E-Z Reader to Model the Effects of Higher Level Language Processing on Eye Movements During Reading.” Psychonomic Bulletin and Review 16, no. 1 (Feb 2009): 1-21. Rice, Keren. Morpheme order and semantic scope. Word formation in the Athapaskan verb. Cambridge University Press, 2000 . Ritter, Elizabeth. “On the Syntactic Category of Pronouns and Agreement.” Natural Language & Linguistic Theory 13, no. 3 (1995): 405-43. —. “Where’s Gender?”. Linguistic Inquiry 24, no. 4 (1993): 795-803. Rivero, María Luisa. “Oblique Subjects and Person Restrictions in Spanish: A Morphological Approach.” Agreement Restrictions (2008): 215-50. Rizzi, Luigi. “The Fine Structure of the Left Periphery. Elements of Grammar: Handbook of Generative Syntax, Ed. By Liliane Haegeman, 281-337.” Dordrecht: Kluwer, 1997. —. The Structure of CP and IP: The Cartography of Syntactic Structures. Vol. 2. Oxford University Press, 2004. Roehm, Dietmar, Ina Bornkessel, Hubert Haider, and Matthias Schlesewsky. “When Case Meets Agreement: Event-Related Potential Effects for Morphology-Based Conflict Resolution in Human Language Comprehension.” Neuroreport 16, no. 8 (2005): 875-78. Rogalsky, C., and G. Hickok. “Selective Attention to Semantic and Syntactic Features Modulates Sentence Processing Networks in Anterior Temporal Cortex.” Cerebral Cortex 19, no. 4 (Apr 2009): 786-96. Rogalsky, Corianne, and Gregory Hickok. “The Role of Broca’s Area in Sentence Comprehension.” Journal of Cognitive Neuroscience 23, no. 7 (2011): 1664-80. Ross, John. R. “On declarative sentences”. Readings in English Transformational Grammar, Ed. By Roderick Jacobs and Peter Rosenbaum, Boston: Ginn & Co, 1970, 222-272. Rossi, S., M. F. Gugler, A. Hahne, and A. D. Friederici. “When Word Category Information Encounters Morphosyntax: An ERP Study.” Neurosci Letters 384, no. 3 (Aug 26 2005): 228-33. Saur, D., B. W. Kreher, S. Schnell, D. Kummerer, P. Kellmeyer, M. S. Vry, R. Umarova, et al. “Ventral and Dorsal Pathways for Language.”

148

Bibliography

Proceedings of the National Academy of Sciences of the United States of America 105, no. 46 (Nov 18 2008): 18035-40. Seghier, M. L. “The Angular Gyrus: Multiple Functions and Multiple Subdivisions.” Neuroscientist 19, no. 1 (Feb 2013): 43-61. Shibata, Midori, Akira Toyomura, Hiroki Motoyama, Hiroaki Itoh, Yasuhiro Kawabata, and Jun-ichi Abe. “Does Simile Comprehension Differ from Metaphor Comprehension? A Functional MRI Study.” Brain and Language 121, no. 3 (2012): 254-60. Shlonsky, Ur. “The Hierarchical Representation of Subject-Verb Agreement.” In ms., University of Haifa, 1989. —. “Subject Positions and Copular Constructions.” Interface Strategies, Ed. By Hans Bennis, Martin Everaert and Eric Reuland, Amsterdam: Royal Netherlands Academy of Arts and Sciences 2000, 325-337. —. “Hebrew as a Partial NullǦSubject Language.” Studia Linguistica 63, no. 1 (2009): 133-57. —. “The Cartographic Enterprise in Syntax.” Language and Linguistics Compass (2010): 417-429. Sigurdsson, H.A. “Remarks on Featues.” In Explorations of Phase Theory: Features and Arguments (Interface Explorations), Ed. by Kleanthes Grohman, Berlin: Mouton de Gruyter, 2008, 21– 52. Sigurdsson, Halldór Ármann. “Language Quarks.” Iberia: An International Journal of Theoretical Linguistics 1, no. 1 (2010). —. “The Syntax of Person, Tense and Speech Features.” Italian Journal of Linguistics 16 (2004): 219-51. Sigurðsson, Halldór Ármann, and Anders Holmberg. “Icelandic Dative Intervention: Person and Number Are Separate Probes.” Agreement Restrictions. Ed. By Roberta D’Alessandro, Susann Fischer and Gunnar Hrafnbjargarson, Berlin, Mouton de Gruyter, 2008, 251-80. Silverstein, Michael. “The Hierarchy of Features and Ergativity.” Grammatical Categories in Australian Languages. Ed. By Robert Dixon, New Jersey: Humanities Press, 1976. Steele, Susan. “Word Order Variation: A Typological Study.” Universals of Human Language. Ed. By Joseph Greenberg, Charles Ferguson and Edith Moravcsik. Redwood, CA: Standford University Press, 4 (1978): 585-623. Steinhauer, Karsten, and John E Drury. “On the Early Left-Anterior Negativity (eLAN) in Syntax Studies.” Brain and Language 120, no. 2 (2012): 135-62. Steinhauer, Karsten, and Michael T Ullman. “Consecutive ERP Effects of Morpho-Phonology and Morpho-Syntax.” Brain and Language, 83 (2002): 62-65.

Functional and Temporal Dissociation of Agreement Mechanisms

149

Swaab, Tamara Y, Kerry Ledoux, C Christine Camblin, and Megan A Boudewyn. “Language-Related ERP Components.” Oxford Handbook of Event-Related Potential Components (2012): 397-440. Tanner, Darren. “On the Left Anterior Negativity (LAN) in Electrophysiological Studies of Morphosyntactic Agreement: A Commentary on ‘Grammatical Agreement Processing in Reading: ERP Findings and Future Directions’ by Molinaro et al., 2014.” Cortex 66 (2015): 149-55. Tanner, Darren, and Janet G Van Hell. “ERPs Reveal Individual Differences in Morphosyntactic Processing.” Neuropsychologia 56 (2014): 289-301. Taraldsen, Knut T. “On Agreement and Nominative Objects in Icelandic”. Studies in Comparative Germanic Syntax. Kluwer, 1995. Taylor, Stephan F, Emily R Stern, and William J Gehring. “Neural Systems for Error Monitoring: Recent Findings and Theoretical Perspectives.” The Neuroscientist 13, no. 2 (2007): 160-72. Tenny, Carol L, and Peggy Speas. “The Interaction of Clausal Syntax, Discourse Roles, and Information Structure in Questions.” Paper presented at the Workshop on Syntax, Semantics and Pragmatics of Questions, ESSLI, Nancy, France, July 18, 2005. Tettamanti, Marco. “Language acquisition and processing: hierarchically organized cognitive processes.” Ph.D Thesis, University of Zurich, 2003. Tettamanti, Marco, Irene Rotondi, Daniela Perani, Giuseppe Scotti, Ferruccio Fazio, Stefano F. Cappa, and Andrea Moro. “Syntax without Language: Neurobiological Evidence for Cross-Domain Syntactic Computations.” Cortex 45, no. 7 (2009): 825-38. Torrego, Esther, and Itziar Laka. “The Syntax of ĭ-Features: Agreement with Plural DPs in Basque and Spanish.” Ibon Sarasola, Gorazarre. Homenatge, Homenaje, Bilbao: UPV/EHU, 2015, 633-646. Van de Meerendonk, Nan, Peter Indefrey, Dorothee J Chwilla, and Herman HJ Kolk. “Monitoring in Language Perception: Electrophysiological and Hemodynamic Responses to Spelling Violations.” Neuroimage 54, no. 3 (2011): 2350-63. Van de Meerendonk, Nan, Herman HJ Kolk, Dorothee J Chwilla, and Constance Th WM Vissers. “Monitoring in Language Perception.” Language and Linguistics Compass 3, no. 5 (2009): 1211-24. Vandenberghe, Rik, Anna C Nobre, and CJ Price. “The Response of Left Temporal Cortex to Sentences.” Journal of Cognitive Neuroscience 14, no. 4 (2002): 550-60.

150

Bibliography

Vigliocco, Gabriella, Brian Butterworth, and Merrill F Garrett. “SubjectVerb Agreement in Spanish and English: Differences in the Role of Conceptual Constraints.” Cognition 61, no. 3 (1996): 261-98. Vigneau, M., V. Beaucousin, P. Y. Herve, H. Duffau, F. Crivello, O. Houde, B. Mazoyer, and N. Tzourio-Mazoyer. “Meta-Analyzing Left Hemisphere Language Areas: Phonology, Semantics, and Sentence Processing.” Neuroimage 30, no. 4 (May 1 2006): 1414-32. Vigneau, M., V. Beaucousin, P. Y. Herve, G. Jobard, L. Petit, F. Crivello, E. Mellet, et al. “What Is Right-Hemisphere Contribution to Phonological, Lexico-Semantic, and Sentence Processing? Insights from a Meta-Analysis.” Neuroimage 54, no. 1 (Jan 1 2011): 577-93. Villa-García, Julio. “To Agree or Not to Agree.” Selected papers from the 39th Linguistic Symposium on Romance Languages (LSRL), Tucson, Arizona, March 2009. Amsterdam/Philadelphia: John Benjamins, 2010, 249-266. Vincenzi, Marica De, Remo Job, Rosalia Di Matteo, Alessandro Angrilli, Barbara Penolazzi, Laura Ciccarelli, and Francesco Vespignani. “Differences in the Perception and Time Course of Syntactic and Semantic Violations.” Brain and Language 85, no. 2 (2003): 280-96. Wagers, Matthew W, Ellen F Lau, and Colin Phillips. “Agreement Attraction in Comprehension: Representations and Processes.” Journal of Memory and Language 61, no. 2 (2009): 206-37. Wechsler, Stephen. “Number as Person.” Empirical Issues in Syntax and Semantics 5 (2004): 255-74. Wernicke, Carl. Der Aphasische Symptomencomplex: Eine Psychologische Studie Auf Anatomischer Basis. Cohn., 1874. Wilson, S. M., S. Galantucci, M. C. Tartaglia, and M. L. Gorno-Tempini. “The Neural Basis of Syntactic Deficits in Primary Progressive Aphasia.” Brain and Language 122, no. 3 (Sep 2012): 190-8.

INDEX

ACC, anterior cingulate cortex, 33, 34, 36, 40, 42, 62, 123 Addressee, 2, 11, 13, 53, 67, 69, 70, 83, 85, 87, 94, 95, 124 AG, angular gyrus, 32, 33, 42, 101, 102, 103, 115, 126 Anchor, interpretive, 11, 12, 13, 14, 36, 43, 44, 46, 47, 56, 57, 58, 61, 65, 66, 90, 91, 95, 96, 100, 103, 104, 106, 108, 110, 114, 117, 118, 119, 120, 121, 122, 124, 125, 126, 130 BOLD signal, 30, 34 Cartography, 1, 5, 15, 46, 88, 117 distinct-cluster representation, 5, 15, 118 Controller, 2, 14, 44, 110, 112 Dorsolateral prefrontal cortex, 33, 102, 115, 123 eLAN, 27, 28, 39, 40, 131 Goal, 4, 43 LAN, 27, 28, 29, 39, 40, 41, 42, 54, 55, 56, 57, 58, 66, 81, 82, 83, 86, 98, 119, 122, 124, 125, 128, 131 LIFG, left inferior frontal gyrus, 31, 32, 39, 40, 41, 61, 62, 64, 65, 66, 122, 123, 127, 128, 129 pars opercularis, 32, 39, 40, 41, 61, 64, 66, 122 pars orbitalis, 61, 62, 64, 101, 123 pars triangularis, 32, 39, 64, 101, 123 Matching, 4 Minimalist Program, 1, 2, 3, 5, 8, 13, 14, 15, 43, 46, 65, 90, 116, 117, 129 single-cluster representation, 5, 46

N400, 22, 23, 24, 28, 39, 40, 42, 56, 57, 58, 63, 66, 83, 86, 96, 98, 99, 119, 123, 124, 125, 128, 133 Number plural, 2, 9, 13, 20, 53, 56, 68, 69, 70, 71, 81, 91, 92, 94, 95, 96, 98, 103, 110, 112, 113, 114, 122, 132 singular, 1, 9, 13, 20, 29, 53, 56, 68, 69, 72, 91, 94, 95, 96, 132 P600, 24, 25, 29, 39, 40, 41, 42, 54, 55, 56, 58, 96, 98, 100, 103, 124 broad/frontal, 25, 26, 28, 58, 59, 83, 85, 86, 123 centro-parietal, 24, 25, 28, 59, 81, 82, 85, 99 diagnosis, 25, 59, 82, 83, 99 error monitoring, 26, 124 repair, 25, 29, 56, 60, 82, 83, 99 Person 1st, 9, 11, 12, 13, 56, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 85, 86, 87, 88, 89, 94, 95, 96, 98, 103, 106, 110, 112, 113, 114, 119, 120, 123, 132 2nd, 8, 9, 11, 12, 56, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 85, 86, 87, 88, 89, 94, 95, 119, 120, 123, 132 3rd, 2, 12, 13, 53, 56, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 85, 86, 87, 88, 89, 94, 95, 96, 103, 119, 120, 124, 132 Probe, 4, 94 Speaker, 2, 8, 9, 11, 12, 53, 67, 68, 69, 70, 79, 82, 83, 85, 89, 94, 95,

152 96, 106, 108, 116, 121, 124, 128, 132 Specifier, 2, 3, 13, 73, 75, 118 Target, 2, 14, 92 Temporal cortex, 31, 33, 39, 40, 41, 61, 62, 63, 64, 66, 100, 101, 102,

Index 103, 115, 118, 119, 122, 123, 126, 127, 128 Unagreement, 29, 36, 44, 82, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 103, 104, 106, 108, 110, 112, 114, 117, 133