Anaphors in Text
Studies in Language Companion Series (SLCS) The SLCS series has been established as a companion se...
220 downloads
1203 Views
2MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
Anaphors in Text
Studies in Language Companion Series (SLCS) The SLCS series has been established as a companion series to Studies in Language, International Journal, sponsored by the Foundation “Foundations of Language”.
Series Editors Werner Abraham
Michael Noonan
University of Vienna
University of Wisconsin, Milwaukee
Editorial Board Joan Bybee
Christian Lehmann
University of New Mexico
University of Erfurt
Ulrike Claudi
Robert Longacre
University of Cologne
University of Texas, Arlington
Bernard Comrie
Brian MacWhinney
Max Planck Institute For Evolutionary Anthropology, Leipzig
Carnegie-Mellon University
William Croft
University of California, Santa Barbara
University of New Mexico
Edith Moravcsik
Östen Dahl
University of Wisconsin, Milwaukee
University of Stockholm
Masayoshi Shibatani
Gerrit Dimmendaal
Rice University and Kobe University
University of Leiden
Russell Tomlin
Ekkehard König
University of Oregon
Marianne Mithun
Free University of Berlin
Volume 86 Anaphors in Text Cognitive, formal and applied approaches to anaphoric reference Edited by Monika Schwarz-Friesel, Manfred Consten and Mareile Knees
Anaphors in Text Cognitive, formal and applied approaches to anaphoric reference
Edited by
Monika Schwarz-Friesel Manfred Consten Mareile Knees University of Jena
John Benjamins Publishing Company Amsterdam/Philadelphia
8
TM
The paper used in this publication meets the minimum requirements of American National Standard for Information Sciences – Permanence of Paper for Printed Library Materials, ansi z39.48-1984.
Library of Congress Cataloging-in-Publication Data Anaphors in text : cognitive, formal and applied approaches to anaphoric reference / edited by Monika Schwarz-Friesel, Manfred Consten and Mareile Knees. p. cm. (Studies in Language Companion Series, issn 0165–7763 ; v. 86) Includes bibliographical references and index. 1. Anaphora (Linguistics) 2. Discourse analysis. I. Schwarz-Friesel, Monika. II. Consten, Manfred. III. Knees, Mareile. P299.A5A65 2007 415--dc22 isbn 978 90 272 3096 6 (Hb; alk. paper)
2007003847
© 2007 – John Benjamins B.V. No part of this book may be reproduced in any form, by print, photoprint, microfilm, or any other means, without written permission from the publisher. John Benjamins Publishing Co. · P.O. Box 36224 · 1020 me Amsterdam · The Netherlands John Benjamins North America · P.O. Box 27519 · Philadelphia pa 19118-0519 · usa
Table of contents Anaphors in text – Introduction Monika Schwarz-Friesel, Manfred Consten and Mareile Knees
vii
section 1 Anaphors in Cognitive, Text- and Discourse Linguistics Indirect anaphora in text: A cognitive account Monika Schwarz-Friesel Indirect pronominal anaphora in English and French: Marginal rarity, or unmarked norm? Some psycholinguistic evidence Francis Cornish Lexical anaphors in Danish and French Lita Lundquist Referential collaboration with computers: Do we treat computer addressees like humans? Alfons Maes, Pascal Marcelis and Frank Verheyen Reflexivity and temporality in discourse deixis Friedrich Lenz The function of complex anaphors in texts: Evidence from corpus studies and ontological considerations Manfred Consten, Mareile Knees and Monika Schwarz-Friesel Metaphorical anaphors: A phenomenon of the semantics-pragmatics interface Helge Skirl
3
21
37
49
69
81
103
Anaphors in Text
section 2 The Syntax and Semantics of Anaphors Accessibility and definite noun phrases Klaus von Heusinger
123
The non-subject bias of German demonstrative pronouns Peter Bosch, Graham Katz and Carla Umbach
145
Anaphoric properties of German right dislocation Maria Averintseva-Klisch
165
Antecedents of diverse types. An investigation of the syntactic and semantic relationships in a wh-relative construction Anke Holler
183
Corpus-based and machine learning approaches to anaphora resolution: A critical assessment Michael Strube
207
section 3 Neurolinguistic Studies Neuroimaging studies of coherence processes Evelyn C. Ferstl and Florian Th. Siebörger
225
Reference assignment in the absence of sufficient semantic content Petra Burkhardt
241
Resolving complex anaphors: Evidence from online comprehension Konstanze Marx, Ina Bornkessel-Schlesewsky and Matthias Schlesewsky
259
Index
279
Anaphors in text – Introduction Monika Schwarz-Friesel, Manfred Consten and Mareile Knees
The roots of this book lie in the workshop “Anaphors in Text” which was held in the framework of the annual meeting of Deutsche Gesellschaft für Sprachwissenschaft at the University of Cologne in February 2005. It is a collection of original research articles on the representation and interpretation of definite descriptions used as anaphors in text and discourse. The investigation of anaphoric reference and different anaphoric relations is of central interest to the understanding of the structure of textual meaning and of text processing. Anaphoric reference is a typical means of textual coherence and continuity and, at the same time, of semantic variability. It is a phenomenon of underspecification since most anaphors lack a lexical meaning that suffice to determine the referent. In order to comprehend anaphors it is not only necessary to activate grammatical but also conceptual knowledge. Thus, anaphoric reference is object of different areas of linguistic research. This will be accounted for by the variety of frameworks presented in this book. The first section comprises cognitive, text- and discourse linguistic approaches, the second syntactic, semantic and a computational linguistic account and the third neurolinguistic research on the reception of anaphoric reference. So the book covers methodically broadly invested approaches concerned about actual questions in the research on anaphora, e.g. different linguistic realisations of indirect and direct anaphors; the anaphora resolution process from cognitive-, neuro- and computational-linguistic perspectives, the role of anaphors in information-packing and their function in text. For a long time, the view on anaphoric expressions has been that anaphors are used to continue a pre-established reference in text, that they point back to a specific antecedent, usually an NP in the preceding text, and thereby sustain the current fo1. The editors would like to thank Konstanze Marx for her participation in the workshop organisation, all contributors of the workshop and the Deutsche Forschungsgemeinschaft (DFG) for their financial support. Furthermore, the DFG funded the editor’s research projekt on complex anaphors (KomplexTex, SCHW 509/6).
Anaphors in Text
cus in text comprehension. Thus, their main function is topic continuity. The prototypical anaphoric case realised by personal pronouns can be seen in example (1):
(1) But I tell you the reputation of our gang leaderi is excellent. […] No joke! Nobody is ashamed to serve himi. Hei does not murder for robbery like us. Hei does not seem to ask for money although hei could have it. And hei even gives hisi legal part to the poor. (Schiller, Die Räuber, II, 3, our translation)
Based on examples like this, anaphors have been described as mere “retrieval cues” or “echoes” of their antecedents. According to this view, anaphora resolution is seen as some kind of search procedure which aims at discovering the best matching antecedent in textual structure. Research of the last years, however, has shown that the analysis of anaphors calls for a more flexible description: In (2) you can see that there are anaphoric expressions which do not fall into the category “semantic echo of their antecedent”. Those progressive and specifying anaphors introduce new information about the discourse referent (see, for example, Skirl, this vol., on metaphorical anaphors).
(2) A burning tyre has been treacherous for a drunken car driveri. The 37-yearsold womani noticed smoke and stopped to check the car.
Definite lexical NPs and pronouns also commonly occur as “indirect anaphors”:
(3) The cari was found in the ditch. The tyresk were punched.
(4) It has beautifully snowedi, and it’s cold enough that itk will settle. (oral comm., cit. from Consten 2003)
The passage in (3) is perfectly coherent although there is no explicit textual antecedent for the definite NP the tyres. In (4), the pronoun itk is easily understood as referring to snow while the referent is not introduced by an NP but implicitly by the verbal phrase (for indirect anaphors see Schwarz-Friesel, this vol., for constraints on the use of pronouns as indirect anaphors see Cornish, this vol.). Both corpus-based and text linguistic studies support the assumption that indirect anaphors are quite common and normal uses of definite reference. Very often, the resolution of both direct and indirect anaphors involves much more than a simple search-and-match-procedure. Their full interpretation requires a cognitive process involving the activation of knowledge structures (for neurolinguistic evidence, see Ferstl and Siebörger, this. vol). Moreover, another form of anaphoric reference is the so-called complex anaphora:
(5) Good linguists are bad football players. This / This fact / This image / This prejudice / This impertinence…
Introduction
Here, the anaphors are not anchored by an NP-antecedent but by a larger, propositionally structured text segment. There are four contributions dealing with complex anaphors in this volume: For textlinguistic and semantic aspects of complex anaphors see Schwarz-Friesel, Consten and Knees, for the differentiation of complex anaphora and discourse deixis see Lenz, for reference to complex entities in wh-relative clauses see Holler, and for the resolution and neurolinguistic evidence see Marx, Bornkessel-Schlesewsky, and Schlesewsky. Typical lexical means for anaphors are definite NPs where various forms (i.e. lexical NPs, personal pronouns, demonstratives) typically characterise different textual functions. However, it is difficult to match anaphoric forms precisely to certain functions. The contributions in this volume reflect the wide variety of the different forms and their functions. See, for example, the papers by Lundquist, by von Heusinger and by Maes, Marcelis and Verheyen for the role of definite lexical NPs in information packing and specifying referents; and Bosch, Katz and Umbach’s contribution on demonstratives. Burkhardt investigates differences in the resolution of syntactically bound and non-bound pronouns. The interpretation of most anaphoric descriptions is highly context-sensitive. In order to access the referential entity described by an anaphor the reader or hearer must seek information beyond the anaphoric NP itself. The knowledge required for the interpretation of definite descriptions can be of various kinds, ranging from textual to conceptual, from specific semantic to general world knowledge. To sum up, the establishment of referential continuity through anaphoric descriptions relies much more both on the context and on the constructive mental activity of the reader than on the grammatical devices in the text. We know by now that both surface information and conceptual information plays an important role in resolving anaphors (for the use of surface information in machine based anaphora resolution see Strube, this vol). Accessibility of anaphoric referents in text or discourse comprehension is a function of building up coherence within mental text world models. Furthermore, the concept of focus and topic and discourse structure play a central role in anaphora resolution (see Averintseva-Klisch, this. vol, for a special kind of anaphoric reference marking topicness). The articles have been organised into three thematic sections in order to gain topical coherence for the reader of this volume. The first section covers cognitive, text- and discourse linguistic approaches with a focus on cognitive semantics and text linguistics. The first two papers deal with indirect anaphors as a central topic of contemporary anaphora research. Monika Schwarz-Friesel gives an overview of indirect anaphors in her contribution Indirect anaphora in text: A cognitive account. Indirect anaphors are definite NPs which have no explicit antecedent in text but which are linked to some previously mentioned element (i.e. anchor) by a cognitive process. Taking a procedural per-
Anaphors in Text
spective, the author presents different types of indirect anaphora. She distinguishes between indirect anaphors which are based on the activation of semantic knowledge in the mental lexicon and those which require more general conceptual knowledge for their interpretation. Thus, she provides a more complex cognitive approach for the interpretation processes involved in the resolution of indirect anaphors (cf. Schwarz 2000). Moreover, she argues that there is no clear-cut distinction between direct and indirect anaphors. So she proposes a unified account for explaining the interpretation of the different types of referring expressions. While the first paper considers primarily definite lexical NPs as indirect anaphors, Francis Cornish investigates in the second article Indirect pronominal anaphora in English and French the use of unaccented third person pronouns as indirect anaphors. Some linguists (e.g. Erkü / Gundel, 1987) argue that this is impossible or highly marked, while others (e.g. Yule, 1982) maintain that it is not only acceptable but commonly used in normal conversational discourse. However, in actual fact both sides in the debate may be correct, if we make a principled distinction between two main types of implicit referent: central or nuclear implicit referents, and peripheral ones. The author hypothesizes that while peripheral implicit referents (which evoke the means or the instrument by which a given state of affairs is established) would not easily be referred to using unstressed (object) pronouns, central or ‘nuclear’ implicit referents would be. The results of two experiments show that pronominal reference to implicit referents caused slower reading times compared to explicit referents only for peripheral referents. The author briefly discusses these results with respect to Gundel, Hedberg and Zacharski’s (1993) Givenness Hierarchy. An issue discussed in the next two papers is the question how anaphors influence the so-called information-packing, i.e. how speakers use anaphors in order to structure information in text. With Lita Lundquist’s paper Lexical anaphors in Danish and French we have a contribution on the language-specific usage of direct anaphors. As the author points out Danish and French show different strategies for assuring anaphoric chains in texts. One sign of this is the predilection in Danish for using pronominal anaphors or lexical repetitions to co-refer to an already existing discourse referent, where French often prefers lexical anaphors consisting of definite or demonstrative descriptions, i.e. “infidel anaphors”, which are lexically varied with respect to the antecedent. The author discusses how the use of different types of anaphors participates in and contributes to more general differences in patterns of information packing, which again reveal differences in the communicative contract between sender and receiver in the two cultures. Finally, she discusses how infidel anaphors in the form of definite and demonstrative descriptions may be seen as the ultimate grammaticalisation of textual relations.
Introduction
The paper Referential collaboration with computers by Alfons Maes, Pascal Marcelis and Frank Verheyen concentrates on cognitively oriented research of discourse reference to predict and explain forms and patterns of referential expressions in discourse. It originates from the question whether language user are sensitive to a human vs. computer nature of the addressee and therefore use different referential devices to ‘satisfy’ their vis-à-vis. The authors discuss the role of some systematic determinants of referential form, which relate to variables in the communicative situation, namely the criticality or importance of the instructive task (critical vs. non-critical), the available feedback facilities (with vs without feedback) together with the nature of the feedback (problematic vs non-problematic), and assumptions on the nature of the addressee (human vs computer addressee). In a controlled production experiment, instructive writers displayed a broad spectrum of redundancy or overspecification strategies. The results show that human addressees trigger more overspecification than computer addressees in a nonfeedback situation. Friedrich Lenz’ contribution Reflexivity and temporality in discourse deixis is based on a concept of discourse deixis which allows a clear distinction between discourse deixis and anaphora (cf. Lenz 1997). Discourse deixis needs to be regarded as a specific form of temporal deixis invariably containing a meta-communicative element. The meta-communicative element in discourse deixis is expressed above all by verbs. Meta-communicative verbs are therefore classified according to their aspect classes since temporal deictic localisation is based on the combination of lexical aspect classes and other temporal expressions, at least one of which has to be deictic. In the framework of discourse deixis deictic expressions (such as obligatory tense and optional adverbs) interact with non-deictic categories (like verb meaning and aspect) in order to identify a segment of the ongoing discourse. The article by Manfred Consten, Mareile Knees and Monika Schwarz-Friesel on The Function of complex anaphors in text is related to the previous one as some of the phenomena dicussed by Lenz are a special type of complex anaphora. In both accounts it is assumed that speakers create meta-discoursive abstract objects. These abstract objects can be categorised according to their ontological types. Following Consten, Knees and Schwarz-Friesel, complex anaphors are nominal expressions that refer to propositionally structured objects (such as propositions, states, facts and events). At the same time, this very act of reference establishes them as unified discourse entities and, sometimes, also evaluates them. The authors explain the resolution of complex anaphors – including ambiguous cases – by enriching a Text World Model with information about the ontological status of discourse objects as well as conceptually based resolving strategies. Thus, they integrate constraints based on an abstractness hierarchy into an anaphora resolution model that combines DRT- and cognitive aspects. The model also covers resolution strategies for indirect complex-
Anaphors in Text
anaphora, a phenomenon not yet described neither in formal nor in cognitive approaches. By means of a corpus study a distributional analysis of complex anaphors is proposed. Since this approach integrates nominal and complex anaphors as well as the distinction of direct and indirect cases, it leads to a general model of anaphora resolution which explains anaphoric reference as a widely applicable means of establishing coherence and progression in texts. The topic of the final paper of this section are Metaphorical anaphors: A phenomenon of the semantics-pragmatics interface. Here, Helge Skirl claims that the resolution of metaphoric anaphors depends on the one hand on the lexical meaning of the anaphor and on the other hand on pragmatic principles. Thus, metaphoric anaphors are considered as a phenomenon of the semantics-pragmatics interface. By means of examples from literary text the author shows that metaphoric anaphors are either direct or indirect. Finally, the author presents how metaphoric anaphors can be interpreted within the framework of text world models. The second section comprises syntactic and semantic approaches and a computational linguistic one. The first two papers of this section are concerned about linguistic realisations of anaphors, the first one with lexical NPs and the second one with demonstrative pronouns. In the third and fourth paper syntactic constructions, namely right dislocation and wh-relativ-clauses, are discussed with respect to their anaphoric character. The final paper of this section gives an overview of automatic anaphora resolution which depends on text-structural and syntactic features. Klaus von Heusinger investigates the interaction of Accessibility and definite noun phrases. He points out that there are many aspects of the coherence structure that determine the conditions under which an anaphoric expression is linked to its antecedent. Most of these parameters are summarized under the general term of “accessibility” (Ariel 1990). In contrast to the commonly held view that only the accessibility structure determines the definiteness status of referring expression, he argues that definite NPs also change the accessibility structure due to their descriptive content. He further assumes that accessible discourse items are ranked sets with respect to the predicate by which they were introduced or activated rather than one single set of ranked elements without further association to the predicate by which they were introduced. So definite NPs change the salience structure of some semantically related set, a process called “salience spreading”. The author also discusses the difference between anaphoric pronouns and anaphoric definite NPs with respect to the accessibility structure. While von Heusinger is mainly concerned with anaphors realised as lexical NPs, the next paper by Peter Bosch, Graham Katz and Carla Umbach deals with demonstrative pronouns, more precisely with The non-subject bias of German demonstrative pronouns. The difference between German personal pronouns (er, sie, es) and German demonstrative pronouns of the type der, die, das is often taken to
Introduction
be a matter of style or register. The authors investigate the differences between demonstrative and personal pronouns in German on the basis of (i) their use as evidenced by their distribution in a large text corpus and (ii) the way they are processed as evidenced by a psycholinguistic reading experiment. Their corpus study and a reading time experiment revealed striking differences in intrasentential distribution and with respect to the antecedent: non-subject demonstrative pronouns appear predominantly in the Vorfeld, non-subject personal pronouns appear in the Mittelfeld; demonstrative pronouns prefer non-subject and personal pronouns prefer subject antecedents. These findings support the hypothesis that demonstratives relate to non-topical referents whereas personal pronouns prefer topical referents. The following paper Anaphoric properties of German right dislocation by Maria Averintseva-Klisch examines the syntactic phenomenon of right dislocation. The author shows that German right dislocation subsumes two distinct constructions, which she labels ‘right dislocation proper’ and ‘afterthought’. These differ in a number of prosodic, syntactic and semantic characteristics, and have also different discourse functional properties. Right dislocation marks a discourse referent as especially salient on the current stage of the discourse. This requires the fulfilment of certain anaphoric demands on the following discourse. Afterthought is a local reference clarification strategy and has no impact on the global discourse structure. As the previous contribution, Anke Holler investigates in her paper Antecedents of diverse types. An investigation of the syntactic and semantic relationships in a wh-relative construction if the anaphoric constructions discussed are syntactically and semantically independent of their host sentences. Object of her study are the referential relations in wh-relative clause constructions. Against existing assumptions (cf. Helbig 1980) the author hypothesises that wh-relative clauses are not generally sentence-related but they can only be regarded as sentence-related with respect to their syntactic properties since they are non-integrated sentences. Semantically, wh-relative clauses can refer to entities of different semantic types (e.g. propositions as well as facts, types of events and events). By means of the wh-anaphor a discourse referent is introduced which is identified with respect to the predicate of the wh-relative clause. The final paper of this section by Michael Strube deals with anaphora from a computational linguistic perspective. Thus, textual features are given priority. In his contribution Corpus-based and machine learning approaches to anaphora resolution: A critical assessment the author gives an overview of state-of-the-art applications on machine based anaphora resolution. Anaphora resolution is an important component of natural language processing applications like information extraction or automatic summarisation. Therefore, anaphors have to be resolved in unrestricted input, which can be done using anaphora resolution algorithms
Anaphors in Text
based on machine learning. Those algorithms learn constraints and preferences automatically from data annotated with information on anaphoric relations. Machine learning based algorithms are mostly implemented as binary classification (e.g. Soon et al., 2001). While this makes the task accessible to standard machine learning techniques, it has the drawback that knowledge about the context is lost (Yang et al., 2004; Luo et al., 2004). In the third section of this book, three approaches from neurolinguistic research on the reception of anaphoric reference are presented. They reflect relevant issues theoretically discussed in section 1. Different kinds of indirect anaphora resolution (cf. Schwarz-Friesel in section 1) and the role of cohesive ties and inferences in coherence building are examined in the paper on Neuroimaging studies of coherence by Evelyn C. Ferstl and Florian Th. Siebörger. Language processing in context is shown to elicit activation in both hemispheres with the left hemisphere being dominant. In contrast to the expectation of the right-dominance during inference processes, the authors find left-medial brain regions to be involved in coherence building. Present claims on anaphora resolution are differentiated on the basis of ERP studies: Direct anaphora processing is described as a an anchoring process located at the text base level that is realised in left fronto-temporal brain regions. In contrast, some indirect anaphors require more extensive integration with general world knowledge, a process that is reflected in right hemisphere activation at the parieto-medial brain region. In her paper Reference assignment in the absence of sufficient semantic content Petra Burkhardt investigates the establishment of referential dependencies during the interpretation of pronouns, namely reflexive pronouns, personal pronouns and logophors (i.e. reflexive pronouns whose distribution patterns with personal pronouns). The quality of the dependency between these pronouns and their antecedents varies. So the author distinguishes between syntactic and discourse-based dependencies which are discussed within the framework of the syntax-discoursemodel (Burkhardt 2005). This model differentiates between two levels of re presentation: the syntactic one that encodes phrase-structural relations and the discourse-based one that holds information about events and discourse referents. Reaction time and ERP experiments as well as real-time studies on the processing of syntax-based coargument reflexive pronouns versus discourse-based logophoric pronouns give evidence for the model. The contribution by Konstanze Marx, Ina Bornkessel-Schlesewsky and Matthias Schlesewsky gets back to complex anaphora that was discussed from a textlinguistic point of view by Manfred Consten, Mareile Knees and Monika Schwarz-Friesel in section 1. It investigates the question of whether the complexation process postulated theoretically shows up in ERP data. Comparing complex anaphors with NP-based
Introduction
anaphors, the authors find an effect supporting the assumption that the establishment of a referent for a complex anaphor induces additional cognitive processing cost.
References Ariel, M. 1990. Accessing Noun-Phrase Antecedents. London: Routledge. Burkhardt, P. 2005. The Syntax-Discourse Interface: Representing and interpreting dependency. Amsterdam: John Benjamins. Consten, M. 2003. Towards a unified model of domain-bound reference. In Deictic Conceptualisation of Space, Time and Person, F. Lenz (ed.), 223–248. Amsterdam: John Benjamins. Erkü, F. and Gundel, J.K. 1987. The pragmatics of indirect anaphors. In The Pragmatic Perspective: Selected papers from the 1985 International Pragmatics Conference, Amsterdam, J. Verschueren and M. Bertuccelli-Papi (eds), 533–545. Amsterdam: John Benjamins. Gundel, J.K., Hedberg, N. and Zacharski, R. 1993. Cognitive status and the form of referring expressions in discourse. Language 69: 274–307. Helbig, G. 1980. Was sind weiterführende Nebensätze? Deutsch als Fremdsprache 17: 13–23. Lenz, F. 1997. Diskursdeixis im Englischen. Sprachtheoretische Überlegungen und lexiko-grammatische Analysen. Tübingen: Niemeyer. Luo, X., Ittycheriah, A, Jing, H., Kambhatla, N. and Roukos, S. 2004. A mention-synchronous coreference resolution algorithm based on the Bell Tree. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics, Barcelona, Spain, 21–26 July 2004, 136–143. Schwarz, M. 2000. Indirekte Anaphern in Texten. Studien zur domänen-gebundenen Referenz und Kohärenz im Deutschen [LA 413]. Tübingen: Niemeyer. Soon, W.M., Ng, H.T. and Lim, D.C. 2001. A machine learning approach to coreference resolution of noun phrases. Computational Linguistics 27: 521–544. Yang, X., Su, J., Zhou, G. and Tan, C.L. 2004. Improving pronoun resolution by incorporating coreferential information of candidates. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics, Barcelona, Spain, 21–26 July 2004, 128–135. Yule, G. 1982. Interpreting anaphora without identifying reference. Journal of Semantics 1: 315–323.
section 1
Anaphors in Cognitive, Text- and Discourse Linguistics
Indirect anaphora in text A cognitive account Monika Schwarz-Friesel The term indirect anaphor (IA) refers to a definite NP which has no explicit antecedent in text and is linked via a cognitive process to some element in prior text which functions as some kind of anchor for the interpretation of IA. According to one popular view, IA are treated as phenomena that can best be explained in terms of associability. In this paper I will argue for a more complex cognitive approach. Taking a procedural perspective, I want to demonstrate, firstly, that there are different types of IA. A distinction will be drawn between IA which are based on the activation of semantic knowledge in the mental lexicon and IA which require more general conceptual knowledge for their interpretation. I will base the discussion on naturally occurring data from German texts. All kinds of indirect anaphora have to be seen as “given-andnew-entities”. Combining both easy accessibility and the establishment of new referential files in text-world models, they serve as “progressive continuity markers”. The second aim of this paper is to illustrate that in many cases there is no clear-cut distinction between direct and indirect anaphora in text, showing that the most typical examples of direct and indirect anaphors may best be regarded as two extremes on a continuum of textual reference, thereby weakening the claim for a strict distinction between the two types and arguing for a unified account in explaining the interpretation of referring expressions.
1. Anaphora in text: basic assumptions The traditional view in textlinguistics has been that anaphora are used to continue a pre-established reference in text, that they point back to a specific antecedent, usually an NP already introduced, and thereby sustain the current focus. Their main function is topic continuity. In this respect, anaphors have been described as mere
Monika Schwarz-Friesel
“retrieval cues” or “semantic echoes of their antecedents”. According to this view, the prototypical anaphoric case can be seen in example (1):
(1) A man (Referent 1)1 entered the room. Then, he/this man (Referent 2)1 looked around.
The referential relationship between anaphor and antecedent is based on coreference, and the link between the expressions denoting coreference can be described in terms of syntactic congruence and semantic compatibility. According to this view, anaphor resolution is seen as some kind of search procedure which aims at discovering the best matching antecedent in textual structure. This mental checking in resolving anaphors in text comprehension always leads to the conclusion that R(eferent)2 is identical with r(eferent)1. My distinction between referent 1 (the text referent first encountered in reading) and referent 2 (and accordingly, referent 3 etc. in longer reference chains) takes into account that in on-line processing the reader must decide in each case of definite reference whether there is referential identity or not. Anaphors which establish topic continuity by merely reactivating the information knot of their antecedent involve pronouns, NPs with the same head noun, hyperonyms, synonyms and certain general expressions. Research of the last years, however, has shown that the analysis of anaphors calls for a more flexible description (Fraurud 1992, Dahl/Hellmann 1995, Cornish 1999, Schwarz 2000a,b, Schwarz-Friesel in press). Establishing the relationship of identity (R2 is identical with R1) which is necessary for successful anaphoric processing, very often demands more than just checking semantic compatibility. Reference chains in natural language texts reveal a lot more indirectness and underspecification. Thus, the establishment of coherence and topic continuity relies very often much more on the constructive nature of the interpretation process of the recipient than on the cohesive devices in text (see Givon 1992, Rickheit/Strohner 1993, Schwarz 2000b). 1. Considering the following examples, one can see that referent 2 in (a) is identical with referent 1, whereas in (b) this is not the case. (a) Klaus K. (R11) murdered his wife. The man (R21) is quite scrupulous. (b) Klaus K. (R11) murdered his wife. The man (R22) who lives across the street, saw everything. 2. Underspecification is an essential feature of the relation between verbal expressions and the conceptual representations they are intended to express. Therefore, underspecification and implicitness in text should by no means be regarded as some kind of textual deviance (as it is implied in many accounts). On the contrary, systematic underspecification is to be regarded as default. In accordance with some principle of cognitive economy, underspecification is achieved by selecting only the really relevant information and thereby avoiding redundancy (Schwarz 2000a: 83, Sanders/Spooren 2001: 4).
Indirect anaphora in text
In example (2), the vehicles involved in the accidents are not mentioned. In order to establish the full conceptual representation of the text structure, the reader has to draw some slot-filling inference and insert the referents into the mental text-world model. Thus, the reader will elaborate the semantic form with a referential default value such as VEHICLES.
(2) Wegen Glatteis mußte die Zoobrücke [...] vorübergehend gesperrt werden. Auf spiegelglatter Fahrbahn hatten sich [...] mehrere Unfälle ereignet. Dabei entstand Blechschaden. (Newspaper Kölner Stadtanzeiger, 4.3.96: 10) (The Zoo-Bridge had to be closed down because of slippery ice. Several accidents had happened on the glassy road. This caused a lot of damage.)
The result of such an activation is quite predictable because it is based on common and shared default assumptions which are part of our conceptual knowledge in long-term memory. No reader of (2) would make an elaboration filling the referential slot with BOOKS or KEYS. Definite NPs also commonly occur as “indirect anaphors” (IA) without ex plicit antecedents in text or discourse. The passage in (3) is perfectly coherent although there is no textual antecedent for the definite NP the tyres.
(3) The car was found in the ditch. The tyres were punched.
The missing overt connectedness is made up by a plausible conceptual interrelationship between the events described in the sentences. In order to complete the semantic form of the text, the reader automatically assigns a mental text-world model to the surface structure, which incorporates both information from the text and information activated through conceptual instantiation and inferential processing. Thus, the text-world model represents a referential constellation of states-of-affairs in working and in episodic memory which is (in most cases) more complex and elaborated than the semantic text basis. From a procedural perspective, semantic enrichment and conceptual elaboration (as part of text-world model building) occur spontaneously and have to be considered as part of our textual competence. The readers’ elaborative capacity is based on their memory structures and the ability to activate parts of these knowledge structures or to draw inferences. In general, the comprehension of indirect anaphors confronts the cognitive system of the reader/listener with the problem of identifying and localizing the referent to which the definite description refers in the text. Thus, we are concerned 3. The text-world model of the text is the mental representation of the referential states-of-affairs. It can be described in terms of conceptual constellations including all participating referents, their relations among each other and the events and situations in which they are performing. For a detailed account on the text-world model approach, see Schwarz (2000a and 2001b).
Monika Schwarz-Friesel
with the accessibility of indirect referents in the text-world (see Ariel 1990, 1994, 1996; Bosch 1988, Gundel et al. 1993). The general idea behind this is that verbal expressions serve as mental processing instructions for the reader (Givon 1992, 1995, Fauconnier 1994). As already pointed out, the text-world model of a text is based on the information of the text itself but elaborated by the readers’ cognitive processing. It is the result of applying knowledge representations to the text base. Building up a text-world model is an automatic process which always takes place in the comprehension process (for slightly different conceptions of text-worlds or mental models see deBeaugrande/ Dressler 1981, VanDijk/Kintsch 1983, Johnson-Laird 1983, Garnham/Oakhill 1992, Givon 1995, Cornish 1999). Concern with the resolution of definite referential expressions which connect sentences and serve as signals for topic continuity leads to the question how readers access and recover the referents of these expressions in text comprehension and what information has to be available in order to establish continuity. One of the main functions of referring expressions is to indicate the degree of accessibility of the mental representations of their referents in a given context (Ariel 1990:16; see von Heusinger 2000 and 2003 for formal aspects of accessibility theory). Accessibility theory has generally treated accessible text referents as mental entities that are retrievable from memory. In this sense, accessible referents constitute given information. Linking accessibility and givenness within the account of our text-world model proposal, referential movement in text comprehension can best be described as a sequence of the following three transitional states: Activation, reactivation and deactivation (see Givon 1992 and Chafe 1994; similar states have been postulated by centering theory, i.e. continuing, retaining and shifting; cf. Grosz 1981, Gordon et al. 1993, Maes 1997). Activation of a text referent takes place when a new mental file (formally represented as a knot in a network) is opened. The referent is activated and stored as a conceptual label, and incoming information about the same referent may be filed under this label. This process is to be characterised as reactivation, since the already established knot in the text-world model is called up again in short-term or in working memory. If a new referent is mentioned in the text, the referent in current focus is deactivated while at the same time a file for the new referent is opened and stored as an additional conceptual knot in the text-world model. In this regard, “given” means that the referent is mentally accessible in text-world model,
Indirect anaphora in text
“new” means that this is not the case. This can be illustrated with the following example from a newspaper text (concentrating on the protagonist entity):
(4) A burning tyre has been treacherous for a drunken car driver (activation). The 37-years-old woman (reactivation) noticed smoke and stopped to check the car. A man (deactivation of WOMAn and activation of new referential file For man) who had observed this, called the police.
The distributive flow of information in text comprehension, then, is to be described as an interplay of retaining continuity and moving/progressive development. Traditionally, definite NPs serving as anaphors are regarded as the main grammatical means to indicate continuity by referring to accessible, given entities. In the following section, however, I shall present a number of definite anaphoric NPs whose referents are seen as accessible even though they would procedurally be viewed as constituting new instead of given information.
2. Indirect anaphors in text: Towards a cognitive classification of given-andnew-entities in text Corpus-based and text-linguistic studies have shown that definite NPs commonly occur as “indirect anaphors”, that is without explicit antecedents in text or discourse (see Schwarz 2000a, b and Consten 2004 for a detailed account of indirect anaphora as a form of domain-bound reference). The prevailing assumption is that there is an associative relationship between a textual trigger and the anaphor which helps to establish a link between the two expressions. Thus, these referential instances often are called “associative anaphors” (e.g. Hawkins 1978, Vater 1984, Heim 1991, Löbner 1996). However, since only a small amount of indirect reference relations can be explained through the activation of associations (cf. Schwarz 2000a: 49), I prefer to use the more general term “indirect anaphor”. The alternative terms “bridging anaphors” (Clark 1977, 1978) and “inferables” (Prince 1981) seem to express that the relationship between trigger and anaphor is always established by inferences (cf. Haviland/Clark 1974, Clark/Haviland 1977). However, an equalisation of inference and indirect anaphora is too broad and too narrow at the same time: First, even incoherent text items can be interpreted by inferences without being acceptable anaphors, second, not all IA are inference based (cf. Schwarz 2000a: 88–90; Consten 2004: 68–70).
Monika Schwarz-Friesel
For a long time, an adequate conception of IA has been a desideratum. In recent years, some results on the resolution of IA from a cognitive point of view (Cornish 1999, Schwarz 2000a,b, Schwarz-Friesel in press, Consten 2003, 2004) have been published. How can this phenomenon best be described? Which characteristic properties do IA possess? Five main characteristics of IA can be made out: 1. There is no explicit antecedent to which the definite NP functioning as IA refers back. Instead, there is some kind of trigger or anchor, that is some element or elements in the preceding text in relation to which IA are interpreted. 2. The relation between the anchor and the IA is not based on coreference (but on some other close semantic link or conceptual relation). 3. There are restrictions for coding IA with pronouns or demonstratives. (See Cornish 1999, 2005 and this volume, Schwarz 2000a, Consten 2004, Consten/ Knees/Schwarz-Friesel, this volume). 4. The resolution of an IA in text comprehension involves much more than a simple search-and-match-procedure. They require for their full interpretation a cognitive process involving the activation of knowledge structures. 5. From the perspective of accessibility, indirect anaphora have to be seen as “given-and-new-entities”, since they combine activation and reactivation processes (see Schwarz 2000a,b). IA fall under the heading of anaphora because the interpretation of the definite NPs depends on the existence of an adequate textual element in the preceding text. Adopting the terminology of Fraurud (1990, 1992), I will call such elements “anchors” and the process of linking IA to these expressions “anchoring”. In many models IA are treated as a deviance from the normal anaphoric case, as “a violation from a felicity condition” (Heim 1982: 371). Both corpus-based and text linguistic studies, however, support the assumption that IA are quite common and normal uses of definite reference (Fraurud 1990, Schwarz 1996, 1997, 2001a,b). The common feature of IA is the lack of an explicit antecedent. The definite article signals accessibility of the referent but in the text structure no such referent can be found. Thus, we are concerned with the accessibility of implicit referents in the text-world. For their full interpretation they require some anchoring process. The examples I am going to discuss illustrate the wide range of indirect anaphors in text. Some linking process is required in all cases but the type of operation needed differs in a number of ways. A distinction between semantic and conceptual types is introduced (for a general discussion on semantic and conceptual knowledge see the papers in Schwarz 1994). Whereas the use and interpretation of semantic IA depends on the activation of knowledge in the mental lexicon, conceptual IA involve the processing of more general world knowledge. Within these two classes, four types of IA can be
Indirect anaphora in text
distinguished. The criteria established for distinguishing these different types are based on the linguistic expressions used as anchor and on the different knowledge structures involved in interpreting the mental relationship between anchor and indirect anaphor. The examples (which are taken from natural language texts in German) concern the most frequent and prototypical types of textual IA. In example (4) the NP a blue Mercedes is the anchor for the IA the wheel and the anchor for the IA the temples is the NP the driver / the man.
(5) Alphons Clenin [...] fand einen blauen Mercedes [...]. Es war ihm, als sei der Fahrer auf das Steuer gesunken. [...] Er bemerkte jedoch im gleichen Augenblick, dass der Mann tot war. Die Schläfen waren durchschossen. (Dürrenmatt: Der Richter und sein Henker, 5) (Alphons Clenin noticed a blue Mercedes. He had the impression that the driver was lying on the wheel. At that particular moment he noticed that the man was dead. His temples (German: the temples) were shot through.)
This semantic type of indirect anaphor is established by the noun-semantic relation between the anchor and the indirect anaphor, there is a part-whole-relationship between them. The knowledge structures needed for the interpretation of the anaphoric expressions are stored in our mental lexicon as part of our semantic knowledge. Example (5) shows an IA based on lexical knowledge, too, more precisely on the thematic roles of the verb in the preceding sentence.
(6) Ich wollte rasch die Haustür aufschließen, weil ich das Telefon klingeln hörte. Der Schlüssel war aber tief unten im Einkaufswagen vergraben. (Janet Evanovich: Einmal ist keinmal, 123) (I wanted to unlock the door quickly, because I could hear the telephone ringing. The key, however, was buried deeply in the trolley.)
The definite NP the key used as indirect anaphor represents the instrument role which is part of the semantic representation of the verb unlock which serves as the anchor. Referential anchoring is established through the assignment of the IA to the specific thematic role INSTRUMENT. In example (6) a conceptual type is seen. The interpretation of the IA is based on the activation of script-knowledge.
(7) Ich kenne ein schönes Restaurant in Refrath. Das Essen ist köstlich, und der Kellner ganz besonders nett.
4. For the German example there is an additional morphological anchorage since Schlüssel “key” and (auf-) schließen “(un-)lock” have the same root morphem. Consten (2004: 86) calls such examples “morphologically based IA”.
Monika Schwarz-Friesel
(I know a lovely restaurant in Refrath. The food is excellent and the waiter is an extremely nice guy.)
Establishing a coherent link between the sentences is to treat the waiter and the food as fulfilling typical roles or default values of the script which is activated in a top-down-process by the anchor-expression restaurant. The recipient has to localize the referent of the indirect anaphor in the mental structure of the script. According to a popular view, indirect anaphora are explained as ‘frame-evoked entities in an implicit focus. In this respect, Sanford/Garrod (1981, 1994a,b, Garrod 1995) distinguish direct and indirect anaphors with the help of explicit and implicit focus. Explicit focus is “the current focus of attention”, the highly activated part of workspace where incoming information is temporarily held, that is shortterm memory which contains representations of active referents explicitly introduced. Implicit focus is the representation of entities evoked implicitly as they form part of a frame retrieved from memory. Whereas the referents of direct anaphora are accessible in explicit focus, the referents of indirect anaphora have to be available in implicit focus, that is as default values of a frame or script. But not all indirect anaphora in a text refer to particular parts of a frame. The IA in example (8) illustrates the second conceptual type which is based on drawing an inference.
(8) Nachts stürzt ein Mann auf die Polizeiwache und berichtet erregt, daß er soeben im Dunkeln an seiner Gartenpforte niedergeschlagen worden sei. Ein Polizist wird beauftragt, am Ort des Überfalls Spuren zu sichern. Kurz darauf kommt er mit einer Riesenbeule am Kopf zurück: “Ich habe den Fall aufge klärt.” “Bravo”, sagt sein Chef, “und wie?” “Ich bin auch auf die Harke getreten!” (Hörzu, 19.4.96, 17) One night a man rushes into the police station and tells the policemen that he has just been knocked down in his garden. One policeman is asked to go and look for traces at the place of the assault. After a short time he returns with a huge swelling at his head and says “I solved the case”. “Bravo”, says his boss, “and how did you do that?” “I stepped on the rake, too.”
The NP the rake is a typical example for an IA based on inference. It can only be understood if the reader/listener draws some complex inference about what has been going on. This inference is triggered by the expression too: there never has been any assault. the man, just like the policeman, stepped on some rake which could not be seen in the dark. This process of inferential anchoring does not necessarily involve the activation of one specific frame or script. However, a lot of inferences are based on scripts. Thus, a strict distinction between the different types is often hard to draw.
Indirect anaphora in text
Still, lexical decomposition, slot-filling or spreading activation, belong to a type of semantic activation processes which should not be equated with using knowledge in inferential problem-solving activities. The term inference should not be used for all kinds of memory processes involved in text comprehension in the sense of “any piece of information that is not explicitly stated in the text” (McKoon/Ratcliff 1992: 440), but only for problem-solving activities, that is the constructive use of world knowledge. Otherwise, it is impossible to describe cognitive processes more precisely and in explicit detail (see Marx et al., this volume). The examples so far discussed show that within the group of IA different types can be made out which rely on different knowledge structures for their interpretation. However, a classification of types of anchoring can be made in gradual terms only (cf. Schwarz 2000a, Consten 2004). In general, anchoring conditions of referents for indirect anaphors can be formulated in the following way: The referent of an indirect anaphor must be either an identifiable part of the semantic structure of the preceding sentences, or it must be a default value of a specific frame or script, or it must be inferable on the basis of cognitive plausibility determined by general world knowledge. All kinds of indirect anaphora have to be seen as “given-and-new-entities”. Combining both easy accessibility and the establishment of new nodes in textworld models, they serve as “progressive continuity markers”. The definite NP denoting an indirect anaphor, refers to an entity being introduced into the text for the very first time. In this respect, indirect anaphora carry new information into the text-world model. Even more, a new node for the new referent has to be established in an activation process in working memory. On the other hand, at the semantic level of the text the definite article signals that (given) information is easily accessible and, thus, demands a reactivation process. Seen as grammatical means with processing instructions, indirect anaphora signal accessibility of the text referent while a new referent is introduced to the text-world model. From a given-new-perspective, thus, indirect anaphora combine both referential continuity and elaborative development in information processing. An activation process is taking place by establishing a new file in short-term memory, while at the same time we have reactivation in long-term memory. More precisely, the cognitive domain of the anchor expression has been in the status of semi-activation (for aspects of activation levels and the status of semi-activation in long-term memory representations of the anchors, see Schwarz 2000a: 136ff.). Hence, reactivation of the semi-activated domain accounts for the ease of processing and the acceptability of indirect anaphora as continuity markers. In this respect, indirect anaphora serve as means to establish coherence in the same way as direct anaphora.
Monika Schwarz-Friesel
4. Progressive anaphors I am going to discuss now a number of anaphors with explicit antecedents which resemble IA in specific ways, since they do not only reactivate old files in textworld model constellations, but activate at the same time new information about the referent in focus. It should be noted that these types of anaphors occur quite frequently in natural language texts and are by no means to be evaluated as some kind of textual deviance. To illustrate these “progressive or moving anaphors”, consider example (9) which depicts specifying anaphors (“Spezifikationsanaphern” in German).
(9) A burning tyre has been treacherous for a drunken car driver (activation of R(eferent)1). The 37-years-old woman (R2 is identical with R1: reactivation and specification of R1 by activation) noticed smoke and stopped to check the car. A man who had observed this, called the police….The mother of 3 children (R3 is identical with R1: reactivation and further specification of R1/2 by activation) had a lot of blood alcohol. The traffic offender (R4 is identical with R1: reactivation and specifying evaluation of R1/2/3) had to submit her drivers’ licence.
In (9), the anaphoric expressions are much more than a mere semantic echo of their antecedent. Several new information chunks are filed successively in the same conceptual node of the text referent. Hence, continued elaboration through activation processes of the current open file is the case. In text comprehension, specifying anaphors include both reactivation and activation of referential information in the text-world model. In cognitive terms, their informational status is both given and new, since the text-world model constellation changes. Examples (10) and (11) show that referents can change dramatically in the course of time (or more precisely in the text-world; see Stenning 1992). (10) Johannes hat sein altes Auto endlich doch noch zum Schrottplatz gebracht. Als er dann aber das zusammengepresste Eisenpaket sah, wurde ihm weh um’s Herz. (John brought his old car (R1) to the scrap yard. But when he saw the little iron package (R2=CHANGED R1), he felt very bad.) (11) Ich muß wohl zwei oder drei Raupen aushalten, wenn ich die Schmetterlinge kennenlernen will. (Saint-Exupéry, Der kleine Prinz, Dt.Übersetzung, 26) (Well, I have to cope with two or three caterpillars (R1), if I want to see the butterflies (R2=FUTURE R1).
Indirect anaphora in text
In both examples, the anaphor contains new information about the referent. The reader has to infer that some change has taken place (in 8) or (as in 9) will take place in order to interpret the anaphor properly and thus, get a coherent representation. In both cases we get an awkward result if we use pronouns or demonstratives instead of the lexical NPs (or the meaning representation is a completely different one). In (12) not a single word or phrase functions as antecedent, but several clauses. (12) Butter und Mehl gut verrühren, bis keine Klümpchen mehr da sind. Milch dazugeben. Die Mehlschwitze in eine Schüssel schütten. (Kochrezept) (Stir butter (R1) and flour (R2), until there are no more lumps. Add some milk (R3). Pour the/this roux (R4= R1, 2 and 3) into a bowl.)
A new text referent (ROUX) is created on the basis of old information. In contrast to (10) and (11) where we already have individual entities as referents which are coded by the anaphors as transformed entities and the reader has to mentally reconstruct this transformation, in (12) the new referent comes out as the result of a process which is presented before our very eyes. The referent can or should be coded by a demonstrative NP. However, if roux was replaced by it, the anaphor would be much more likely to be assigned to some milk. So far we have seen, that in many cases the processing of anaphoric devices is not only constrained by their lexical content, but strongly by the activation of conceptual knowledge representations and the performance of referent-creating operations (In “complex anaphors”, referent-creating mechanisms can be seen, as well, see Schwarz-Friesel et al. 2004, Consten et al., this volume). Besides, the examples discussed show that the resolution of many direct anaphors involves much more than a simple search-and-match-procedure. Their full interpretation requires a more complex cognitive process. Resolving anaphors with the strategy (R2 is identical with R1) may involve specification, combination, classification, and/or evaluation mechanisms. From a procedural perspective, many anaphors combine given and new information for the reader, and therefore can be characterized as “given-and-new-entities” in the text. Felicity of all those anaphors depends crucially on some plausibility criterion which is dependent both on long-term memory knowledge and on the text-world model at hand. This is quite evident in (13): (13) John brought his old car, which had served him well for many years, to the scrap yard. But when he later saw the little iron package #the cheese/#the flower/#the dog/#the ice cream, he felt very bad.
On the basis of world knowledge (on scrap yards, cars are sometimes pressed into iron packages), the definite NP the iron package is regarded as an anaphoric device although the anaphor and the antecedent car are not semantically
Monika Schwarz-Friesel
equivalent or compatible. Drawing the plausible inference, however, that the car has been crushed by a steel press, continuity and co-reference can be established. Obviously, this does not work with definite NPs (as indicated by the #) whose referents do not fit into the text-world model constellation. In many cases the interpretation of anaphoric devices is only weakly constrained by their lexical content, but strongly by the activation of conceptual knowledge representations and the performance of referent-creating operations compatible with the text-world model at hand. Besides, anaphors with explicit antecedents may not only introduce new information about the textual referent. They may even trigger new referents into the text-world model. Hence, the resolution of many direct anaphors involves much more than a simple search-andmatch-procedure. A mental representation of the textual referent must be either available or inferable in memory or text structure at the time the anaphoric item is processed. Given this similarity between direct and indirect anaphors, the interpretation of direct and indirect anaphors seems to obey similar rules. A unified theory of anaphora has to specify the conditions under which the various uses of anaphora can be interpreted and how the interpretation depends on the meaning of the antecedents/anchors. In order to provide a full account of how indirect and direct anaphors are interpreted, a cognitive theory of reference resolution must include both representational and procedural factors. So far I have concentrated on the various instances of anaphors in text and the mental knowledge structures needed for their interpretation. In the next section I want to discuss briefly some of the linguistic and cognitive constraints on indirect anaphora, more precisely on the associability in the anchoring process.
5. Constraints on associability According to one popular view, antecedentless definite NPs are described as associative anaphors with an associative relation underlying anchor and indirect anaphor. Obviously, associability which is ultimately based on our knowledge stored in long term memory does play a fundamental role for the use and interpretation of IA. Associability, however, is controlled by a number of linguistic restrictions and cognitive constraints. Example (14) illustrates that even a strong association between two items is sometimes not a sufficient criterion for the use of an indirect anaphor. (14) Es floss viel Blut in diesem Kampf. #Das Rot/#Die Röte stach den Kriegern in die Augen. (A lot of blood was shed in the battle. #The red thrust the warriors in the eyes.)
Indirect anaphora in text
Although we strongly associate blood with the colour red (in fact, this relation is stored in our mental lexicon as part of our semantic-conceptual knowledge), it is not possible to use blood as an anchor and red as an indirect anaphor in example (14). Interestingly, this constraint does not hold for all mass nouns as anchors and/ or colour terms as anaphors. Consider, for instance: (15) Als wir erwachten, lag alles voller Schnee. Das Weiß blendete unsere Augen. (When we got up, there was snow all over. The white was simply overwhelming.)
Obviously, contextual information plays a crucial role in anaphor resolution. Reading (15) we activate a text-world model in which the spatial representation of the white snow is dominant (covering the ground completely). Thus, in (15) the plausibility and saliency structure of the referential representation of the state-ofaffairs allows for the use of the white as an indirect anaphor. Accordingly, within the context of a specific text-world model, such as the setting of a novel or a fairy tale, we accept the red as an indirect anaphor: (16) Es war einmal […] eine Königin […]. Und wie sie so nähte und nach dem Schnee aufblickte, stach sie sich mit der Nadel in den Finger und es fielen drei Tropfen Blut in den Schnee. Und weil das Rote im weißen Schnee so schön aussah […]. (Schneewittchen, Grimms Märchen) (Once upon a time there was a queen. And as she sewed and looked up at the snow she pricked her finger with her needle and three drops of blood fell into the snow. And as the red on the white snow looked so beautiful [...]. Snowwhite, Grimms Fairytales)
Concluding, the mental state and focus structure of the current text-world model (including the representation of referential entities and spatial relations) determines and constrains the acceptability of anaphoric items (see Schwarz 2001b). Example (17) illustrates another well-known constraint imposed on the use of indirect anaphors: Although the broken pieces fulfill a necessary part of the semantic-conceptual representation of break, it is not possible to refer to these parts with a demonstrative NP. (17) Sie hat die Vase zerbrochen. #Diese Scherben liegen... (She broke the china vase. #These broken pieces still lie...)
This is also the case with a lot of direct anaphoric devices. Coming now to the cognitive constraints determining the associability conditions of IA. First, unambiguous associability in the relevant anchoring domain is needed. 5. In contrast, demonstratives are fine for some kinds of indirect complex anaphor, s. Consten et al., this volume.
Monika Schwarz-Friesel
(18) On Monday we went to a Chinese Restaurant and to an ice cream parlour. #The waiter was a nice guy. (19) Es sind viele schöne Möbelstücke im Zimmer zu sehen. Die alten Stühle gefallen mir besonders gut. / Der alte Tisch / #Der alte Stuhl gefällt mir besonders gut. (You can see a lot of lovely furniture in the room. The old chairs I like best. / The old table / #The old chair I like best.)
This restriction follows from the general claim that definite reference requires identifiability of the referent. In (19), a referent for the old chair would be problematic to identify since the recipient expects more than one chair due to a prototypical ROOM-FURNITURE-concept that is part of his conceptual knowledge. In contrast, it is most plausible that there is one and only one table in the room, which agrees with the semantic information of the definite NP that the referent is unique in the relevant domain. Thus, absence of referential ambiguity is a general condition for definite reference and has to be ruled out in cases of direct anaphora, too. (20) On Monday we went to see professor Martin and professor Landau. #The professor was very nice.
Second, associability within one anchoring domain is required. See (21) which illustrates that a change from one domain to another in text structure interferes with the principle of associability. (21) Ella feierte fröhlich Geburtstag. Erst gestern war sie aus dem Krankenhaus entlassen worden. Die Krankenschwestern hatten sich rührend um sie gekümmert. #Die Geschenke wurden mit großem Hallo geöffnet. (Ella was celebrating her birthday. Only yesterday did she come from hospital. The nurses had taken very good care of her. #The presents were opened with Hurrah.)
The presents is an indirect anaphor which is no longer associable with the anchor birthday because of the shifted focus to the hospital-domain. The same constraint is valid for direct anaphors, consider (22): (22) Ella feierte fröhlich Geburtstag. Sie hatte viele Geschenke bekommen. Erst gestern war sie aus dem Krankenhaus entlassen worden. Die Krankenschwes tern hatten sich rührend um sie gekümmert. #Die Geschenke wurden mit großem Hallo geöffnet. (Ella was celebrating her birthday. She got a lot of presents. Only yesterday did she come from hospital. The nurses had taken very good care of her. #The presents were opened with Hurrah)
Indirect anaphora in text
Third, and most important, plausibility of associability within the anchoring domain is needed. (23) I wanted to open the door. The key/#The debit card did not work.
In (23) the associability principle rules out the debit card if the recipient lacks the knowledge that debit cards can be used as a tool to open doors. However, in (24) which is a text from a detective novel, cognitive plausibility determines the anchoring process. Obviously, associability depends not only on the relation between anchor and indirect anaphor but also seems to vary within particular contexts. (24) “Wie sollte ich bloß ohne Schlüssel in die Wohnung kommen? Die EC-Karte musste herhalten!” (Einmal ist keinmal, Goldmann-Roman, 19) (How could I get into the apartment without a key? I had to try with the debit card.)
Thus, felicity of indirect anaphors depends on the plausibility of the required inference within the text-world model. The same holds true for many direct anaphors (see example (10)): To sum up: The anchoring of IA is determined by the principle of associability. It turns out, however, that in many cases the principle of associability is restricted by linguistic and conceptual constraints. A strong link between anchor and IA does not per se guarantee proper and successful anchoring. The resolution of indirect anaphors is governed by the overall principle of plausibility through an interaction between the information in the anchor and the anaphor, the mental accessibility of the implied referent within a specific context and the focus state of the prior and current textual representation of the text-world model. And it is the interaction between these variables that deserves more attention in future research.
6. Conclusion Considering the knowledge structures used for the interpretation of IA as the basis for a more general explanation of how people process referring expressions, four main types of IA have been distinguished within a classification that differentiates between semantic IA and conceptual IA. Some cognitive process is required in all cases but the type of mental operation needed differs in a number of ways. I have further shown that some anaphoric occurrences with explicit antecedents in text apparently share some characteristics of IA. Besides, it has been shown, that the traditional view of anaphora as a means of referential continuity with topical function and the availability of given information has to be revised. Taking a procedural perspective concerning the thematic development and information
Monika Schwarz-Friesel
distribution in text, I have illustrated that both direct and indirect anaphors may add new information to the mental files of the referents in the text-world model. I have argued against attempts to explicate IA merely in terms of associability conditions. Instead, the principle of conceptual plausibility is the most dominant constraint which determines the process of anaphora resolution. IA should be approached within the text-world model approach which takes contextual, representational and procedural factors of the plausibility criterion into account. There is no clear-cut distinction between direct and indirect anaphora in text, showing that the most typical examples of direct and indirect anaphors may best be regarded as two extremes on a continuum of textual reference, thereby weakening the claim for a strict distinction between the two types. This should lead to a more unified account of anaphora in text.
References Ariel, M. 1990. Accessing Noun-Phrase Antecedents. London: Routledge. Ariel, M. 1994. Interpreting anaphoric expressions: A cognitive versus a pragmatic approach. Journal of Pragmatics 30: 3–42. Ariel, M. 1996. Referring expressions and the +/– coreference distinction. In Reference and Referent Accessibility, T. Fretheim and J. Grundel (eds), 13–25. Amsterdam: John Benjamins. Beaugrande, R.A. de and Dressler, W.U. 1981. Einführung in die Textlinguistik. Tübingen: Niemeyer. (English translation: Beaugrande, R.A. de and Dressler, W.U. 1994. Introduction to Text Linguistics (7th ed.). London: Longman). Bosch, P. 1988. Representing and accessing focussed referents. Language and Cognitive Processes 3: 207–231. Chafe, W. L. 1994. Discourse, Consciousness, and Time. The flow and displacement of conscious experience in speaking and writing. Chicago IL: Chicago University Press. Clark, H.H. 1977. Bridging. In Thinking. Readings in cognitive science, N. Johnson-Laird and P. Wason (eds), 411–420. Cambridge: CUP. Clark, H.H. 1978. Inferring what is meant. In Studies in the Perception of Language, W.J.M. Levelt and G.B. Flores d’Arcais (eds), 295–322. New York NY: Wiley. Clark, H.H. and Haviland, S. 1977. Comprehension and the given-new contract. In Discourse Production and Comprehension, R. Freedle (ed.), 1–40. Norwood NJ: Ablex, Consten, M. 2003. Towards a unified model of domain-bound reference. In Deictic Conceptualisation of Space, Time and Person, F. Lenz (ed.), 223–248. Amsterdam: John Benjamins. Consten, M. 2004. Anaphorisch oder deiktisch? Zu einem integrativen Modell domänengebundener Referenz [LA 484]. Tübingen: Niemeyer. Cornish, F. 1999. Anaphora, Discourse, and Understanding. Evidence from English and French. Oxford: Clarendon. Cornish, F. 2005. Degrees of indirectness: Two types of implicit referents and their retrieval via unaccented pronouns. In Anaphora Processing: Linguistic, cognitive and computational modelling, A. Branco, T. McEnery and R. Mitkov (eds), 199–220. Amsterdam: John Benjamins.
Indirect anaphora in text
Dahl, Ö. and Hellmann, C. 1995. What Happens When we use an Anaphor? Paper, Stockholm University, Dept. of Linguistics. Fauconnier, G. 1994. Mental Spaces: Aspects of meaning construction in natural language. Cambridge MA: Bradford. Fraurud, K. 1990. Definiteness and the processing of NPs in natural discourse. Journal of Semantics 7: 395–434. Fraurud, K. 1992. Processing Noun Phrases in Natural Discourse. PhD dissertation, Stockholm University. Garnham, A. and Oakhill, J. (eds). 1992. Discourse Representation and Text Processing. A special issue of Language and Cognitive processes. Hove: Lawrence Erlbaum Associates. Garrod, S. 1995. Distinguishing between explicit and implicit focus during text comprehension. In Focus and Coherence in Discourse Processing, G. Rickheit and Ch. Habel (eds), 3–17. Berlin: Mouton de Gruyter. Givon, T. 1992. The grammar of referential coherence as mental processing instructions. Linguistics 30: 5–55. Givon, T. 1995. Coherence in text vs. coherence in mind. In Coherence in Spontaneous Text, M.A. Gernsbacher and T. Givon (eds), 59–115. Amsterdam: John Benjamins. Gordon, P.C., Grosz, B.J. and Gilliom, L.A. 1993. Pronouns, names, and the centering of attention in discourse. Cognitive Science 17: 311–347. Grosz, B. 1981. Focusing and description in natural language dialogues. In Elements of discourse understanding, A. Joshi, B. Webber and I. Sag (eds), 84–105. Cambridge: CUP. Gundel, J.K., Hedberg, N. and Zacharski, R. 1993. Cognitive status and the form of referring expressions in discourse. Language 69: 274–307. Haviland, S.E. and Clark, H.H. 1974. What’s new? Acquiring new information as a process in comprehension. Journal of Verbal Learning and Verbal Behavior 13: 513–521. Hawkins, J. 1978. Definiteness and Indefiniteness: A study in reference and grammaticality prediction. London: Groom Helm. Heim, I. 1982. The Semantics of Definite and Indefinite Noun Phrases. PhD dissertation, University of Massachussets, Amherst (Schriftenreihe des SFB 99, Linguistik, Nr. 73, Universität Konstanz). Heim, I. 1991. Artikel und Definitheit. In Semantik. Ein internationales Handbuch der zeitgenössischen Forschung, A. von Stechow and D. Wunderlich (eds), 487–535. Berlin: Walter de Gruyter, Heusinger, K. von. 2000. Anaphora, antecedents, and accessibility. Theoretical Linguistics 26: 75–93. Heusinger, K. von. 2003. The double dynamics of definite descriptions. In Meaning in the Dynamic Turn. J. Peregrin (ed.), 150–168. Amsterdam: Elsevier. Johnson-Laird, P.N. 1983. Mental Models. Towards a cognitive science of language, inference, and consciousness. Cambridge MA: Harvard University Press. Löbner, S., 1996. Associative anaphors. Ms. University of Düsseldorf. Maes, A. 1997. Referent ontology and centering in discourse. Journal of Semantics 14: 207–235. McKoon, G. and Ratcliff, R. 1992. Inference during reading. Psychological Review 99: 440–466. Prince, E.F. 1981. Toward a taxonomy of give-new-information. In Radical Pragmatics, P. Cole (ed.), 223–255. New York NY: Academic Press, Rickheit, G. and Strohner, H. 1993. Grundlagen der kognitiven Sprachverarbeitung [UTB 1735]. Tübingen: Francke. Sanders, T. and Noordman, L.G. 2000. The role of coherence relations and their linguistic markers in the text processing. Discourse Processes 29(1): 37–60.
Monika Schwarz-Friesel Sanders, T. and Spooren, W. 2001. Text representation as an interface between language and its users. In Text Representation: Linguistic and psycholinguistic aspects, T. Sanders, J. Schilperoord and W. Spooren (eds), 1–25. Amsterdam: John Benjamins. Sanford, A.J. and Garrod, S.C. 1981. Understanding Written Language: Explorations in comprehension beyond the sentence. Chichester: Wiley. Sanford, A.J. and Garrod, S.C. 1994a. Resolving sentences in a discourse context. In Handbook of Psycholinguistics, M.A. Gernsbacher (ed.), 675–698. San Diego CA: Academic Press. Sanford, A.J. and Garrod, S.C. 1994b. Selective processing in text understanding. In Handbook of Psycholinguistics, M.A. Gernsbacher (ed.), 699–719. San Diego CA: Academic Press. Schwarz, M. (ed.). 1994. Kognitive Semantik/Cognitive Semantics. Tübingen: Narr. Schwarz, M. 1996. Lexikalische und konzeptuelle Restriktionen beim Verstehen direkter und indirekter Anaphern. In Lexical Structures and Language Use, Vol. II., F. Hundsnurscher and E. Weigand (eds), 399–407. Tübingen: Niemeyer. Schwarz, M. 1997. Anaphern und ihre diversen Antezedenten: Koreferenz und Konsorten. In Sprache im Fokus, C. Dürscheid, K.H. Ramers, M. Schwarz, (eds), 445–456. Tübingen: Niemeyer. Schwarz, M. 2000a. Indirekte Anaphern in Texten. Studien zur domänengebundenen Referenz und Kohärenz im Deutschen [LA 413]. Tübingen: Niemeyer Schwarz, M. 2000b. Textuelle Progression durch Anaphern – Aspekte einer prozeduralen Thema-Rhema-Analyse. Linguistische Arbeitsberichte 74: 111–126. Schwarz, M. 2001a. Kohärenz – Auf den materiellen Spuren eines mentalen Phänomens. In Gesprochene Sprache – transdisziplinär. Festschrift für Gottfried Meinhold, M. Bräunlich, B. Neuber and B Rues (eds), 151–160. Frankfurt am Main: Peter Lang. Schwarz, M. 2001b. Establishing coherence in text. Conceptual continuity within the domain of text-world models. Logos and Language II(2): 15–23. Schwarz-Friesel, M. In press. Text comprehension as the interface between verbal structures and cognitive memory processes: The case of resolving direct and indirect anaphora. In Memory and Language, C. Zelinsky-Wibbelt (ed.). Berlin: Mouton de Gruyter. Schwarz-Friesel, M., Consten, M. and Marx, K. 2004. Komplexanaphern. In Stabilität und Flexibilität in der Semantik, K.P. Konerding and I. Pohl (eds), 67–86. Frankfurt am Main: Peter Lang, Stenning, K. 1992. Representations of individuals and the processing of reference change. Quaterly Journal of Experimental Psychology. Human Experimental Psychology 45A(3): 373–398. Van Dijk, T.A. and Kintsch, W. 1983. Strategies of Discourse Comprehension. New York NY: Academic Press. Vater, H. 1984. Referenz und determination im text. In Sprache und Pragmatik,Lunder Symposium 1984 [Lunder Germanische Forschungen 54], I. Rosengren (ed.), 323–344. Zelinsky-Wibbelt, C. 2000. Discourse and the Continuity of Reference: Representing mental categorization. Berlin: Mouton de Gruyter. Zelinsky-Wibbelt, C. (ed.). 2003. Text, Context, Concepts. Berlin: Mouton de Gruyter.
Indirect pronominal anaphora in English and French Marginal rarity, or unmarked norm? Some psycholinguistic evidence Francis Cornish While for certain linguists (e.g. Erkü & Gundel, 1987) and psycholinguists (e.g. Sanford et al., 1983), using unaccented third person pronouns to refer to implicit referents is impossible or highly marked, for other linguists (e.g. Yule, 1982) and psycholinguists (e.g. Greene et al., 1994), this is not only acceptable but common in normal conversational discourse. If we draw a principled distinction between two main types of implicit referent (central or nuclear referents, and peripheral ones), then both sides in the debate may be correct. The results of two reading experiments in both English and French conducted to test this distinction showed indeed that object pronominal reference to implicit referents only caused slower reading times compared to explicit referents for peripheral referents.
1. Introduction The details of the two experiments reported on in this paper (in particular, the statistical analyses of their results) are presented in full in Cornish et al. (2005).
1. This paper was presented at the 27th Annual Meeting of the German Linguistic Society, 23–26 February 2005: Workshop n° 1, “Anaphora in Texts”, held at the University of Cologne. I am very grateful to the audience at this presentation for the useful discussion which followed it, as well as to the organizers of this Workshop, Monika Schwarz-Friesel and Manfred Consten, for inviting me.
F.rancis Cornish
The reader is invited to consult that article for further information, since there is no space in the present paper to include the details. I take a discourse-cognitive view of anaphora, whereby the anaphor retrieves, not necessarily the referent associated with its associated antecedent (as illustrated in (1a) below) at the point of occurrence, but a salient discourse representation compatible with what is predicated of the anaphor’s potential referent and available when the anaphor is uttered. Minimally, the antecedent’s discourse representation will have accrued (and/or shedded) properties as a result of the predication of which its referent is a part, and of any subsequent predications in which it is involved. The anaphoric or “host” predication plays a crucial orienting role in determining the potential range of interpretations of an anaphor, not only by filtering out incompatible candidate referents, but also by targeting a certain salient discourse representation as the one yielding the highest degree of coherence following the integration of the anaphoric clause with its immediate discourse context. “Indirect” pronominal anaphora offers a particularly clear insight into the discourse-functioning of anaphora as a whole, since it necessarily involves recourse to inference, to cognition, to context, and to stereotypical and encyclopaedic knowledge to a high degree. I will start by distinguishing direct and indirect anaphora, and will then present several linguistic and psycholinguistic hypotheses on the possible retrieval of indirect, implicit referents via pronouns. Following this, I will present the two reading time experiments in both French and English which I conducted recently in collaboration with two teams of psycholinguists, in order to test our own hypothesis of the existence of two types of indirect referent evoked via a given antecedenttrigger: namely, central and peripheral subtypes, and the possibility of object pronouns’ felicitously retrieving the former, but not the latter. I will end by evaluating our results in the light of one influential theory of cognitive accessibility (that of Gundel, Hedberg & Zacharski, 1993).
2. What is (in)direct anaphora? Anaphora functions exophorically (as illustrated in (1b) below) in exactly the same way as ‘endophora’ (co-textual anaphora: see (1a)). (See Cornish, 1999: Ch. 4 for arguments for this position.) Both are illustrations of ‘direct’ anaphora: the intended referent is directly evoked co-textually in (1a), and situationally in (1b), where the interlocutors are not only party to the scene at issue, but are actually focusing on the goat’s actions. In both cases, the relevant discourse ‘field’ (co-text and situation of utterance) directly evokes a salient discourse representation in terms of which the pronouns operate.
Indirect pronominal anaphora in English and French
DIRECT ANAPHORA
(1) a A young goat suddenly entered by the half-open front door; but no-one could guess exactly what it was looking for.
b [Context: a young goat suddenly enters by the half-open front door. Speaker, observing the scene with fascination:] What do you think it’s looking for, exactly? (Cornish et al., 2005: 364)
Turning now to ‘indirect’ anaphora, this is illustrated by the attested examples (2ac) below. (2a) from French involves the stereotypically close relation between a work (here a film) and its author (the film-maker). (2b) from English works in terms of the illocutionary point of the woman’s question (to elicit from the man the reason(s) why he did not send her any letters) as well as the lexical semantics of the predicate “write” (in its “correspond” sense). Finally, (2c) from German operates in terms of the stereotypical function of carparks and the lexical-semantic structure of the predicate “parken”. In all three examples, the indirect referent at issue is retrieved via an unaccented 3rd person pronoun. INDIRECT ANAPHORA
(2) a “Ah dis donc maman tu t’souviens Cinéma Paradiso, ben il a fait un nouveau film” ‘Oh I say, Mum, you remember ‘Cinema Paradiso’, well, he made a new film’ (example (65) cited in Reichler-Béguelin, 1993) b Woman: “Why didn’t you write to me?” Man: “I did…, started to, but I always tore ’em up.” (Extract from the film Summer Holiday) c “Ich wäre wunschlos glücklich, wenn Sie nicht immer auf dem Lehrerparkplatz parken würden. Das nächste mal lasse ich ihn abschleppen.” (Example taken from the TV soap opera Beverly Hills 90210, German RTL, 31.8.95, cited in Consten, 2001)
Indirect anaphora, then, requires a semi-automatic ‘bridging’ inference to get from what is explicitly mentioned or focused upon to a referent which is in some way associated with it – via a part-whole, token-type or metonymic relation of some kind. For the retrieval to be felicitous, there must be a potential “aboutness” or topical relation between the indirect referent and the subsequent discourse.
2. The conversational Spanish corpus presented in Blackwell (2003) includes several naturally-occurring examples of indirect pronominal anaphora of this kind in extended text.
F.rancis Cornish
3. Linguistic and psycholinguistic hypotheses on possible retrieval of implicit referents by pronouns Several linguists have put forward severe restrictions on the occurrence of unstressed pronouns with indirect anaphoric reference. For example, according to Dik (1978: 20), “Anaphoric reference to implicit referents is the exception rather than the rule.” For their part, Erkü & Gundel (1987) state that “Indirect anaphora may not be realized via (unstressed) pronouns”. Witness their property III, p.539:
(3) “Neither type of indirect anaphora may be pronominal.”
The two impossible types of indirect pronominal anaphora that Erkü & Gundel had in mind here are “inclusive” (associative) anaphora, as in (4), and “exclusive” (complementary) anaphora, as in (5): (4) I couldn’t use the box you gave me. The bottom/#it fell out. (Erkü & Gundel, 1987, ex (1))
(5) The ant daubs part of her burden onto a cocoon and passes the rest/#it to a thirsty larva. (Erkü & Gundel, 1987, ex (6))
In (4), the antecedent-trigger evokes a frame (‘the box’) of which the anaphor’s referent is a part (‘the bottom of the box’). Clearly, the parts of a whole will not be activated psychologically upon mention of that whole, just as the residue of a set or a mass (as in (5)) is not made salient when the other part is evoked. So unaccented pronouns will clearly not be able to target anaphorically such ‘background’ referents. See the further strictures of Gundel, Hedberg & Zacharski (2000: 93), who claim that “Indirect pronominal and demonstrative anaphors are relatively infrequent in relation to indirect anaphors introduced by a definite or indefinite article.” In their conclusion (p.100) they further state that “Indirect anaphors, typically, may not be coded via pronominal or demonstrative phrases…” In the psycholinguistics field, Sanford & Garrod (1981) draw a processing distinction between two distinguishable areas within the mental workspace: explicit focus, which is claimed to house representations of referents which have been introduced textually and are thus located in short-term memory; and implicit focus, which is located within the wider area of working memory, which includes short-term memory as a proper part. This section would house implicit, stereotypical entities forming part of the background scenario evoked in order to sustain the more foregrounded domain of explicitly mentioned entities. In this early work, Sanford & 3. Recall in this respect the discussion of the strong constraints on so-called anaphoric islands first put forward by Paul Postal in the late 1960’s.
Indirect pronominal anaphora in English and French
Garrod posit a ban on antecedentless pronouns: references to implicit entities such as the waiters in a restaurant scenario, or the nurses in a hospital one, must be carried out via lexically more explicit phrases than unaccented pronouns. In a later publication (Sanford et al., 1983), the authors reported on a series of experiments carried out in collaboration with several colleagues: first, an on-line self-paced reading time test was administered to 30 subjects, using several sets of texts of the type illustrated in Table 1; these texts appeared sentence by sentence on a computer screen. Table 1. Stimuli from Sanford et al. (1983), Experiment 1
(6) a Matching Gender/ Number
(6) b No Match
Sentence 1
Explicit antecedent
Implicit Antecedent
Roland parted his long hair with a comb.
Ronald parted his long hair.
Sentence 2 (Target)
It was twisted with many teeth missing.
Sentence 3
He had had it since childhood.
Question
Did Ronald part his hair with a brush? (No)
Sentence 1
Being arrested by the police was embarrassing for Andy
Being arrested was embarrassing for Andy.
Sentence 2 (Target)
They took him to the station in a van.
Sentence 3
He was charged with breach of the peace.
Question
Was Andy embarrassed by his arrest (Yes)
(Cornish [et al.], 2005: 365)
After the standard Reading Time test, they gave the subjects tested a set of written texts (assembled in a booklet) made up of each of the conditions represented in Experiment 1 (see Table 1 for a sample), and assigned them the following (off-line) rewriting task: “Re-write these texts if you consider it necessary, in such a way as to make them ‘sound better’.” Predictably, only 7% of texts with explicit antecedents were re-written, but between 83% and 92% of those with implicit ones were. These results are a reflection of the influence of this off-line, “reflective” context and the implicitly normative effect of the instruction, where the implication conveyed is that “these texts may not all sound very good”.
F.rancis Cornish
However, in their on-line reading time experiment 1, where the very same materials were used (see again Table 1 for a sample), only a non-significant difference of 139 msecs was found between the mean reading times for the target pronominal utterances in both the Explicit and Implicit Antecedent conditions, in the “non-matching” materials, as in (6b) within Table 1. “Non-matching” materials are those where there is no “alternative” explicit textual antecedent in Sentence 1 to which the pronoun could “bond”, in terms of identical gender and/or number features. This is the case in (6a) in Table 1, where the pronoun it in the critical, target sentence 2 could “bond” (inappropriately) to the referent ‘Roland’s long hair’, since the NP his long hair in sentence 1 of this item is both singular and neuter (inanimate) in gender. So the conclusion drawn from the off-line rewriting test results by the authors, to the effect that “This study clearly demonstrates an unprompted judgement that pronouns without antecedent are unacceptable – at least in materials of the type used here” (Sanford et al., 1983: 313), cannot in fact be upheld. It is really a question of the distinct genres corresponding to informal, colloquial speech, on the one hand, and formal written prose, on the other, and the listener’s or reader’s expectations based on their knowledge of the contexts appropriate to these genres, which is at issue here. The contrasting pairs of examples shown in (7)-(9), with pronouns (the (b) examples in (8) and (9)) and definite lexical NPs (in (7b) and (9a)) as anaphors, provide the evidence on which Sanford & Garrod (1981) based their original distinction between explicit and implicit focus spaces, and their ban on pronouns’ retrieving referents from within the implicit focus space. (7) a Stated antecedent b Implied antecedent Mary put the baby’s clothes on. Mary dressed the baby. The clothes were made of pink wool. The clothes were made of pink wool. (8) a Mary put the baby’s clothes on. They were made of pink wool.
b Mary dressed the baby. #They were made of pink wool.
(9) a Harry drove to London. The car broke down half-way.
b Harry drove to London. #It broke down half-way.
4. Predictions to be tested via the French and English experiments Comparing examples (2a-c) with (4) and (5) on the one hand, and with those in (6b), (8b), and (9b) on the other, we observe the following range of implicit referent indirectness:
Indirect pronominal anaphora in English and French
1 Central or Nuclear indirect referents The pronouns with implicit (indirect) antecedents in (2a-c) could be easily processed, since the intended referent was central within the discourse representation targeted by the anaphoric predication as a whole. The film-maker is a central factor in the making of a film in (2a), and the concepts of ‘a letter’ and ‘a car’ are nuclear arguments of the predicates corresponding to the verbs write and parken in (2b) and (2c), respectively. 2 Associative (‘part-whole’) or complementary indirect referents In examples (4) and (5), on the other hand, pronouns could not retrieve the intended implicit referent, since it is inactive psychologically within the target discourse representation – that is, in the extreme background of consciousness. 3 Peripheral, ‘semi-active’ indirect referents In (6b), (8b) and (9b), the intended ‘implicit’ referents were evoked via the relevant trigger predicate as the means by which the situation described was (partially) realised. They are thus ‘peripheral’, ‘semi-active’, but not central participants in these states of affairs. However, in Sanford & Garrod’s experiments, this crucial distinction between ‘nuclear’and ‘peripheral’ implicit referents was not taken into account. Our experiments using French- and English-language materials took specific account of the first and third of these types of implicit referents (‘central’ and ‘peripheral’). We wanted to determine whether the degree of conceptual centrality of given indirect referents could be the factor accounting for the apparently conflicting claims of these linguists or psycholinguists. See Cornish (2005: 208–211) for arguments in favour of the ‘nuclear’ vs. ‘peripheral’ distinction in connection with indirect referents. The target pronouns in our experiments were all object ones, unlike in Sanford et al.’s (1983) experiments. Thus, either they followed the verb (in English), without being interpretable independently of it, or they preceded it (in French) as proclitic pronouns – which as such are not independent of their host, either accentually or semantically. In no case did we use materials where the implicit referent was inferable in terms of a morphological connection between the antecedent-trigger and the expression in terms of which the anaphor would receive its interpretation.
4. Though not of linguists such as Yule (1979) or psycholinguists such as Greene et al. (1994).
F.rancis Cornish
4.1
General Prediction…
That, contrary to the claims of Dik (1978), Erkü & Gundel (1987), Sanford & Garrod (1981) and Sanford et al. (1983), a (non-subject) pronoun CAN felicitously retrieve an implicit referent without increasing processing cost – but on condition that it is nuclear and not peripheral. 4.1.1 More specifically… When the referent is nuclear, no processing difference (no significant difference in reading times) was expected between explicit and implicit conditions. But when the referent is peripheral, reading times in the explicit conditions should be faster than in the implicit ones. Similarly, in the implicit conditions, reading times should be faster with a nuclear referent than with a peripheral one, while no difference in reading times was expected in the explicit conditions.
5. Design, materials and method adopted in the two experiments The aim of the study was to test experimentally the psychological reality of the distinction between ‘central’ and ‘peripheral’ implicit referents. To this end, two selfpaced reading time experiments were devised, prepared and implemented. The first, in French, was set up in collaboration with M. Fossard University of Laval Quebec, Canada, F. Rigalleau and V.André (University of Poitiers, France); and the second, in English, in conjunction with A.Garnham (University of Sussex, U.K.) and H.W. Cowles (University of Florida, USA). Each experiment tested 20 subjects. 5.1
Design of the study
24 experimental texts (short two-part dialogues) were devised in both French and English (see Tables 2 and 3 below for a sample of the materials used), interspersed with 24 filler (distractor) texts in order to avoid subjects’ developing comprehension strategies. The two main factors were Centrality (nuclear vs. peripheral referens) and Explicitness (implicit vs. explicit referents). These factors were crossed so as to yield four conditions for each text:
version 1: nuclear * implicit; version 2: nuclear * explicit; version 3: peripheral * implicit; version 4: peripheral * explicit.
Indirect pronominal anaphora in English and French
See Tables 2 and 3 below for illustrations of these four conditions. Two-part dialogues were chosen, of the type Statement-Reaction or Question-Answer, due to the greater frequency of indirect pronominal anaphora in informal speech than in more formal written prose. The second dialogue turn containing the object pronoun was the target utterance (whose reading times were recorded and measured), and Statements were presented at the end of each dialogue for assessment as TRUE or FALSE with respect to the situation described. The Statements for the Implicit conditions related to the target (pronominal) utterance, and were always TRUE; whereas those for the Explicit conditions related to the initial, antecedent-trigger utterances, and were always FALSE. They were included in order to ensure that subjects read the dialogues for comprehension. Table 2. Experiment 1 (French): Example Stimuli Nuclear
Speaker 1
Speaker 2 (Target)
Peripheral
Explicit antecedent
Implicit Antecedent
Cet artiste a peint toute la journée en plein air hier. Ses tableaux ont vivement impressionné une passante très riche. That artist painted all day in the open air yesterday. His pictures greatly impressed a very wealthy lady passing by.
Cet artiste a peint toute la journée en plein air hier. That artist painted all day in the open air yesterday.
Oui, et il les a vendus à bon prix en plus. Yes, he sold them for a good price as well.
Statement
L’artiste a peint des tableaux dans l’atelier. (FAUX) The artist painted pictures in the studio. (FALSE)
L’artiste a pu vendre ses tableaux. (VRAI) ������ The artist was able to sell his pictures (TRUE)
Speaker 1
Cet artiste a peint toute la journée en plein air hier. Ses pinceaux étaient nombreux et de tailles différentes. That artist painted all day in the open air yesterday. His brushes were numerous and of different sizes.
Cet artiste a peint toute la journée en plein air hier. That artist painted all day in the open air yesterday.
Speaker 2 (Target) Statement
(Cornish et al., 2005: 369)
Oui, et il les a tous utilisés, du plus fin au plus épais. Yes, and he used them all, from the finest to the thickest. L’artiste a peint juste une partie de la journée. (FAUX) The artist only painted for part of the day. (FALSE)
L’artiste a bien utilisé tous ses pinceaux. (VRAI) The artist did use all his brushes. (TRUE)
F.rancis Cornish
The first utterance by Speaker 1 was the same in all four conditions, and introduced the predicate which evoked the targeted referent. In the explicit conditions, there was a second initial utterance which explicitly mentioned the critical argument of the predicate (whether central or peripheral) – always in subject position. The second turn of the dialogues (uttered by Speaker 2) consisted of an utterance that referred back to the target argument from Speaker 1’s turn via a non-subject pronoun. The content of this second utterance oriented the reference towards the target nuclear or peripheral referent. Table 3. Experiment 2 (English): Example Stimuli Explicit antecedent Nuclear
Speaker 1
Speaker 2 (Target)
Peripheral
Implicit Antecedent
Have you noticed that Mark Have you noticed that Mark isn’t shaving? His straggly beard isn’t shaving? makes him look like a tramp. Yes, in fact he’s really allowing it to grow now.
Statement
Mark is sharply reducing the length of his beard. (FALSE)
Mark does seem to be growing a beard. (TRUE)
Speaker 1
Have you noticed that Mark isn’t shaving? His disposable razors have all completely disappeared.
Have you noticed that Mark isn’t shaving?
Speaker 2 (Target) Statement
Yes, he tells everyone he’s thrown them all away. Mark always uses an electric razor for shaving. (FALSE)
Mark has clearly decided to get rid of his razors. (TRUE)
(Cornish et al., 2005: 372)
6. Results The average percentage of incorrectly answered TRUE/FALSE Statements was 9.3% in Experiment 1 (French), and 8.75% in Experiment 2 (English) – with no significant differences in response accuracy as between conditions. This suggests that in the Implicit conditions, Statements referring to peripheral referents did not cause a higher error rate than their nuclear counterparts. The average reading times for the critical second dialogue turn are shown in Figure 1 for the French Experiment 1, and in Figure 2 for the English experiment 2. As is evident from
Indirect pronominal anaphora in English and French
these Figures, the target utterances were read slightly more slowly in the Nuclear Implicit conditions than in the Nuclear Explicit ones, resulting in a statistically non-significant difference. There was no difference (as expected) between the reading times in the two Explicit conditions. But the target utterances in the Peripheral Implicit condition were read much more slowly than either those of the corresponding Explicit condition or of the Nuclear Implicit condition. These factors (Explicitness and Centrality) interact, since as noted, in the Implicit conditions, the peripheral target utterance was read more slowly than when it was explicitly mentioned. 3800 3600 3400
Reading time (in msec)
3200 3000
Referent Centrality Peripheral
2800
Nuclear
2600 2400 2200 2000 1800 Implicit
Explicit
Referent Explicitness
Figure 1. Experiment 1: Reading times for target dialogue turn (Speaker 2): French version (Cornish et al., 2005: 371)
F.rancis Cornish 3800 3600 3400 3200 Reading time (in msec)
3000
Referent Centrality Peripheral Nuclear
2800 2600 2400 2200 2000 1800 Implicit
Explicit
Referent Explicitness
Figure 2. Experiment 2: Reading times for target dialogue turn (Speaker 2): English version (Cornish et al., 2005: 372)
7. Conclusion: Evaluation of the results Our predictions were largely borne out by the results obtained both in French and in English: clearly, (non-subject) pronouns are capable of retrieving an implicit referent, but only on condition that it is ‘nuclear’ and not ‘peripheral’. Experiment 2 replicated the results yielded by Experiment 1, showing that these were not something specific to French anaphor resolution. So reference centrality is indeed conceptual and not purely linguistic in nature. Our results show why there seems to be evidence both FOR and AGAINST the use of unaccented pronouns in indirect anaphora. Using these results regarding the existence of a distinction between central and peripheral indirect referents, in conjunction with Erkü & Gundel’s (1987) examples of associative or complementary anaphora illustrated in examples (4) and (5), we may posit a scale of conceptual centrality for indirect referents, as follows:
nuclear > peripheral > associative/complementary referents
Indirect pronominal anaphora in English and French
This is systematized in Table 4, which includes the correlations noted between the degree of conceptual centrality of the targeted indirect referent, and the indexical expression type(s) that may be used to signal it. Table 4. Three types of indirect referent and their targetability via pronouns and definite lexical NPs Type of Anaphor
3rd person pronoun Definite lexical NP
Type of Referent Nuclear >
Peripheral >
Associative/complementary
+ +
+/– +
– +
1 Associative referents: Ex.(4): “I couldn’t use the box you gave me. The bottom/#it fell out.” 2 Peripheral referents: Exs.(7b) and (8b): “Mary dressed the baby. The clothes/#they were made of pink wool.” 3 Nuclear referents: Ex.(2b): W: “Why didn’t write to me?” M: “I did…, started to, but I always tore’ em (/the letters) up.” Table 4 reveals a situation of partial complementary referential distribution as between the possible uses of 3rd person pronouns and definite lexical NPs (this recalls Levinson’s, 2000 pragmatic account of the use of these two expression types). At the top of the scale, we find the nuclear indirect referents, and at the bottom, the associative/complementary ones, with the middle position being occupied by the peripheral indirect referents. Starting from the bottom of this scale, it is clear, as we have seen, that associative or complementary indirect referents, such as ‘the bottom of the box’ in (4), cannot be accessed via unaccented pronouns, since this referent type is psychologically inactive at the point of retrieval. Only definite NPs with a lexical component can achieve this. But this point requires testing experimentally. Next up the scale are indirect referents which are peripherally involved in setting up a given situation, as means, instrument or habitual accompaniment. This referent type is semi-active and not central within the mental representation of this situation. Such indirect referents can only be retrieved with difficulty via unaccented pronouns, as our results have shown. This difficulty is not as absolute as 5.
See also the experimental results achieved as well as reported on by Lucas et al. (1990).
F.rancis Cornish
with the associative or complementary background referents illustrated in (4) and (5). Examples are ‘the paintbrushes’ used by the artist in the set of French experimental items given in Table 2, or ‘the razors’ used by Mark in the English ones in Table 3. These referents are easily retrievable via definite lexical NPs, but only marginally via pronouns. Finally, at the very top of the scale, implicit nuclear referents are almost as easy to retrieve via non-subject unaccented pronouns as their explicit counterparts are (see the attested examples (2a-c) as illustration). So pronouns and definite lexical NPs are almost equivalent as means of anaphoric access to such an indirect referent type. Now, according to Gundel, Hedberg & Zacharski (1993, 2000), 3rd person pronouns canonically target referents with the most restrictive type of cognitive status targetable by a range of indexical forms. See their Givenness Hierarchy presented as Table 5 below. Table 5. Gundel et al.’s (1993) Givenness Hierarchy in focus it
>
activated that/this this N
>
familiar that N
>
uniquely > referential > type identifiable identifiable the N
indef. this N
aN
(Gundel et al., 1993: 275)
Each pronoun or determiner type is claimed to code one or other of the six cognitive statuses recognized by the scale. “In-focus” at the top is the most restrictive search space for the expression type’s referent, and “type-identifiable” at the bottom, the least. The hierarchy is implicational in character. At any position on the GH, the corresponding lexical item type is said to code that status as well as implying all the lower statuses (those to its right on the hierarchy). Moreover, the use of an expression containing a lexical item type occurring to the right of a given status only implicates that the leftward one doesn’t apply. As such, its use is sufficient to suggest a change of referent. Now, according to the results of our two experiments, unaccented pronouns may not only signal the status “in-focus” of their intended referent, but also the status “activated” (the status immediately below it on the GH, to its right). Yet this possibility is not predicted by the authors: for unstressed pronouns cannot, according to the way in which the GH is claimed to operate, be used to signal a status to the right of “in-focus”. Donna Byron (2000) provides a range of attested as well as constructed examples, a number of which correspond to indirect ana-
6. As confirmed in a personal communication by Nancy Hedberg.
Indirect pronominal anaphora in English and French
phora, where 3rd person pronouns do in fact felicitously retrieve referents bearing the cognitive status “activated” but not “in-focus”. In fact, she argues that pronouns can go all the way down the GH to “type-identifiable” at the bottom right (so long as the semantic content of the anaphoric clause is sufficient to warrant this: see her notion of “semantically enhanced pronouns”). This latter situation would characterize the use of the generic indefinite pronoun they. The next position down the GH is “familiar”, a status said to be coded by the distal demonstrative determiner that in English, as in Do you remember that summer we spent together two years ago?. The demonstrative NP that summer in the example retrieves a representation of the intended referent from shared long-term memory. According to the results of our two experiments, peripheral indirect referents (which are only targetable with difficulty via pronouns) could well correspond to the status “familiar” on the GH. This status constitutes the limit of indirectness of potential referents retrievable via 3rd person pronouns. I would suggest that this is the reason why subjects found it significantly more difficult to resolve the reference of object pronouns in this way in our two experiments. This is incidentally also the cut-off point for topical referents, according to Lambrecht (1994): for to be considered a potential topic, a referent must be both identifiable and activated – something which peripheral indirect referents are clearly not. Unaccented 3rd person pronouns, as is well known, are sensitive to the topical status of their potential referents.
References Blackwell, S. E. 2003. Implicatures in Discourse. The case of Spanish NP anaphora. Amsterdam: John Benjamins. Byron, D. 2000. Semantically Enhanced Pronouns. Ms, Department of Computer Science, University of Rochester, Rochester NY. Consten, M. 2001. Indirect Anaphora and Deixis in Texts: Domain-bound reference and coherence. Poster presentation at the International Workshop on Reference, Utrecht/NL, 11th January 2001. Consten, M. 2004. Anaphorisch oder deiktisch? Zu einem integrativen Modell domänengebundener Referenz. Tübingen: Max Niemeyer. Cornish, F. 1999. Anaphora, Discourse and Understanding. Evidence from English and French. Oxford: Clarendon Press. Cornish, F. 2005. Degrees of indirectness: Two types of implicit referents and their retrieval via unaccented pronouns. In Anaphora Processing: Linguistic, cognitive and computational
7. And which she classifies as having the status “activated”, due to the inference necessarily involved.
F.rancis Cornish modelling [Current Issues in Linguistic Theory 263], A. Branco, T. McEnery and R. Mitkov (eds), 199–220. Amsterdam: John Benjamins. Cornish, F., Garnham, A., Cowles, H. W., Fossard, M. and André, V. 2005. Indirect anaphora in English and French: A cross-linguistic study of pronoun resolution. Journal of Memory and Language 52(3): 363–376. Cote, S. 1998. Ranking forward-looking centers. In Centering Theory in Discourse, M. A. Walker, A. K. Joshi and E. F. Prince (eds), 55–69. Oxford: Clarendon Press. Dik, S. C. 1978. Stepwise Lexical Decomposition. Lisse: Peter de Ridder Press. Erkü, F. and Gundel, J. K. 1987. The pragmatics of indirect anaphors. In The Pragmatic Perspective: Selected papers from the 1985 International Pragmatics Conference, Amsterdam 1997, J. Verschueren and M. Bertuccelli-Papi (eds), 533–545. Amsterdam: John Benjamins. Greene, S.B., Gerrig, R., McKoon, G. & Ratcliff, R. 1994. Unheralded pronouns and management by common ground. Journal of Memory and Language 33: 511–526. Gundel, J. K., Hedberg, N. and Zacharski, R. 1993. Cognitive status and the form of referring expressions in discourse. Language 69(2): 274–307. Gundel, J. K., Hedberg, N. and Zacharski, R. 2000. Statut cognitif et forme des anaphoriques indirects. Verbum XXII(1): 79–102. Levinson, S.C. 2000. Presumptive Meanings. The theory of Generalized Conversational Implicature. Cambridge MA: The MIT Press. Lucas, M. M., Tanenhaus, M. K. and Carlson, G. N. 1990. Levels of representation in the interpretation of anaphoric reference and instrument inference. Memory & Cognition 18(6): 611–631. Mauner, G. and Tanenhaus, M. K. 1995. Implicit arguments in sentence processing. Journal of Memory and Language 34: 357–382. Mauner, G., Melinger, A., Koenig, J.-P. and Bienvenue, B. 2002. When is participant information encoded? Evidence from eye-monitoring. Journal of Memory and Language 47: 386–406. Reichler-Béguelin, M.-J. 1993. Anaphores associatives non-lexicales: Incomplétude macrosyntaxique? In Complétude et incomplétude dans les langues romanes et slaves. Actes du V1è colloque international de linguistique romane et slave, Cracow, 29 Sept.- 3 Oct. 1991, S. Karolak and T. Muryn (eds), 327–379. Sanford, A. J. and Garrod, S. C. 1981. Understanding Written Language. Explorations in comprehension beyond the sentence. New York: John Wiley. Sanford, A. J., Garrod, S. C., Lucas, A. and Henderson, R. 1983. Pronouns without explicit antecedents? Journal of Semantics 2: 303–318. Schwarz, M. 2000. Indirekte Anaphern in Texten. Studien zur domänengebunden Referenz und Kohärenz im Deutschen. Tübingen: Max Niemeyer. Walker, C.H. and Yekovitch, F.R. 1987. Activation and use of script-based antecedents in anaphoric reference. Journal of Memory and Language 26: 673–691. Yule, G. 1979. Pragmatically-controlled anaphora. Lingua 49: 127–135. Ziv, Y. 1996. Pronominal reference to inferred antecedents. Belgian Journal of Linguistics 10: 55–67.
Lexical anaphors in Danish and French Lita Lundquist The article treats lexically varied noun-phrase anaphors, so-called ‘unfaithful anaphors’, which are found more frequently in French than in Danish texts. The difference is related to more fundamental differences between the two languages, and types of language, at different linguistic levels. Thus, French lexicalises more semantic content in nouns than in verbs, and vice versa in Danish, the former language preferring exocentric ordering at the syntactic level compared to Danish endocentric ordering. Much of the information of unfaithful anaphors being new, French can be said to communicate presupposed information, where Danish ‘poses’, i.e. states new information. And finally, in French the semantic content of an unfaithful anaphor often entertains a rhetorical relation with its host predicate, where Danish sets the relation explicitly.
1. Introduction The present study focuses on lexical anaphors, i.e. anaphors consisting of NPs with a lexical head, the so-called head-noun nominal anaphors. It further narrows its scope to the type of anaphoric expression which, while lexically varied, still entertains a co-referential relation to the antecedent, as exemplified in (1)–(3):
(1) Luc Ferry1 pour une autonomie des universités. […] L’annonce de cette réforme […] tombe à pic pour redonner du crédit à un ministre accusé de préférer le verbe à l’action. En un an, le philosophe de la rue de Grenelle1 n’est pas parvenu à s’imposer.
i. ‘Luc Ferry1 is for more autonomy for the universities. […] The announcement of this reform is welcome in order to give more credit to a minister who is accused of preferring the word for action. In one year, the philosopher from la rue de Grenelle1, has not succeeded in imposing himself.’
Lita Lundquist
(2) “Suivez-moi, je ne vais pas vous tromper”, avait promis Carlos Menem1 en arrivant au pouvoir en juillet 1989. Le caudillo péroniste de La Rioja (Nord)1 avait pourtant très vite troqué ses promesses populistes…
i. ‘“Follow me, I am not going to cheat on you”’ promised Carlos Menem1, when he1 took power in July 1989. Yet, the peroniste caudillo from La Rioja (North)1 quickly changed his1 populistic promises…’
(3) Qui est Nestor Kirchner1? Le gouverneur de la lointaine et riche province pétrolière de Santa Cruz (Patagonie) 1 reste un mystère.
i. ‘Who is Nestor Kirchner1? The governor from the remote and rich oil province in Santa Cruz (Patagonia) 1 remains a mystery.’
The study compares the use of such anaphors in two different languages, Danish and French, showing first and foremost – in line with other studies comparing lexical anaphors in Germanic and romance languages – that the use of these distinct head-noun nominal anaphors is especially frequent in French and much less so in Danish, which in contrast shows a predilection for identical head-noun or pronominal anaphors. The investigation, however, brought to light other interesting and quite unexpected differences, which, as I account for below, can be related to other more fundamental differences between the two languages – and two language types. Indeed, dissimilarities in the way the two languages lexicalise nouns and verbs, syntactically order information and express rhetorical relations in texts all contribute to the different patterns in using lexical anaphors. For this reason I propose that variation in the use of lexical anaphors in Danish and French can be traced back to typological differences between the two languages.
2. “Unfaithful anaphors” The use of lexically varied anaphors in French compared to identical anaphors in Danish is illustrated by (4) and (5) below. The examples stem from two texts treating the same topic: a French politician, Dominique Ambiel, who was caught in fragrant délit with a child prostitute.
(4) Ambiel1 er ikke hvem som helst. Han1 var kommunikationsdirektør – med andre ord spindoktor – for selveste regeringschefen, Jean-Pierre Raffarin […] Han1 var Raffarins nærmeste og mest fortrolige rådgiver… (Information, April 30, 2004)
i. ‘Ambiel1 is a name to reckon with. He1 was Director of Communications – in other words, spindoctor – for the Prime Minister himself, Jean-Pierre
Lexical anaphors in Danish and French
Raffarin […] He1 was Raffarin’s closest and most confidential adviser, a very influential person…’
(5) Après deux années passées auprès du premier ministre, dont il1 était le très fidèle conseiller en communication, l’ancien producteur de télévision Dominique Ambiel1 teste bien malgré lui la « télé-réalité ». Soupçonné d’avoir eu recours à la prostitution d’une mineure, ce fidèle de Jean-Pierre Raffarin1 est obligé d’interrompre immédiatement sa carrière dans les coulisses du pouvoir. Après les téléfilms historiques […] et enfin le pouvoir, cet infatigable homme de l’ombre1 est soudainement placé sous les projecteurs des faits divers… (Le Monde, April 30, 2004)
i. ‘After two years with the Prime Minister whose loyal counsellor he1 was, the former TV producer Dominique Ambiel1 himself is now testing a “reality show”. Accused of having sought recourse to a child prostitute, this faithful friend of Jean-Pierre Raffarin1 is forced to immediately interrupt his career behind the political scene. After historical TV films […] and at last power, this untirering man from the shadow1 is all of a sudden placed in the limelight…’
The Danish text uses only pronouns or proper nouns to refer to the discourse referent, whereas the French uses only distinct head-noun anaphors in order to maintain the co-referential chain. These lexically varied anaphors in French are preceded either by the demonstrative determiner as in (5) ce fidèle de Jean-Pierre Raffarin, cet infatigable homme de l’ombre, or by the definite determiner as shown in examples (1)–(3). Though there is a difference between the use of definite descriptions and demonstrative descriptions (Corblin 1987, LePesant 1998), I shall treat them here under one heading. The type of anaphoric expression used in the French texts is characterised by two features: first, they may contain – and introduce – rather heavy semantic information via their modifiers and complements, and second, the definite description – being an autonomous referential expression as compared to the pronouns used in the Danish text – could semantically refer to and introduce another discourse referent. Thus, in the following (constructed) text sequence, the underlined referring expressions may be co-referring, just as they may be referentially disjoint, i.e., referring to two different discourse referents (the latter being the case):
(6) Dominique de Villepin a présenté, mercredi, le programme de stabilité des finances publiques françaises. Le ministre de l’Intérieur a fixé jeudi un objectif de baisse de 3% de la délinquance en 2006. Voilà quelques titres tirés du journal Le Monde le 12 janvier 2006.
Lita Lundquist
i. ‘Wednesday Dominique de Villepin presented the stability programme for the French public finances. Thursday, the Minister of Internal Affairs set a goal of reducing delinquency with 3% in 2006. These are some of titles of the French newspaper Le Monde from January 12, 2006´.
This potential referential ambiguity of the definite description has given way to its name in French: “anaphore infidèle” (Lundquist 1990, 2005a, 2005b). As there is no canonical term for this phenomenon in English, we shall here use the name of ‘unfaithful anaphor’, in spite of its inopportune sexual connotations, albeit these are also present in the French term. It has been made known, in other contrastive studies, that romance languages such as French and Italian seem to prefer the use of unfaithful anaphors to maintain co-referential chains, compared to the preference of Germanic languages such as Danish for lexical repetitions or pronominal anaphors (Korzen & Lundquist 2003). This difference has often been ascribed to rhetorical norms and stylistic traditions that prevail in the two types of cultures, but my contention is that the reason may also be found in more fundamental differences between the two language families when it comes to basic patterns within lexicalisation and syntax.
3. Lexicalisation patterns in Danish and French Let us start with the lexicalisation of nouns in Danish and French and with the fact we can observe in the examples above, namely that unfaithful anaphors in French might carry a dense load of semantic information. It has been shown that romance languages, as a general rule, show a tendency to lexicalise more semantic features in nouns than Germanic languages. Or, in other words, that romance languages lexicalise nouns at a more concrete and subordinate level compared to the more abstract, semantically void nouns in Germanic languages. An example of this is given in table 1, in which for each noun in Danish (‘jug’, ‘chair’, ‘track’ respectively) there is a whole series of corresponding nouns in French: Table 1. Lexicalisation of nouns and verbs in Danish and French Kande/’jug’
{cruche, broc, pichet, pot …}
Stol/’chair’ Bane/’track’
{chaise, fauteuil, stalle, chaire, tribune …} {court, piste, champ, terrain, couloir …}
Mettre/’put’
{lægge, stille, sætte, putte …}
1.
A hyperonymic term is not easy to find in English either.
Lexical anaphors in Danish and French
Whereas Danish lexicalises the function only [something to sit on] via stol (‘chair’) at a hyperonymic level, French lexicalises more semantic features specifying the type, constitution and form of the object (Pustejovsky 1995); a cruche/ ’pitcher’, e.g., is big, has two handles and is made of earthenware. French nouns are more specified than Danish nouns in the context of nouns referring to artefacts. The difference in level of lexicalisation of nouns has a curious effect on the translation between Danish and French of associative anaphors, in that the translation into French imposes a choice between the French co-hyponyms (Lundquist 2000, 2003), but it also has an impact, I maintain, on the use of unfaithful anaphors in the two languages. If we now turn to the lexicalisation of verbs in the two languages, we observe that Danish verbs are more precise compared to the more abstract verbs in French. This is especially noticeable in verbs of movement which obligatorily lexicalise the semantic feature of manner. Thus Danish has no equivalent to the French aller/ English go, as Danish necessarily has to specify the manner of aller, i.e. gå (D: gå/ walk > F: aller/go). It is also true for verbs in the paradigm of mettre/’put’ as in table 1, which has to be specified in Danish in accordance with the object in question. Thus mettre un livre/un vase sur la table (‘to put a book/a vase on the table’) would correspond to at lægge en bog på bordet (‘to lay a book on the table’) and at sætte en vase på bordet (‘to sit a vase on the table’) respectively.
4. Endocentric and exocentric ordering of Danish and French sentences This general difference in lexicalisation patterns with Danish lexicalising more semantic features in verbs, and French more features in nouns, has an effect on syntax too, since it yields different sentence patterns in Danish and French, as explained below: …whereas the Danish clause appears as a concrete relation (verbs are precise) between abstract entities, the French clause is an abstract relation (verbs are general) between concrete entities. In this sense Danish and French belong to two different linguistic types: endocentric languages, where the information is concentrated in the verb, the centre of the clause, and exocentric languages, where the information rather than being localised in the clause-centre is distributed onto the nouns. (Herslund & Baron 2003: 31).
The tendency observed within the two languages to concentrate information in the sentence arguments in French and in the sentence verb in Danish is shown in figure 1.
Lita Lundquist
Danish: noun
VERB
noun
endocentric language
French : NOUN
verb
NOUN
exocentric language
Figure 1. Sentence patterns in Danish and French
The predilection for semantically dense nouns in exocentric French may facilitate the use as in (1)–(3) and (5) of rich and compacted nominal anaphors in French texts. For Danish, on the other hand, the endocentric character, i.e. the centre being the verb of the sentence, draws or pulls so to speak the information from the arguments to the verb, which might explain the semantically light anaphoric expressions in Danish. In fact, a close comparison of parallel texts shows that semantic content encoded in unfaithful anaphors in French is often expressed in the finite, tensed, verb of a full copula predicate in the parallel Danish text. This is the case in the examples in table 2 below, where the content expressed in unfaithful anaphors in French is expressed more or less equivalently in Danish via full copula expressions which explicitly link the subject to a predicate. Table 2. Reduced predications in French and complete predications and Danish French – “unfaithful” anaphors l’ancien producteur de télévision Dominique Ambiel ’the former TV producer’
ce fidèle de Jean-Pierre Raffarin ‘this faithful friend of JPR’ cet infatigable homme de l’ombre ’this untiring man from the shadow’
Danish – copula (like) constructions
Ambiel der blev stenrig som ejer af flere tv produktionsselskaber ‘Ambiel, who became extremely wealthy as the owner of several televison production companies’ han var kendt som Raffarins nærmeste rådgiver, ‘he was known as Raffarin’s closest and most confidential adviser’ (Ambiel)… er kendt som en magtglad person ‘(Ambiel)… is known to lust for power’
In other words, what is expressed in these examples via a finite verb predicate in Danish is condensed in unfaithful anaphors in French. Or vice versa, the information presented in French in an unfaithful anaphor is presented in Danish in a primary predicate of a full proposition. Hence, an unfaithful anaphor can in this re-
Lexical anaphors in Danish and French
spect be regarded as a predicational component of a secondary rank, presenting information in a reduced proposition. This contrast between the use of full propositions in Danish and reduced propositions in French is found in many other linguistic constructions in the two languages. One of the most prevailing is the existence and use in French of non-finite verb phrases such as the gérondif, and present and past participles, which present information in a subordinate syntactic ordering. An example of such a subordinate, hypotactic ordering of information is expressed via a past participle in (5’): (5’) Soupçonné d’avoir eu recours à la prostitution d’une mineure, ce fidèle de Jean-Pierre Raffarin est obligé d’interrompre immédiatement sa carrière. ‘Accused of having sought recourse to a child prostitute, this faithful friend of Jean-Pierre Raffarin is forced to immediately interrupt his career behind the political scene.’
Danish does not have a gérondif, and Danish does not use participles as a device to grammaticalise subordinate sentence structure to the same degree as French. Danish, as a general order, seems to prefer paratactic sentence ordering via finite verbs as we have seen above. This means, at yet another level of linguistic description, that information is presented differently in the two language types (Korzen 2003).
5. Presenting information in Danish and French In French, unfaithful anaphors may – as we have seen – present not only a large amount of semantic information, but also information which is new. This contrasts with the fact that unfaithful anaphors often occupy the role of grammatical subject, in the first position of the sentence. Linking up with the preceding sentence, the first position constitutes the backward looking centre of the sentence, generally encoding for old or known information. Unfaithful anaphors thus have the peculiarity of presenting new information as known, or at least as presupposed knowledge. Let us return to example (1) and the expression, le philosophe de la rue de Grenelle/’the philosopher from la rue de Grenelle’, which is supposed to co-refer to Luc Ferry. Choosing to express an anaphoric relation by this particular definite description, the sender presupposes that the reader knows – or can infer – that Luc Ferry is (or rather was) a philosopher in la Rue de Grenelle (a metonymy for the ministry of education). Danish, in contrast, would in similar cases prefer to present such new information in full-blown finite predicates, as those shown in table 2, and in so doing, Danish would assert the information instead of presenting it as presupposed. It is my contention that the use of unfaithful anaphors in French and of faithful ana-
Lita Lundquist
phors in Danish reveals two different ways of presenting information, two different tactics which can be subsumed by the concept presupposition as opposed to asserting, presupposition prevailing in French as opposed to assertion in Danish.
6. Contract of communication This brings me to the concept of epistemic contract, or contract of communication, introduced by Givon (1984). In fact, the use of unfaithful as opposed to faithful anaphors, in my opinion, does have an effect even at this more general level of communication. Givon describes the difference between assertion and presupposition in the epistemic contract in the following terms: Assertion of P: hearer is unfamiliar with P; P is open to challenge by the hearer and speaker is more willing to tolerate challenge from hearer. Presupposition of P: hearer is familiar with P; P is uncontested knowledge; P is not open to challenge by hearer, and speaker is less willing to tolerate challenge from hearer (Givón 1984: 252 ff).
Continuing this line of thought, a conclusion could be that a typical French way of ordering information is via presupposing constructions, such as unfaithful anaphors and non-finite, subordinate clauses, and that this strategy imposes the information as uncontestable and non-negotiable for the hearer. Danish, in contrast, via its assertive way of presenting information and its paratactic way of ordering information in texts leaves more space for negotiating and evaluating what is presented as new knowledge, and also for successively integrating new knowledge into old. I shall not continue this line of thought here as I do not want to risk overgeneralisation. Rather I shall turn to another aspect of presenting information and creating relations between pieces of information in respect to which unfaithful anaphors seem to play a major role.
7. Unfaithful anaphors and rhetorical relations We shall now address the so-called rhetorical relations, which can be either expli citly encoded between propositions in texts, e.g. via connectors, or implicitly inferred. I shall demonstrate below how also unfaithful anaphors can call rhetorical relations into being. Let us briefly recall the theoretical framework of rhetorical relations advocated by Mathiessen and Thompson (1988). They suggest that relations between textual units can be usefully viewed as a relation between a nucleus
Lexical anaphors in Danish and French
and one or more satellites. The nucleus is the textually central part, whereas the satellites are circumstantial expansions such as explanations and causes. Returning to our examples in table 2, reproduced in table 3 below, we observe that the unfaithful anaphors (shown in italics on the left) create a rhetorical relation with their respective predications (on the right): Table 3. Unfaithful anaphors and rhetorical relations Satellite
Nucleus
l’ancien producteur de télévision D. Ambiel ce fidèle de Jean-Pierre Raffarin
teste bien malgré lui la « télé-réalité ». est obligé d’interrompre immédiatement sa carrière dans les coulisses du pouvoir.
cet infatigable homme de l’ombre
est soudainement placé sous les projecteurs des faits divers…
‘the former TV producer Dominique Ambiel ’‘this faithful friend of Jean-Pierre Raffarin
himself is now testing a “reality show” is forced to immediately interrupt his career behind the political scene’
‘this untiring man from the shadow
is all of a sudden placed in the limelight’
The rhetorical relations can be determined on the basis of the predicational component contained in the anaphoric expressions on the one hand and the semantic content of the primary predicate on the other. In all three cases in the table above, the rhetorical relation is one of opposition: the opposition between Ambiel in his former role, and in his present situation. In example (1), repeated below, the relation can be interpreted as one of ‘cause’ or explanation: Luc Ferry has had no success as a minister of education because he is a philosopher: Satellite Nucleus (1’) le philosophe de la rue de Grenelle n’est par parvenu à s’imposer ‘the philosopher from la rue de Grenelle has not succeeded in imposing himself’
I think that examples such as those above that abound in French texts in general give us good reason to regard unfaithful anaphors as a kind of grammaticalisation of textual relations, on a par with those found in non-finite verb forms which also introduce some kind of rhetorical relation between a satellite and nucleus (Herslund 1998). This was shown in example (5’): soupconné d’avoir eu recource à la prostitution d’une mineure, …/’Accused of having sought recourse to a child prostitute…’ I am not using the concept grammaticalisation here in its core theoretical sense as defined by e.g. Hopper and Traugott:
Lita Lundquist
When a content word assumes the grammatical characteristics of a function word, the form is said to be ‘grammaticalized’. (Hopper and Traugott 1993: 4).
“Grammaticalisation” is used here as a description of the phenomenon that a lexical anaphor in the role of grammatical subject may, via its semantic content, create a rhetorical function with respect to its predicate. If this is so, it can even be claimed that unfaithful anaphors constitute the ultimate form of grammaticalisation of rhetorical relations, since the rhetorical relation is established between the subject and the predicate.
8. Conclusion Let me wind up by admitting that I have not treated the lexical anaphor in itself in any detail; among the many aspects which I have omitted is the difference in anaphoric and textual function between lexical anaphors preceded by a demonstrative determiner as compared to a definite determiner. My aim here has instead been to draw attention to the relation between the use of unfaithful anaphors and other linguistic characteristics of the two languages in question, Danish and French. Thus it seems plausible to link the frequent use of unfaithful anaphors in French to its general pattern of lexicalising specifying semantic features in nouns, to its syntactic distribution of information onto the noun phrases of the clause, and to its preference for condensing information in reduced, hypotactic propositions. The preference in Danish for pronominal or repeated lexical anaphors may, on the other hand, be related to its pattern of lexicalising only few and very abstract semantic features in nouns, to its distribution of informative weight onto the verb of the clause, and to its general tendency to present information in paratactic propositions with finite verbs. I have also pointed out the more overarching effects that unfaithful anaphors have on the rhetorical structuring of texts and on the communicative contract between text producer and text receiver in the two language communities.
References Corblin, F. 1987. Indéfini, défini et démonstratif. Constructions linguistiques de la reference. Genève: Droz. Givòn, T. 1984. Syntax. A functional-typological introduction, Vol. I. Amsterdam: John Benjamins. Herslund, M. 1998. Grammaticalisation of textual relations. Causal structures in French. In Clause Combining and Text Structure [Copenhagen Studies in Language 22], I. Korzen and M. Herslund (eds), 9–21. København: Samfundslitteratur.
Lexical anaphors in Danish and French
Herslund, M. and I. Baron. 2003. Language as world view. Endocentric and exocentric representations of reality. In Language and Cultures. [Copenhagen Studies in Language 29], I. Baron (ed.), 29–43. København: Samfundslitteratur. Hopper, P. J. and Traugott, E. 1993. Grammaticalization. Cambridge: CUP. Korzen, I. 2003. Hierarchy vs. linearity. Some considerations on the relation between context and text with evidence from Italian and Danish. In Language and Cultur,. [Copenhagen Studies in Language 29], I. Baron (ed.), 97–110. København: Samfundslitteratur. Korzen, I. and L. Lundquist (eds.). 2003. Sprogtypologi og oversættelse. Endocentriske og exocentriske sprog. HHK: Samfundslitteratur. Le Pesant, D. 1998. Utilisation des propriétés des anaphores dans la définition de certaines relations lexicales. Internet. Lundquist, L. 1990. Un cas d’ambiguïté référentielle. Aspects pragmatiques. In L’anaphore et ses domaines, G. Kleiber and J.-E. Tyvaert (eds), 229–249. Paris: Klincksieck. Lundquist, L. 2000. Translating associative anaphors. A linguistic and psycholinguistic study of translation from Danish into French. In Argomenti per una linguistica della traduzione, I Korzen and C. Marello (eds). 111–129. Alessandria: Ed. Dell’Orso. Lundquist, L. 2003. L’anaphore associative en danois et en français – sur quoi roule-t-elle? Etude contrastive et expérimentale. In Aspects linguistiques de la traduction, M. Herslund (ed.), 105–124. Bordeaux : Presses universitaires de Bordeaux. Lundquist, L. 2005a. Noms, verbes et anaphores infidèles. Pourquoi les Danois sont plus fidèles que les Français. Langue française 145. 73–91. Lundquist, L. 2005b. L’anaphore infidèle. Regard(s) unificateur(s) sur le texte? In L’unité texte: Actes du colloque: ‘Regards croisés sur l’Unité texte / Conjoint Perspectives on Text’, S. Porheil et D. Klinger (ed.). Pleyben: Association Perspectives. Matthiessen, C. and Thompson, S. 1988. The structure of discourse and ‘subordination’. In Clause Combining in Grammar and Discourse, J. Haiman and S. Thompson (eds), 275–329. Amsterdam: John Benjamins. Pustejovsky, J. 1995. The Generative Lexicon. Cambridge MA: The MIT Press.
Referential collaboration with computers Do we treat computer addressees like humans? Alfons Maes, Pascal Marcelis and Frank Verheyen “Is the referential behavior of language users sensitive to the human vs. computer nature of the addressee? In a production experiment participants were asked to help a distant partner to solve a problem by identifying objects on a computer screen. The addressee was either a distant human or a distant computer who/ which was in need of the information. The addressees either provided feedback or not. The results show that human addressees trigger more overspecified referential expressions than computer addressees in a non-feedback situation. In the feedback conditions, the nature of the addressee had no effect. The results suggest differences between humans and computers, not only in terms of users’ evaluations, attitudes and motivation, but also in terms of assumptions on cognitive capacities.”
1. Introduction Humans have always been inventive in finding ways to communicate with other humans on a spatial or temporal distance. A large number of communication tools, such as Morse signs, fax, telephone or letters show how important distant communication is for humans. Recently, computers gradually take over this communicative function: digital technology enables humans to bridge time and space and creates ever more variants of computer mediated communication (CMC), i.e., text or speech based, human-human or human-computer communication enabled or mediated by computers. Typical of computers is that they gradually turn communication into a multimodal experience. They enable humans to enjoy or simulate each other’s co-presence and to represent or simulate a shared workspace on the screen. Chat programs for example enable more and more the combination of text, speech and streaming video; learning environments introduce embodied agents to simulate
Alfons Maes, Pascal Marcelis and Frank Verheyen
face-to-face interaction with learners; and collaborative software creates sophisticated shared workspaces for participants, entailing visual or animated representations of objects which are relevant in the communicative interaction. The technological evolution develops the computer not only into a sophisticated multimodal and interactive communication medium, but gradually also into a knowledgeable communication partner. The question which will be answered in this article is how communication with computers changes the linguistic behavior of language users, thereby focusing on one particular aspect of linguistic behavior, i.e., the way in which humans refer to objects relevant to the communicative exchange. In interactive communication, the referential behavior of humans is an excellent and sensitive measure of the communicative conditions applicable: the presence or absence of the addressee and the objects to be identified determines whether and how we use language and other modalities to create successful reference. Our referential behavior reflects finegrained assumptions we have about the communicative conditions applicable to producer and addressee, about the (assumed) relationship or common ground between participants and the knowledge and referential needs of the addressee. In this article, we focus on the referential behavior of humans in distant communication mediated by computers. We want to know whether language users are sensitive to the human vs. computer nature of the addressee. Is there a difference in the way in which they conceptualize objects for an assumed human or computer addressee? Do they use different referential ingredients or a different number of ingredients to ‘satisfy’ a human or computer addressee? In a production experiment participants were asked to help a distant partner to solve a problem by identifying objects on a computer screen. The addressee was either a distant human or a distant computer who/which was in need of the information. The addressees either provided feedback or not. The results show that human addressees trigger more overspecification than computer addressees in a non-feedback situation. In the feedback conditions, the assumed nature of the addressee had no effect. The results suggest differences between humans and computers, not only in terms of users’ evaluations, attitudes and motivation, but also in terms of assumptions on cognitive capacities and establishing common ground. In the next sections, relevant research is discussed with respect to the computer as a communicative partner and it is discussed why referential expressions are a suitable measure of the language user’s assessment of the (computer vs. human) nature of the addressee.
Referential collaboration with computers
2. Computers as communicative partners If we think of computers as communication partners, there is a difference to make between communication via and with computers. If we use a computer, we consider it in the first place as a more or less intelligible interactive medium via which we perform communicative tasks. We react to questions like do you really want to close this file?, we receive a beep tone indicating that we tried an illegal action, or we politely follow the request of the little Word dog or paper clip to type in our help question in the predesigned frame presented on the screen. In these cases, we only communicate via the computer, as these interactions are instrumental and subordinate to the interaction we have with the real addressees, i.e., the readers of our text, the chat or e-mail partner etc. But gradually we also use computers as addressees to communicate with. For example, we can call automated help desks to know which train to take or we can communicate with a virtual agent in a learning environment, knowing very well that on the one hand this application is ultimately man made, but also that this particular communication exchange is not directly supervised by humans. Most of the everyday natural language processing and generation systems, as well as speech recognition and synthesis systems and dialogue managers have a fairly restrictive scope, and suffer from a lack of flexibility to simulate real natural language. But the number of virtual applications is growing as well as the flexibility of the systems. This evolution certainly requires technological progress in different fields, but also more ‘simple’ knowledge about how humans in different situations produce and process natural language, how they combine linguistic and audiovisual signals to produce natural messages, and also, how a ‘virtual partner’ affects the linguistic behavior of humans. There is reason to believe that computer mediation changes the linguistic behavior of humans. As a technological medium, the computer comes together with a multilayered and multisensorial overhead or interface with its own contingencies and constraints, which undoubtedly results in new types of linguistic interaction. For example, computer writing (like chat or email) is more interactive and direct than traditional writing, because of the rapid exchange and feedback options provided. But at the same time, the computer interface makes computer communication less interactive than real human-human communication, for example because of synchronization or sensorial coordination problems. Hancock and Dunham (2001) demonstrate for example that one simple coordination device in chat, i.e., a turn marker, has a pervasive influence on the linguistic behavior of chat participants in an instructive task. Apart from these situational or ergonomic influences, one may expect humans to ‘look’ differently to computers as compared to human partners. A unique characteristic of humans is that they are able to understand the ‘other’ as an inten-
Alfons Maes, Pascal Marcelis and Frank Verheyen
tional agent (Tomasello & Rakoczy, 2003). This capacity is the basis of understanding intentions, beliefs and perspectives of others, which turn is a prerequisite for participating in cultural and communicative activities. This capacity also explains that humans adapt their communicative behavior to what they think their addressees know and need. But what happens when instead of communicating with real humans, we interact and communicate with computer addressees? Do we consider them as normal intentional agents? Do we display the same or similar psycho-social reactions as what we do vis-à-vis real humans? Do we adapt our linguistic behavior on the basis of different assumptions we have about computers, or different ways to establish common ground? As far as we know, research into the effect of computer mediation thus far hardly focused on the linguistic consequences of communicating with the computer, but rather on the socio-psychological influence and the way it changes the organizational dynamics of communication (e.g., Herring, 1996; Newlands, Anderson, & Mullin, 2003; Walther, Anderson, & Park, 1994; Walther & Burgoon, 1992). Worth mentioning in this respect is the work of Reeves and Nass and their colleagues (Nass, Fogg, & Moon, 1996; Nass & Lee, 2001; Reeves & Nass, 1996). They show how the computer as a medium and as a communication partner affects social responses related to communication and communicative partners. In a series of experiments Reeves and Nass (1996) show that computers as a medium evoke social responses which are highly similar to the social behavior displayed in human interaction. For example, they found that humans not only tend to be polite and careful vis-à-vis other humans, but also vis-à-vis computers. They asked participants for example to execute a task on a computer. Afterwards they had to fill in an online evaluation form about the task and about the way the computer had helped them to execute the task. Half of the participants had to fill in the form on the same computer, the other half had to move to another computer. Participants in the same computer condition were significantly more positive about the computer than participants in the other computer condition. They judged the computer as more competent, nice, honest and helpful, which reflects a politeness behavior which is known from human-human interaction. Similar results were obtained with flattery, humor and self criticism. More recently, Lee and Nass (2002) asked participants to interact on social dilemma’s with humans (represented on the screen as text, stick figures or real characters) or with computer agents. In general, and as expected, the social pressure to conform to norms of others appeared to be stronger vis-à-vis humans than vis-à-vis computer agents. Moreover, participants attributed greater competence, social attractiveness, and trustworthiness to partners represented by anthropomorphic characters than those represented by textboxes or stick figures. This may mean that humans treat computers differently, attribute them other values than humans.
Referential collaboration with computers
Typical of these studies is that they are primarily interested in social and psychological, instead of linguistic reactions. The dependent variables used in these studies are scores on evaluative questions. This way, it remains unclear if and how humans adapt their normal linguistic behavior to the nature of the addressee. Furthermore, the communicative cases used in Lee and Nass (2002) are not really the first area of communicative tasks we want computer addressees to be experts in. More relevant is that computers are able to interact in normal referential communication, for example in giving instructions on everyday tasks and actions. Therefore, we will focus on referential communication in this article and confine ourselves on one typical and important ingredient of it, i.e., the production of referential expressions.
3. Referential expressions and communicative conditions Interacting via and with computers apparently results in interesting social behavior, responses and attitudes between communicative partners. It is less clear, however, if and how computer mediation influences linguistic behavior as well. An important aspect of linguistic behavior is the way in which humans refer to entities which are relevant in the communicative setting and thus the way in which they establish referential coherence. Referring to objects and establishing referential coherence is a matter of producing and interpreting adequate referring expressions, like proper nouns (Pieter), definite or indefinite descriptions (the pope, a large green bag), demonstratives (this or this project) or pronouns (it or she). Research in the field of cognitive psychology, computer linguistics and discourse studies makes it clear that the adequacy of referential expressions is highly dependent on assessments and assumptions on the part of the producer about the physical and mental conditions applicable in the communicative setting. This results in language users being sensitive to these conditions by displaying sophisticated referential strategies when they produce and interpret referential expressions. Speakers and hearers make a ‘contract’ with each other about how to conceptualize objects (Metzing & Brennan, 2003), or which entities to consider given (e.g., he or the boy) or new (e.g., a boy) (Clark, 1996; Clark & Haviland, 1977). They also negotiate and collaborate on the construction of reference in interaction (Clark & Wilkes-Gibbs, 1986). Writers and readers know that referential expressions in unilateral communication express a particular degree of mental activation or accessibility of entities (e.g., Ariel, 1994; Gundel, Hedberg, & Zacharski, 1993). Thus, by using an indefinite noun phrase (e.g., a boy) the writer by default indicates that the entity is not accessible at that point in the discourse, whereas the use of a lean expression like it ‘means’ that the
Alfons Maes, Pascal Marcelis and Frank Verheyen
entity is highly accessible and focal, an instruction which guides the search for the intended entity. In the research literature on reference, a large number of conditions and communicative variables have proven to be relevant in guiding referential behavior and shaping referential expressions. Most intrusive are the perceptual or medium richness conditions applicable, basically determined by the degree in which participants share each other’s presence and share a common visible situation or workspace. In medium rich, face-to-face interaction speakers tend to use referentially underspecified or ambiguous expressions knowing that the addressee can and will interrupt when needed (Clark & Wilkes-Gibbs, 1986). Moreover, the speaker’s presence (visually or auditive) also enables users to use a gamut of audiovisual tools in referring to objects, like intonation, gestures and facial expressions (Beun & Cremers, 1998; Clark & Krych, 2004; Van der Sluis & Krahmer, 2004). These referential strategies clearly differ from non face-to-face communicative situations, and situations without feedback facilities, in which speakers or writers often tend to use more linguistic material than needed to identify objects, because they feel distantly responsible (e.g., Maes, Arts, & Noordman, 2004; Pechmann, 1989). The role of a joint situation or workspace is equally important for reference. The well-known dichotomy between deictic and non-deictic referential expressions is based on the situational availability of things to be referred to. Research within the visual world paradigm shows how sophisticated language users make use of shared knowledge of the environment in producing and interpreting referential expressions (e.g., Barr & Keysar, 2002; Hanna & Tanenhaus, 2004). Research using the map task paradigm (Anderson et al., 1991) makes it clear how the referential behavior changes depending on whether both participants have a workspace which is either identical or incongruent (in this case a geographical map with either the same or different landmarks on it). Similarly, production experiments in Maes et al. (2004) showed that writers used completely different referential strategies depending on whether they assumed their distant addressees to have either a functionally and perceptually identical workspace or a functionally identical and perceptually different workspace. Still, these ‘visible’ communicative conditions, which can easily be experimentally manipulated and investigated, are only the top of the iceberg when it comes to establishing referential coherence. In producing referential expressions, humans make much more assumptions on ‘invisible’ communicative conditions. Referential expressions indicate many different aspects of the assumed common ground between communicative partner (familiarity with the entity, familiarity with the addressee, hierarchical relations between partners) as well as assumptions on the conceptual structure of discourse (changes in perspective, referential domain, scenario etc.). The research literature makes it clear that referential expressions are
Referential collaboration with computers
highly sensitive to assumptions humans derive from the communicative conditions applicable, and this makes them an excellent case for the study of the influence of computer mediation on the linguistic behavior of humans. In conclusion, research shows that referential expressions are sensitive to communicative conditions. This means that their format and their variation somehow bear witness of these conditions. To that end, they can show variation in the types of expressions (e.g., it or the ball), the degrees of specificity (e.g., it or the ball or the large white ball or the large white ball on top of the shelf, etc.) and the way in which referents are conceptualized (e.g., the white button or the alarm button or the left button or the button you just pressed etc.). This characteristic of referential expressions is used in this article to investigate the relevance of one specific communicative condition, i.e., the nature of the addressee. In a production experiment, we ‘measured’ the referential behavior of humans instructing human or computer addressees. The next section discusses the communicative context and the referential variables focused on in the experiment.
4. Instructing human vs. computer addressees: a series of production experiments 4.1
Communicative context – the independent variables
We opted for a communicative context which enabled us to naturally and reliably implement the main independent variable (human vs. computer addressee) in combination with the most important addressee oriented variable, i.e., feedback (without feedback vs. with problematic or unproblematic feedback). Although early instances of human-computer interaction, like the virtual psychiatrist Eliza produced expressive and relational discourse, the bulk of human-computer interaction is located and to be expected in more down to earth communication tasks, especially in referential and instructive communicative exchanges. Therefore, we focused on an instructive setting as a natural environment for human-computer interaction. Instructive communication is typically addressee oriented, even in the unilateral context of user manuals or online help texts, in that an instructor aims at causing the addressee to display a particular behavior, execute specific actions etc. This makes instructions particularly suitable as a context to investigate the referential behavior in relation to addressee variables. Furthermore, instructions have proven to be a grateful context for experiments on reference, due to the controllable relationship between referential expressions, actions to be executed and the addressee’s feedback.
Alfons Maes, Pascal Marcelis and Frank Verheyen
In the experiment, we opted for a text based, ‘medium lean’ communicative context: we asked instructors to identify one out of four objects on the screen by typing in a referential expression on the screen. Instructors could not see or hear their (assumed human or computer) addressee. If they received feedback from the addressee, it invariably took the form of a visual screen of objects with the solution object marked. This communicative situation reflects technically the most practicable mode of human-computer communication, and for computer users the most conceivable one. But it also guaranteed the different conditions to be completely identical in terms of the computer’s interface and reactions: all participants received exactly the same interface, input modality and feedback from the computer. The only difference was the way in which the assumed addressee was presented and conceptualized in the experimental task. That means that differences in the writer’s referential behavior were caused exclusively by writer’s assumptions about the (human vs. computer) addressee, not by perceptual differences which would inevitably emerge when we had decided to use human and virtual voices or faces on the screen. On the basis of this context, three human addressee and three computer addressee conditions were constructed, based on different feedback facilities available (no-feedback, unproblematic feedback, problematic feedback). 4.2
Referential behavior – the dependent variables
Basically, language users can select one of three types of referential procedures: they can take the so-called endophoric or anaphoric procedure and use ‘underdetermined’ linguistic signals that enable addressees to find the intended entity via discourse internal interpretative mechanisms (Locate the black button and press it twice). They can choose the deictic procedure and choose for ‘underdetermined’ linguistic signals that enable addressees to find the intended entity via discourse external perceptual strategies (Press that (button) twice). Or they can semantically conceptualize entities in a way which enables addressees to retrieve the intended entity from memory on the basis of (semantic or encyclopedic) knowledge (Press the alarm button twice) In the experiment, we focused on a situation in which objects had to be identified on the basis of their perceptual characteristics, i.e., a combination of strategy two and three. The experimental set up allowed participants to vary the degree of specification of the object, i.e., they could produce minimal expressions (the square object) as well as different degrees of overspecified expressions (the small / red / square / object / at the top / left). Furthermore, participants could opt for different conceptualization strategies, in that they could use object attributes (the small, red, square object) or location (the object at the top left) as conceptualization source. We
Referential collaboration with computers
did not ask them to produce real discourse, thereby blocking the use of anaphoric expressions or other discourse internal referential strategies. We did not give them the opportunity to see their addressees in a shared workspace, thereby blocking the use of underspecified deictic expressions such as this here or that there. Opting for an object reference task instead of the production of referentially coherent discourse has the disadvantage that a rich array of especially anaphoric strategies can never show up in the referential behavior of the participants. On the other hand, unlike a free writing task, the object reference task allows for a reliable control of the output of the participants. Furthermore, in previous experiments studying the behavior of instructive writers on a distance (Maes et al., 2004), we discovered that instructive writers hardly ever use anaphoric expressions and almost always prefer to re-identify task relevant objects in full format. The set up of the experiment enabled us to enlarge and control the referential strategies used to conceptualize objects semantically, in particular the use of minimal vs. overspecified expressions, and locative vs. attributive expressions. The hypothesis we start from is that humans may make different assumptions vis-à-vis human and computer addressees as to the amount (minimal vs. overspecified) and type (locative vs. attributive) of referential information they need. The degree of (over)specification is a well known indicator of the communicative conditions applicable. Clark and Wilkes-Gibbs (1986) introduce the principle of mutual responsibility, which states and explains that participants in a face-toface setting minimize their collaborative referential effort and often use underspecified expressions, knowing that the partner can always ask for more or indicate when referential problems arise. Problematic feedback results in extra effort to overcome communication problems. For distant and non-feedback communication, Clark and Wilkes-Gibbs propose the principle of distant responsibility, which explains why writers on a distance and without feedback facilities are more referentially explicit. Experiments in Arts (2004) and Maes et al (2004) indeed show that distant instructors in a non-feedback situation tend to use more referential material (i.e., overspecified expressions) than necessary in an attempt to be extremely clear for the distant addressee, even in situations in which simple anaphors would perfectly be able to do the referential job. Moreover, these experiments showed that the degree of overspecification in a non-feedback situation crucially depended on the importance of the task and the assumed goal of the addressee: instructors produced more referential material as the task was more im1. We blocked the deictic strategies for ecological and technological reasons. The use of deictic reference is not unthinkable in human-computer communication, for example in game contexts in which robots on the screen can be instructed by using pointing devices combined with deictic utterances. But this referential strategy is highly unusual and would require a technologically much more advanced set up.
Alfons Maes, Pascal Marcelis and Frank Verheyen
portant to the addressee and his goal was not only to execute but also learn the task. This makes overspecification an important measure of the criticality of the communicative task and perhaps also of the role of the variable we are interested in here, i.e., the (human vs. computer) nature of the addressee. As a careful prediction, humans may be less distantly responsible towards computer than human addressees, which may result in humans showing less overspecification towards computers than towards humans, and also less overspecification in the case of problematic feedback coming from the computer instead of humans. The difference between locative and attributive expressions is less clear in relation to the nature of the addressee. A careful prediction here may be that humans do not look at computer addressees as having the same perceptual/visual capacities as human addressees, which may result in humans producing less location oriented references vis-à-vis computers than vis-à-vis humans. 4.3
Method
4.3.1 Materials For the experiment, a web based internet application was created, related to a database. This enabled us to run the experiment in the laboratory as well as via internet. The application enabled participants to execute the task and fill in an online questionnaire, while all data were automatically transferred to a database. For the technical organization of this experimental environment, see Marcelis (2003). The instructive task embedded in this environment consisted of identifying an object in 30 successive screens by typing in an appropriate referential expression. For this task, thirty object screens were constructed, one of which is presented in Figure 1. Each object screen consisted of four objects. One of them was marked with an x. At the bottom of the computer screen, there was a small screen where participants were to type in a referential expression that identified the marked object. In the feedback conditions, participants only received visual feedback from their addressee, i.e., a similar object screen as is illustrated in Figure 2, with one object marked in a greyscale, which was the object chosen by the addressee on the basis of the referential expression of the instructor. Each of the objects could be identified by three attributive and two location oriented characteristics. The shape of an object was round, square, triangular or rectangular; the size large or small, and the colour white or grey. The object was positioned on the vertical axis (top or bottom) and on the horizontal axis (left or right). The construction of the screens was such that each of the 30 objects to be identified enabled not only the use of attributive and locative referential strategies but also different degrees of overspecification (the rectangle object – the small rectangle object – the small, white, rectangle object – the small, white, rectangle
Referential collaboration with computers
object at the top – the small, white, rectangle object at the top left). In sum, the perceptual environment enabled us to define overspecification as the number of elements which are not necessary to identify the object unambiguously in the perceptual context of the screen. Apart from that, for each object, a full attributive description and a full locative description could be defined, i.e., the combination of shape, size and color representing the ‘gestalt’ and the combination of the vertical and horizontal axis description respectively.
Figure 1. The input screen of the instructive task
Figure 2. The feedback screen
2. See for the relationship between degrees of overspecification and objects on the screen Arts (2004).
Alfons Maes, Pascal Marcelis and Frank Verheyen
4.3.2 Design The conditions in the experiment were based on different conceptualizations of the instructive task. In each condition a fictitious instructive situation was presented to the participants, including a human or computer addressee and a particular type of feedback facility (no, unproblematic, problematic). This resulted in a 2 by 3 between subjects design. In the no-feedback human condition, the panel was presented as a control panel in a long-distance surgery procedure. The objects were buttons to be selected by a distant human surgeon as part of a surgery action. The distant human surgeon had the same screen available, without the element marked. The surgeon fully depended on the instruction of the participant to execute the tasks correctly. The participant was asked to inform the surgeon which button to select. In the no-feedback computer condition, the panel was presented as a control panel in a distant computer controlled nuclear plant. The objects were buttons to be selected by a distant computer as part of a safety procedure in the nuclear plant. The distant computer had the same screen available, without the element marked. The computer fully depended on the instruction of the participant to execute the tasks correctly. In the feedback human condition, the panel was presented as part of an experiment which was being carried out simultaneously by a distant person elsewhere on the university campus. The objects were buttons to be selected by the distant person. The participant was asked to inform the distant addressee which button to select. The distant person had the same screen available, without the element marked and depended fully on the instruction of the participant to carry out the experiment correctly. After each screen, the participant received visual feedback, which was presented to the participant as the solution chosen by the distant person on the basis of the verbal input of the participant. Part of the participants received unproblematic feedback (i.e., the correct object selected), the other received problematic feedback, i.e., 26 correct feedback screens and four incorrect feedback screens (after screen 4, 5, 12 and 21). In the feedback computer condition, the panel was presented as part of the evaluation of a computer programme, which was designed to localize and identify geometric forms. The objects were buttons to be selected by a distant computer. The computer had the same screen available, without the element marked. The computer fully depended on the instruction of the participant to execute the tasks correctly, and test its ‘intelligence’.
Referential collaboration with computers
After each screen, the participant received visual feedback, which was the solution chosen by the computer on the basis of the verbal input of the participant. Part of the participants received unproblematic feedback (i.e., the correct object selected), the other received problematic feedback, i.e., 26 correct feedback screens and four incorrect feedback screens in the series of 30 object screens (after screen 4, 5, 12 and 21). Apart from these conditions, there was a neutral (baseline) condition within the same digital environment, but without mention of an explicit addressee or an instructive situation. The data of this condition are intended to show the sensitiveness of the dependent variables to the communicative conditions in the experiment (addressee, communicative goal). In the neutral condition, the panel was presented as a configuration containing four objects. The participants were asked to type in at the bottom of the website page which element had been marked with an x. They were told to do this in such a way that the marked element could not be confused with the other three elements in the configuration. No mention was made about an instructive situation or an assumed addressee. There was no-feedback involved. 4.3.3 Procedure In all conditions, the structure the procedure was as follows: – participants were instructed to carry out an experimental task, consisting in typing in descriptions of objects; – they were asked to read and enter into the communicative situation of the task; – they executed the task: identifying an object in 30 subsequent screens; – they were asked to answer an online questionnaire. Details of the procedure varied over the conditions. This was partly due to the different conceptualizations of the instructive task. Partly, it was caused by a secondary ambition of the experiment, i.e., to find out how reliable it is to carry out production experiments like these via internet. In the internet mode, carefully selected participants received an invitation via email to participate in the experiment. They were instructed about the goal and the procedure. At the end, they could click the internet address which started the web application. The application automatically logged the time spent on the task, the type of internet connection, screen resolution and the operating system, and of course the object references produced by the participants. In the laboratory mode, participants were instructed face-to-face about the goal and procedures, afterwards they started the same web application.
Alfons Maes, Pascal Marcelis and Frank Verheyen
The neutral condition was run partly in the laboratory, partly via internet. The comparison of the results could be used to measure the reliability of the internet mode. The human conditions were run in the laboratory. Especially the Human Feedback Condition required extra procedural steps as to make the feedback situation realistic. After having explained the goal of the experiment, the experimenter so-called called the distant person to synchronize the two experimental events. The computer conditions were all run via internet. 4.3.4 Participants Overall, 377 participants took part in the experiments, 181 males and 196 females. The participants were not distributed evenly over conditions. The differences were mainly due to the unpredictability of the number of internet participants. The response via internet was much more than expected (273 participants). 4.3.5 Instrumentation and data analysis All data were collected via the same web application. This application was connected to a database. The data of each participant were automatically stored in the database (mysql). Data were selected by using the query language of the database, and a Perl script ‘read’ and counted the output automatically in terms of the dependent variables we were interested in (see for details on the programme code Marcelis (2003)). Pp no#
Condition
Sex
Age
Screen1
117
3
1
23
The rectangular button at the top left
Screen2 The black triangle at the top right
Figure 3. Example of a database record
3. 27 in the neutral lab condition (11 male, 16 female), 93 in the neutral internet condition (48 male, 45 female), 26 in the human no-feedback condition HNF (9 male, 17 female), 110 in the computer no-feedback condition CNF (61 male, 49 female), 22 in the human unproblematic feedback condition HUF (8 male, 14 female), 30 in the human problematic feedback condition HPF (17 male, 12 female), 36 in the computer unproblematic feedback condition CUF (18 male, 18 female) 33 in the computer problematic feedback condition CPF (9 male, 25 female). 4. See for a demo (in Dutch) cclab.uvt.nl/research.html
Etc.
Referential collaboration with computers
A
Select count(*) from where screen1 like ‘%circle%’ or screen1 like ‘% rectangle%’ or screen1 like ‘%square%’ or screen1 like ‘%triangle%’ or <etc.>
B
If (A==1) { $vorm=1 };
Figure 4. Example of database query calculating the presence of shape in the description
For example, Figure 3 contains part of the database record of participant 117, containing the output given for the different screens. By using data queries, all relevant entities in the participants’ descriptions were recognized and counted, i.e., the elements used to identify objects, and the degree of specification used. The data query in Figure 4, for example, counts all ‘shape’ elements used. This enabled us to count the elements of all semantic categories (shape, color, size, vertical location or horizontal location) and to measure the proportion of attributive and locative elements. The degree of (over)specification of each description was determined on the basis of a list of so-called minimal descriptions for each object, i.e., the combinations of elements needed for a minimal unambiguous identification of the object, see for details Arts (2004). The number of overspecified elements was obtained by subtracting the minimal number from the total number of elements used. 4.4
Results
For the analysis of the data a repeated measures ANOVA was used with picture as within factor and nature of the addressee and feedback facility as between factor. The results of the no-feedback conditions are presented in Table 1, the feedback conditions results are represented in Table 2. Table 1 shows that the neutral, human and computer condition differ in the number of referential units (F(2, 253)=13,21, p<.001, η2=.095), overspecified units (F(2, 253)= 10,70, p<.001, η2=.078) and full locative expressions (F(2, 253)=13.86, p<.001, η2=.099). There is no difference in the number of full attributive (gestalt) expressions (F(2, 253)=1.73, p=1.79). Pairwise comparisons show that participants are referentially more specified in the addressee conditions compared to the neutral condition (referential units: neutral-human t(144)=4.90, p<.001; neutral-computer t(228)=3.66, p<.001; overspecified units: neutral-human t(144)=4.55, p<.001; neutral-computer t(228)=3.20, p<.005; full attributive expressions (one tail): neutral-human t(144)=1.86, p<.05; full locative expressions: neutral-human t(144)=5.34, p<.001; neutral-computer t(228)=4.80, p<.001). There was no difference in the number of full attributive expressions between the neutral and computer condition.
Alfons Maes, Pascal Marcelis and Frank Verheyen
Furthermore, participants in the human condition are referentially more specific than in the computer condition; they produce more referential units (t(134)=2.13, p<.05), overspecified units (t(134)=2.00, p<.05) and full locative expressions (one tail): t(134)=1.70, p=05). There was no difference in the number of full attributive expressions. Table 1. Mean number of referential units and overspecified referential units per expression, the proportion of full attributive ‘gestalt’ and full locative expressions per participant, in relation to the no-feedback communicative context (neutral, human instructive and computer instructive) neutral context5 n= 120
no# referential units per expression no# overspecified units per expression proportion of full attributive expressions (‘gestalt’) per participant (shape + color + size) proportion of full locative expressions per participant (vertical + horizontal axis)
addressee/instructive context human addressee HNF n=26
computer addressee CNF n=110
2.18 1.13 .202
3.18 2.10 .333
2.66 1.58 .243
.139
.538
.373
Table 2 shows first of all that overspecification increases when feedback is problematic. In the problematic feedback conditions, participants produce more referential units (F(1,119)=14,99, p<.001, η2=.112), more overspecified units (F(1,119)=15,49, p<.001, η2=.115) and full locative expressions (F(1,119)=5,47, p<.05, η2=.044) than in the unproblematic feedback conditions. The proportion of full attributive expressions did not differ (F<1). In the unproblematic feedback conditions, human addressee participants do not differ from computer addressee participants in the number of referential units (F<1), overspeficied units (F<1), full attributive expressions (F<1) and full locative expressions (F(1,56)=5,24, p=.206). In the problematic feedback conditions, human and computer addressee participants did not differ (F<1) except in the number of full attributive expressions (F(1,61)=4.50, p<.05, η2=.07). 5. In the neutral condition, the results of the laboratory and internet participants were taken together as they did not differ at all, which indicates that both procedures produced equally reliable data.
Referential collaboration with computers
Finally, the instructive no-feedback conditions differed from the unproblematic feedback conditions. In the no-feedback conditions, participants produce more referential units (F(1,192)=33.25, p<.001, η2=.148), more overspecified units (F(1,192)=27.81, p<.001, η2=.127), more attributive expressions (F(1,192)=6.53, p<.05, η2=.033) and full locative expressions (F(1,192)=15.91, p<.001, η2=.077) than in the unproblematic feedback conditions. Table 2. Mean number of referential units and overspecified referential units per expression, the proportion of full attributive ‘gestalt’ and full locative expressions per participant, in relation to the type of feedback (human-unproblematic, computer-unproblematic, human problematic, computer-problematic). unproblematic feedback human addressee n=22
computer addressee n=36
human addressee n=30
computer addressee n=33
1,95 0,90
1,74 0,69
2,15 1,35
2,14 1,31
.156
.113
.237
.094
.195
.081
.233
.340
no# units per expression no# overspecified units per expression proportion of full attributive expressions (‘gestalt’) per participant (shape + color + size) proportion of full locative expressions per participant (vertical + horizontal axis)
4.5
problematic feedback
Discussion and Conclusions
overspecification and communicative task The results first of all show that the experimental set up is sensitive enough to register the effect of communicative conditions on the referential behaviour of language users. In the no-feedback conditions, language users increase their referential effort when the communicative task is embedded in a conceivable instructive scenario (addressee/instructive > neutral context). The difference between the neutral and the instructive version is more than the assumed presence of an addressee alone. The instructive scenario entails a clear task perspective, assumes an overall goal and suggests a high level of task importance and as such creates contextual conditions which are absent or at least fully implicit in the neutral context. All these factors can trigger the increased referential effort as well.
Alfons Maes, Pascal Marcelis and Frank Verheyen
In the feedback conditions, participants increase their referential efforts when they are confronted with problematic (i.e., incorrect) as opposed to unproblematic feedback from the assumed addressee. This result is congruent with the idea that more referential material is used in situations in which referential problems are perceived. Finally, in the instructive conditions participants receiving no-feedback are referentially more specific than participants receiving correct feedback from the assumed addressee. This result is compatible with Clark’s difference between joint and distant responsibility (e.g., Clark & Wilkes-Gibbs, 1986). Feedback is known to decrease the specification level of referential expressions. In a no-feedback condition, language users feel the need to be more specific, because addressees can not signal possible problems. Although these effects were to be expected, two characteristics of the experimental set up make them remarkable, however. First, the referential task in the experiment was ridiculously simple, which could have resulted in participants always using the minimum effort needed to identify the objects. Second, participants had to produce ‘unconnected’ referential expressions, and were not allowed to produce subsequent referential speech acts or full discourse. That way, the overspecification produced by the participants could never be caused by discourse internal considerations, e.g., the presence of competing candidate expressions in previous sentences. overspecification and the nature of the addressee Strong evidence for the relevance of the human vs. computer addressee is provided in the no-feedback instructive condition. The assumption of a human addressee induces more referential effort than the assumption of a computer addressee. In both cases, the task orientation and the importance of the task (distant surgery and nuclear power plant control) are very high. One explanation may be that the computer is less conceivable as an addressee than the human addressee. The surgeon is easily conceived of as a person somewhere in a distant hospital, whereas the computer in the nuclear plant may be imagined and conceptualised less easily. This explanation would predict that the difference between human and computer addressees would decrease when computers would be given human characteristics, such as a name (for example Eliza), gestures (like the paper clip in Word), a face, a voice etc. In the feedback conditions, there is hardly any difference between human and computer addressees. This may be caused by the narrow freedom of expression of the participants and hence the narrow scope of the dependent variables, which – connected with the simplicity of the task - may have been not sensitive enough to register significant differences in addressee assumptions. On the other hand, the very presence of feedback after each trial (identical in the human and computer addressee condition) may have acted as a personalization trait, as a means of presenting the computer as a communicative partner and making the computer more
Referential collaboration with computers
conceivable and present, and may result in an equal referential behaviour in the human and computer addressee condition. There is one more difference between the no-feedback and feedback instructive conditions, which may have been relevant in the differential effect. In the nofeedback condition, participants were presented with a purely fictitious communicative situation. Participants acted ‘as if ’ they were instructing a surgeon or a nuclear power plant computer. In the feedback conditions the participants believed they were testing a computer or helping a distant participant. Participants evaluated these feedback communicative tasks as (equally) believable in the online questionnaire. This may have increased the sense of reality and conceivability of the addressee in the human as well as the computer condition. In sum, if we consider the specification level of referential expressions as a measure of reader orientation and responsibility, the results indicate that computer addressees generate a lower degree of reader awareness than human addressees in fictitious situations in which there is no interaction with the assumed addressee. The results in the feedback conditions suggest that this gap may be bridged if we make the computer addressee and the communicative situation more imaginable, personal, realistic and interactive. And that is exactly what digital technology is trying to.
References Anderson, A. H., Bader, M., Gurman Bard, E., Boyle, E., Doherty, G., Garrod, S., Isard, S., Kowtko, J., McAllister, J., Miller, J., Sotillo, C., Thompson, H. S., & Weinert, R. 1991. �������� The HCR Map Task Corpus. Language and Speech 34(4): 351–366. Ariel, M. 1994. Interpreting anaphoric expressions: A cognitive versus a pragmatic approach. Journal of Linguistics 30: 3–42. Arts, A. 2004. Overspecification in Instructive Texts. Tilburg:Tilburg University. Barr, D. and Keysar, B. 2002. Anchoring and comprehension in linguistic precedents. Journal of Memory and Language 46: 391–418. Beun, R.J. and A.H.M. Cremers. 1998. Object reference in a shared domain of conversation. Pragmatics & Cognition 6: 121–152. Clark, H.H. 1996. Using Language. Cambridge: CUP. Clark, H.H. and Haviland, S.E. 1977. Comprehension and the given-new contract. In Discourse production and comprehension, R.O. Freedle (ed.), 1–40. Norwood NJ: Ablex. Clark, H.H. and Krych, M.A. 2004. Speaking while monitoring addressees for understanding. Journal of Memory and Language 50(1): 62–81. Clark, H.H. and Wilkes-Gibbs, D. 1986. Referring as a collaborative process. Cognition 22: 1–39. Gundel, J.K., Hedberg, N. and Zacharski, R. 1993. Cognitive status and the form of referring expressions in discourse. Language 69(2): 274–307. Hancock, J.T. and Dunham, P.J. 2001. Language use in computer-mediated communication: The role of coordination devices. Discourse Processes 31(1): 91–110.
Alfons Maes, Pascal Marcelis and Frank Verheyen Hanna, J. E. and Tanenhaus, M.K. 2004. Pragmatic effects on reference resolution in a collaborative task: Evidence from eye movements. Cognitive Science 28: 105–115. Herring, S.C. 1996. Two variants of an electronic message schema. In Computer-Mediated Communication. Linguistic, social and cross-cultural perspectives [Pragmatics and Beyond 39], S.C. Herring (ed.), 81–106. Amsterdam: John Benjamins. Lee, E.-J., and Nass, C. 2002. Experimental tests of normative group influence and representation effects in computer-mediated communication: When interacting via computers differs from interacting with computers. Human Communication Research: Official Publication of the International Communication Association 28(3): 349–381. Maes, A., Arts, A. and Noordman, L. 2004. Reference management in instructive discourse. Discourse Processes 37(2): 117–144. Marcelis, P. 2003. Overspecificatie in Instructieve Mens-computerinteractie. Een experimenteel onderzoek naar de manier waarop mensen computers instrueren. MA thesis, Universiteit van Tilburg. Metzing, C. and Brennan, S. E. 2003. When conceptual pacts are broken: Partner-specific effects on the comprehension of referring expressions. Journal of Memory and Language 49: 201–213. Nass, C., Fogg, B. J. and Moon, Y. 1996. Can computers be teammates? International Journal of Human-Computer Studies 45: 669–678. Nass, C. and Lee, K.M. 2001. Does computer-synthesized speech manifest personality? Experimental tests of by an recognition, similarity-attraction, and consistency-attraction. Journal of Experimental Psychology: Applied 7: 171–181. Newlands, A., Anderson, A. H. and Mullin, J. 2003. Adapting communicative strategies to computer-mediated communication: An analysis of task performance and dialogue structure. Applied Cognitive Psychology 17: 325–348. Pechmann, T. 1989. Incremental speech production and referential overspecification. Linguistics 27: 89–110. Reeves, B. and Nass, C. 1996. The Media Equation: How people treat computers, television, and new media like real people and places. Cambridge: CUP. Tomasello, M. and Rakoczy, H. 2003. What makes human cognition unique? From individual to shared to collective intentionality. Mind & Language 18: 121–147. Van der Sluis, I. and Krahmer, E. 2004. The Influence of Target Size and Distance on the Production of Speech and Gesture in Multimodal Referring Expressions. Paper presented at the 8th International Conference on Spoken Language Processing (ICSLP), Jeju Island, Korea. Walther, J.B., Anderson, J.F. and Park, D.W. 1994. Interpersonal effects in computer-mediated interaction: A meta-analysis of social and antisocial communication. Communication Research: An International Quarterly 21: 460–487. Walther, J.B. and Burgoon, J.K. 1992. Relational communication in computer-mediated interaction. Human Communication Research: Official Publication of the International Communication Association 19: 50–88.
Reflexivity and temporality in discourse deixis Friedrich Lenz In this paper discourse deixis and anaphora are treated as two separate phenomena. It is argued that discourse deixis is a special case of metacommunication on the one hand and of time deixis on the other. Since discourse deixis focuses on a part of the discourse itself, it is a form of reflexive language use. In order to identify a segment of the ongoing discourse, time deictic expressions have to be used. This combination of both a reflexive and a time deictic element is not inherent in anaphora.
1. Introduction The distinction between deixis and anaphora has been a controversial issue since the origins of grammatical theory (cf. Lenz 1997, 8ff.). Unlike Consten (2004) I will take a non-integrative stance in this paper. Deixis in general and discourse in particular will be regarded as a linguistic phenomenon of its own. Discourse deixis can primarily be described in terms of reflexivity and temporality. It needs to be regarded as a specific form of temporal deixis invariably containing a meta-communicative element. Text deixis or discourse deixis (I will use the term discourse deixis referring to deictic reference in both written and spoken discourse) has been included in most recent treatments of deixis. Although there are no specific discourse-deictic expressions, it has even sometimes been given the status of a deictic dimension of its own. It seems to meet most definitions of deixis according to which – along the lines of Bühler’s pointing metaphor – deixis is a way to identify entities in the communicative context of the participants. As the pointing metaphor is not applicable to all entities, including those of discourse deixis, it is more appropriate to treat deixis as a special form of definite reference (cf. Hanks 1990). As the referents of discourse deixis are temporal entities, it will be argued that discourse deixis is not a separate deictic dimension, but a special case of time deixis. Since it also contains a meta-communicative element, both a temporal and a meta-communicative rela-
Friedrich Lenz
tion have to be expressed. Some of the lexico-grammatical means by which this is done will be discussed in the second part of the paper.
2. Reflexivity If discourse deixis were an independent deictic dimension, it would be possible to isolate specific relations on which the identification of a definite referent could be based. In the traditional dimensions, deictic expressions are used to identify spatial, temporal or personal entities. The identification depends on spatial, temporal or speech role relations which exist between the origo and the entity in question. If we treat discourse deixis as a separate dimension relying on traditional deictic expressions, we at least have to know to what extent the entities referred to differ from those of the traditional dimensions and how they can be characterised. According to the definitions of discourse (or text) deixis, the answer seems to be easy: we are dealing with some “portion“ (Levinson, 1983), “aspect“ (Fillmore 1997) or “segment“ (Rauh 1983) of the discourse itself. In this way discourse deixis is a form of talk about communication and inconsequently, reflexive language use. Unlike extra-communication – to use a term by Wunderlich (1970) – which is talk about other people’s discourse (and a gossip’s favourite preoccupation, for instance) discourse deixis is meta-communicative. In meta-communication the participants speak of their own discourse, i.e. of the units of the discourse they are currently engaged in. Whereas extra-communication, no matter whether it is descriptive-explicative (as in the case of discourse analysis) or gossipy (cf. Lenz 1997a), refers to outside, usually past events, metacommunication is reflexive language use within an ongoing situation. If we assumed discourse deixis to be an independent deictic dimension, discourse deictic relations would then exist between two discursive entities one of which has to be attached to the origo, and this could only be the deictic expression itself. Thus, discourse deixis is very close to anaphora, which is also characterised by a relational link between two expressions in a discourse. As many (discourse) deictic expressions can be used anaphorically as well, the distinction between dis1. The paper is a modified version of Lenz (1999). I would like to thank the Narr Verlag for giving copyright permission. 2. Unfortunately, the term meta-communication has given rise to a great deal of confusion. First, it has not been distinguished from extra-communication. Second, it has not been kept apart from the general phenomenon of meta-language (vs. object language). Third, meta-communication has been associated with communication about the socio-emotional relationships between the participants. Psychologists in particular follow Watzlawick’s usage and assume that the so-called ‘relationship aspect’ is dealt with in meta-communication (cf. Watzlawick et al. 1967, 53ff.).
Reflexivity and temporality in discourse deixis
course deixis and anaphora has been one of the main topics (if not the main topic) in the discussion of discourse deixis. The specific criterion to distinguish it from anaphora – there seems to be a consensus among the authors advocating a discourse deictic dimension – is its reference to linguistic entities rather than to the (identical) referents of antecedent expressions (cf. Fillmore 1997, Lyons 1977, Levinson 1983) Whereas an anaphoric expression refers to a non-discursive entity, an entity which is not itself an element of the discourse, a discourse deictic expression refers to an entity within the discourse. Discourse deixis is a phenomenon of reflexive language use (cf. Lucy 1993). The reflexive character of discourse deixis requires us to switch from the ‘ordinary’ object level to the linguistic level. We use deictic, that is linguistic, expressions to refer to linguistic entities. This view is implicitly taken by most authors and explicitly, for instance, by Lyons (1977, 667) when he generally states that in discourse deixis “deictic expressions refer to linguistic entities of various kinds in the co-text of the utterance.“ According to this view all linguistic entities could be the (metalinguistic) referent. But it is important to note that in the framework of discourse deixis these entities must have been brought about in the co-text or co-discourse. Reference to linguistic entities as such is not necessarily discourse deictic. In Saussurean terms we could say that discourse deixis is directed to the level of parole. The referent is a unique instance of a linguistic entity, an entity of the parole. A problematic case is represented by reference to linguistic items that occur in the discourse, but are merely mentioned (for example for the illustration of linguistic phenomena in a linguistics book) in contrast to being used. This reference is not discourse-deictic even if deictic expressions are employed. For instance, in the passage ‘John is in the house. This is a declarative sentence’ the demonstrative refers to the example sentence mentioned, but this sentence does not refer to anything. One could say that the antecedent is a quote which does not originate from a text or discourse, but from the langue. In this way, it is part of the meta-language and the relation between this entity and the this would be on the same level, thus constituting what one could call ‘meta-linguistic anaphora’. We do not have to switch between two levels as is required in discourse deixis. The anaphoric relation is also indicated by the fact that the this can be substituted by is antecedent. ‘John is in the house’ is a declarative sentence is synonymous with our previous passage, even though the deictic expression is eliminated and there is no relation to the deictic origo. In discourse deixis we do not refer to linguistic entities as entities of the language system, Lyons’ equivalent to Saussure’s langue. But it seems that Lyons has exactly these entities in mind when he claims that his text (a book on semantics) was full of text-deictic reference. This also becomes evident from Lyons’ (1977, 667) only (and often quoted) explicit example: “(X says) That’s a rhinoceros (and Y responds) A what. Spell it for me.“ What is referred to by Y (by means of the it) is not a token, as
Friedrich Lenz
it is realised in the dialogue, but a linguistic type (cf. Conte 1981). It is the graphemic, i.e. time- and discourse-independent form that is taken up by the pronoun. Interestingly, it is not treated as a deictic pronoun in most accounts of deictics anyway. Elements of the parole are concrete and in this sense Fillmore’s somewhat imprecise formulation when he speaks of “portions“ of the discourse seems to be quite appropriate. What we refer to in the framework of discourse deixis are not the abstract types, but the tokens of linguistic entities as they are realised in the course of a discourse. That we know of the type-token identities and use the type-terms to identity the tokens is part of the language user’s knowledge and the normal referential process, but it does not mean that we refer to the types. Types are independent of time. A discourse develops in time and so do its portions that we can refer to. The relations between these portions and the egocentric orientation point are temporal, and it is these relations that are responsible for the identification of the entities.
3. Temporality 3.1
Temporal and local deictics
Given the temporal character of its referents, it is hardly surprising that we primarily find time-deictic expressions in discourse deixis. Local deictics are only found (with very few exceptions) in writing and even there we transform the three-dimensional space into a one-dimensional one, thus recreating the same linear order in which we conceive of time. The three spatial dimensions are usually reduced to the vertical. We would refer to the portions of a written text as to the parts ‘above’ or ‘below’, but only in very rare cases to parts to the right or the left as well, and revealingly these cases are only found when we have diagrams, tables, pictures and the like in our publication. The sagittal dimension (front/back) is not found in our western writing tradition. We somehow still seem to conceptualize our writings in terms of paper rolls. In this way, even a three-dimensional book – a volume – appears as a one-dimensional entity. When we refer to parts above or below within a book, the use of the vertical dimension is due to our writing conventions, but it is ultimately not important which of the three spatial dimensions is selected. What is important is the fact that it is only one. Only one-dimensional entities can be treated as temporal entities. They appear on the time line. 3. From a cognitive perspective on reference and deixis, one could say that it is an instantiation of a cognitive category rather than a linguistic entity that we refer to. For a detailed discussion of deixis from a cognitive point of view see Janssen (1995). 4. In Hausa, for instance, the sagittal dimension can be used for discourse-deictic reference. I would like to thank Clifford Hill (Columbia University) for drawing my attention to this fact.
Reflexivity and temporality in discourse deixis
In discourse, and that is what I am talking about here, temporal-deictic expressions are almost exclusively used. One of the few exceptions is the deictic adverb here which I have just used when I said: … in discourse, and that is what I am talking about here … In some cases its use can simply be interpreted as local-deictic, i.e. independent of the discourse (not of the speech situation, of course), and our example is one such border-line case. It could mean the place where I am presently located. In this situational understanding, the adverb is not used discoursedeictically at all, but as an autodeictic spatial, i.e. a three-dimensional expression referring to a local entity around the standpoint of the speaker (cf. Harweg 1990, 182). This interpretation is likely when we add, for instance, in this heat to our example (… in discourse, and that is what I am talking about here in this heat …). The actual size of the speaker region referred to depends on the current situation and topic, for example, whether we are indoors or outdoors or whether we talk about temperatures in different towns or countries. However, the more likely interpretation of our original example (… in discourse, and that is what I am talking about here …) is discourse-deictic. It refers to a portion of the discourse, and the actual local space where the discourse takes place seems to be irrelevant. Again, here would be used autodeictically, that is the relation that identifies the entity referred to would include the standpoint of the speaker. But what is this standpoint or the origo like? A spoken discourse is only conceivable as constituted in time. Our deictic relation can then only be specified by two temporal data. The first would be the time of the utterance of the deictic, and neither the deictic as such nor the place where the (originally local) deictic is uttered. The second would be the time of the entity referred to and the temporal relation between the two data would be the source of the identification process. Although we are concerned with time, we inevitably use local metaphors (cf. Radden 1997). Our data (the relatum and the theme, in the terminology of Klein 1994) have to be conceived of as places on the time line. As in any instance of time deixis, a referent is located on the time line and has some extension. In our example, however, the extension is not exactly specified. The extension is left open, parallel to the openness of the local-deictic interpreta5. It is not easy to find a specific context where the target place precisely equals the speaker’s location. Perhaps when someone says I’m standing here and don’t move an inch s/he refers exactly to the place where s/he is. S/He even abstracts from the vertical dimension, which is irrelevant with regard to the particular area s/he is standing on. The sentence is still true if s/he moves her/his head or shoulders. In any case, we inevitably have to take the current situation and topic into account. I think we have to assume something like a local Gricean relevance maxim on the basis of which the speaker region is interpreted. The relevant speaker region depends on the situation and topic of the conversation, and according to those, co-operating partners can presuppose an adequate region (see Lenz 2001).
Friedrich Lenz
tion of here. It can only be retrieved through the context, if at all. When I said that I was talking about something here, I could have meant at that very moment, in that passage or in my entire talk. Within the framework of discourse deixis, here means the same as now. Here and now serve to specify a position, not an extension on the time line and this can only be done in relation to a temporal anchoring point, the time of utterance. 3.2
Deictic and meta-communicative expressions
Discourse deixis, even when local deictic expressions are (metaphorically) used, is clearly time-deictic. It is a special case of time deixis, however, since both the time of utterance and the referent are limited to the time span of the discourse. What is at least equally important is the restriction which applies to the type of possible referents. In time deixis the referents are entities which occur or take place in time: states, processes, activities, events. This ‘in principle’ also applies to discourse deixis, but it is further restricted to discursive entities. The meta-communicative character of the referent also has to be expressed. It is therefore not only a deictic expression, but also a meta-communicative expression that indicates discourse deixis. The necessity of this combination has rarely been seen. The paradigm cases for discourse deictic expressions that we find in the literature are, besides now and here, earlier and later (the equivalent of above and below) and this and that. Examples of discourse deictic utterances of sentence size are very rare. This restriction in the literature has probably prevented us from seeing the meta-linguistic component. It is not the use of isolated deictic words, but full utterances in which we refer discoursedeictically. And even the few non-sentential isolated examples of discourse deixis could already have led in the right direction. Fillmore gives the following discoursedeictic formula: the preceding X and the following X and replaces the X in this only example with paragraph. In principle, X could be any expression that denotes something that could be a portion of a discourse: a preceding argument, description, explanation, joke, utterance, sentence etc. Preceding or following could, of course, also be replaced by this and that, the former, the latter, the last or the next. In any case, we have a combination of a time deictic and a meta-communicative expression. The meta-communicativity is expressed by a noun and the deicticity by an adjective. Nominal phrases are only one type of discourse deictic expression. Particularly in everyday conversation they seem to be rather the exception than the norm, and it is rather revealing that its head is usually a deverbal noun (argument, description, explanation, etc.). In most cases the deictic and the meta-communicative element are combined in the verbal phrase. Let us look at our preceding example again: in discourse, and that is what I am talking about here (or now). It is the verb
Reflexivity and temporality in discourse deixis
through which the meta-communicative component is expressed. Moreover, it is not only the adverb, but also the tense of the verb that is deictic. In fact, the adverb is not even necessary and we can easily omit it in our example: in discourse, and that is what I am talking about would still be discourse-deictic according to our and most other definitions of discourse deixis. Such an utterance does clearly refer to a portion of the ongoing discourse. We have to acknowledge the fact that not only nominal phrases can be referring expressions. Davidson (1967) has shown that this applies to every ‘action sentence’, one type of which are meta-communicative sentences. The utterance is deictic on the basis of the deictic function of the tenses. They localise events. The lexical content of the verb (and its arguments) specifies the event type or situation type and the tense localises a token of this type with regard to the time of the utterance as anchoring point. The (temporal) localising function of tense is the same for time and discourse deixis. Before we come back to this function, we will have to elaborate on the notions of temporal schemata and situation types. 3.3
Temporal schemata and discourse situation types
In the framework of discourse deixis, situation types must be discourse situation types. Due to the meta-communicative and temporal character of discourse deixis, the types of potential referents, i.e. the situation types, are restricted. We have already seen that there are only temporal entities that can be referred to in time deixis: states, activities, processes and events with certain sub-classifications. I follow the usual terminology and call these entities situations. Vendler (1967) pointed out that verbs (to be more precise: verbs with their arguments) exhibit different internal temporal schemata according to which we can distinguish between different classes of situations. For discourse deixis we can exclude those situations that are not temporally bounded (i.e. not confinable to the time of the discourse) or those that are independent of the participants. Discourse situations are brought about by the participants and must be dynamic and intentional so that states and processes, though temporal entities, cannot be potential referents in the framework of discourse deixis. In terms of Vendler’s classification there are three classes of temporal entities in discourse: activities and two classes of events, namely achievements and accomplishments. The three classes can be differentiated by a combination of two features: durativity and telicity. Telic situations are events with some sort of climax or culmination point that result in a change of state. Accomplishments are durative and telic, whereas achievements are telic, but not durative. Activities are durative, but non-telic situations. As was said before, these situation classes (or Vendler classes as they are often called) are reflected by the lexical content of verbs, or rather verb constellations (I borrow this
Friedrich Lenz
term from Smith 1997) since one has to go beyond Vendler’s consideration of isolated verbs (for a detailed discussion see Lenz 1997). In our case, these verb constellations denote some discourse situation in the sense of a situation type (not to be confused with situation class) and an instance of that type is a discourse situation with a unique location on the time line. As the situations are intentionally brought about by the participants of a discourse, it is not surprising that discourse situations can be subsumed under what has been treated as speech act types or speech act patterns in speech act theory. However, one has to be aware that not only those verbs which can be used in performative formulae or those denoting some specific illocution fit into our framework. There are complex discourse situation types that comprise more than one sentence, utterance or illocution. When we explain or narrate something, it will usually take more than one sentence. Some complex discourse situations involve not only a number of utterances, but speaker change as well, for example when we negotiate or quarrel. In any case, there are meta-communicative verbs that denote such situation types, and neither the verbs nor the types are covered in traditional speech act theory. Table 1 shows how some meta-communicative verbs (or verb constellations) fits into the three relevant classes: Table 1. Classification of meta-communicative verbs classes
durative
telic
examples:
activities
+
–
talk, converse, chat, gossip, lecture, discuss, debate, negotiate, bargain, argue, quarrel … about something (and/or) with someone
accomplishments
+
–
explain, describe, narrate, relate, report, comment on, prove, reply, summarize … something
achievements
–
+
agree, apologise, regret, promise, order, command, recommend, affirm, claim, approve, name, appoint, say …
Let us now see how the time schemata expressed in the examples interact with (discourse-) deictic elements. The verbs listed in the table are apparently speech act verbs in a very broad sense (cf. Wierzbicka 1987) and by no means only performative verbs (in the narrow sense of Searle 1969). When they are used discoursedeictically, they have to localise one token of the situation type described by the lexical content of the verb within the time of the discourse. This is done by time deictic expressions, especially tenses and temporal adverbials (for lists and a de-
Reflexivity and temporality in discourse deixis
tailed discussion see Lenz 1997). The grammatical category of tense is decisive for the localisation though it interacts with the aspect category. The localisation is based on the basic temporal order relations with regard to the time of utterance. 3.4
Present, past and future discourse situations
The present tense expresses the simultaneity of the time of utterance with the situation. Again, our example can illustrate the point: discourse, and that is what I am talking about. In this case, the discourse situation is marked as a present situation by the present tense. As talking about something is an activity with some duration and the time of utterance is conceived punctually, it is not astonishing that the progressive (imperfective) aspect has to be used in our example. The time of utterance is included in the situation, and that means that the situation has to be within the discourse. The same is true when I, you or some other participant utters that he or she is chatting, gossiping, discussing, arguing, quarrelling about something or with someone. Achievements are non-durative situations; they are conceptualized as practically instantaneous, as if initial and final point coincided. Smith (1997: 58) defines achievements as having simultaneous initial and final points, but concedes that “discernible fractions of time may be involved”. The language user treats non-durative situations as happening within “the psychological now” (cf. Schopf 1984). He or she abstracts from the actual extension of this minimal time span in the same way as he or she abstracts from the extension that the production of a sentence takes (cf. Klein 1994). Discourse situations which are simultaneous with the time of utterance but do not exceed (overlap) it would then be present achievements. In performative utterances such as I apologise, I agree, I approve (of something), I promise (to do something) we refer to exactly the event that we simultaneously perform. In this sense, performative utterances are discourse deictic. Reference is made to a speech event that coincides with the time of the utterance. As it is a closed situation, we can expect the use of the present simple (perfective). As far as tense and aspect are concerned, performative formulae are not really exceptions as some grammars treat them. Since accomplishments are durative, the time of utterance is included in the present situation. Like activities they have to be expressed by the present progressive (imperfective). We only find examples like I am explaining something and not I explain something. When accomplishment verb constellations are used in the present tense, i.e. are located at the time of utterance, the speaker only refers to a part of the accomplishment token because the situation is only in progress and the result state has not yet been reached; in fact, it might not be reached at all. In any case, what is referred to is a portion of the discourse. In the present, all speech act
Friedrich Lenz
verb constellations (in the broad sense) with participants as the subject are discourse deictic. The temporal order relation for the past is temporal anteriority. The situation is located before the time of utterance. Without further specification it is not clear whether the situation referred to is within the discourse or not, no matter which situation class is involved. When I say that I or you talked about something, I or you apologised or I or you explained something, it could well have happened many years ago with different participants from those of the ongoing discourse. What we need is to have another temporal anchor, something like Reichenbach’s reference time to which the situation can be hooked up (Reichenbach 1947). Klein’s notion of topic time seems to be appropriate as it does not overstrain the term reference and conveys the idea that the situation time cannot be calculated without contextual information. According to Klein (1994), topic time is the time being talked about, which does not necessarily coincide with the time of the situation. In the case of discourse deixis, the time being talked about has to be within the discourse as well as the situation time. It has to be contextually given. It is either attached to some other situation which has already been talked about (in previous utterances) or specified by some time adverbial, no matter whether it is deictic or not, for example five minutes ago, at noon, at three o’clock on the third of December 2006, when John came in. If we know that the time specified is located within the discourse, the situation described by the lexical content of the verb constellation is hooked up to this time in a similar way as it is to the time of utterance in the case of the present, i.e. with the same restrictions on situation classes and aspect. Future discourse situations can in principle be referred to along similar lines, but as time is scarce and every discourse is a temporally bounded entity, I will not go any further into the complicated details of tenses, aspect and topic time specification, which is a far more tricky process than could be sketched here (for a detailed discussion see Lenz 1997).
4. Conclusion Discourse deixis cannot be regarded as a separate deictic dimension. It lacks deictic expressions of its own and its referents are related to the origo in the same way as in other dimensions. But the referents share a specific feature. They are entities which are brought about within the ongoing discourse. It is this feature which complicates discourse deictic reference and justifies why it should be treated as a specific deictic phenomenon. On this basis, it was argued that discourse deixis is a special case both of time deixis and of meta-communication.
Reflexivity and temporality in discourse deixis
It is meta-communication in the specific sense that the referent cannot be a linguistic entity as such, but has to be an instance or subsituation of the communicative situation as a whole. The deicticity is specific in the sense that the referent has to be located within the time of the discourse. Apart from the exceptional metaphorical use of local deictics, the temporal relation between the origo and the referent is expressed by time deictic expressions. Both the meta-communicative and the deictic element can be expressed within a nominal phrase, but it is the verbal phrase that is the more likely candidate for discourse deixis since it combines tense, an obligatory time deictic expression which interacts with other temporal expressions, with the discourse situation type that is coded in its lexical content. We have to investigate lexico-grammatical means of this kind if we want do draw a clear picture of the ways in which discourse deixis is expressed.
References Bühler, K. 1934/1965. Sprachtheorie. Die Darstellungsfunktion der Sprache. Stuttgart: Fischer. Consten, M. 2004. Anaphorisch oder deiktisch? Zu einem integrativen Modell domänengebundener Referenz. Tübingen: Niemeyer. Conte, M.-E. 1981. Textdeixis und Anapher. Codikas/Code 3(2): 121–132. Davidson, D. 1967. The logical form of action sentences. In The Logic of Decision and Action, N. Rescher (ed.), 81–95.Pittsburgh PA: The University of Pittsburgh Press. Fillmore, C. 1975. 1997. Lectures on Deixis. Stanford CA: CSLI. Hanks, W. 1990. Referential Practice. Language and lived space among the Maya. Chicago IL: University of Chicago Press. Harweg, R. 1990. Studien zur Deixis. Bochum: Brockmeyer. Janssen, T. 1995. Deixis from a cognitive point of view. In Meaning as Explanation: Advances in linguistic sign theory, E. Contini-Morava and B. Sussman Goldberg (eds), 245–270. Berlin: Mouton de Gruyter. Klein, W. 1994. Time in Language. London: Routledge. Lenz, F. 1997. Diskursdeixis im Englischen. Sprachtheoretische Überlegungen und lexiko-grammatische Analysen. Tübingen: Niemeyer. Lenz, F. 1997a. Speaking of speech acts. Erfurt Electronic Studies in English. EESE. http://webdoc.sub.gwdg.de/edoc/ia/eese/artic97/lenz/5_97.html Lenz, F. 1999. The temporal dimension of discourse deixis. Arbeiten aus Anglistik und Amerikanistik (AAA) 24(1): 3–14. Lenz, F. 2001. Here is hier, there is dort, but where is da? Contrastive Reflections on English and German Demonstrative Adverbs. Arbeiten aus Anglistik und Amerikanistik (AAA) 26(1): 39–52. Levinson, S.C. 1983. Pragmatics. Cambridge: CUP. Lucy, J. 1993. Reflexive language and the human disciplines. In Reflexive Language. Reported speech and metapragmatics, J. Lucy (ed.), 9–32. Cambridge: CUP. Lyons, J. 1977. Semantics. Cambridge: CUP. Radden, G. 1997. Time is space. In Human Contact through Language and Linguistics, B. Smieja M. Tasch (eds), 147–166. Frankfurt: Lang.
Friedrich Lenz Rauh, G. 1983. Aspects of deixis. In Essays on Deixis, G. Rauh (ed.), 9–60. Tübingen: Narr. Reichenbach, H. 1947. Elements of Symbolic Logic. New York NY: Academic Press. Schopf, A. 1984. Das Verzeitungssystem des Englischen. Tübingen: Niemeyer. Searle, J. 1969. Speech Acts. An essay in the philosophy of language. Cambrigde: CUP. Smith, C. 1997. The Parameter of Aspect. Dordrecht: Kluwer. Svartvik, J. and Quirk, R. 1980. A Corpus of English Conversation. Lund: Gleerup. Vendler, Z. 1967. Verbs and times. In Linguistics in Philosophy, Z. Vendler, 97–121. Ithaca CA: Cornell University Press. Wierzbicka, A. 1987. English Speech Act Verbs. A semantic dictionary. Sydney: Academic Press. Watzlawick, Paul et al. 1967. Pragmatics of Human Communication. A study of interactional patterns, pathologies, and paradoxes. New York NY: Norton. Wunderlich, D. 1970. Die Rolle der Pragmatik in der Linguistik. Der Deutschunterricht 22: 5–41.
The function of complex anaphors in texts Evidence from corpus studies and ontological considerations Manfred Consten, Mareile Knees and Monika Schwarz-Friesel Complex anaphors are nominal expressions that pick up sentential or longer antecedents referring to propositionally structured referents such as events and states. Concerning their textual function, complex anaphors do not only serve as a means of continuity but also effect the progression of the information flow in text, i.e. they are thematic and rhematic at the same time. The complexation process establishes the pre-mentioned referent as a unified discourse entity. On the basis of natural language corpus data, we show that complex anaphors have a central function in establishing coherence in texts. We distinguish type shifting and non type shifting complexation processes where the former is subject to ontological based constraints. In order to show how ontological features are involved in disambiguation we discuss an example of ambiguous complex anaphora.
Introduction Complex anaphors are nominal expressions referring to propositionally structured referents (such as propositions, states, facts and events) while introducing them as unified entities into a text world model. On the basis of natural language corpus data, we will show that complex anaphors have a central function in establishing coherence in texts. We will distinguish between different anaphoric complexation processes that are subject to ontological based constraints. These constraints are confirmed by corpus analysis and will be used to describe anaphoric 1. This paper has been written within the context of the Research project „KomplexTex“, granted by the Deutsche Forschungsgemeinschaft (SCHW 509/6–2). We would like to thank Maria Averintseva-Klisch, Tübingen, for her innumerable inspiring questions and helpful comments, and Marlies Schleicher, Jena, for her unfailing work with the Tiger-Corpus.
Manfred Consten, Mareile Knees and Monika Schwarz-Friesel
resolution within a resolution model based on semantic as well as conceptual structures. Following this model, referents of complex anaphors are established as discourse objects during the anaphoric process by (re-)activating entities at the text-semantic level. Thus, complex anaphors do not only serve as a means of continuity but also effect the progression of the information flow in text. In the first section, we discuss our definition of complex anaphors and we present the findings of a natural language corpus study. Section 2 will provide a description of anaphoric complexation processes in terms of ontological categories. Due to the lexical content of the anaphoric expression the ontological status of the referent can change during the anaphoric process. This process is subject to ontological constraints, cf.:
(1) Young drivers usually drive too fast. *This event / *These events…
Section 3 will explain this complexation process and provide a resolution model which takes semantic and conceptual features into account. We will discuss an example of ambiguous complex anaphora in order to show how ontological constraints are involved in disambiguation. In section 4 we will consider indirect complex anaphora. Within this context the textual function of complex anaphors as means both of textual coherence and referential progression is illustrated, i.e. they are thematic and rhematic at the same time.
1. Complex anaphora as a basic phenomenon in texts 1.1
Definition: Complex anaphor
Complex anaphors are nominal expressions referring to propositionally structured referents (such as propositions, states, facts and events) while introducing them as unified entities into a discourse. Additionally, they can classify or evaluate the referent (see section 3).
(2) Young drivers usually drive too fast. This/ this fact/ this image/ this impertinence …
Researchers have referred to complex anaphors heterogenously, e.g. abstract object anaphora (Asher 1993, 2000), labelling (Francis 1994); as well as extended reference and reference to fact (Halliday/ Hasan 1976), sentence-related reference (Koeppel 1993), proposition-related anaphora (Greber 1993), situational anaphora (cf. Frau-
2. “Let us here use ‘referent’ for the discourse entity referred to, regardless of its level of representation“ (Fraurud 1992: 26). For levels of representation see section 3.
The function of complex anaphors in texts
rud 1992, Dahl/ Hellmann 1995), discourse deixis (Webber 1991, Eckert 1998) or shell nouns (Schmid 2000). We, however, prefer the term complex anaphor since complexitiy is the defining characteristic feature of these anaphors. Our notion of complex anaphor includes two criteria that do not completely depend on each other: First, the antecedent has to be a complex linguistic entity, which means that it consists of (at least) a clause. Second, the referent has to be a conceptually complex item, a “second- or third-order entity” in terms of Lyons (1977, 1989). In contrast to first-order entities (like things and persons), second order entities are defined in terms of both space and time: events and processes are defined as taking place at a certain location with a certain temporal extension (Lyons’ second-order entities), while facts and propositions are mere mental constructions derived from such temporal-spatial items, but abstract from space and time (Lyons’ third order entities) (see section 2., cf. Zacks / Tversky 2001; van Lambalgen / Hamm 2005). Example (3) shows that there can occur inconsistencies between the first criterion that reflects the grammatical structure of the preceding text, and the second criterion that deals with the conceptual structure of the referent. (3) Ende der achtziger Jahre hatte ein Schelm die Wohlstandvöllerei auf den Punkt gebracht: “Immer schöner kaufen, immer weiter fliegen, immer schneller vögeln.” Viele Monate lang verzierte dieses Graffito die Wände eines alten Toilettenhäuschens im Hamburger Yuppi-Stadtteil Winterhude-Süd. (Tiger-Corpus, 32269f) At the end of the eighties a rogue characterised the decadent gluttony: “Buy nicer and nicer, fly further and further, shag faster and faster. ”During several months this graffito ornamented the walls of an old comfort station at Hamburg’s Yuppie district Winterhude-South.
The referent for the anaphor this graffito is introduced by the two preceding sentences, the first referring to the act of characterising that resulted in the graffito as a written form of utterance, the second quoting the content of the graffito. Thus, the claim for a complex antecedent is fulfilled. But the referent GRAFFITO itself is not a complex entity but a mere first-order entity, namely a special formation of paint on a wall. Consequently, we do not categorise this anaphor as a complex one. 1.2
Findings of the corpus study
On the basis of natural language corpus data we show that complex anaphors have a central function in establishing coherence in texts. For our data analysis we used the 3. In contrast to Webber and Eckert’s discourse deixis, there is also a narrow notion that is not congruent with complex anaphora (cf. Consten 2004: 29–33; Lenz 1997). Here, discourse deixis is defined as meta-textual, i.e explicitly ‘pointing’ at the preceding text (the joke I made before, the hypothesis suggested in the first chapter). 4. Or, in case of indirect complex anaphors, the anchor. See section 4.
Manfred Consten, Mareile Knees and Monika Schwarz-Friesel
Tiger-Corpus which consists of approximately 40,000 sentences (700,000 token) taken from German newspaper texts (Frankfurter Rundschau). The corpus is tagged with syntactic structures, so we could use it in order to systematically determine different grammatical types of complex anaphors. We distinguish between pronominal and lexical complex anaphors. For the group of pronominal complex anaphors there are three different part-of-speech-tags (POS) which have to be accounted for since they could contain a pronoun functioning as complex anaphor: weak and strong demonstrative pronouns like das, dies, dem, dessen (“that”, “this”, “whose”), indefinite pronouns like beides, letzteres, solches (“both”, “the latter”, “such”) and the personal pronoun es (“it”). For the group of lexical complex anaphors there are no specific POS-tags since they are whole syntactic phrases. Thus, the amount of data to be analysed is very large as every definite NP could function as an anaphor. Nonetheless, we assume that NPs containing a demonstrative determiner like diese Aktion (“this action”) are more likely to function as complex anaphors than other definite NPs. We also checked whether certain lexemes like Problem (“problem”) and Entwicklung (“development”) are preferentially used as complex anaphors due to their semantic content. Our results for the pronominal group show that dies (“this”) and das (“that”) frequently function as complex anaphors whereas es (“it”) hardly occurs as complex anaphor. We analysed 50 occurrences of all three types. Out of 50 occurrences of each pronoun type we found 45 (90%) respectively 34 (68%) complex anaphors for dies (“this”) and das (“that”) as subjects in the Vorfeld while there were no such occurrences (0%) of es (“it”). We only analysed subject occurrences of es, das and dies as the only possibility for es to be in the Vorfeld is being the subject. We will further investigate this issue in section 3. With respect to the lexical group we could confirm our observation that NPs containing a demonstrative determiner (demNP) are always anaphoric and more likely to occur as complex anaphors as other definite NPs (defNP). We analysed 250 occurrences of demNPs and 250 of defNPs and we found 54 (21,6%) complex anaphors in the demNP set whereas the set of defNPs only contained 11 (4,4%) complex anaphors. Moreover, our analysis of lexems shows that certain lexems occur more often as complex anaphors than others, s. table 1. In section 2.1.3, we will further discuss specific lexemes with respect to their potential to change the ontological type of the referent. Such an ontological change is only one of several types of the complexation process that are discussed in the next section. 5. There is no exact correspondence between German dies / das and English this / that, however, both pairs seem to indicate different grades of proximity. Cf. Lenz (1997). 6. As the following example shows es can be the subject but not the object in the Vorfeld: 1. Das Kind ging zum Kaufmann. 2. Es kaufte sich ein Eis. 2.’ *Es habe ich gesehen. (1. The child went to the shop. 2. It bought ice-cream. 2.’ *It I saw.) where in 2. es can neither refer to the child nor to the event described by the previous sentence.
6 32%
19
no anaphor
total
2
0 0%
2 100%
0 0%
dem
9
2 22%
2 22%
5 56%
def
6
0 0%
3 50%
3 50%
dem
Prozess “process“7
48
26 54%
10 21%
12 25%
def
14
0 0%
11 79%
3 21%
dem
Entwicklung “development“
21
14 67%
1 5%
6 29%
def
2
0 0%
2 100%
0 0%
dem
Zustand “state“
45
18 40%
7 16%
20 44%
8. German die Sache corresponds with the whole thing referring to a complex referent.
3
0 0%
2 67%
1 33%
dem
Problem “problem“ def
Prozess in the sense of “court case, trial” (36 defNP and 2 demNP items) was not accounted for.
7 37%
complex anaphor
7.
6 32%
nominal anaphor
def
Aktion “action“
Table 1. The anaphoricity of certain lexems
8
8 100%
0 0%
0 0%
def
7
0 0%
7 100%
0 0%
dem
Erkenntnis “insight”
24
7 29%
9 38%
8 33%
def
2
0 0%
1 50%
1 50%
dem
Sache8 “thing”
The function of complex anaphors in texts
Manfred Consten, Mareile Knees and Monika Schwarz-Friesel
2. Types of complexation processes 2.1
Ontological types
Discussing semantic aspects of complex anaphors requires the analysis of (1) the status of the referents referred to by the anaphor, (2) semantic features of the anaphoric expression, (3) the relationship between (1) and (2). Complex referents are propositionally structured entities, that have been topic of several detailed analyses: The Davidsonian notion of an “event argument” (Davidson 1967 and Parsons 1990) is adopted by several current semantic theories, especially DRT (cf. Asher 1993, 2000, Higginbotham 2000, Maienborn 2003, 2004, Kratzer 1995, forthc.). There is no final agreement on the ontological categorisation of such referents as events, states, processes or situations (cf. Vendler 1967, Dowty 1979, Kim 1969, 1976, as well as Asher 1993, 2000, and Maienborn 2003). Nonetheless, we get the following classification showing the increasing abstractness of the proposed ontological types. degree of ontological abstractness category high low
proposition (pp) fact (f) state (s) process (p) event (e)
[dependent on world] [-dynamic, -telic /dependent on world and time] [+dynamic, -telic] [+dynamic, +telic]
Figure 1
The anaphoric accessibility of these propositionally structured referents is a criterion for their ontological status (cf. Maienborn 2003: 59–62). In order to distinguish events and processes from states it is tested whether a complex anaphor can be combined with the verb happen (s. (4) and (5)):
(4) Arthur played the piano. This happened while…
(5) Arthur owned a bike. / The apples cost 3 Euro. *This happened while…
9. A further distinction is made between Kimian (K-) and Davidsonian (D-) states: Davidsonian states are expressed by verbs like sit, lie, hang as well as sleep, wait, stick. Kim’s states are expressed by know, believe, love, possess, cost, resemble. K-states are abstract objects but in contrast with facts, they are bound to an experiencer and time (cf. Maienborn 2003: 121; Asher 2000: 133f).
The function of complex anaphors in texts
In general, situational arguments are accessible by complex anaphors. Nonetheless, the discussion of the examples like the following (Maienborn 2003: 111–113) shows that in many cases the ontological status of the complex referents only becomes explicit by the complex anaphoric reference, cf:
(6) Arthur sleeps / lies in bed. This shows / proves that he played the whole afternoon…
(7) Arthur sleeps. This has already lasted for the whole morning.
Due to the verbal context the complex anaphor this in (7) denotes an entity with a temporal extension. It cannot refer to a fact since facts are not temporally bound by definition (they are only bound to a world, namely the world in which the proposition is true). Maienborn (2003: 112) describes (7) as reification of the fact because of the temporal dimension. That means that the ontological status of the referent in (7) gets changed by the complex anaphor. In contrast to this view, we assume that an ontology-changing complexation does not occur in (7) but in (6). The antecedent sentences in (6) and (7) denote states – and these states feature an inherent temporal dimension which in (7) gets picked up by the complex anaphor whereas in (6) the complexation process evokes an abstraction. By means of this abstraction the speaker makes the referent lose its temporal dimension and become a fact. In 3. we will postulate that this abstraction is an essential function of anaphoric complexation while the opposition (i.e. making the referent less abstract) cannot occur. In order to investigate these assumptions, let us first have a closer look at the complexation process. We distinguish between three types of complex anaphoric reference (cf. 2.2, 2.3, and 2.4): 2.2
Maintenance by neutral anaphors
The anaphor itself is neutral with respect to ontological types. For this reason, the discourse entity established by the anaphoric process usually keeps the ontological type denoted by the antecedent which, of course, has to be compatible with the context of the anaphor.10
(8) zneutral ≈ x
In most of these cases the anaphor is a demonstrative pronoun which is ‘neutral’ due to its semantic weakness. However, there are lexikal NPs like the whole thing 10. ‘≈’ assigns a complex referent (x) to an anaphor (z) (cf. Asher 1993: 145). “zneutral” stands for an anaphor that does not denote any ontological type due to its lexical meaning, whereas in (15) “zx” indicates an anaphor with specific ontological meaning in terms of fig. 1. “x” and “y” are objects of respective ontological type.
Manfred Consten, Mareile Knees and Monika Schwarz-Friesel
that are used like pronominal complex anaphors as flexible expressions with a vague descriptive content. We will give corpus examples for three different ontological types of referents:11 (a) Events (9) Mehrere kleine Anteilseigner gaben aber Widerspruch zu Protokoll und kündigten Anfechtungsklagen an. Dies könnte die Eintragung in das Handelsregister verzögern. (Tiger-Corpus, 3718f) [Several minor chairholders filed an objection and announced a legal challenge.]e [This]n might delay the record in the commercial register.
(b) Processes (10) Das beste Denkmal für den Toten ist, sein Werk des Friedens fortzusetzen. Dies kann schwierig werden. (Tiger-Corpus, 8372f) The best memorial to the dead person is [to continue his works of freedom]p. [This]n can be difficult.
(c) States (11) Auch gab es Leute aus meinem Umfeld, die mich auf der Straße nicht mehr gegrüßt haben. Das begann bereits, als ich 1990 in Potsdam mit der Auflösung des Stasi-Archivs betraut war. (Tiger-Corpus, 4901f) [There were also people among my acquaintances who didn’t say hello to me any longer.]s [This]n has already begun in 1990 when I was responsible for the termination of the Stasi archive in Potsdam.
Even though the anaphor is neutral with respect to ontological types, there are cases where a different type is fixed within the context provided by the sentence the anaphor is part of. In (12) and (13) the antecedent denotes an event respectively a state while the referent for the neutral anaphor this is understood as a fact due to the lexical meaning of the verb: if an entity serves as a proof or as a reason for another entity it must be factual. (12) [The Americans tried to invade the building but were forced back by shots from the top floor.]e [This]n proves that the situation in Bagdad isn’t under control yet. (13) Panic nahm die Gelegenheit wahr, um eigene Aktien im Wert von 13 Millionen Dollar abzugeben. Dies brachte ihm zahlreiche Aktionärsklagen ein, die zum Teil heute noch anhängig sind. (Tiger-Corpus, 1547f)
11. Most of our examples are taken directly from the Tiger-Corpus as indicated. Those without any reference are based on natural language examples from the Tiger-Corpus or other newspaper texts, but modified in order to show range of possible and impossible ontological change.
The function of complex anaphors in texts
[Panic grabbed the opportunity to sell his shares worth 13 millions dollars.]e [This]n resulted in several claims of share holders which have been still pending.
Similarly, in (14) the referent is denoted as a proposition by the verb assume that creates an attitude context: (14) [The amount of industrial jobs decreases, while the importance of the service sector is growing at the same time.]p At least [this]n is assumed by some economists.
2.3
Maintenance by lexical anaphors
With this type of complex anaphors, the ontological status of the referent stays the same during the anaphoric process since the antecedent and the anaphor denote the same ontological type. (15) zx ≈ x
As mentioned above in 1.2, certain nouns seem to be specific with respect to an ontological category due to their lexical meaning. As a simple example, NPs like this event, this incident, this action, this accident are likely to denote an event-type object since a dynamic element as well as boundness to a certain actor, time and place seem to be part of their semantic structure. However, for most of the nouns it is anything but easy to fix an ontological category independent from context. Like in 2.2.,we present here some corpus examples for events, processess and states. (a) Events (16) Ermittler des Bundeskriminalamts (BKA) in Wiesbaden haben 1992 mehrere Tonnen sogenannter weicher Drogen wie Haschisch und Marihuana in die Niederlande eingeführt, um einer internationalen Drogenmafia auf die Spur zu kommen. Nach niederländischen Angaben waren es 30 Tonnen. Im Anschluß an diese Aktion seien 13 tatverdächtige Drogenhändler festgenommen worden, sagte BKA-Sprecher Willi Fundermann am Mittwoch in Wiesbaden. (TigerCorpus, 11111) [Investigators of the Federal Criminal Police Office (BKA) in Wiesbaden imported 1992 several tons of so called soft drogs like hashish and marihuana into the Netherlands in order to get onto an international drug mafia.]e According to Dutch authorities there were 30 tons of imported drugs. After [this action]e 13 drug dealers under suspect have been arrested, said BKA spokesman Willi Fundermann in Wiesbaden on wednesday.
The event referent denoted by the antecedent sentence is picked up by the anaphor this action denoting an event as well. In contrast, the sentence between the ante-
Manfred Consten, Mareile Knees and Monika Schwarz-Friesel
cedent and the anaphor represents a state. For reasons that are discussed in section 3.1, its referent is not accessible for the anaphor. (b) Processes (17) Nach den langen Jahren der Unterdrückung durch die kommunistischen Macht haber suchten die Kirchen nun „in der neuen Situation der Freiheit“ nach „eigenen Wurzeln“ und wollten ihren Weg nicht nach westlichem Muster gestalten. Die EKD müsse sich dieser Entwicklung „dringend“ stellen, mahnte der Präses der Evangelischen Kirche im Rheinland, Peter Beier. (Tiger-Corpus, 8235f) [After a long period of oppression by the communistic powers, the churches are looking for their “own roots in the new situation of liberty” and they do not want to arrange their way following the western example.]p The EKD has to face up to [this development]p urgently, Peter Beier, chairman of Rhineland Evangelic Church, reminded.
In (17) the antecedent sentence denotes a process since the duration of the referent is focussed and it is presented as unaccomplished. This corresponds with the anaphoric expression this development. (c) States (18) Auf einmal befanden wir uns in einem großen Raum der Freiheit, in dem wir unbewußt unter dem Eindruck standen, daß das Leben keine Banden mehr hätte. Ich vergleiche diesen Zustand oft mit der Gemütsverfassung eines Menschen, der nach einer längeren Haft plötzlich freigelassen wird. (Tiger-Corpus, 11412f) [All at once we were in a big room of freedom in which we unconsciously got the impression that life is unlimited.]s I often compare [this state]s with the mood of someone who got set free suddenly after a long period of arrest.
Here, the complex referent picked up anaphorically is the state of being in a big room of freedom. In all these cases the ontological type assigned by the antecedent corresponds with the ontological type marked by the anaphor. 2.4
Ontology-changing complexation
This is the most interesting of our three categories. Like 2.3, it deals with full-NPanaphors which specify ontological types due to their lexical meaning. In contrast to 2.3, the anaphoric expression denotes another ontological type than its antece dent. Consequently, the anaphorical process changes the ontological type of the referent and in doing so it creates a new kind of discourse object. (19) zx ≈ y
The function of complex anaphors in texts
Cf. some selected examples: (a) Event becomes state (20) [Über die TV-Sendung „Presseclub“:] Nur zweimal kamen in diesem Jahr frauenpolitische Themen auf den Tisch und die auch nur dank der außerordentlichen Ereignisse wie UN-Frauenkonferenz und des berüchtigten Quotenurteils der EuGH-Richter. Als die Journalistin Charima Reinhardt diesen Zustand anläßlich der Diskussion über die damals unmittelbar bevorstehende UN-Frauenkonferenz in Peking kritisierte, stellte der Presseclub-Moderator Fritz Pleitgen zwar vage eine Änderung in Aussicht [...] (Tiger-Corpus, 32323f) [On “Presseclub”, a TV discussion forum:] [This year feminist politics were only discussed twice just because of special events like the UN women conference and the notorious judgement on the contigent of women by the European Court of Justice.]e When the journalist Charima Reinhardt criticised [this state]s on the occasion of the discussion about the forthcoming UN women conference in Peking, “Presseclub” presenter Fritz Pleitgen announced some changes […]
Here, the antecedent sentence denotes two incidents, namely two discussions. As these referents are bound to certain actors and spatiotemporally limited to a certain place, a beginning and an end, they have to be categorised as events. However, the anaphor this state seems to refer to a more abstract referent; it focuses on the lack of interest in women rights which is meant to be an enduring state of the TV programme. Thus, the referent is subject to an abstraction process evoked by anaphoric reference. (b) Event becomes fact (21) Ghosh hat aus der enormen Stofffülle einen virtuos erzählten Roman über die im Westen so gut wie unbekannte, dennoch höchst dramatische Geschichte Birmas verfasst. Diese Tatsache ist mindestens genauso verblüffend wie die Kultur jenes südostasiatischen Landes, das nicht nur seiner unzähligen Pagodendächer wegen im wörtlichen und metaphorischen Sinn das „goldene Land“ genannt wird. (taz 20.3.2001, p.15 ) [From the enormous abundance of the material Gosh composed with virtuosity a novel about the extremely dramatic history of Birma which is fairly unknown in the West.]e [This fact]f is at least as amazing as the culture of this south east Asian country that is literally and metaphorically named the “golden country”not only because of its endless pagoda roofs. (22) [The Americans tried to invade the building but were forced back by shots from the top floor.]e Rumsfeld had to explain the consequences resulting from [this fact]f during a press conference in the afternoon.
Manfred Consten, Mareile Knees and Monika Schwarz-Friesel
Like in (12), the event referred to in the antecedent sentence is picked up as a fact; however, in (21) and (22) the anaphor itself initiates this change by its lexical meaning. Again, in some cases we have to consider the context. Here, the lexical meaning of the complex anaphor has to fit to the context: (23) …*[This fact]f happened yesterday while Mr. Rumsfeld visited Bagdad.
With this (fictitious) example it is evident that fact does not allow an event-reading that, on the other hand, the verb happen requires. (c) State becomes fact
(24) Gleichwohl: Branchen- und Industriepolitik an sich sind nicht zu kritisieren und aus nordrhein-westfälischer Sicht kann gesagt werden: Sie ist nicht nur wegen des Zieles einer ausgewogenen industriellen Struktur, sondern auch aus Gründen des sozialen Zusammenhalts der Gesellschaft unverzichtbar. Diese Erkenntnis ist aber weder neu noch originell, und wer Strukturwandel betreiben will, weiß aus Erfahrung, daß sozialdemokratische Wirtschaftspolitik weit mehr ist und sein muß als Branchenpolitik. (Tiger-Corpus, 16614f) [Anyhow: Industrial and trade policy has not to be inherently criticised and from Nordrheinwestphalian point of view it can be said: It is not only essential because of the goal of balanced industrial structures but also because of reasons of strong social solidarity.]s [This insight]f is neither new nor inventive and anyone who wants to practise structural changes knows from experience that social democratic economic policy has to be much more than industrial and trade policy.
(d) Process becomes fact / negatived fact / proposition (25) [The amount of jobs decreases, while the importance of the service sector is growing at the same time.]p [This fact / This insight]f / [This misbelief]negf / [This assumption]p determined economical sciences of the 20th century. The state of being essential in (24) as well as the processes of decreasing and growing in (25) are marked as factual by using the anaphor fact (it would be contradictional to say This fact isn’t true). The opposite logical value is assigned by misbelief in (25), while an assumption (as well as an estimation, cf. the next example) can be either true or false. For this reason, we regard epistemic nouns like assumption and estimation as denoting a proposition that by definition is not determined to be true or false. (e) State becomes proposition (26) In der Gemeinschaft Unabhängiger Staaten (GUS) fehlt es an vielen notwendigen politisch-rechtlichen und ökonomisch-institutionellen Voraussetzungen für effektivere Hilfe und Kooperation des Westens. Hier wiederum mangelt es an Opferbereitschaft sowie der Fähigkeit, Anstrengungen vernünftig zu bündeln und gezielt
The function of complex anaphors in texts
auf die für den Systemwandel in der Ex-UdSSR entscheidenden Schwerpunkte zu konzentrieren. Mit dieser Einschätzung warnt das Bundesinstitut für ostwissenschaftliche und internationale Studien vor zu hochgesteckten Erwartungen an den Sieben-plus-eins-Gipfel der Regierungschefs der führenden Industrieländer und des russischen Präsidenten Boris Jelzin. (Tiger-Corpus, 3108) [In the Commonwealth of Independent States (CIS) many necessary politicaljudical and economic-institutional conditions for more effective support and cooperation from West are missing. Here [in the West] in turn the readiness to make sacrifices and the ability to sensibly bunch endeavours and concentrate them on crucial emphases in order to bring about the change of the political system in the Ex-UdSSR are lacking.]s With [this estimation]pp the Federal Institute for eastern and international studies warns about placing too high hopes on the Seven-plus-One-summit conference attended by the heads of government of the leading developed countries and by the Russian president Boris Jelzin.
3. Processing complex anaphors 3.1
Constraints on ontology changing complexation
The examples in section 2.1.3 are evidence for our claim that anaphoric complexation processes can change the ontological type of a prementioned referent and, in doing so, establish a new kind of discourse object. This process is subject to a constraint that has to be mapped out in terms of categories of abstractness, cf.: (27) [The earth turns about the sun.]p [This process]p / [This state]s will presumably last for 7-109 years. [This fact]f is well known since the Middle Ages. Researchers of the Vatican were not allowed to examine [this possibility].pp / *[This event]e…
As the example shows, anaphorical complexation can shift referents of any ontological type to a discourse entity of either the same ontological type or an ontological type that is more abstract. They cannot be shifted to a discourse entity that is less abstract. Thus, anaphorical complexation can be a process of increasing abstractness (in terms of fig. 1). (28) *zy ≈ x if x > y (“if x is higher on abstractness scale than y“)
This ‘abstractness-constraint’ can serve to explain ontological based resolution of ambiguous complex anaphors: (29) [Schröder promoted Mehdorn;]e [after all, Mehdorn is Schröder’s buddy]s.
Manfred Consten, Mareile Knees and Monika Schwarz-Friesel
(a) [This crude act]e distressed the Opposition. / (b) [This friendship]s distressed the Opposition.
In (a), it is our intuition that the complex anaphor this crude act refers to the evente specified in the first part of the antecedent.12 But (as shown in (b)) the second clause (referring to a states) is structurally accessible by complex anaphoric reference, too. Disambiguation is based on the lexical meaning of the complex anaphor: (a) this act characterizes an event (whereas (b) this friendship characterizes a state). Therefore, due to the abstractness constraint the complex anaphor in (a) is assigned to the first antecedent clause that refers to an event as well. These kinds of disambiguation are difficult to explain in terms of purely structural constraints (like DRT-approaches). 3.2
A resolution model
As we have seen so far, ontological features of referents seem to be essential for anaphoric complexation processes and have an important effect on the resolution of complex anaphors. In the resolution model given in this section we will integrate such features as well as procedural aspects in using a combination of (simplified) DRS and cognitive Text World Models (Schwarz 2000a, 2001).13 According to our definition in 1., we assume that complex anaphors condense propositionally structured referents and thereby establish new discourse entities on the text-world level. Thus, they differ from (direct) nominal anaphors as the latter refer to objects already introduced as discourse entities. DRT approaches do not reflect this difference as they assume that each incidence of an anaphor integrates a new discourse referent at the DRS (cf. the critical remarks in Löbner 1985: 320, Cornish 1999: 186, and Consten 2004: 61). Like DRT, we distinguish between different levels in our model: the text-semantic level and the text-world level. Textual structures are introduced at the text-semantic level; they establish discourse entities at the text-world level by activating the 12. With a causal reading for the after all clause, it is also possible to assign the following structure, since the referent of the second clause, giving a reason for the event, is considered part of the event: [Schröder promoted Mehdorn; [after all, Mehdorn is Schröder’s buddy]s.]e However, this doesn’t make any difference for our analysis. Furthermore, an event-reading of friendship as “friendly turn” shoul not be considered with our example. 13. Schwarz (2000a: 45-47) defines Text World Model from a receptional point of view as a mental representation, that is constituted by the proceeding referencialisation and elaboration of the text and the activation of conceptual knowledge. The concept of Text World Models is similar to Johnson-Laird’s (1994) “mental models”, but, in contrast, differentiates propositional (“text semantic”) and mental (“text world”) levels.
The function of complex anaphors in texts
corresponding concept in long term memory (phase 1) (for a more detailed description of this process s. Schwarz 2000a: 46f). Initially, anaphors do not establish discourse entities at the text-world level but are interpreted at the text-semantic level where the appropriate part of the textual structure is (re)activated. In case of complex anaphors these textual parts are propositionally structured (phase 2). In phase 3 the complex anaphor establishes the referent as new complex discourse entity. Once the complex referent is established as a unified discourse entity by a complex anaphor, the discourse entity is accessible by personal pronouns (as it in the 3rd sentence),14 whereas the use of personal pronouns in the Vorfeld as a complex anaphor (as it in the 2nd sentence) is restricted (cf. Hegarty/ Gundel/ Borthen 2002, Hegarty 2003).15 This assumption is supported by the findings of our corpus study in section 1.2: the personal pronoun es (it) cannot function as complex anaphor at the German Vorfeld position whereas complex anaphoric das and dies (this and that) are fine at this position.16 (30) [The earth turns about the sun.]p [This process]p / [This]n / *[It] will presumably last for 7–109 years. [It] might, however, terminate a few years earlier.
Let us now look at example (29a) again. In phase 1, the nominal expressions Schröder, Mehdorn and the opposition introduce referents at the text-semantic level (w, x... in (31)). Moreover, they directly establish discourse entities in the textworld level (W, X…). In contrast, propositional expressions introduce complex referents (like events, states etc.) only into the text-semantic level (e1, s1…) but they do not establish discourse entities at the text-world level. Due to its lexical meaning, the complex anaphor z (act) of type e (“event”) denotes an event type referent; but in this phase, like propositional expressions, they do not establish a 14. It in the 3rd sentence is not a complex anaphor since it is not assigned to a propositionally structured antecedent but to a NP-antecedent (This process / This) by which a unified discourse entity has already been established. 15. Hegarty (2003: 1f) assumes that events introduced by a clause are immediately accessible by personal pronouns since they are in focus merely due to their ontological status. However, some of our data does not support his claim. We have no evidence that ontological states of referents are determinants of a salience hierarchy. 16. There are still problems in defining the interaction between personal pronouns as “weak“ referential means (versus d-pronouns, cf. Consten 2004) and the Vorfeld as a position that is assumed to be preferred for topics, cf. Eckert 1998. While there are few (but somehow idiomatic) examples of complex anaphoric es (“it”) as accusative objects in the Mittelfeld, complex anaphoric es as subject in the Mittelfeld seems to be as odd as in the Vorfeld, cf.: “Diese Defizite abzubauen, wird erkennbar schwierig und langwierig sein. Gleichwohl ist dies [??es] die wichtigste Aufgabe fortschrittlicher Politik.” (Tiger-Corpus 18153f). [To abolish this deficiency]pp will be difficult and tedious. However, [this]n [??it] is the most important task of modern politics.
Manfred Consten, Mareile Knees and Monika Schwarz-Friesel
discourse entity. In phase 2, the anaphor (ze) (following the abstractness-constraint (28) (ze ≈ e1; *ze ≈ s1)) is assigned to the adequate prementioned referent e1 and (re)activates it. Finally, the anaphor establishes e1 as a unified discourse entity E1 at the text-world level (phase 3). At the same time, this discourse entity is evaluated due to the descriptive content of the anaphor (phase 3’). (31) Resolution model W
text-world level
Schröder (w)
textsemantic level
X
establish
e1 – promote (w, x) Mehdorn (x) s1 – buddy_of (x, w)
W X Y *z e ≈ s1 z e ≈ e1
(re)activates e1 – promote (w,x) crude act (e1)
crude act (z e ) e2 – distress (y, z
Y
e)
e2 – distress (y, e1) Opposition (y)
W X Y E1
establishes ze
1
2 (re s o lv ing the complex anaphor)
evaluates
e1 – promote (w,x)
e1 – promote (w,x)
crude act (e1)
crude act (e1)
e2 – distress (y, e1) Opposition (y)
Opposition (y) P ha s e
W X Y E1 (val: CRUDE)
3 (establishing e1 as discourse object)
e2 – distress (y, e1) Opposition (y) 3’ (evaluating E1)
Legend: x, y: nominal referents on text-semantic level e1, s1: complex referents on text-semantic level indicated as “event”, “state” etc. z e , z s : anaphors indicated as “event”, “state” etc. W, X: nominal discourse entity established in text-world level E, S: complex discourse entity indicated as “event”, “state” etc., established in text-world level
4. Continuity and progression – Textual function of complex anaphors As we have seen from our model in 3., complex anaphors serve basically as means of establishing new discourse objects by re-activating referents on the text-semantic level. Conceptual based cognitive strategies, especially those based on ontological categorising, have to be applied in order to extract the proper object from the preceding text represented at the text-semantic level. Accordingly, complex anaphors serve a double purpose as a means of textual continuity and progression: Firstly, by condensing prementioned referential structures they stabilise the coherence structure that is already established – which means, they are thematical; secondly, by setting up new entities at the text-world level they contribute to the progression of information flow and the incremental constitution of the text world model. Hence, they are rhematical at the same time.
The function of complex anaphors in texts
In this respect complex anaphors are similar to so-called indirect anaphors. These anaphors, as a means of underspecified reference (Schwarz 2000a, 2001; Consten 2003, 2004), go without an explicit coreferential antecedent, but have a “systematical relationship to entities of the preceding text structure” (Schwarz 2000a: 98), called anchors. Thus, resolving indirect anaphors requires the accessibility of a suitable textual anchor on the text-semantic level as well as enriching the text world model with a plausible link between a prementioned anchor and the anaphor. Cf. the following example: (32) Nach dreijähriger Vorarbeit zeigt die Londoner Academy eine überwältigende [...] Ausstellung, die zum Ausgangs- und Schwerpunkt des weitgespannten Festivals Africa 95 [...] geworden ist. Der Titel klingt schlicht: Afrika – Die Kunst eines Kontinents. In dem Katalogwälzer mit gut drei Kilo Informationen macht der aus Ghana stammende Harvard-Historiker Kwame Anthony Appiah darauf aufmerksam, daß er zwei Begriffe enthält, deren Definition unklar und deren Verknüpfung problematisch ist: [...] (Tiger-Corpus 24141) After three years of preparation the London Acadamy presents an overwhelming exhibition that has become basis and main focus of the broadly based Africa 95 festival. The title sounds simple: Africa – The Art of a Continent. In the huge catalogue containing about three kilogrammes of information Harvard historian Kwame Anthony Appiah, born in Ghana, points out that it comprises two terms with ambiguous definition and problematic relationship.
The definite NPs the title and the huge catalogue… have to be considered as indirect anaphors, since they are conceptually related to the preceding expression exhibition which serves as an anchor: it is part of our extra-linguistic, cultural know ledge that exhibitions usually have titles and that there are catalogues offered to the audience. Thus, readers have to make the inference that the title / the huge catalogue refer to the TITLE / CATALOGUE OF THE PREMENTIONED EXHIBITION. It is also worth mentioning that the referent TITLE, once introduced by the indirect anaphor, can be picked up again by the personal pronoun it in the last clause, whereas the use of pronouns (personal pronouns as well as demonstrative ones) as indirect anaphors is restricted (cf. Cornish, this volume). This issue marks another analogy to complex anaphors. Consequently, some researchers have categorised complex anaphors as indirect ones (Conte 1996), whereas Dahl / Hellmann (1995) classifiy them as direct anaphors (cf. Schwarz 2000a: 129f, who regards complex anaphors as a “mixed type”.) However, our corpus data shows that coreference as a criterion for distinguishing direct from indirect anaphors is generally applicable to nominal based as well as complex anaphors, that means, within the category of complex anaphors there
Manfred Consten, Mareile Knees and Monika Schwarz-Friesel
are both direct and indirect anaphors (cf. Schwarz-Friesel/ Consten/ Marx 2004: 81f). Our examples of complex anaphors discussed so far refer to propositional referents already mentioned in the pretext, that is, there is an explicit antecedent the anaphor is coreferent with. In contrast, the following example lacks such a coreferential relationship, rather the referent for the anaphor has to be inferred from a preceding anchor by using conceptual strategies. (33) FBI-Agent Anderson kennt seine Heimat [Mississippi] nur zu gut. Im Sommer 1964 sucht er mit seinem Kollegen Ward im Yessup County nach drei vermissten Bürgerrechtlern – und trifft dabei auf eine Wand des Schweigens. Vor Ort regiert der Ku-Klux-Klan, der sogar vom Sheriff und seinen Leuten unterstützt wird. Nur Mrs. Pell, die Frau des Hilfssheriffs, zeigt sich kooperativ. Damit macht sie sich keine Freunde... Am 21. Juni 2005, 41 Jahre nach der Tat, wurde der 80-jährige Edgar Ray Killen [...] des dreifachen Totschlags für schuldig befunden und zu 60 Jahren Haft verurteilt. (TV guide „TV Spielfilm “about the movie „Mississippi Burning“) FBI-Agent Anderson knows his homeland only too well. In the summer of 1964 he and his colleague Ward are searching for three missing civil rights activists at Yessup County. There he faces a wall of silence. At the face, the Klu Klux Klan rules and is supported even by the sheriff and his men. Only Mrs. Pells, the deputy’s wife, is willing to cooperate. Thus, she doesn’t make herself popular… On 2005–06–21, 41 years after the deed, 80-years old Edgar Ray Killen […] was found guilty of three homicides and sentenced to 60 years imprisonment.
This is an example of an indirect complex anaphor based on inference. In the preceding text it is not told explicitly that the three missing men have been slaughtered, but this fact becomes clear when resolving the anaphor the deed which is referring just to this act of slaughter. (34) Auf mein Drängen hin entschied sich Bethe, nach Los Alamos zu kommen, [...] jedenfalls verstand ich ihn so. Währenddessen rief Oppenheimer an und lud uns zu sich nach Princeton ein. Ich sagte zu Bethe: “Nach diesem Gespräch werden Sie nicht mehr kommen.“ (Kipphardt, Heinar. In der Sache J. Robert Oppenheimer. Frankfurt am Main: Suhrkamp, 1964. Cited from Koeppel 1993: 87.) Bethe decided to come to Los Alamos at my urging, at least I think that was what he meant. Meanwhile, Oppenheimer called and invited us to stay with him at Princeton. I told Bethe: “After this interlocution you won’t be willing to come any more.”
In this second example the referent for this interlocution has not been explicitly prementioned, but can be derived from the fore-text. Readers will easily bridge the referential gap the text leaves between the invited-sentence and this interlocution since
The function of complex anaphors in texts
the latter results from the former due to a conceptual INVITATION-scheme. Thus, the complex anaphor this interlocution is interpreted as INTERLOCUTION BETWEEN OPPENHEIMER, BETHE AND THE SPEAKER DURING THEIR STAY AT PRINCETON by re-activating the INVITATION-referent that has just been introduced on the text-semantic level. Since both related referents – the event of inviting people to come and the event of having an interlocution – are propositionally structured and, furthermore, the anchor expression is not a mere NP, but a sentence, this example fulfills our criteria for complex anaphors discussed in section 1. As the example shows, the criteria “nominal based“ versus “complex“ and “direct“ versus “indirect” exist independent from each other. Another insight our data has delivered is that the strict dichotomy of thematic versus rhematic entities in texts has to be abandoned (Schwarz 2000b). Complex anaphors – regardless of directness or indirectness – are means of textual continuity and progression at the same time, which is a function they have in common with indirect nominal based anaphors. Furthermore, even ‘simple’ direct anaphors with nominal antecedents sometimes contribute to textual progression by enriching the text world model with additional information about a prementioned referent, cf.: (35) Yesterday I met Mr. Miller. The old-fashioned socialist was singing “the Internationale” again until some citizens made the old man shut up.
Similar to (35), some lexical complex anaphors evoke due to their lexical meaning an additional evaluation (cf. (36)) or a meta-discursive specification (cf.(37)) of the referent while establishing it as a discourse object. (36) Ratzinger has been elected pope. This fortune / This catastrophe… (cf. Schwarz 2000a,b; Schwarz-Friesel/ Consten/ Marx 2004) (37) The earth turns about the sun. This thesis / This claim / This proposal / This blasphemous misbelief / This joke Johannes Kepler made when he was drunk… (cf. Consten 2004: 34)
Consequently, the coincidence of thematic and rhematic status is a feature that can occur with all types of anaphors. Indirect anaphors act in a rhematic way by introducing new referents, these can be nominal structured like in (32) as well as complex like in (34). In case of complex anaphors, the rhematic function is based on condensing complex referents on the text-semantic level in order to establish them as manageable objects at the text-world level.
Manfred Consten, Mareile Knees and Monika Schwarz-Friesel
5. Summary We have defined complex anaphors as anaphors that condense propositional referents establishing them as unified discourse entities. In doing so, they fulfill a double textual function: On the one hand, they are an economical means of referential continuity and stability, since they are interpreted by reactivating and resuming prementioned antecedents or anchors. On the other hand, they contribute to informational progression, since they create new discourse objects by shifting the ontological status of the referent and / or by evaluating and categorising it. Consequently, anaphoric complexation has been described as a potential process of increasing abstractness with respect to ontological categories. Thus, we suggested a differentiation between neutral and ontology changing complexation and proposed an “abstractness constraint” which serves to explain the resolution of certain kinds of ambiguous complex anaphora not solved by current approaches. Our sketch of a pro cess model of anaphoric complexation is able to integrate cognitive aspects of language processing into a semantic framework. The aim of such an integration is to explain the wide variety of anaphora as an important means of establishing coherence in texts.
6. References Asher, N.1993. Reference to Abstract Objects in Discourse. Dordrecht: ������������������ Kluwer. Asher, N. 2000. Events, facts, propositions and evolutive anaphora. In Speaking of Events, J. Higginbotham, F.Pianesi and A.C. Varzi (eds),123–150. Oxford: OUP. Consten, M. 2003. Towards a unified model of domain-bound reference. In Deictic Conceptualisation of Space, Time and Person. F. Lenz, 223–248, Amsterdam/Philadelphia: John Benjamins. Consten, M. 2004. Anaphorisch oder deiktisch? Zu einem integrativen Modell domänengebun dener Referenz [LA 484]. Tübingen: Niemeyer Conte, M.E. 1996. Facts, Events, Propositions in Anaphoric Encapsulation. Talk at Indirect Anaphors-Workshop in Lancaster (UK), July 1996. Cornish, F. 1999. Anaphora, Discourse, and Understanding: Evidence from English and French. Oxford: Clarendon. Dahl, Ö. and Hellmann, C. 1��� 995. What Happens When We Use An Anaphor? Ms. Dept. of Linguistics, Stockholm. Davidson, D. 1967. The logical form of action sentences. In The Logic of Decision and Action. Pittsburgh Press, N. Resher (ed.), 81–95. (Reprint in Davidson, 1980. Essays on Action and Events, 105–122.Oxford: Clarendon Press). Dowty, D.R. 1979. Word Meaning and Montague Grammar: The semantics of verbs and times in ������������������ Reidel. generative semantics and in Montague’s PTQ. Dordrecht: Eckert, M. 1998. ������������������������������������������������ The German topic position and null anaphora. In Proceedings of the 2nd Colloquium on Discourse Anaphora and Anaphor Resolution, S. Botley and T. McEnery (eds). Lancaster: University of Lancaster.
The function of complex anaphors in texts
Francis, G. 1994. Labelling discourse: An aspect of nominal-group lexikal cohesion. In Advances in Written Text Analysis, M. Coulthardt (ed.), 83–101. London: Routledge. Fraurud, K. 1992. Situation ������������������������������� reference. What does ‚it’ ������������������ refer to? In Processing Noun Phrases in Natural Discourse, K. Fraurud. PhD dissertation, Stockholm University. Greber, E. 1993. Zur Neubestimmung von Kontiguitätsanaphern. Sprachwissenschaft 18(4): 361–405. Halliday, M. and Hasan, R. 1976. ���� Cohesion in English. London: Longman. Hegarty, M. 2003. Type shifting of entities in discourse. Presentation at the First International Workshop on Current Research in the Semantics-Pragmatics Interface, Michigan State University. Hegarty, M. Gundel, J. and Borthen, K. 2002. Information structure and the accessibility of clausally introduced referents. Theoretical Linguistics 27(2–3): 163–186. Higginbotham, J. 2000. On events in linguistic semantics. In Speaking of Events, J. Higginbotham (ed.), 49–80. New York NY: OUP. Johnson-Laird, P. 1994. ���� Mental models and probabilistic thinking. Cognition 50: 189–209. Kim, J.�������������������������������������������������������������� 1969. Events and their descriptions. Some considerations. In Essays in Honor of Carl G. 198–215. Dordrecht: Reidel. Hempel, N. Resher et al.,���������������������������� Kim, J.����������������������������������������������� 1976. ���������������������������������������������� Events ���������������������������������������� as property exemplifications. In Action Theory. Proceedings of the Winnipeg �������������������������������������������������������� Brand and D. Walton (eds),159–177. Dordrecht: Reidel. Conference on Human Action, M. Koeppel, R. 1993. Satzbezogene Verweisformen: Eine datenbankgestützte Untersuchung zu ihrer Distribution und Funktion in mündlichen Texten, schriftlichen Texten und schriftlichen Fachtexten des Deutschen [Tübinger Beiträge zur Linguistik 386]. Tübingen: Narr. Kratzer����������������������������������������������������������� , A. 1995. ������������������������������������������������ Stage-level and individual-level predicates. ��� In The Generic Book, G.N. Carlson and F.J. Pelletier (eds),������������������������������������������������� 125–175. Chicago ��������������������������������������� IL: The Univ. of Chicago Press. Kratzer, A. Forthcoming. The Event Argument and the Semantics of Verbs. Cambridge MA: The MIT Press. Lenz, F. 1997. Diskursdeixis im Englischen [LA 369]. Tübingen: Niemeyer. Löbner, S. 1985. Definites. Journal of Semantics 4: 279–326. Lyons, J. 1977. Semantics. Cambridge: CUP. Lyons, J. 1989. Semantic ascent: A neglected aspect of synactic typology. In Essays on Grammatical Theory and Universal Grammar, D. Arnold, 153–186. Oxford: Clarendon. Maienborn, C. 2003. Die logische Form von Kopula-Sätzen. Berlin: Akademie-Verlag. ��� Existence: Syntax and semantics, I. Maienborn, C. 2004. On Davidsonian and Kimian states. In Comorovski and K. von Heusinger (eds). Dordrecht: Kluwer. Parsons, T. 1990. Events in the Semantics of English: A study in subatomic semantics. Cambridge MA: The MIT Press. Schmid, H.-J. 2000. English Abstract Nouns as Conceptual Shells. From corpus to cognition, Berlin: Mouton de Gruyter. Schwarz, M. 2000a. Indirekte Anaphern in Texten. Studien zur domänen-gebundenen Referenz und Kohärenz im Deutschen. Tübingen: Niemeyer. In Prosodie – Struktur – InterpretaSchwarz, M. 2000b. Textuelle Progression durch Anaphern. ��� ������������������������������������������������ Dölling and T. Pechmann (eds), 111–126. ����� Leiption [Linguistische Arbeitsberichte 74], J. zig: Institut für Linguistik, Universität Leipzig. Schwarz������������������������������������������������������������������������������������ , M. 2001. Establishing coherence in text. Conceptual continuity and text-world mod������������� 15–24. ������ els. Logos and Language II(1): Schwarz-Friesel, M. Consten, M. and Marx K. 2004. Semantische und konzeptuelle Prozesse bei der Verarbeitung von Komplex-Anaphern. In ��� Flexibilität und Stabilität, I. Pohl (ed.). Frankfurt am Main: Peter Lang. Vendler, Z. 1967. Linguistics in Philosophy. Ithaca NY: Cornell University Press.
Manfred Consten, Mareile Knees and Monika Schwarz-Friesel Webber, B. 1991. Structure and ostension in the interpretation of discourse deixis. Language and Cognitive Processes 6: 107–135. Van Lambalgen, M. and Hamm, F. 2005. The Proper Treatment of Events. Malden MA: Blackwell. Zacks, J. M. and Tversky, B. 2001. Event structure in perception and cognition. Psychological Bulletin 127(1): 3–21.
Metaphorical anaphors A phenomenon of the semantics-pragmatics interface Helge Skirl Metaphorical anaphors are a special case of textual anaphors which occur in stylistic expressive texts. They have not yet been sufficiently investigated either by theories of metaphor or by theories of anaphors. In my paper I show that the identification and comprehension of metaphorical anaphors has to be characterized as a phenomenon of the semantics-pragmatics interface. They can best be described within the cognitive theory of text-world models. I also show different types of metaphorical anaphors and what authors do to facilitate their comprehension. The procedural costs of metaphorical anaphors, which has been shown by psycholinguistic research, are generally counterbalanced by the rich content they convey, but the specific features of that content have to be studied in more detail.
1. Introduction Most theories of textual anaphors do not deal with metaphors. Likewise, most theories of metaphor are not concerned with textual anaphors. But an anaphoric and a metaphorical use of an expression are not mutually exclusive. Metaphorical anaphors can be found quite often in natural language texts, especially in stylistic expressive texts of literature and journalism. Hence, both a general study of anaphors and one of metaphors should deal with metaphorical anaphors in order to provide a complete survey. A detailed examination may reveal important insights into the comprehension of textual anaphors. An example of a metaphorical anaphor is given in (1):
(1) A man lies on the sofa. The elephant is probably dreaming.
The NP the elephant in the second sentence of (1) signals by its definite article that its referent should be easily accessible in the reader’s mental representation of the text-
Helge Skirl
world. But in the first sentence, no referent compatible to the descriptive content of the elephant is mentioned. The indefinite NP a man introduces a singular referent man, who belongs to another ontological category than a literally possible referent of the NP the elephant. To establish a coherent reading, the reader will nevertheless take the elephant as a direct anaphor, referring to the same referent as the NP a man in the first sentence and thus establishing co-reference. So, a man is the antecedent for the anaphor the elephant. The reader will understand the descriptive content of the elephant as a metaphorical attribution of properties to the referent man. By means of the anaphor, the referent man is characterized as resembling, at least in some specific respect, an elephant. The resemblance could be one of appearance, size, behavior, mental disposition, etc. or of any other property attributed to an elephant. Note, that the attribution can include properties which are perhaps not suitable for the real animal, but only for an imagined, manlike elephant. It would depend on a larger linguistic co-text or communicative context, to get a more specific reading of the descriptive metaphorical content of the elephant in (1). The example shows that a metaphorical anaphor serves at the same time to establish referential continuity by re-mentioning a referent and to convey specific information about that referent. Note that, the latter is true only if the metaphorical anaphor itself introduces the metaphor. If the metaphor is already introduced, as in the second sentence of (2), the metaphorical anaphor merely establishes coreference, but it does not convey new information:
(2) A man lies on the sofa. He is an elephant. The elephant is probably dreaming.
In my paper I want to show that the identification and comprehension of metaphorical anaphors has to be characterized as a phenomenon of the semanticspragmatics interface. They can best be described within the cognitive theory of text-world models. I also want to show different types of metaphorical anaphors and what authors do to facilitate their comprehension. My aim is to highlight the importance of dealing with metaphorical anaphors in order to provide a full account of textual anaphors.
2. Metaphorical anaphors within anaphor research Common accounts of anaphors deal in extenso with cross-sentential anaphors, which establish co-reference without attributing specifying information. Consider (3):
(3) A man lies on the sofa. He/The man/The guy/The person is probably dreaming.
A man serves as antecedent and the possible anaphors are the pronoun he, the NP with the same head noun the man, the synonym the guy (which slightly differs in stylistic
Metaphorical anaphors
level, because it is colloquial) and the hyperonym the person. The function of the anaphors is solely to mention again a referent which is already established in the reader’s mental text-world. They therefore make a contribution to the referential continuity. But they don’t add new information to the mental representation of the referent. Detailed accounts of textual anaphors also deal with more informational complex anaphors such as specifying anaphors and indirect anaphors (for a survey see Schwarz 1997, 2000, Schwarz-Friesel, in press; for another type of anaphor, complex anaphor, see this volume). The former are anaphors which contribute specific information about the referent by their semantic content, as in (4):
(4) A man lies on the sofa. The salesman/The neat 47-year old family man and pet lover is probably sleeping.
The referent man in (4) is characterized by his profession by the NP the salesman. The extensive NP The neat 47-year old family man and pet lover adds even more and different knowledge to the mental representation of the referent. Indirect anaphors are definite NPs which have no antecedent to which they co-refer. But nevertheless, they are textual anaphors, because they are understood due to a close association to an anchor expression in the preceding text. Most frequently, this association is a part-whole-relation in which the referent denoted by the indirect anaphor is a part of the referent denoted by the anchor expression. Indirect anaphors thus mention implicit referents and contribute thereby to the referential continuity of the text (Schwarz 2000 provides a comprehensive theory). As example (5) shows, their definite article signals rightly that the referent is easily accessible within the mental text-world model of the reader (see 3 below):
(5) A man lies on the sofa. The face is deathly pale./The arm-rest is dirty./The pillows are vermilion.
In (5), a man is the anchor expression for the comprehension of the indirect anaphor the face, the sofa is the anchor expression for both the resolution of the indirect anaphors the arm-rest and the pillow. The face is easily understood as denoting a part of the referent introduced by a man, and the arm-rest and the pillow are likewise understood as denoting parts of the referent introduced by the sofa. If one relates the two types of anaphors—specifying and indirect anaphors—to metaphorical anaphors, the result is as follows: Firstly, metaphorical anaphors can doubtlessly be seen as a particular case of specifying anaphors, because they add specific knowledge to the representation of the referent (provided, that they introduce the metaphor). And secondly, examples of indirect metaphorical anaphors, like the trunk in (6), often occur in expressive texts (cf. 3 below):
(6) A man lies on the sofa. He is a dreaming elephant. The trunk makes a lot of noise.
Helge Skirl
Indirect metaphorical anaphors can be a part of a complex combination of several metaphors in a text, which is called an “extended metaphor”. Those chains of metaphors are very common (cf. Tirrell 1989, White 1996). So, the important types of specifying and of indirect anaphors occur as metaphorical anaphors, too. However, as with general accounts of textual anaphors, the otherwise extensive research on metaphors has barely considered the phenomenon of metaphorical anaphors. Neither metaphorical specifying anaphors nor indirect metaphorical anaphors have been investigated sufficiently.
3. Metaphorical anaphors within metaphor research Accounts of metaphor mostly only deal with predicative metaphors and are not concerned with referential metaphors, which could also act as metaphorical anaphors. The most frequently documented examples—beside metaphorically used verbs and adjectives—are indefinite NPs used metaphorically in a predication as in the second sentence of (7):
(7) A man lies on the sofa. The man is an elephant.
In (7), it is asserted that the referent of the NP the man is a member of the category denoted by the NP an elephant. Though an elephant denotes an animal, it is used to predicate properties to a human being. To get a plausible reading of the sentence, the reader will understand the NP an elephant metaphorically. Note, that only the man is used referentially in the second sentence, an elephant is just the main part of the predication. The difference between the predicative and referential metaphorical use of an NP can clearly be seen in examples of apposition. Consider the examples (8) to (11):
(8) Mr. Smith, an elephant, lies on the sofa.
(9) Mr. Smith, the elephant, lies on the sofa.
(10) An elephant, Mr. Smith, lies on the sofa. (11) The elephant, Mr. Smith, lies on the sofa.
In (8), the apposition an elephant is an elliptic predication. This reading is triggered by the indefinite article. The full form of the predication would be the subordinate clause who is an elephant. One cannot refer to an already established referent by an indefinite NP: If it is not possible to read the given NP as an elliptic predication, one would consider it to be the introduction of a new referent. In contrast to (8), a referential reading of the apposition is triggered by the definite article in (9). The definite NP the elephant serves as a direct anaphor to Mr.
Metaphorical anaphors
Smith. (Because my paper is about cross-sentential anaphors, in the following I will leave out examples where the metaphorical anaphor occurs as an apposition.) But the indefinite article does not necessarily trigger a predicative reading as it is shown in (10). Here, the indefinite NP is not part of an elliptic predication, but it serves as a metaphorical introduction of the referent and the apposition Mr. Smith, denoting the referent by his name, makes the identification of the appropriate referent possible. In (11), the definite NP the elephant is also used referentially as in (10). It also introduces the referent and is therefore the antecedent for the anaphor Mr. Smith. The distinction between predicative and referential metaphors is fundamental, but often ignored (see below). The referential metaphors which are functioning as anaphors are the cases a theory of anaphors has to deal with. But also the theories of metaphor should be concerned with this mode of metaphorical use. Up to now, there are hardly any accounts of metaphor which are pointing at the problem of metaphorical anaphors. Exceptions are the semantic approaches of Kittay (1987) and Tirrell (1989) and their critical discussion by Leezenberg (2001: 236–238). Another indication of the topic is given by Asher and Lascarides (2001) (see 2 below). But neither of these approaches provides a sufficient survey; they all tackle the problem rather marginally. If most accounts of metaphor predominantly deal with the predicative use of metaphorical expressions, Kittay’s account (1987) is just the other way round: She only emphasizes the referential use of metaphorical expressions by treating all metaphors as anaphora. For instance, she describes an elliptical predication like “a performing seal” as a referring expression (Kittay 1987: 305 f.). So her view of metaphor is very limited, because it doesn’t treat predicative metaphors adequately, “and it is hard to see how such predication could be reduced to anaphoric reference” (Leezenberg 2001: 236). As regards the nature of metaphorical anaphors, Kittay argues that they belong to an anaphoric chain of several expressions denoting the same referent. She describes the reference of the metaphorically used expression as a case of “anaphoric reference, in which some member of an antecedent chain refers literally” (1987: 303). So, as Leezenberg puts it, “Kittay reduces metaphorical reference to a literal reference to the same object by some antecedent expression” (Leezenberg 2001: 236). The account offers no explanation how readers identify and understand a metaphorical anaphor. Kittay premises the understanding of the anaphoric relation. But she nowhere discusses the fact that the anaphoric and metaphorical reading of the definite NP is not the only possibility. The definite article marks the referent of the NP as easily accessible within the textworld model. But the NP can nevertheless introduce a new referent, as in (12), if the topic of the text is for instance ‘The circus on the fairground’:
Helge Skirl
(12) The man lies on the bed of nails. The elephant is dreaming.
Hence, Kittays account gives no satisfying answer to the question of how metaphorical anaphors are identified nor to the question of what kind of information they convey. The information expressed by a metaphorical anaphor is what Tirrell’s account is mostly concerned with (Tirell 1989). She emphasizes the stylistic effects of referential metaphors and on their ability to build up complex chains of multiple metaphors. Leezenberg (2001: 238) criticizes her emphasis on the stylistic values of metaphor for its lack of semantic insight. However, with respect to the comprehension of metaphorical anaphors, the problem with Tirrel’s account is not her stylistic approach but the fact that she only states an analogy between anaphors and extended metaphors, not seeing that referring metaphorical definite NPs can in fact functioning as anaphors. This is due to her very limited view of anaphors. She only discusses pronouns as anaphors and their mere function of establishing co-reference to a referent already introduced by an NP. Hence, her account gives no insight into the functioning of metaphors as anaphors. Leezenberg summarizes his critical examination of Kittay’s and Tirrell’s account by saying that the “few earlier discussions of the anaphoric or discourse effects of metaphorical interpretation […] have remained rather inconclusive” (Leezenberg 2001: 236). But his own approach ignores entirely the phenomenon of metaphorical anaphors: Though he emphasizes “the fact that metaphorically interpreted expressions may maintain ‘anaphoric’ relations with earlier or subsequent expressions” (Leezenberg 2001: 233), his account is only confined to cases, where metaphorically understood NPs serve as antecedent for a pronominal anaphor. With respect to the differences between referential and predicative metaphors in general, Leezenberg claims that referential metaphors “do not serve to ascribe a property to a specific individual, but rather to identify or refer to that individual. [...] they refer to a specific individual in virtue of the content they metaphorically express [...]” (Leezenberg 2001: 234). In my opinion, this is a misleading distinction: The referential aspect and the aspect of property attribution of a metaphorical anaphor belong together. Consider again (1):
(1) A man lies on the sofa. The elephant is probably dreaming.
If the content of the mentioned NP the elephant cannot be understood metaphorically in a way which fits the referent already introduced by the NP a man, the NP will not be understood as an anaphor, but as the (literal) introduction of a new referent. If the NP the elephant is understood as metaphorical anaphor co-referring to a man, it adds specifying information to the mental representation of the
Metaphorical anaphors
referent man. The referent is then understood as elephant-like in some way which has to be determined in a given context. Thus, the resolution of metaphorical anaphors is always a complex cognitive process which involves both the identification of the appropriate referent and the determination of the metaphorical descriptive content. This complexity is highly reflected in the results of several psycholinguistic studies concerning metaphorical anaphors (Gibbs 1990, Onishi/Murphy 1993, Noveck/Bianco/Castry 2001, Budiu/ Anderson 2002, Lemaire/Bianco 2003). In the studies, the comprehension of metaphorical anaphors has been tested. The general results are that metaphorical anaphors are more difficult to understand than literal anaphors and that they are more difficult to understand than predicative metaphors. A plausible explanation of the latter finding is that, in the case of predicative metaphor, the reference is already established and only a metaphorical property attribution has to be performed. In the case of metaphorical anaphors, the reference for the definite NP has to be established and the metaphorical descriptive content has to be understood. A theory of metaphorical anaphors has to recognize those empirical results of the psycholinguistic research and should explain them convincingly in detail.
4. Metaphorical anaphors as a phenomenon of the semantics-pragmatics-interface The resolution of metaphorical anaphors cannot be understood and explained within the theoretical framework of sentence semantics, because they exceed the boundaries of the sentence level. They are a textual phenomenon par excellence. However, even in the greater framework of discourse semantics, the pragmatic component is needed to cope with metaphorical anaphors. They are basically a phenomenon of the semantics-pragmatics-interface, because their resolution depends on both the semantic information given explicitly in the text and on further pragmatic knowledge and principles of text comprehension. In case of literal direct anaphors, it is the traditional view that “referential relationship between anaphor and antecedent is based on coreference […], and the link between the expressions denoting coreference can be described in terms of grammatical congruence and semantic compatibility” (Schwarz-Friesel, in press). In the case of metaphorical anaphors, there is no semantic compatibility between the anaphor and its antecedent (which holds true for several literal direct anaphors as well; see Schwarz 2000). But readers understand those referential NPs metaphorically in order to get a consistent and coherent comprehension of the text. This metaphorical reading of a referential NP is thus pragmatically motivated, because the asking for consistence and coherence is still a general pragmatic princi-
Helge Skirl
ple of text interpretation. Schwarz-Friesel calls it the “principle of conceptual plausibility” and describes it as “the dominant constraint which determines the reconstruction of coherence within which resolving anaphors occurs” (SchwarzFriesel, in press). Negatively said, a metaphorical anaphor is recognized, when another reading of the definite expression is simply not plausible. The reader has always to decide whether the mentioned definite NP introduces a new referent or is an anaphor for an already established referent. In the case of metaphorical anaphors, this decision is still a complex one. Because to regard the NP as an anaphor means at the same time to find its antecedent which refers to the appropriate referent and to determine the metaphorical descriptive content of the NP. At the sentence level, it is not possible to identify a metaphorically used referential NP without an actual context or co-text respectively. If at all, a violation of the restriction rules of the sentence verb is detectable. Consider (13), where the NP the stone violates the restriction rules of the verb dream because dream requires a subject which belongs to the category human (or animal): (13) The stone is probably dreaming.
Though this violation of restriction rules is a good indication for a metaphorical reading, it is not at all clear, whether the subject or the verb (or even both) should be interpreted metaphorically. In (14), the co-text triggers a metaphorical, anaphoric reading of the NP the stone, in (15), it triggers the metaphorical reading of the verb dream and in (16), it triggers a metaphorical reading of both the NP and the verb: (14) The apathetic person lies on the sofa. The stone is probably dreaming. (15) The pebble lies in the creek. The stone is probably dreaming. (16) The corpse lies on the deathbed. The stone is probably dreaming.
Though the violation of semantic restriction rules can be an indication for a metaphorical reading of at least one word in a sentence, it is no guarantee for such a reading. A special type of text may also suggest a quite literal reading. Note, that the characteristics of the text type are part of the pragmatic knowledge which readers activate in text comprehension. This knowledge is not a part of the semantic text structure. But it influences the construction of a coherent reading. Consider again (13). If it is for example used in a textbook of logic, a metaphorical reading is rather unlikely. If the sentence is explicitly shown as an example for a false assertion, then a metaphorical reading is even excluded: (13) The stone is probably dreaming.
Metaphorical anaphors
Literal readings are also possible, when the sentence is part of a text, which does not describe reality, but an imaginary world, like in a fantasy story (17) or in a fairy-tale (18): (17) The refrigerator is playing a guitar. The stone is probably dreaming. (18) The rabbit won’t wake up his best friend. The stone is probably dreaming.
Another problem for semantic accounts is, that the restriction rules of the verb don’t have to be necessarily violated in case of metaphors, as the second sentence of (19) shows: (19) The apathetic person is probably dreaming. The stone lies on the sofa.
Only because of the first sentence and the absence of further co-text or context, can the NP the stone be understood as a metaphorical anaphor to the antecedent the apathetic person. Apart from the identification of metaphorical anaphors, the comprehension of their descriptive content is also a phenomenon which goes beyond the sentence border. Consider (20) and (21) where the following sentence determines in each case the descriptive content: (20) Mrs. Smith’s husband lies on the sofa. The teddy-bear is probably sleeping. He is the most apathetic person imaginable. (21) Mrs. Smith’s husband lies on the sofa. The teddy-bear is probably sleeping. He is the most sensitive person imaginable.
In (20) and (21), the metaphorical descriptive content of the NP the teddy-bear, which functions as an anaphor to the antecedent Mrs. Smith’s husband, depends on the co-text of the last sentence. In (20), the teddy-bear is used to convey a negative evaluation of the husband’s trait, and in (21), it conveys a positive evaluation. Asher and Lascarides (2001) propose an account, where restrictions for the metaphorical use of an NP are even codified in the lexicon. They call those restrictions the “Metaphorical Lexical Rules”. However, as they admit, the metaphorical descriptive content cannot be codified in the lexicon, because it depends on the semantic co-text and pragmatic context, in which the metaphor occurs. The Metaphor Lexical Rule serves to codify rather conventional metaphorical uses. It tries to “predict which aspects of a lexical entry can vary in metaphorical interpretation, and which cannot” (Asher/Lascarides 2001: 274). For example, for the NP the ocean, the Metaphor Lexical Rule would probably state that it cannot be metaphorically used for an entity which is clearly defined concerning its shape. But novel uses are always possible. And authors can enrich the NP with attributive information, which triggers a reading that is not licensed by the rule, as one can see in (22): (22) A man lies on the sofa. The human ocean is probably sleeping.
Helge Skirl
Though the ocean is not a good conventional candidate for a metaphorical anaphor which co-refers to the antecedent a man, the adjective human can trigger such a reading. The metaphorical descriptive content of the human ocean could maybe characterize the energy, appearance or movements of the man. But it has to be specified in a larger co-text and context. As the examples (13) to (22) have shown, the identification and comprehension of metaphorical anaphors is a phenomenon of the semantics-pragmatics interface. Their resolution depends both on the explicit, thus semantic information given in the text and on the implicit, thus pragmatic principles and knowledge, which readers use to build up a full and convincing understanding of a text. The explicit, semantic information is given by the NP itself, by the sentence in which it occurs and by the further co-text, i.e. the preceding (and following) sentence(s). This information is the explicit, underspecified basis upon which the resolution of a metaphorical anaphor can operate. The implicit, pragmatic information consists of pragmatic principles, like the principle of conceptual plausibility, and of contextual factors, i.e. knowledge about the text type, the communicative situation and general world-knowledge, which readers need to enrich the information given by the text (cf. Schwarz-Friesel, in press).
5. The comprehension of metaphorical anaphors within the text-world model approach In modern text linguistics, text comprehension is described “as a complex cognitive process which is dependent on both the surface structure of the text and on the mental activity of the reader” (Schwarz-Friesel, in press). The cognitive theory of text-world models (Schwarz 2000: 39–46) provides a framework in which the text comprehension in general as well as the resolution of metaphorical anaphors in particular can best be described and explained. The text-world model is a complex mental representation, in which readers integrate the text-internal semantic information and the text-external pragmatic knowledge they activate to get a coherent understanding of the text. Thus, the text-world model is an enriched mental representation of the referents and the situation described in the text. With respect to the resolution of metaphorical anaphors in the reading process, the already established text-world model is the relevant background against which readers decide whether a definite NP can be identified as a metaphorical anaphor or as just a literal introduction of a new referent. The comprehension of a metaphorical anaphor is successful if it can be integrated plausibly in the textworld model, contributing in such way to the text coherence.
Metaphorical anaphors
As already shown, the text type plays an important role in text comprehension. It influences from the beginning the construction of the text-world model. Provided, for example, that a text is dealing with the topic ‘animals in the forest,’ readers will activate quite different knowledge and expectations, if the text type is a fairy-tale or if it is a scientific paper. Beside the text type, the text topic itself plays an important role in the construction of the text-world model. The text topic, which is often given in the text title, can be the key factor in the identification and comprehension of metaphorical anaphors as in (23): (23) FAILURE IN THE CONCERT HALL The pianist doesn’t strike a key. The bird won’t fly today. (24) LETHARGIC CREATURES The pianist doesn’t strike a key. The bird won’t fly today.
In (23), the title triggers a metaphorical anaphoric reading of the bird, because it is the only plausible reading within the reference situation of a classical concert. In (24), the title names a hyperonym of the entities shown in the text. Therefore, the bird will be understood as the literal introduction of a new referent. Note, that the definite article of the NP the bird is convenient because of the text title. If the already established text-world model suggests identifying a definite NP as a metaphorical anaphor, then the easiness of comprehension depends on the conventionality of the metaphorical use. The more conventional the metaphor is, the easier it will be understood. Consider (25): (25) A man lies on the sofa. The couch potato/The bookworm/The sad dog is probably dreaming.
In (25), understanding the definite NPs the couch potato, the bookworm, and the sad dog as direct anaphors to the man is easy, because these are highly conventionalized metaphors, which can denote human beings quite literally. The formerly metaphorical meaning of the word is now one of the literal meanings of the polysemous word. So, these examples are functioning as literal specifying anaphors. In (26), the metaphoricity of the examples is much more noticeable, because the metaphorical reading is not yet lexicalized: (26) A man lies on the sofa. The elephant/The stone/The ice block is probably dreaming.
Though the metaphorical use of the definite NPs the elephant, the stone, the ice block is not yet lexicalized, it is also not novel. On a scale of metaphoricity, those metaphors take an intermediate position between lexicalized, thus conventional metaphors and novel metaphors. According to Asher and Lascarides (2001: 282),
Helge Skirl
the metaphorical use of these definite NPs would be licensed by the Metaphorical Lexical Rule. The metaphorical descriptive content of the NPs is that of metaphorically used adjectives, which can be attributed to the literally used NP. However, the specific descriptive content has to be determined by the co-text or context. Novel metaphorical anaphors are much more difficult to understand if they are not even incomprehensible: (27) A man lies on the sofa. ?The microbe/?The telephone/?The ball-pen/?The ocean is probably dreaming.
In (27), the referent is introduced in the first sentence by the indefinite NP a man. In the second sentence, the verb denotes an activity, which is apt for a human being like a man. But the given subjects of the second sentence, the microbe, the tele phone, the ball-pen and the ocean are very difficult to understand as metaphorical anaphors which co-refers to a man. That is because the metaphorical descriptive content, they would have to convey remains odd without further information in the foregoing or following co-text. As already shown by example (22), it is possible to trigger an identification of a metaphorical anaphor, although its metaphorical descriptive content remains enigmatic. As in (22), the addition of the adjective human to the NPs triggers a reading as a metaphorical anaphor: (28) The man lies on the sofa. The human microbe/The human telephone/The human ball-pen/The human ocean is probably dreaming.
Note, that in (27), the heads of the NPs, microbe, telephone, ball-pen and ocean, are understood metaphorically. The adjective human, in contrast, is understood literally. It simply indicates the appropriate referent for the definite NP, so that it corefers to a man. The strategy of adding information to the NP can be used for novel metaphorical anaphors, but also for slightly simpler cases, as in (29): (29) SIE schlägt ihre Streich- und Blasinstrumente und die schweren Notenhefte den Leuten in die Rücken und Vorderfronten hinein. In diese Speckseiten, die ihr die Waffen wie Gummipuffer zurückfedern lassen. (E. Jelinek, Die Klavierspielerin, 19) SHE bangs into people’s backs and fronts with her stringed instruments and her heavy musical scores. Her weapons bounce off these people, whose fat is like a rubber buffer. (E. Jelinek, The Piano Teacher, 15)
In the English version, the resolution of the metaphorical anaphor her weapons, coreferring to her stringed instruments and her heavy musical scores, is made less difficult due to the personal pronoun her. (Note, that in the German original, the personal pronoun is missing. The metaphorical anaphor is just the definite NP die Waffen.)
Metaphorical anaphors
The adding of further information to the NP can facilitate the identification of metaphorical anaphors, which means that their appropriate referent within the text-world model can be found. For a complete integration into the text-world model, their metaphorical descriptive content has to be understood, too. In normal cases, the text will give enough information for the comprehension of that content. Consider (29) from a modern novel: (30) Erika, die Heideblume. Von dieser Blume hat diese Frau den Namen. Ihrer Mutter schwebte vorgeburtlich etwas Scheues und Zartes dabei vor Augen. Als sie dann den aus ihrem Leib hervorschießenden Lehmklumpen betrach tete, ging sie sofort daran, ohne Rücksicht ihn zurechtzuhauen, um Reinheit und Feinheit zu erhalten. (E. Jelinek, Die Klavierspielerin, 27) Erika, the meadow flower. That’s how she got her name: erica. Her pregnant mother had visions of something timid and tender. Then, upon seeing the lump of clay that shot out of her body, she promptly began to mold it relentlessly in order to keep it pure and fine. (E. Jelinek, The Piano Teacher, 23)
The antecedent for the metaphorical anaphor in (29) is something timid and tender, which refers to a newborn. The metaphorical anaphor is the lump of clay, which is explained by the relative clause that shot out of her body. (In the German original, this information is integrated in the NP itself: den aus ihrem Leib hervorschießenden Lehmklumpen.) The metaphorical descriptive content is quite clear in this case. The explicitly expressed contrast relation between something timid and tender and the lump of clay in accordance with further co-textual information triggers a very negative evaluation of the newborn (to the importance of contrast relations see Asher/Lascarides 2001: 283–285). In summary, it can be ascertained that the resolution of metaphorical anaphors is performed against the background of the text-world model, but at the same time, this resolution adds information to the text-world model. The re-mentioning of a referent conveys to the conceptual continuity of the text. And the metaphorical descriptive content adds specifying information to the mental representation of the referent.
6. Procedural constraints of metaphorical anaphors The examples (25) to (30) have shown, that the integration of metaphorical anaphors into the text-world model can be facilitated, if the metaphor is rather conventional or if information indicating the appropriate referent is added to a novel metaphor. However, psycholinguistic research has proved that even the simple cases of metaphorical anaphors complicate the reader’s understanding (cf. Gibbs 1990, Onishi/
Helge Skirl
Murphy 1993, Noveck/Bianco/Castry 2001, Budiu/Anderson 2002, Lemaire/Bianco 2003). Whether a metaphorical anaphor is rather conventional or whether a novel metaphorical anaphor is supported by further information: A significant difference to the resolution of literal anaphors remains. Metaphorical anaphors are always more difficult to understand, which means that the processing effort, as reflected in reading times, is bigger. If the processing effort is nearly comparable, then the metaphorical anaphors aren’t correctly understood (Budiu/Anderson 2002). As already said, an explanation to these findings is that in the case of metaphorical anaphors the cognitive system must solve two tasks at the same time: The reference for the definite NP has to be established and the metaphorical descriptive content has to be understood. The interesting question is why metaphorical anaphors are used, although they are not as easy to understand as literal anaphors. Based on relevance theory (cf. Sperber/Wilson 21995, Carston 2002), Noveck, Bianco and Castry (2001) argue, that the procedural costs of metaphorical anaphors have to be outweighed by some benefits and that those benefits have to be further investigated. The examples discussed so far show that the procedural costs of metaphorical anaphors are generally counterbalanced with their contribution to referential continuity and the rich content they convey by just a very succinct expression. The recurrence to an already established referent—or to an implicit referent in the case of indirect metaphorical anaphors—aids the conceptual continuity and therefore the coherence of the text-world model. This use is quite expressive because the anaphor is not a literal, but a metaphorically used NP. The metaphorical descriptive content of a metaphorical anaphor adds much information to the mental representation of the referent. The specific features of that content have to be studied in more detail. But it is clear that the informational complexity of a metaphorical anaphor is much due to its combination of two concepts: A target concept—the concept, to which the metaphorical anaphor refers— is characterized metaphorically by a source concept—the concept, which the definite NP denotes, when it is used literally. This aspect of concept combination is a general characteristic of metaphors (see Lakoff/Turner 1989; cf. Schwarz-Friesel 2004). Consider again (1):
(1) A man lies on the sofa. The elephant is probably dreaming.
The target concept is man, which is characterized by the source concept elephant. As already suggested, the conciseness and informational richness of the metaphorical anaphor the elephant results in its expressive combination of reference and property attribution. Because the definite NP the elephant is identified as a direct anaphor, it refers—like its antecedent a man—to the target concept man. But it denotes literally the source concept elephant, which is now used meta-
Metaphorical anaphors
phorically to characterize the concept man, presupposing that there are analogies between a man and an elephant in at least some respect. This combination of two concepts and the wide mental scope of the presupposed analogies is surely a fundamental reason for the processing effort metaphorical anaphors requires and for the rich information content they can add to the text-world model. The contribution of metaphorical anaphors to the complexity of the text-world model can also be seen in the case of indirect metaphorical anaphors. In this specific case, the presupposed analogous structure of two complex concepts or conceptual frames guides the resolution of the metaphorical anaphors. The readers understand the definite NPs because they refer to variables of a conceptual frame, which is mostly introduced by an anchor expression (cf. Schwarz 2000, SchwarzFriesel, in press). Consider again (6):
(6) A man lies on the sofa. He is a dreaming elephant. The trunk makes a lot of noise.
The anchor expression for the resolution of the indirect metaphorical anaphor the trunk is the indefinite NP a dreaming elephant, which is a predicative metaphor. This metaphorical assertion establishes an analogy between the referent of both the NP a man and the pronominal anaphor he and an elephant. So, the NP the trunk names a part of the elephant’s body. As an indirect metaphorical anaphor, it will be understood as referring to the analogous part of the man’s body, hence his nose. The presupposed analogy between two conceptual frames can be used for a complex of different indirect metaphorical anaphors, as in (31): (31) The old professor is an eagle. The claws and the beak are still very dangerous. But he cannot move the wings anymore.
In (31), the predicative metaphor an eagle is the anchor expression for the indirect metaphorical anaphors the claws, the beak and the wings. Note that, in contrast to example (6), where the trunk refers to the analogous part of the man’s body, such a clear reference to a part of the professor’s body is rather unlikely. Here, it is more likely, that the indirect metaphorical anaphors will be understood as referring to personal attributes of the professor, which are rather abstract concepts. Thus, in the comprehension process, the conceptual frame with its parts, to which the different indirect metaphorical anaphors refer, has to be actively constructed by the reader. With respect to those chains of several indirect anaphors, the reference to rather abstract concepts which are never mentioned literally is quite common (cf. Brooke-Rose 1958). It complicates the reader’s effort to resolve those metaphorical anaphors. But a successful comprehension contributes strongly to the informational richness of the text-world model.
Helge Skirl
In (6) an (31), we have seen examples, where the anchor expression for the indirect metaphorical anaphor is itself a (predicative) metaphor. But there are also examples, where the anchor is a literally used expression and the metaphor is introduced by the indirect metaphorical anaphor, like in (32): (32) SIE wendet sich mehrmals nach der vollkommen desorientierten Dame um, bevor SIE einen vertrauten Weg in ein vertrautes Zuhause einschlägt. SIE grinst die Dame dabei an, vergessend, daß SIE in ein paar Minuten unter der heißen Flamme des mütterlichen Schneidbrenners zu einem Häufchen Asche verbrennen wird, weil sie zu spät gekommen ist. (E. Jelinek, Die Klavierspielerin, 27) “SHE peers back several times to the completely disoriented woman before setting off on the familiar road to her familiar home. SHE smirks at the woman, forgetting that a few minutes from now, SHE will feel the hot flame of her mother’s blow-torch and she will be burned to a pile of ashes because SHE is late in getting home.” (E. Jelinek, The Piano Teacher, 23)
In (32), the anchor expression is her family home. The indirect metaphorical anaphor introduces the metaphor of the hot flame of her mother’s blow-torch. To facilitate the understanding, the NP is enriched by the genitive attribution of her mother. Without that addition, the link between the anchor her family home and the indirect metaphorical anaphor would be difficult. In this case, the reference of the metaphorical anaphor is triggered by the co-textual information: The metaphorical anaphor is probably referring to the rage and hatred of the mother. To sum up, in all the discussed cases of this section, the resolution of the metaphorical anaphors is constructed against the background of the text-world model. The metaphorical anaphors are contributing to the conceptual continuity, hence the coherence of the text-world model. And they add rich information to the mental representation of the referents.
7. Conclusion Metaphorical anaphors are a special case of textual anaphors. They particularly occur in stylistic expressive texts. They have not yet been sufficiently investigated by either theories of metaphor or theories of anaphors. It has been shown that metaphorical anaphors are a phenomenon of the semantics-pragmatics-interface. In addition, it has been argued that they can best be described and explained within the cognitive theory of text-world model. It has been demonstrated that there are various different forms of metaphorical anaphors. Concerning the psycholinguistic findings that metaphorical anaphors always complicate the reader’s understanding, it has been pointed out that the procedural costs of met-
Metaphorical anaphors
aphorical anaphors are generally counterbalanced with the rich content they convey, but that the specific features of that content have to be studied in more detail. Metaphorical anaphors are still a challenge for theories dealing with anaphors. They have to be further investigated in order to provide a complete and convincing account of textual anaphors.
References Asher, N. and Lascarides, A. 2001. Metaphor in discourse. In The Language of Word Meaning, P. Bouillon and F. Busa (eds), 262–290. Cambridge: CUP. Brooke-Rose, C. 1958. A Grammar of Metaphor. London: Secker & Warburg. Budiu, R. and Anderson, J. 2002. Comprehending anaphoric metaphors. Memory & Cognition 30: 158–165. Carston, R. 2002. Thoughts and Utterances: The pragmatics of explicit communication. Malden MA: Blackwell. Gibbs, R. 1990. Comprehending figurative referential descriptions. Journal of Experimental Psychology: Learning, Memory and Cognition 16: 56–66. Kittay, E.F. 1987. Metaphor. Its cognitive force and linguistic structure [Clarendon Library of Logic and Philosophy. Metaphor]. Oxford: Clarendon. Lakoff, G. and Turner, M. 1989. More than Cool Reason. A field guide to poetic metaphor. Chicago IL: Univ. of Chicago Press. Leezenberg, M. (2001). Contexts of Metaphor. Amsterdam et al.: Elsevier. (= Current Research in the Semantics/Pragmatics Interface. 7) Lemaire, B. and Bianco, M. 2003. Contextual effects on metaphor comprehension: Experiment and simulation. In Proceedings 5th International Conference on Cognitive Modelling (ICCM), Bamberg, Germany, F. Detje, D. Dörner and H. Schaub (eds), 153–158. Noveck, I. A., Bianco, M. and Castry, A. 2001. The costs and benefits of metaphor. Metaphor and Symbol 16(1/2): 109–121. Onishi, K. H. and Murphy, G.L. 1993. Metaphoric reference: When metaphors are not understood as easily as literal expressions. Memory & Cognition 21(6): 763–772. Schwarz, M. 1997. Anaphern und ihre diversen Antezedenten: Koreferenz und Konsorten. In Sprache im Fokus. Festschrift für Heinz Vater zum 65. Geburtstag, C. Dürscheid, K. H. Ramers and M. Schwarz, 445–455. Tübingen: Niemeyer. Schwarz, M. 2000. Indirekte Anaphern in Texten. Studien zur domänengebundenen Referenz und Kohärenz im Deutschen [Linguistische Arbeiten 413]. Tübingen: Niemeyer. Schwarz-Friesel, M. 2004. Kognitive Linguistik heute – Metaphernverstehen als Beispiel. Deutsch als Fremdsprache 41. Jg., 2. Quartal, H. 2, 83–89. Schwarz-Friesel, M. In press. Text comprehension as the interface between verbal structures and cognitive memory processes: The case of resolving direct and indirect anaphora. In Memory and Language. C. Zelinsky-Wibbelt (Ed.). Berlin: Mouton de Gruyter. Sperber, D. and Wilson, D. 1995. Relevance: Communication and cognition, Oxford: Blackwell. Tirrell, L. 1989. Extending: The structure of metaphor. Noûs XXIII: 17–34. White, Roger M. 1996. The Structure of Metaphor. The way the language of metaphor works. Cambridge MA: Blackwell.
section 2
The Syntax and Semantics of Anaphors
Accessibility and definite noun phrases Klaus von Heusinger Accessibility is a dynamic discourse structure that is the sum of the salience structure of particular referential expressions. On the one hand, accessibility is continuously updated by referential expressions, while on the other, it determines the reference of definite expression. In this paper, I show that not only indefinite noun phrases, but also definite noun phrases update the accessibility structure of a discourse by promoting the actual referent to the most salient of its kind and some supersets. This dynamic potential of definite noun phrases establishes text coherence by means of salience spreading.
1. Introduction The interpretation of a discourse essentially depends on the referential link between the linguistic expressions. In this paper I discuss in particular the interaction of definite (full) noun phrases and the accessibility structure, which is assumed to be a network of semantically related sets of ranked discourse items associated with the predicates that have introduced them, rather than one single set of ranked items. I contend that not only does the accessibility structure determine the definiteness status of referring expressions, but also that definite noun phrases change this accessibility due to their descriptive content. The definite noun phrase the small bird refers not only to the most accessible small bird in the context, but in general also raises the accessibility of that bird with respect to the predicate small birds and some hyperonyms such as birds and animate objects. I will show that this interaction between accessibility and definite noun phrases is one of the central principles that constitutes the coherence of a discourse. 1. The paper is the revised version of a talk at the workshop “Anaphorik in Texten” at the 27th annual meeting of the DGfS (German Society of Linguistics) held in Cologne, February, 23–25, 2005. I would like to thank the audience for a constructive discussion and their valuable comments. In particular I would like to express my gratitude to Monika Schwarz-Friesel, Manfred Consten and Konstanze Marx for having organized this inspiring event and for editing the present volume.
Klaus von Heusinger
In section 2, I discuss this interaction in four types of common phenomena: (i) disjoint reference of different occurrences of the same definite noun phrase; (ii) the alternation between definite noun phrases and pronouns in anaphoric chains; (iii) the type of anaphoric relation; and (iv) the conditions for a “first-mentioned” definite noun phrase. In section 3, I examine the different aspects of accessibility (i) as activation, (ii) as relation, (iii) as hierarchies, and (iv) as salience. In section 4, I discuss different concepts of accessibility as activation and discuss two accessibility hierarchies: Ariel’s Accessibility Marking Scale and Gundel & Hedlands & Zacharski’s Givenness Hierarchy. I show the particular role of definite noun phrases in these two scales and maintain that both scales neglect some essential features of definite nouns, namely that they are dynamic expressions. The comparison of Discourse Representation Theory and Centering shows that there are different notions of accessibility: a categorical and a graded one. Section 5 discusses the notion of salience in the accessibility structure and in anaphoric relations. Salience denotes a relation between different (semantic) objects of the same kind, which are defined by the descriptive content of the expression. The most salient object is the most-accessible object for an anaphoric term. Referring expressions, such as indefinite and definite noun phrases, raise the salience of their referent with respect to the set associated with their descriptive content. Furthermore, they also change the salience structure of some semantically related sets, a process that I term “salience spreading”, accounting for the dynamic processes in the accessibility structure. Section 6 gives a brief summary and further research prospects,
2. Discourse coherence and definite noun phrases In the late 70s increasing interest in discourse analysis led to a new generation of semantics that investigated discourses rather than isolated sentences. The research was initiated in disciplines as different as philosophy (Stalnaker 1978), logic (Lewis 1979), semantics (Heim 1982; Kamp 1981), the structuralism of the Prague School (Sgall & Hajičová et al. 1973; 1986), discourse pragmatics (Chafe 1976; Prince 1981), and artificial intelligence (Webber 1979). All these approaches were mainly interested in the way the coherence of texts or discourses is linguistically established. Anaphoric and definite expressions are in the center of interest, since they are the main means to creating coherence between sentences. Thus, discourse anaphora reflect the interaction between the discourse structure and the interpretation of a particular linguistic expression. The interpretation depends on discourse structure, as noted by Grosz et al. (1990, 445):
Accessibility and definite noun phrases
Discourse context affects the interpretation of individual phrases and clauses within a single utterance. The meaning of pronouns and definite descriptions is quite obviously influenced by the context in which they are used, among the most widely studied problems in discourse processing are those concerning the influences of context on the processes of generating and interpreting such phrases.
In this paper I will focus on definite noun phrases, on contextual parameters that determine their reference, but also on their unique context change potential. It will be shown that the descriptive content of definite noun phrases changes the accessibility structure of the context. This influence of definite noun phrases on the accessibility is illustrated by the following four phenomena: (i) disjoint reference of different occurrences of the same definite noun phrase; (ii) the alternation between definite noun phrases and pronouns in anaphoric chains; (iii) the type of anaphoric relation; and (iv) the conditions for a “first-mentioned” definite noun phrase. 2.1
Disjoint reference of different occurrences of the same definite noun phrase
Different occurrences of the same definite noun phrase can refer to distinct objects or discourse items, as illustrated in (1). The noun phrase two sharks introduces two discourse items. One is picked up by the one and continuously referred to by his, the shark and the shark. The other discourse item is picked up in the next paragraph by the other shark, but then referred to by his, his, he, his, him and the shark. By mentioning the second discourse item and ascribing predicates to it, its salience or accessibility has been raised such that the last occurrence of the shark refers to it. Similar examples are well-known from the literature (see Lewis 1979 for a comparable example with cats):
(1) The two sharks1⊕2 closed together and as he saw the one1 nearest him open his1 jaws and sink them into the silver side of the fish, he raised the club high and brought it down heavy and slamming onto the top of the shark1’s broad head. He felt the rubbery solidity as the club came down. But he felt the rigidity of bone too and he struck the shark1 once more hard across the point of the nose as he slid down from the fish. The other shark2 had been in and out and now came in again with his2 jaws wide. The old man could see pieces of the meat of the fish spilling white from the corner of his2 jaws as he2 bumped the fish and closed his2 jaws. He swung at him2 and hit only the head and the shark2 looked at him and wrenched the meat loose. (Hemingway, The Old Man and the Sea, 64)
Such disjoint references of one and the same definite noun phrase cannot be explained in static semantics with a strong uniqueness condition for definiteness. We need rather a more flexible condition for definiteness, namely salience, and a dy-
Klaus von Heusinger
namic interpretation of definite noun phrases, which accounts for their context change potential. 2.2
Alternation between definite noun phrases and pronouns in anaphoric chains
The indefinite description a small bird introduces a new discourse item that is subsequently picked up by the anaphoric pronoun he or the anaphoric definite noun phrase the bird. The distribution of these two alternatives is not well understood. In the example below, it seems that the full definite noun phrase is necessary to bring the discourse item back into the topic function. However, there is a general observation that we often find anaphoric chains such that a full noun phrase introduces an item which is subsequently picked up by pronouns. After two or three pronouns another full noun phrase seems necessary to keep the activation level of the item high. So we generally find anaphoric chains of the following type: (in)definite noun phrase1, pronoun2, pronoun3, (… pronounn,), definite noun phrasen+1, pronounn+2 etc. with n not much greater than 3. Again, a discourse semantics has to account for this distributional behavior of pronouns vs. full definite noun phrases.
(2) A small bird1 came toward the skiff from the north. He1 was a warbler and flying very low over the water. The old man could see that he1 was very tired. The bird1 made the stern of the boat and rested there. Then he1 flew around the old man’s head and rested on the line where he1 was more comfortable. (Hemingway, The Old Man and the Sea, 28)
2.3
The type of anaphoric relation
The prototypical anaphoric relation consists of an antecedent and an anaphoric expression where the anaphoric expression contains less information than the antecedent one, such as a bird... he or a small bird... the bird in (3a). This corresponds to the function of an anaphoric expression, namely to establish a link to the already introduced discourse item. The semantic relation between two such expressions can be described as a “superset”-relation, i.e. the descriptive content of the anaphoric term is associated with a superset of the set that is associated with descriptive content of the antecedent term, as in (3a). Anaphoric pronouns do not contain content except for gender and number information; so they are associated with sets of all (male, female, inanimate) objects. However, it is also possible to pick up a discourse item with the same expression as it was introduced, as in (3b). This seems to be a marked option, but it can be appropriate in certain contexts.
Accessibility and definite noun phrases
Here the anaphoric relation corresponds to the semantic relation of identity between the sets associated with the descriptive content of the two expressions involved. These two types of anaphoric relations, i.e. (3a) and (3b), are also known as semantic anaphora, since their relations are encoded in the lexical semantics of the corresponding descriptive terms (or to the composition of such terms).
(3) Types of anaphoric links and corresponding semantic relations
a b c d
anaphoric link a bird... he, a small bird... the bird a bird... the bird a small bird... the beautiful animal a bird... the small bird
semantic relation superset identical sets intersection subset
However, we also find anaphoric relations like (3c) and (3d), which are known as pragmatic anaphora, since additional knowledge is necessary to establish the anaphoric link. In (3c), the anaphoric term expresses additional content and can therefore be linked only to the antecedent if there is either no competitor in the context, or the reference is established by other means. The corresponding semantic relation is intersection (which might be – in extreme cases – empty). Similarly in (3d), where the link is only possible if additional knowledge is supplied, the semantic relation corresponds to the subset-relation. Such pragmatic anaphora are used to introduce background information by forcing the hearer to accommodate the new information into the background. They are often used in newspapers, where it seems necessary to integrate much (background) information into little text. A theory of dynamic accessibility has to account for these different anaphoric relations and must make a distinction between the semantic and pragmatic type. 2.4
Conditions for a “first-mentioned” definite noun phrase
It was observed (see Fraurud 1990, Poesio & Vieira 1998, Poesio 2003 and others) that we find many definite noun phrases being used to introduce new discourse items, even though dynamic theories say that only indefinite noun phrases (and proper names) can introduce new discourse items, while definite noun phrases refer to already established ones. These “first- mentioned” definite noun phrases can be divided into different classes depending on the conditions under which they are licensed. The definite noun phrase the government in (4) is licensed by our knowledge that Koreans, like other people, have a government (this is the “broader situational use” of Christophersen 1939, 29). Other cases include different types of bridging, such as in (5), where the relation between the antecedent term and the anaphoric term is a functional relation, here of authorship (this corresponds to the “implicit contextual basis” of Christo-
Klaus von Heusinger
phersen 1939, 29). Such indirect anaphoric expressions can also be licensed by frames, as in (6), or other kinds of linguistic or encyclopedic knowledge:
(4) For the Parks and millions of other young Koreans, the long-cherished dream of home ownership has become a cruel illusion. For the government, it has become a highly volatile political issue. (Poesio & Vieira 1998, 198)
(5) John bought a book. The author is well known.
(6) John entered a restaurant and sat at a table. The waiter brought the menu.
Other than in the examples in (3), the licensing term (or anchor) and the definite term are not co-referential (this is the reason we do not call them antecedent and anaphoric). The interaction between the descriptive content of two noun phrases and the accessibility structure of the text is more intricate. Still, a dynamic theory of accessibility has to pay attention to these data. Before we develop a dynamic theory that accounts for this interaction, we first present a detailed presentation of the different aspects of accessibility discussed in the literature.
3. Aspects of Accessibility The notion of accessibility plays a crucial role in the discussion of discourse structure and referential behavior of linguistic expressions. The general picture of an anaphoric relation is illustrated in Figure 1. The anaphoric relation is not a direct relation between the antecedent and the anaphoric expression, nor a direct relation to some object in the world, but a mediated one at the level of discourse representation (see Karttunen 1976 and subsequent work). A referring expression (the antecedent) is associated with a mental object, a discourse item or discourse referent. In some theories, this item is additionally linked to an object in the “world” or in a model representing the world by particular rules. A subsequent referring expression, the discourse anaphor, is associated with the discourse item establishing the anaphoric link.
2. See Schwarz 2000, who distinguishes between the following kinds of indirect anaphors: (i) semantic type, (ii) encyclopedic and (iii) inferential; Consten 2004 extends these domains by others such as pictorial information.
Accessibility and definite noun phrases
referring expression antecedent
discourse
discourse representation or mental objects
referring expression discourse anapher
discourse item / mental entity
modeltheoretic semantics “the world”
object
Figure 1. Discourse representation and anaphoric relations
As presented above, the relation between the antecedent and the discourse anaphor is determined by the accessibility (structure) of the discourse, i.e. the access of the anaphoric term to the discourse item, on the one hand, and the ranking of different discourse items, on the other. This intuitive notion can be separated into several aspects of accessibility, which are not, however, always clearly distinguished in the literature. Figure 2 attempts to illustrate five different aspects of accessibility: activation [1], the accessibility relation [2], accessibility hierarchy [3], accessibility structure [4], and the salience [5] of the objects in some model (or the “world”). These different, though closely related, aspects have characteristic features, which are presented in the following subsections.
discourse
discourse representation or mental objects
modeltheoretic semantics “the world”
referring expression antecedent [1] ACTIVATION • linguistic type • descriptive content
referring expression discourse anapher
[2] ACCESSIBILITY RELATION • distance • unity • context knowledge • encyclopedic knowledge
discourse item/mental entity [4] ACCESSIBILITY STRUCTURE • syntactic structure • discourse structure
object
[5] SALIENCE • competition
Figure 2. Five aspects of accessibility
[3] ACCESSIBILITY HIERARCHY • informativity
Klaus von Heusinger
3.1
Accessibility status or activation
A referring expression introduces, activates or evokes a mental object or discourse item. This entity is assigned a certain accessibility status or activation based on the linguistic or lexical type (pronoun, proper name, full noun phrase), the descriptive content and syntactic function of the associated expression. For instance, an entity associated with a subject receives a higher accessibility than one associated with an object, which amounts to one of the central rules in Centering Theory (see below). A definite noun phrase or a proper name activates its associated referent in a different way than an indefinite noun phrase. Inherent properties, such as animacy, of the referred objects are sometimes also included in this activation. However, little attention is paid to the descriptive content’s role with respect to activation. It is generally assumed that the descriptive content restricts the class of elements to which the activation applies. The mirror of this is discussed below in section 3.5 under the heading “salience”. 3.2
Accessibility relation
The accessibility relation holds between an already activated or established element in the discourse and another referring expression that is associated with it, the discourse anaphor. This anaphoric expression “accesses” the discourse entity. The access is determined, among other factors, by the distance and the syntactic structure between the anaphoric expression and the antecedent that has activated or introduced the entity. Furthermore, contextual and encyclopedic, as well as inferential knowledge, may constrain the relation. This relation can be understood as an instantiation of the more general reference relation. If the reference relation holds between a linguistic expression and an object in the immediate context or situation, we call it a deictic reference. If the expression refers to an object that is available in our encyclopedic knowledge, we call it an encyclopedic reference. Only if the relation holds between an expression and an entity that was introduced by the preceding discourse is the accessibility relation adequate. Furthermore, any reference to discourse items has two different incarnations: activating or accessing. If the expression refers for the first time to an entity, it evokes or activates it; while a reference to an already accessible or activated entity is described by accessing this entity. A discourse semantics should be able to describe both aspects – activation and accessibility relation, by the same mechanism.
3.3
Accessibility and definite noun phrases
Accessibility hierarchies
The accessibility hierarchy is reflected in the choice of the anaphoric expression. If the intended entity is easily accessible, the anaphoric expression need not be very informative (e.g. a pro-from or an unstressed pronoun), while less accessible entities need more information in the anaphoric expression (full noun phrases, full names). This reflex of the anaphoric expression with respect to the status of the intended referent is described in various hierarchies, such as in Ariel’s (1990) Accessibility Marking Scale or in the Givenness Hierarchy of Gundel et al. (1993). One interesting question relating to these hierarchies is whether they only reflect the degree of accessibility of the associated entity or whether they also show the degree of activation for a newly activated entity. In other words, does the hierarchy express only the degree of context dependency of the referring expression, or does it also possess the potential for changing the context? We discuss this in more detail in section 4. 3.4
Accessibility structure
The accessibility relation depends on a variety of factors, one of which is the structure of the discourse. Thus, we can assign to a discourse domain, segment or structure an ordered set of accessible entities. The accessibility structure can be understood as a property of a discourse which yields an ordered set of accessible items (with respect to a certain structural position). This notion is predominant in the discussions of Centering Theory and Discourse Representation Theory, and is illustrated in Figure 3: a / the F antecedent activates discourse item
the G discourse anapher link to the most activated element
set of ranked discourse items {a > b > c > ...}
Figure 3. Accessibility to one set of ranked elements
3.5
Accessibility and properties: Salience
The description of an accessibility structure for a discourse captures important grammatical parameters, however it misses to account for one of the most important semantic contributions to accessibility: the descriptive material that intro-
Klaus von Heusinger
duces an entity. Accessibility in this sense is an instantiation of the more general principle of salience (see Lewis 1979, von Heusinger 1997). Salience is a property of a set associated with descriptive material expressed in a referring expression. The accessibility structure of a discourse is formed by listing the salient items of each set associated with some predicate used in that discourse, and the relations between the listed sets, such as hyponymy etc., as illustrated in Figure 4. This different view of accessibility can be illustrated mainly by the behavior of definite and indefinite noun phrases, since they exhibit a considerable descriptive material. Pronouns or proper names do not, and form therefore an unlikely testing field for this view. Nevertheless, most theories focus on the use of pronouns and their antecedents, but not on definite description, thus missing interesting generalizations. a / the F antecedent activates discourse item with respect to F
the G discourse anapher link to the most activated element of G
set of ranked discourse items with respect to sets F: {a > b > c > ...}; G:{d > e > f > ...} + relation between F and G (by lexical or contextual information)
Figure 4. Accessibility to sets of ranked elements with respect to predicates
4. Accessibility and descriptive content One approach to discourse structure and coherence in discourse focuses on the activation of mental entities which are taken to be associated with linguistic expression. In these theories, the mental entities are activated by uttering an associated expression. The choice of a subsequent expression associated with the same entity reflects the activation or accessibility status of the mental entity. Referring expressions can be ordered according to their informativity. The more informative they are, the less accessible the associated entity. These accessibility hierarchies (or informativeness hierarchies) come in different guises. Prince (1981) is one of the first hierarchies of this sort, in which the following types of referential statuses are distinguished: evoked {textually, situationally} > unused > inferable > brand-new. I will discuss and compare the two further developed hierarchies of Ariel (1990) and Gundel et al. (1993) with respect to the five aspects of accessibility in the last section. A brief comparison between Discourse Representation Theory and
Accessibility and definite noun phrases
Centering Theory will show that there are different concepts of accessible discourse items – a categorical and a graded one. 4.1
Accessibility Marking Scale
Ariel (1988; 1990, 73) proposes the Accessibility Marking Scale (7), according to which the speaker selects an expression to refer to an accessible entity. If the entity is highly accessible a pro-form or a pronoun is used, whereas definite noun phrases are employed for less accessible items. The speaker must make sure that the degree of accessibility marked by the referring expression matches the degree of accessibility associated with the mental entity by the addressee.
(7) Accessibility Marking Scale
Low Accessibility (of the antecedent): (a) Full name + modifier (b) Full („namy“) name (c) Long definite description (d) Short definite description (e) Last name (f) First name (g) Distal demonstrative + modifier (h) Proximal demonstrative + modifier (i) Distal demonstrative (+ NP) (j) Proximal demonstrative (+ NP) (k) Stressed pronoun + gesture (l) Stressed pronoun (m) Unstressed pronoun (n) Cliticized pronoun (o) Extremely High Accessibility Markers (gaps, including pro, PRO and wh traces, reflexives, and Agreement) High Accessibility (of the antecedent)
This hierarchy is the result of empirical observations and merges three different kinds of contexts: the encyclopedic context, which is mainly responsible for proper names and unique definite descriptions; the situational context, which licenses deictic expressions; and the linguistic context, which determines the anaphoric expressions. Since Ariel is not only interested in the anaphoric use of referring expressions, she subsumes all these kinds of contexts under one general scale. Ariel (1990, 28–29) lists four main parameters which govern these scales. (i) The Distance between the antecedent and anaphor, what we have subsumed under the accessibility relation in
Klaus von Heusinger
section 3.2. (ii) The Competition is determined by the number of potential antecedents. According the Ariel (1990, 28) “the more competitors the less accessible will be the referenced entity.” However, we have seen that it is not just the number of competitors, but the number of competitors with the same properties. (iii) The Saliency expressing whether an antecedent is topical or not. This corresponds to our concept of activation in 3.1, and not to what we have termed salience in 3.5. (iv) The Unity, which expresses different aspects of the discourse structure, including whether the antecedent is in the paragraph or discourse segment as an anaphoric term. This corresponds to our accessibility structure in 3.4. The activation aspect is described by Ariel (1990, 22) as depending on two main factors, namely salience (in our sense described in 3.5) and the relation between them: The degree of accessibility of the mental entity serving as linguistic antecedent varies according to two main factors. First, some antecedents represent mental entities inherently more salient (...) The second factor that determines the degree of accessibility of a specific antecedent (or rather its mental representation) at the time when the potentially anaphoric expression is being processed, is the nature of the relation between the two. Thus, the distance between the last mention of the potential antecedent and the putative anaphoric expression determines how accessible the antecedent is to the addressee when interpreting the potentially anaphoric expression.
We will come back to these two notions: salience is discussed in section 5, and the “distance” is worked out in theories like Centering Theory or DRT, which are briefly characterized in section 4.3. 4.2
Givenness Hierarchy
Gundel, Hedland & Zacharski (1993) present the Givenness Hierarchy (8), which ranks the different linguistic expressions in an order of statuses. The Givenness Hierarchy confines the domain of potential antecedents in the following way: since pronouns are more likely to relate to highly activated elements, definite noun phrases are only used if one refers to less activated elements (by some general conversational principles).
(8) Givenness Hierarchy (Gundel, Hedland & Zacharski 1993) in focus > activated > familiar > it
that, this
more accessible
that N this N
uniquely type > referential > identifiable identifiable the N
indefinite this N
aN less accessible
Accessibility and definite noun phrases
Unlike Ariel, they do not list a set of parameters that determine the accessibility, but give a particular description of each of the statuses, which makes strong reference to the cognitive status of the referent in some mental domain. In Focus means that the referent is at the current center of attention. Activated expresses that the referent is represented in current short-term memory. Uniquely Identifiable: the addressee can identify the speaker‘s intended referent on the basis of the nominal alone. Referential: the speaker intends to refer to some particular object or objects. Type Identifiable: the addressee is able to access a representation of the type of object described by the expression. This scale is simpler than Ariel’s, but it is surprising that indefinite noun phrases are included in it. Indefinite noun phrases refer to predicates rather than to accessible discourse items. Therefore, there are some problems with the description of these statuses, which are defined by very different concepts as activation, reference and identification. Reference is a relation between an expression and an object in the “real world”, it is not clear whether it is helpful to explain the activation status of a mental or cognitive entity. Both hierarchies are very similarly designed and differ only in some minor points. Both theories conceive accessibility as a gradable property of a mental or “cognitive” entity, which can be more or less accessible. The linguistic expression associated with this entity must match the degree of accessibility associated with the intended entity. The two approaches differ in whether the categories in the hierarchy are mutually disjointed (Ariel) or whether less accessible statuses include more accessible (Gundel et al.). They agree, however, in the view that the accessible items are listed simply at some level of the discourse structure. They do not assume a more complex accessibility structure, nor do they say more about the discourse structure in general. 4.3
Discourse structure and accessibility
Discourse structure restricts accessibility in two respects: (i) whether or not an item is accessible, and (ii) how accessible it is. Discourse Representation Theory (DRT) formulates the rules for (i), while Centering Theory is more concerned with (ii). The two approaches differ in whether accessibility is a gradable or non-gradable notion, and whether it is applied to discourse domains or discourse segments. A discourse domain contains all expressions that have an equal relation to discourse-domain-creating operators like negation, modals or verbs of attitudes. Discourse domains are nested and represented as DRSs or boxes in DRT. Discourse segments, on the other hand, are suprasentential units of two or more connected 3. Hegarty & Gundel & Borthen (2001) combine the Giveness Hierarchy with Information Structure. However, they only compare demonstrative anaphors with pronominal ones.
Klaus von Heusinger
sentences. They determine a local domain in which certain rules for pronominal reference are defined. In the DRT of Kamp & Reyle (1993), or more recently Kamp & von Genabith & Reyle (to appear), both indefinite and definite expressions introduce new discourse referents. The discourse referent of a definite or anaphoric expression must be identified with an already established discourse referent to meet the familiarity condition. Hence, the anaphoric relation is represented as an identification of the new discourse referent with an accessible one. There are structural restrictions on the accessibility of discourse referents, which are encoded in the construction rules for DRSs. For instance, a discourse referent can only be linked to another one that is represented in the same discourse domain or box. The discourse referents form a set of accessible antecedents with respect to a discourse domain, as illustrated in the schematic representation of a discourse consisting of sentences S1, S2, S3, and S4, with a logical form P(a), R(he,b), Q(c,a), T(d,him) (think of sentences like: Arno comes. He greets Benno. Chris looks at Arno. Dora talks to him). DRT incrementally translates these forms into DRSs (boxes), in which the accessible items (discourse referents) are listed simply. S1 P(a)
S2 R(he,b)
S3 Q(c,a)
S4 T(d,him)
a
a, b, he
a, b, c, he
a, b, he, d, him
P(a)
P(a) he=a R(a,b)
P(a) he=a R(a,b)
P(a) he=a R(a,b) Q(c,a)
Q(c,a)
him=a T(d,a)
Figure 5. Incremental translation into DRSs with sets of accessible discourse referents
Accessibility in DRT is a non-gradable notion and the property of the discourse domain. The problem is that there is no additional “fine-tuning” between different accessible discourse referents in the same domain, a fact already noted by Bosch (1988, 207): Although a limitation of the search to discourse domains is certainly a step in the right direction, and although the assumption of discourse domains seems useful also for a number of other purposes, they are still too large to serve as search spaces for reference resolution. It is highly implausible, for instance, that a per-
Accessibility and definite noun phrases
sonal pronoun like she should be interpretable unambiguously only in a discourse where precisely one woman has occurred. A more promising alternative is the notion of temporary salience of particular discourse referents and a corresponding dynamic salience structure or focus struc ture, which may or may not be superimposed on discourse domains.
In Centering Theory, an alternative family of approaches to the representation of discourses (cf. Grosz et al. 1995), the discourse structure has three components: linguistic structure, intentional structure and an attentional state. The attentional state represents the availability of discourse referents at any given point in the discourse. Centering Theory assumes a fine-tuning among accessible discourse items, which is mirrored in the ranking of accessible items. The subject of the first sentence is the only item, and therefore the pronoun can refer back to it. In the third sentence, however, the actual subject becomes the highest ranking element; and in sentence four, the referent b has already disappeared, as sketched in Figure 6: S1
S2
S3
S4
P(a)
R(he,b)
Q(c,a)
T(d,him)
{a}
{a>b}
{c>a>b}
{d>a>c}
Figure 6. Ranking of accessible discourse referents in Centering Theory
This approach to the local coherence of discourse segments provides strategies for finding antecedents for anaphoric pronouns. However, it does not account for the antecedent of a definite noun phrase, which depends on a global accessibility. Centering Theory distinguishes between local and global accessibility or focusing, but describes only the local interaction with pronouns (Gordon et al. 1993, 312): Two kinds of focusing were distinguished: global focusing and immediate focusing. It was claimed that global focusing affected the production and interpretation
4. See Grosz et al. (1995) for a comprehensive presentation of Centering Theory, Gundel (1996) for a comparison with Relevance Theory, and Beaver (2004) for an OT-version of Centering and an evaluation. 5. Passonnneau (1998) gives an account of definite anaphoric noun phrases in Centering Theory. She discusses the principles according which such a anaphoric expression selects one out of many potential antecedents. However, she does not discuss cases in which two occurrences of the same expression can refer to different antecendents.
Klaus von Heusinger
of definite descriptions, whereas immediate focusing affected the production and interpretation of pronouns.
In summary, DRT perceives accessibility as a function of the structure of a discourse domain (box) that yields a set of discourse referents. The structure is defined by the construction rules of DRSs. Expressions like negation or modals create subdomains, and referring expressions like proper names and indefinites differ in whether they introduce their discourse referents in the current domain or in the main domain, reflecting the traditional concept of scope interaction. The set of accessible discourse referents constantly increases in a progressive discourse. However, there is no “competition” or ranking between different discourse referents in the same domain. Centering Theory, on the other hand, assumes that accessibility is a function of a discourse segment that yields an ordered set of discourse items. Linguistic expressions introduce discourse items into the set and rank them. Subsequent expressions not only introduce new elements but also change the ranking. Once a referent is not mentioned for a certain span of sentences, the associated discourse referent is eliminated. This model gives only constraints on pronominalization in local domains, but no rules for accessibility of anaphoric expression in general. Both approaches assume one set of accessible items for each discourse, which they describe in structural terms: DRT in scope interaction, and Centering Theory in grammatical roles. However, they neglect the role of the descriptive content of the expression by which the discourse item is introduced.
5. Salience and accessibility structure In this section I focus on the analysis of anaphoric expressions with descriptive material, i.e. definite noun phrases, and maintain that accessibility is not the function of a discourse that yields a single set of accessible elements; it is rather a function of the discourse that yields various ordered sets, each set corresponding to the descriptive content of a referring expression. 5.1
Uniquely accessible referent
In order to distinguish this notion of accessibility from the gradable and nongradable one, I call the accessible referent of the set s associated with a predicate P (representing the descriptive content) the “most-accessible” or the “uniquely accessible” referent of P. The phenomena which are captured by this notion are il6. See also Roberts (1998) and Cornisch (1999) for a more detailed comparison between DRT and Centering.
Accessibility and definite noun phrases
lustrated by the discourse in (9), which has the schematic structure (9a) and the representation (9b). Two discourse referents dn and dk are introduced by two occurrences of the indefinite noun phrase a student. The two discourse referents for the two occurrences of the anaphoric noun phrase the student are unambiguously related to one of the two already introduced referents. However, in DRT there is no formal tool to decide why to identify the discourse referent dl as the second definite noun phrase the student, with the discourse referent dk standing for the second indefinite noun phrase, and not with dm representing the first indefinite noun phrase. In Centering Theory, there is no principle to relate definite noun phrase s to their indefinite counterparts.
(9) The dean is very busy these days: this morning a student complained about his exam. The dean had to talk to the student for more than two hours. Then a student came to talk about his neighbors who play the trumpet every night. The dean moved the student to a different place.
(9a) P1(a student)... P2(the student)...P3(a student)... P4(the student) (9b) {..., dn,..., dm,... dk,..., dl...|... P1(dn)... P2(dm) & dn = dm.... P3(dk)... P4(dl) & dk = dl....}
While (9) is a constructed example and somewhat artificial, example (1), repeated as (10), is taken from the novel The Old Man and the Sea by Ernest Hemingway. Here, as in (9), different occurrences of one and the same definite noun phrase refer to different objects. (10) The two sharks1⊕2 closed together and as he saw the one1 nearest him open his1 jaws and sink them into the silver side of the fish, he raised the club high and brought it down heavy and slamming onto the top of the shark1’s broad head. He felt the rubbery solidity as the club came down. But he felt the rigidity of bone too and he struck the shark1 once more hard across the point of the nose as he slid down from the fish. The other shark2 had been in and out and now came in again with his2 jaws wide. The old man could see pieces of the meat of the fish spilling white from the corner of his2 jaws as he2 bumped the fish and closed his2 jaws. He swung at him2 and hit only the head and the shark2 looked at him and wrenched the meat loose. (Hemingway, The Old Man and the Sea, 64)
5.2
Salience change
In the view defended here, the first indefinite noun phrase a student changes the accessibility structure of the discourse in the following way: first, it introduces a new discourse referent a, and second, it makes this discourse referent the most salient one for the predicate student. This is illustrated in (10), here the first step (i).
Klaus von Heusinger
(10) represents only the salience change for the set of students. No other set is represented and neither are the predications in the sentence. For a formal reconstruction, see von Heusinger (2003) and Peregrin & von Heusinger (2004). The definite noun phrase the student refers to this student since it is the only student introduced so far, see (ii). The indefinite a student in the third sentence introduces a new discourse referent b and makes this discourse referent the most salient one for the set of students, as in (iii). Here I assume that the first introduced discourse referent is not overwritten, but only pushed back one place. This is a controversial issue, but does not play a role for the given example. We now can account for the reference of the occurrence of the student in the fourth sentence: it refers to the most salient student, who is associated with the discourse referent b, as in (iv). (11) Salience Change This morning a student complained about his exam. The dean had to talk to the student for more than two hours. Then a student came to talk about his neighbors who play the trumpet every night. The dean moved the student to a different place. P1(a student)... P2(the student)... P3(a student)... P4(the student) (i) (ii) (iii) (iv) [Students: {a}] [Students: {a}] [Students: {b>a}] [Students: {b>a}]
5.3
Salience spreading
The formalism given in the last section must be modified in order to catch, in addition, the salience-change potential of definite expressions. In the last section, it was assumed that definite noun phrases do not exhibit a salience-change potential since they would raise to salience an object that was already salient. However, the example (12), a slightly modified version of (2), clearly shows that definite expressions can change the actual accessibility of a discourse. The definite the small bird refers to a small bird. The subsequent definite the bird refers to the same one. We can explain this by assuming that an expression not only changes the most-accessible element of the set introduced, but also that of some relevant supersets of this set – a behavior which I have termed “salience spreading” (see von Heusinger 2003). The definite the small bird changes the most-accessible element of the set of small birds and that of the set of birds into the same element. Furthermore, the set of animate objects are also assigned the very same referent, accounting for the use of the pronoun he. This salience spreading is illustrated in (13) by the downward arrows, while the upward arrows indicate the licensing conditions for the different definite expressions: he refers to the small bird, since it is the most salient of the
Accessibility and definite noun phrases
animate objects; while the bird refers back to the original discourse item, since it is the most salient or accessible referent, etc. (12) The small bird1 came toward the skiff from the north. He1 was a warbler and flying very low over the water. The old man could see that he1 was very tired. The bird1 made the stern of the boat and rested there. Then he1 flew around the old man’s head and rested on the line where he1 was more comfortable. (Hemingway, The Old Man and the Sea, 28) [the small bird…]1
[he…]2
[Small_Birds: {a}]
[Small_Birds: {a}]
superset
superset
[Birds: {a}]
[Birds: {a}]
superset [animate objects: {a}]
[the bird…]3
superset
1=2
[animate objects: {a}]
3
Figure 7. Salience spreading
The box indicates a small fragment of the whole accessibility structure of the discourse. It also tries to indicate the dynamic change of the accessibility structure, which in this example is not very large. The noun phrase the small bird changes the most salient referent of the associated set and some supersets, the anaphoric pronoun he does not change the accessibility structure, however the anaphoric noun phrase the bird activates the referent back again to the set of birds and animate objects. As long as no clear competition exists, this update does not change the most salient element, but it still seems necessary to raise the activation above a certain level. In (12) the salience spreads along the lexically encoded relation of hyponymy (superset relation). However, in sections 2.3 we also discussed instances of pragmatic anaphora, where the relation is of a different type, e.g. the subset relation (the bird …. the beautiful bird). We also saw that such an anaphoric link needs additional plausibility and cannot be reconstructed solely by the descriptive material. I affirm, therefore, that such cases cannot be covered alone by the given mechanism on salience spreading. However, I assume that the given mechanism can be extended to certain cases of indirect anaphora, as discussed in section 2.4 in exam7. McCoy & Strube (1999) give an account for the choice of anaphoric pronouns vs. anaphoric full noun phrases in terms of distance, or more adequate: in terms of the temporal structure of the discourse.
Klaus von Heusinger
ples (4)-(6), repeated as (13)-(15). We can account for the anchoring of the indirect anaphoric expression by assuming the additional information is included in the accessibility structure. For (13) we need the encyclopedic knowledge that people are associated with governments, for (14) the lexical knowledge that a book has an author, and for (15) the frame that a restaurant has a waiter and a menu. (13) For the Parks and millions of other young Koreans, the long-cherished dream of home ownership has become a cruel illusion. For the government, it has become a highly volatile political issue. (Poesio & Vieira 1998, 198) (14) John bought a book. The author is well known. (15) John entered a restaurant and sat at a table. The waiter brought the menu.
6. Summary Interpreting a discourse we link referential expressions to each other, thus reconstructing the coherence of the discourse. There are many aspects of the coherence structure that determine the exact conditions under which an anaphoric expression is linked to its antecedents. Most of these parameters are summarized under the general term of “accessibility” or “accessibility structure”; this paper has investigated how accessibility and definite noun phrases interact. It is generally assumed that the definiteness status is determined by the accessibility structure, while definite noun phrases are interpreted as static terms, i.e. as terms that change the accessibility. In contrast to this commonly held view, it was shown that definite noun phrases dynamically change the accessibility structure. In order to model this dynamic behavior, we first need to revise our representation of accessible discourse items. These are sets of ranked discourse items with respect to the predicate by which they were introduced or activated, rather than one single set of ranked elements without further association to the predicate by which they were introduced. This model allows us to show the dynamic potential of definite noun phrases by the process of salience spreading, which also accounts for the difference between anaphoric pronouns and anaphoric definite nouns.
References Ariel, M. 1988. Referring and accessibility. Journal of Linguistics 24: 65–87. Ariel, M. 1990. Accessing Noun-Phrase Antecedents. London: Routledge. Beaver, D. 2004. The optimalization of discourse anaphora. Linguistic and Philosophy 27: 3–56.
Accessibility and definite noun phrases
Bosch, P. 1988. Representing and accessing focussed referents. Language and Cognitive Processes 3(3): 207–231. Chafe, W. 1976. Givenness, contrastiveness, definiteness, subjects, topics, and point of view. In Subject and Topic, C. Li (ed.), 25–55. New York NY: Academic Press, Christophersen, P. 1939. The Articles. A study of their theory and use in English. Copenhagen: Munksgaard. Consten, M. 2004. Anaphorisch oder deiktisch. Zu einem integrativen Modell domänengebundener Referenz. Tübingen: Niemeyer. Cornish, F. 1999. Anaphora, discourse, and understanding. Evidence from English and French. Oxford: Clarendon Press. Fraurud, K. 1990. Definiteness and the processing of noun phrases in natural discourse. Journal of Semantics 7:395–433. Gordon, P., Grosz, B. and Gilliom, L. 1993. Pronouns, names and the centering of attention in discourse. Cognitive Science 17: 311–397. Grosz, B., Joshi, A. and Weinstein, S. 1995. Centering: A framework for modeling the local coherence of dciscourse. Computational Linguistics 21, 203–225. Gundel, J. 1996. Relevance theory meets the giveness hierarchy. An account of inferrables. In Reference and Referent Accessibility, T. Fretheim and J. Gundel (eds), 141–153. Amsterdam: John Benjamins. Gundel, J., Hedberg, N. and Zacharski, R. 1993. Cognitive status and the form of referring expressions in discourse. Language 69: 274–307. Hajičová, E., Hoskovec, T. and Sgall, P. 1995. Discourse modelling based on hierarchy of salience. The Prague Bulletin of Mathematical Linguistics 64: 5–24. Hajičová, E. and Sgall, P. 1987. The ordering principle. Journal of Pragmatics 11: 435–454. Hegarty, M., Gundel, J. and Borthen, K. 2001. Information structure and the accessibility of clausally introduced referents. In NP Interpretation and Informantion Structure, K. von Heusinger and K. Schwabe (eds). Berlin: Mouton de Gruyter (Theoretical Linguistics 27(2/3): 163–186). Heim, I. 1982. The Semantics of Definite and Indefinite Noun Phrases. PhD dissertation, University of Massachusetts, Amherst. Ann Arbor MI: University Microfilms. Hemingway, E. 1962. The Old Man and the Sea. In Three Novels of Ernest Hemingway. New York NY: Scribners von Heusinger, K. 1997. Salience and definiteness. Prague Bulletin of Mathematical Linguistics 67: 5–23. von Heusinger, K. 2003. The double dynamics of definite descriptions. In Meaning in the Dynamic Turn, J. Peregrin (ed.), 150–168. Amsterdam: Elsevier. Kamp, H., van Genabith, J. and Reyle, U. To appear. Discourse representation theory. In Handbook of Philosophical Logic (2nd ed.,Vol. 13), D. Gabbay and F. Günthner (eds). Dordrecht: Springer. Kamp, H. [1981] 1984. A theory of truth and semantic interpretation. In Truth, Interpretation and Information, J. Groenendijk, T. M. V. Janssen and M. Stokhof (eds), 1–41. Dordrecht: Foris. Kamp, H. and Reyle, U. 1993. From Discourse to Logic. Introduction to modeltheoretic semantics of natural language, formal logic and discourse representation theory. Dordrecht: Kluwer. Karttunen, L. 1976. Discourse referents. In Syntax and Semantics 7: Notes from the linguistic underground, J. McCawley (ed.), 363–385. New York NY: Academic Press. Lewis, D. 1979. Scorekeeping in a language game. In Semantics from Different Points of View, R. Bäuerle, U. Egli and A. von Stechow (eds), 172–187. Berlin: Springer.
Klaus von Heusinger McCoy, K.F. and Strube, M. 1999. Generating anaphoric expressions: Pronoun or definite description. In Proceedings of the Workshop on the Relation of Discourse/Dialogue Structure and Reference held in conjunction with ACL ‘99. June 21, 1999, 63–71. Passonneau, R. 1998. Interaction of discourse structure with explicitness of discourse anaphoric noun phrases. In Centering Theory in Discourse, M. Walker, A. Joshi and E. Prince (eds), 327–358. Oxford: Clarendon. Peregrin, J. and von Heusinger, K. 2004. Dynamic semantics with choice functions. In Context Dependence in the Analysis of Linguistic Meaning, H. Kamp and B. Partee (eds), 309–329. Amsterdam: Elsevier. Poesio, M. 2003. Associative descriptions and salience: A preliminary investigation. In Proceedings of the ACL Workshop on Anaphora, Budapest, April 2003. Poesio, M. and Vieira, R. 1998. A corpus-based investigation of definite description use. Computational Linguistics 24: 183–216. Prince, E. 1981. Toward a taxonomy of given-new information. In Radical Pragmatics, P. Cole (ed.), 223–255. New York NY: Academic Press. Roberts, C. 1998. The place of centering in a general theory of anaphora resolution. In Centering Theory in Discourse, M. Walker, A. Joshi and E. Prince (eds), 359–399. Oxford: Clarendon Press. Schwarz, M. 2000. Indirekte Anaphern in Texten. Studien zur domänengebundenen Referenz und Kohärenz im Deutschen. Tübingen: Niemeyer. Sgall, P., Hajičová, E. and Benesová, E. 1973. Topic, Focus and Generative Semantics. Kronberg/ Taunus: Scriptor. Sgall, P., Hajičová, E. and Panevová, J. 1986. Meaning of the Sentence in its Semantic and Pragmatic Aspects (Ed. by J. Mey). Dordrecht: Reidel. Stalnaker, R. 1978. Assertion. In Syntax and Semantics 9: Pragmatics, P. Cole (ed.), 315–332. New York NY: Academic Press Webber, B. 1983. So what can we talk about now? In Computational Models of Discourse, M. Brady and R. Berwick (eds.). Cambridge MA: The MIT Press.
The non-subject bias of German demonstrative pronouns Peter Bosch, Graham Katz and Carla Umbach This paper presents evidence from both corpus studies and psycholinguistic experiments regarding the referential preferences of German personal pronouns (er, sie, es) and demonstrative pronouns (der, die, das). Both types of pronouns are naturally used in German to refer to both animate and inanimate referents, but they appear to have different preferences with respect to their choice of antecedents: While personal pronouns show a slight preference for antecedents which are grammatical subjects, demonstrative pronouns show a strong bias for non-subject antecedents.
1. Introduction Linguistic constraints on potential antecedents for anaphoric expressions have been investigated in syntactic theory under the rubric of Binding Theory (Chomsky 1981) and in more semantic settings in discourse representation theory (Kamp 1981) and centering theory (Grosz, Joshi and Weinstein 1995), among others. It is well known that syntactic, semantic, and pragmatic factors all play a role in resolving the question of which of the many potential antecedents for a particular use of a pronoun is intended by the author as the antecedent. Since Hobbs (1978), computational linguists have sought to investigate these preferences (Lappin and Leass 1994; Preuss et al. 1994; Mitkov 1998). In this paper we investigate a preference for antecedents that appears to be a lexical property of the pronominal element itself. The anaphoric expressions in question, the little-studied class of German demonstrative pronouns, appear to evidence a bias for non-subject antecedent noun phrases. German makes use of two sets of pronominal elements, the er (sie, es) series, which we will term personal pronouns, and the der (die, das) series, which we will term demonstrative pronouns. Both varieties of pronouns can be naturally used to refer to animate antecedents, but they appear to have slightly different preferences
Peter Bosch, Graham Katz and Carla Umbach
with respect to which antecedent is preferred. In particular while the personal pronouns show a familiar slight preference for antecedents which are grammatical subjects, demonstrative pronouns show a strong bias for non-subjects. In this paper we will present an investigation of these contrasting preferences, with evidence taken from both large scale corpora and from two psycholinguistic studies. Our results seem to confirm the results from Dutch and Finnish reported by Kaiser and Trueswell (2004; in press), and the analogous results reported for Italian by Carminati (2002). German personal and demonstrative pronouns appear at first blush to be synonymous, differing, if at all, only in register or affect. In (1), for example, either pro-form will be understood as referring back to the subject of the first sentence. (1) Ein Mann sprach mich an. {Der/Er} wollte wissen, wo die Kolpingstraße ist. [A man was addressing me. {DEM/He} wanted to know where Kolpingstraße is.]
But this synonymy does not extend to all contexts. In (2), for example, we find different interpretation preferences for the two pronouns: (2) Paul wollte mit Peter laufen gehen. Aber {er / der } war erkältet. [Paul wanted to go running with Peter. But {he /DEM} had a cold.]
Here, the pronoun er is most naturally interpreted as referring to Paul, the grammatical subject of the prior sentence, while the demonstrative der has a preference for Peter, the grammatical object. Subtle pragmatic factors can in some cases override the preferences, but absent these, the lexical preferences remain. The studies we report here support the intuition that such a preference exists.
2. Linguistic background The der (die, das) series demonstrative pronouns are for the most part homophonous with the definite determiners in German. While it might be suspected that these forms are to be analyzed as full definite noun phrases with elided head nouns, and that the difference evident in (2) is derivable from these properties, there are reasons to set this concern aside. First, the interpretation of demonstrative pronouns depends on salience information to a degree that we cannot find in the interpretation of full definite NPs (Gundel, Hedberg and Zacharski 1993). Secondly, 1. Only after we finished this paper we became aware of Comrie (1997), who investigated the opposition between personal and demonstrative pronouns for Dutch with results that are entirely compatible with our results for German and who was probably the first to carry out a systematic corpus study of this opposition. 2. In the English glosses we represent the German demonstrative simply by “DEM”.
The non-subject bias of German demonstrative pronouns
while many forms of the demonstrative pronouns look like definite determiners, there are forms, such as the genitive form and the dative plural, in which a difference becomes apparent. For example, the dative plural demonstrative pronoun is denen (3a) but the definite article is den (3b). (3) a. Denen ist es egal. b. Den Frauen ist es egal. [It doesn’t matter to them / the women.]
We have in much of the work limited our attention to the masculine singular form der (accusative den) because the other forms present problems which make direct investigation difficult. The feminine singular form die is syncretic with the plural prounoun and the neuter form das is used to refer to abstract objects such as facts and events (cf. also Consten et al., this volume). For a language like German with grammatical gender, this is less of a limitation than it would be for a natural gender language such as English. We have also ignored the morphologically more complex (and significantly rarer) demonstrative paradigms dieser, jener, derselbe, (‘this’, ‘that’, ‘the same’) and their likes. Here it is not clear that their uses are similar to the simple demonstrative pronouns, however. Traditional grammatical description of the difference between personal pronouns and demonstrative pronouns is surprisingly cursory. Duden (1984) suggests that personal and demonstrative pronouns are in free variation, with only a difference in tone to distinguish them. This appears to have to do with a difference in evaluative attitude toward the intended referent. In the case of (1), for example, the use of the demonstrative der appears to distance the speaker in some way from the man being referred to and is somehow pejorative. We have nothing to say about this potential pragmatic difference here. Our investigation is more semantically focused. The IDS Grammar (Zifonun et al. 1997) suggests, in this regard, that the semantic difference is that personal pronouns are used if the antecedent relates to a topic which has been established before whereas the demonstrative pronoun is used if the antecedent is part of the rheme or is in contrast with other topics. It is also a common observation that the use of demonstrative pronouns is limited to or primarily to be found in spoken language. This observation is one which we were able to test directly. We compared in two different corpora, one of spoken German and one of written German texts, the number of demonstrative uses of der with the number of uses of er (the corresponding personal pronoun). While in our newspaper corpus fewer than 7% of the masculine pronoun tokens 3. The two corpora were the Verbmobil corpus (http:// verbmobil.dfki.de/), a corpus of spoken appointment-scheduling dialogues and the NEGRA corpus (http://www.coli.uni-sb.de/ sfb378/negra-corpus/negra-corpus.html), a collection of daily newspaper text.
Peter Bosch, Graham Katz and Carla Umbach
were demonstratives, in our spoken language dialog corpus approximately 80% were. This is a dramatic difference, but one which is likely as much attributable to text genre as to the spoken/written contrast. While this is an interesting finding in its own right, it tells us little about the factors that contribute to the choice of one pronominal form over the other, given that in both spoken and written language, both forms are used. In the next sections we will investigate one of the factors that appears to be crucial in determining how demonstrative pronouns are used, namely the grammatical function of the intended antecedent. We sought to investigate the contrast using two complementary methodologies. First we examined the distribution of demonstrative and personal pronouns in text corpora, looking both at the sentence internal distribution and the pronoun antecedent relation. The corpus studies are reported in Section Three. We also carried out two psycholinguistic studies, these are reported on in Section Four. Our goal here is not to account for the observed differences but to establish what the differences are that need to be accounted for. In the concluding section we will offer some speculations about how to explain the observed differences.
3. Corpus investigation We made use of a hand-parsed corpus of newspaper texts to investigate the distributional patterns associated with personal and demonstrative pronouns. In this study we investigated two different aspects of pronoun use, their distribution of the pronominal element within its own sentence and the distributions of the antecedents for these elements in prior sentences. In both cases there were striking differences between the two classes of pronouns. 3.1
Sentence-internal distribution
The first question we investigated was whether there were obvious differences in the sentence-internal distribution of demonstrative and personal pronouns in simple German sentences. As is well known, German is a verb-second language, with a distinguished topic position preceding the tensed verb of the main clause. This pre-verbal position can be filled by any of the grammatical arguments of the verb, or by adverbial material. This means that in simple transitive clauses there are four different potential positions that a pronoun can occupy: pre-verbal sub4. The corpus used was the NEGRA corpus. NEGRA is a POS-tagged and syntactically annotated corpus of German of 355,000 words, a subset of the Frankfurter Rundschau corpus, which is also available as a tree bank. See http:// www. coli. uni-sb.de/sfb378 /negra-corpus/negra-corpus.html.
The non-subject bias of German demonstrative pronouns
ject, post-verbal subject, post-verbal object and pre-verbal object. These are illustrated in (4). (4) a. {Er/Der} hat Hans gesehen. [{He/DEM} has Hans seen] b. Hans hat {er/der} gesehen. [Hans has {he/DEM} seen] “He has seen Hans” c. Hans hat {ihn/den} gesehen. [Hans has {him/DEM} seen]
% of total number of tokens of each pronoun type
d. {Ihn/den} hat Hans gesehen. [{Him/DEM} has Hans seen] “Hans has seen him” 100 90 80 70 60 demonstrative
50
personal
40 30 20 10 0 pre-verbal subject
pre-verbal object
post-verbal subject
post-verbal object
Figure 1. The distribution of demonstratives and personal pronouns over different syntactic positions
In this study we compared the number of token uses of personal pronouns in each of the four positions to those of demonstrative pronouns. The results are indicated in Figure 1. The first thing to note is the clear dominance of subject uses over object uses for both types of pronouns (in both cases about four to one). In contrast to personal pronouns, demonstrative pronouns have a clear preference for preverbal positions. Approximately 93% of all demonstrative tokens occur in this position, either as subjects or objects. Personal pronouns used as subjects are evenly
Peter Bosch, Graham Katz and Carla Umbach
divided between pre- and post-verbal uses, while personal pronouns used as objects are overwhelmingly post-verbal (not even 0.5% are used pre-verbally). This contrasts strikingly with demonstrative pronouns, where pre-verbal object uses far dominate post-verbal object uses. We might summarize by saying that the pre-verbal subject position is a natural position for both demonstrative and personal pronouns, while the post-verbal subject position is an unnatural position for demonstrative pronouns and the preverbal object position is an unnatural position for personal pronouns. The preferred syntactic positions of a pronoun can, of course, give only a very indirect hint at the discourse parameters governing its use. Even if we believe that the choice of syntactic position is to some extent a function of information structure, we expect the pronoun-antecedent relation to be of central importance to this contrast. A natural next step was to compare the antecedents of demonstrative pronouns with those of personal pronouns. 3.2
Preference for antecedent types
For this study, in order to investigate the potential differences in antecedent-pronoun relations between demonstrative and personal pronouns we hand-annotated a corpus of German newspaper texts with pronoun antecedent relations. For practical reasons, we limited ourselves to cases in which the antecedent for a pronoun appeared in the immediately preceding sentence. Even in these cases there are a number of aspects of the pronoun-antecedent relation that one could choose to investigate. The grammatical function of the antecedent, the distance of the antecedent measured in syntactic nodes, the depth of embedding of the antecedent, the semantic class of the antecedent, syntactic parallelism between anaphora and antecedent are all factors that have been suggested to play a role in anaphora resolution (cf. Lappin and Leass 1994). As a first step, we chose to investigate the grammatical function of the antecedent NP. We compared the number of uses of demonstrative pronouns with antecedent NPs that were grammatical subjects of their sentences with the number of uses of in which the antecedents were objects. The same comparison was made for personal pronouns. The four types of discourses illustrated in (5) and (6) were compared, where we have underlined the pronouns and their antecedents. (5) a. Hans reparierte den Rolladen. Der hatte sich verklemmt. [Hans was repairing the roller blind. DEM had got stuck.] 5. We used again the NEGRA corpus. As one might expect, in very few cases was the intended antecedent unclear. These cases were discarded. The counts include singular masculine and feminine and plural forms. Neuter forms were excluded.
The non-subject bias of German demonstrative pronouns
b. Hans reparierte den Rolladen. Er fluchte ununterbrochen. [Hans was repairing the roller blind. He kept swearing all of the time] (6) a. b.
Hans reparierte den Rolladen. Der fluchte ununterbrochen. [Hans was repairing the roller blind. DEM kept swearing all of the time] Hans reparierte den Rolladen. Er hatte sich verklemmt. [Hans was repairing the roller blind. It had got stuck.]
% of total number of tokens of each pronoun type
The results, displayed in Figure 2, show a strong preference of demonstratives for object NP antecedents; Personal pronouns, on the other hand, show a complementary behaviour and prefer subject NPs as their antecedents. In other words, discourses such as (5a) and (5b) are common, while those such as (6a) and (6b) are rare. The frequencies displayed in Figure 2 reflect quite well native speakers’ intuitions about the relevant uses. The discourses in (5) are judged natural, while those in (6), in particular the one in (6a), appear less natural. Our corpus studies showed us that there was indeed a striking difference in use between demonstrative and personal pronouns. The typical use of a demonstrative is as a pre-verbal subject referring back to an NP object of the prior sentences, while the typical use of a pronoun is as a pre-or post-verbal subject referring back to the NP subject of the prior sentence. We sought to confirm that this difference in use is correlated with a processing difference by carrying out two psycholinguistic studies. These are described in the next section.
100 90 80 70 60
demonstrative
50
personal
40 30 20 10 0
subject antecedents
object antecedents
Figure 2. The distribution of demonstratives and personal pronouns over two types of antecedents
Peter Bosch, Graham Katz and Carla Umbach
4. Experiments The two psycholinguistic experiments we carried out were designed to complement each of our corpus studies. The first experiment investigated the naturalness of demonstrative and personal pronouns in the various sentence-internal syntactic positions, while the second investigated the naturalness of the pronoun-antecedent relation for various positions of the antecedent. 4.1
Experiment one: sentence-internal parameters
Thirty students of the University of Osnabrück participated in the first experiment (9 female and 21 male). Experimental materials consisted of a total of 64 sentences. Thirty-two of these, 16 containing a personal pronoun and 16 containing a demonstrative pronoun, were extracted form the Frankfurter Rundschau corpus. The sentences were chosen so that the position of the pronoun varied along the parameters pre-verbal/post-verbal position and subject/object, as illustrated in (7): (7) a. pre-verbal subject: b. pre-verbal object: c. post-verbal subject: d. post-verbal object:
Der stellte der Hausverwaltung ein Ultimatum. [DEM put an ultimatum to the property administration.] Doch den könnte die Gemeinde verteilen. [But the council could distribute DEM.] Ein Buch wollte er nicht geschenkt haben. [A book he didn’t want for a present.] Man konnte ihn nicht mehr erneuern. [One couldn‘t renew it any more.]
This was the naturally occurring set. The other 32 were generated by substituting a demonstrative for each personal pronoun and conversely a personal pronouns for each demonstrative. The intent was to determine whether the naturally occurring set was judged more normal sounding by the participants than the artificially generated set, the assumption being that the sentence-internal factors that influenced the original author to use the form that he did use would make the other form in some way unnatural. This we investigated via a self-paced reading study and a follow-up questionnaire study. In the self paced reading study, participants were asked to read 16 of the naturally occurring sentences and 16 artificial sentences. The stimuli were evenly divided among the different sentence types. Participants were asked to read the sentences as they appeared on a computer monitor carefully and with a normal speed. 6. The Frankfurter Rundschau Corpus is a large (41 Mio words) collection of German national daily newspaper text.
The non-subject bias of German demonstrative pronouns
The sentences were displayed in a word by word fashion, with only one word displayed at a time. The reading times for each word were recorded. After each sentence a comprehension question was asked. An analysis of the results showed no overall significant difference between the time course for reading the naturally occurring sentences and the artificially generated sentences. Also an analysis of variance with z-transformed reading times (so that the reading times of different subjects become comparable) showed no significant differences for the reading times of the target pronoun or the following word. While we had hoped to find that the reading times of the substituted pronouns was elevated, indicating recognition of a disfluency, the lack of significant elevation may only indicate that the effect is too small to be detected by this method. After the completion of the self-paced reading study, subjects were presented with a list of all 64 sentences in pairs of original and substitute and were told that one sentence in each pair had actually appeared in a newspaper while the other one had not and they were asked to guess which was which. If a subject marked a sentence as “appeared in print” this was interpreted as an intuitive preference for the relevant choice of pronoun. The analysis of the responses, shown in Figure 3, revealed a clear and significant preference for the personal pronoun in the postverbal position of the sentence. The evident preference of the demonstrative for the pre-verbal positions was suggestive, but not significant. 100 90 80
Percent
70 60
demonstrative
50 40
personal
30 20 10 0 pre-verbal subject
pre-verbal object post-verbal subject post-verbal object
Figure 3. Questionnaire results. Preference for personal and demonstrative pronouns in different syntactic positions
While the result of the questionnaire study is congruent with the quantitative corpus study reported above, there is very little that can be concluded.
Peter Bosch, Graham Katz and Carla Umbach
4.2
Experiment two: discourse parameters
While the first experiment looked at sentences in isolation, the second experiment takes both discourse context and background knowledge into account. We saw already in example (2) that the interpretation preferences for personal and demonstrative pronouns are sensitive to the way the potential referents are introduced in preceding sentences. And, as always, interpretation preferences may be influenced by what a listener or reader expects on the basis of general world knowledge. Consider (8) (8) Paul spielte Tennis mit Peter. {Er / Der} hat ihn gewinnen lassen. [Paul was playing tennis with Peter. {He /DEM} let him win.]
It is plausible to assume that any knowledge of the reader of (8) about how good Peter and Paul are at tennis would influence the interpretation of the pronoun. If the reader knows that Paul is an international champion while Peter plays tennis only once in a while, the likely interpretation is going to be that Paul let Peter win – never mind the type of pronoun. So we may expect a competition between intrinsic interpretation preferences of the pronoun type, as they are suggested by our corpus data, and world-knowledge-based preferences. The basic idea in experiment two is to make use of this competition. If an antecedent which is preferred due to factors relating to grammar is unlikely due to world knowledge, interpretation load should be increased. Accordingly, we designed the stimulus materials taking both factors into account. Each stimulus set consisted of four lines of text. The first line (headline) introduces a situation and the second (context sentence) brings in two protagonists. The third line (target sentence) refers to one of the protagonists by either a demonstrative or a personal pronoun, and the final line (completion sentence) contains a gap, which subjects are asked to fill in with a noun so that this line would paraphrase their understanding of the target sentence. For example: 1. Im Krankenhaus [At the hospital] 2. Der Oberarzt untersucht den Notfallpatienten. [The senior doctor is examining the emergency patient.] 3. Er ist gerade erst gekommen. [He has only just arrived.] 4. Der ____ ist gerade erst gekommen. [The ___ has only just arrived.]
In this example world knowledge implies no preference for the doctor or the patient as a referent for the target pronoun. The intrinsic interpretation preferences
The non-subject bias of German demonstrative pronouns
for the personal pronoun in the target sentence, however, may still lead to an interpretation that takes the doctor as the referent. Stimulus sets were varied with respect to the parameters world knowledge bias and pronoun type. In each stimulus set world knowledge would support either the first or the second protagonist, that is, either the subject or the object of the context sentence as the referent of the target pronoun, or be neutral in this respect. The pronoun would either be a demonstrative or a personal pronoun. So we had six versions of the stimulus set with identical headline and context sentence, differing with respect to the target sentence and the completion sentence: 1. Im Krankenhaus [At the hospital.] 2. Der Oberarzt untersucht den Notfallpatienten. [The senior doctor is examining the emergency patient.]
Version one (world knowledge supporting the subject; personal pronoun)
3. Er ist nämlich Herzspezialist. [He is a heart specialist.]
4. Der ___ ist nämlich Herzspezialist. [The ____ is a heart specialist.]
Version two (world knowledge supporting the subject; demonstrative pronoun)
3. Der ist nämlich Herzspezialist. [DEM is a heart specialist.]
4. Der ___ ist nämlich Herzspezialist.
Version three (world knowledge supporting the object; personal pronoun)
3. Er muß sofort operiert werden. [He must be operated on at once.]
4. Der ___ muß sofort operiert werden. [The ___ must be operated on at once.]
Version four (world knowledge supporting the object; demonstrative pronoun)
3. Der muß sofort operiert werden. [DEM must be operated on at once.]
4. Der ___ muß sofort operiert werden.
Version five (world knowledge support neutral; personal pronoun)
3. Er ist gerade erst gekommen. [He has only just arrived.]
4. Der ___ ist gerade erst gekommen. [The ___ has only just arrived.]
Version six (world knowledge support neutral; demonstrative pronoun)
3. Der ist gerade erst gekommen. [DEM has only just arrived.]
4. Der ___ ist gerade erst gekommen
Peter Bosch, Graham Katz and Carla Umbach
Seven such sets were constructed: Five in which the context sentence is in SVO (subject-verb-object) order (e.g. Der Oberarzt untersucht den Notfallpatienten) and two with context sentences in OVS (object-verb-subject) order (e.g. Den Notfallpatienten untersucht der Oberarzt). This yields a total of 42 stimulus sets. To prevent participants from getting used to the routine of choosing antecedents we interspersed additional stimulus sets as distractors, in which there was a different completion task in line four. 69 students of the University of Osnabrück participated in the experiment (23 female and 46 male). Participants were randomly assigned to one of six groups so that each participant only saw one version of each block of stimulus sets. Participants read the stimulus sets line by line as they were displayed in the upper left corner of a computer monitor. Each set was started by a star displayed at the position of the first character of the first line to focus the participant’s attention before the stimulus sets were displayed. Only one line of the stimulus set was displayed at a time. When the line with the gap was shown, participants had to type in their completion. By pressing the enter key the stimulus set was finished and the next one was started. The reading times for the target sentence were measured and the answers that participants typed into the gap were recorded. 4.2.1 Reading time results The reading times for the target sentence are as presented in Figures. 4 and 5. The columns for ‘subject support’ represent the averaged reading times in milliseconds for those targets in which world knowledge would support an interpretation of the pronoun with reference to the subject of the context sentence and the columns for ‘object support’ give the reading times when world knowledge suggests the object of the context sentence as the appropriate referent. The columns given as ‘unsupported’ are for those cases where world knowledge does not induce any preference. As Figure 4 shows for context sentences with SVO structure, reading times are longer for the demonstrative pronoun when the pronoun is interpreted in reference to a preceding subject protagonist. Thus in the context of (9) continuation (a) takes more time to read than (b): (9) Der Oberarzt untersucht den Notfallpatienten. a. Der ist nämlich Herzspezialist. b. Er ist nämlich Herzspezialist.
The non-subject bias of German demonstrative pronouns
SVO Structure
Time ms
4000 3500 3000 2500
demonstrative
2000 1500
personal
1000 500 0 Subject support Object support
Unsupported
Figure 4. Reading times with SVO context sentences
OVS Structure 4000 3500
Time ms
3000 2500
demonstrative
2000
personal
1500 1000 500 0 Subject support Object support
Unsupported
Figure 5. Reading times with OVS context sentences
And the same holds for context sentences with OVS structure, as shown in Figure 5. Also here (a) takes significantly longer than (b). Cf. (10).
7. The reading times displayed in Figs. 4 and 5 are not directly comparable because the number of stimulus sets was different. They were measured for SVO structure on 30 different stimulus sets, while the OVS structure only had 12 different stimulus sets. So the results for the subject verb object structure are more reliable.
Peter Bosch, Graham Katz and Carla Umbach
(10) Den Notfallpatienten untersucht der Oberarzt. a. Der ist nämlich Herzspezialist. b. Er ist nämlich Herzspezialist
The remaining reading time results for object support and the unsupported cases unfortunately show too much variation among participants to count as significant. 4.2.2 Completion results The results of the completion study, i.e., the answers that were filled into the gap in the fourth line of the stimulus sets are as in Figs. 6 and 7 (demonstrative pronouns are displayed as ‘DEM’ and personal pronouns as ‘PERS’):
Percent
SVO structure
100 90 80 70 60 50 40 30 20 10 0
Subject Object
DEM
PERS
subject support
DEM
PERS
object support
DEM
PERS
unsupported
Figure 6. Choices for the subject or the object of the SVO context sentence as referent for the target pronoun.
The central observation is that wherever there is a world knowledge bias for the interpretation of the pronoun, world knowledge is clearly the dominating factor in pronoun comprehension. In the unbiased cases we find a clear preference for one interpretation only where the demonstrative pronoun refers to the context sentence object in SVO structure (cf. Figure 6, unsupported, DEM), while in the remaining cases (the other three column pairs of the unbiased cases) we can only detect tendencies: Preference for the object of the context sentence as antecedent for the demonstrative also in OVS structure and preference for the subject of the context sentence as antecedent for the personal pronoun in both SVO and OVS structure.
The non-subject bias of German demonstrative pronouns
Percent
OVS structure 100 90 80 70 60 50 40 30 20 10 0
Subject Object
DEM
PERS
subject support
DEM
PERS
object support
DEM
PERS
unsupported
Figure 7. Choices for the subject or the object of the OVS context sentence as referent for the target pronoun.
We can further observe that in the condition where world knowledge supports the subject of the context sentence, more participants disregard world knowledge and choose the object when the pronoun is a demonstrative than in those cases where the pronoun is a personal pronoun. Analogously, although less clearly, when world knowledge has a bias for the object of the context sentence more participants disregard world knowledge and choose the subject when the pronoun is a personal pronoun than in those cases where the pronoun is a demonstrative. In sum, the completion study strongly supports the results from the reading time study: In both SVO and OVS conditions, the natural preference for a demonstrative would seem to be a reference to the object of the context sentence and for a personal pronoun the preference would be the subject of the context sentence. When these preferences agreed with the world-knowledge bias, reading times were shortest. 4.2.3 Questionnaire study After the experiment participants were given a questionnaire with a list of all sentences they had seen and they were asked to recall which type of pronoun they had seen in the target sentence during the experiment. If we assume that participants would not normally be able to remember which pronoun they had seen, their answers may be taken to reflect their preference for a particular pronoun type, given a particular interpretation and context. Since in the ‘unsupported’ cases the choice of pronoun is not constrained by any interpretation, the answers were expected to be arbitrary. The results are given in Figures 8 and 9.
Peter Bosch, Graham Katz and Carla Umbach
Percent
SVO Structure 100 90 80 70 60 50 40 30 20 10 0
match no match
P_DEM
P_PERS
subject support
P_DEM
P_PERS
object support
P_DEM
P_PERS
unsupported
Figure 8. Recall of pronoun use in after-experiment questionnaire
Percent
OVS Structure 100 90 80 70 60 50 40 30 20 10 0
match no match
P_DEM
P_PERS
subject support
P_DEM
P_PERS
object support
P_DEM
P_PERS
unsupported
Figure 9. Recall of pronoun use in after-experiment questionnaire
‘P_DEM’ means that the sentence was presented with the demonstrative pronoun der during the reading time study. ‘P_PERS’ means that the sentence was presented with the personal pronoun er. The columns show matches between what the participants saw during the experiment and what they ‘recalled’ when filling in the questionnaire. In the results for the SVO structure in Figure 8 we find a preference for the personal pronoun to refer to the subject and the demonstrative to refer to the object of the context sentence, as we should have expected from the reading time and completion study. The OVS structure displayed in Figure 9, however, is less clear. The tendency is indeed the same as in the SVO cases, with the exception of those targets that used the personal pronoun for reference to the context sentence object, which is re-
The non-subject bias of German demonstrative pronouns
membered largely correctly and against the hypothetic preference for a demonstrative under these conditions. This may be due to a number of factors, but we have to be very careful at this point, because, as already mentioned, the OVS results rest on a fairly small number of stimulus sets (12 vs. 30 sets for SVO).
5. Conclusion The corpus study as well as the experimental study show a clear preference of demonstrative pronouns to refer to object antecedents. These findings, however, raise the question of whether the grammatical function of the potential antecedent is the only factor biasing the choice of the antecedent for demonstrative pronouns, or whether this is correlated with another factor such as recency, salience, nontopicality, rheme, or accenting. Recency is ruled out by our findings because in the OVS condition the subject is mentioned only after the object and still the object is the preferred antecedent of the demonstrative pronoun. Our findings are consistent with the Complementarity Hypothesis put forward in Bosch et al. (2003), which claims that demonstrative pronouns prefer non-topical referents while personal pronouns prefer referents which are established as discourse topics, assuming that topics are established by nominative NPs. Our findings are also consistent with the results of Winter (2003) who investigated demonstrative pronouns in the framework of Centering Theory (Grosz et al. 1995) showing that demonstratives choose their antecedents by inducing the least coherent transition relation (i.e. rough shift > smooth shift > retain > continue). Again, the underlying assumption is that subject antecedents are more salient than object antecedents. In general, our findings are consistent with the widely accepted idea that lower-ranked referring expressions refer to less salient antecedents (cf., e.g., Gundel et al. 1993), assuming that demonstrative pronouns rank lower than personal pronouns, and that objects are less salient than subjects. For Dutch personal and demonstrative pronouns Kaiser and Trueswell (2004) show that in the case of personal pronouns the subject is preferred over the object while in the case of demonstrative pronouns the object is even more preferred over the subject. Where English is concerned, examples like (11) seem to suggest that German demonstrative pronouns pair with stressed personal pronouns in English. This as8. Actually, in Dutch two forms of personal pronouns have to be distinguished, a reduced form (unstressed) and a full form (mostly stressed). Different from what would be predicted by the correspondence of salience of antecedents and rank of expressions, Kaiser and Trueswell found that both reduced and full pronouns show a clear preference for the subject as the antecedent.
Peter Bosch, Graham Katz and Carla Umbach
sumption is disproved, however, by cases like (12) and (13). In (12a) the stressed pronoun refers to John whereas the demonstrative in the corresponding German sentence in (12b) does not and is actually hard to interpret. Similarly, in (13) translating the demonstrative pronoun by a stressed personal pronoun makes the sentence infelicitous because it is unclear what the antecedent of the stressed pronoun should be. The examples in (11a) and (12a) stem from Kameyama (1999) who claims that English stressed pronouns take the complementary preference of the unstressed counterpart where preferences include constraints given by world knowledge. This predicts that a stressed pronoun is interpreted as referring to an antecedent unexpected from the point of view of world knowledge; in (12a), for example, the preference for the subject is due to the fact that the agent of a hitting event is less likely to be injured. In contrast, in the case of German demonstrative pronouns preference is clearly guided by grammatical factors. World knowledge may of course interfere, strengthening or weakening an interpretation preference, but it is not an intrinsic factor as in the case of English stressed pronouns. (11) a. Paul called Jim a Republican. Then HE insulted HIM. HE = Jim b. Paul nannte Jim einen Betonkopf. Dann beleidigte der ihn. der = Jim (12) a. John hit Bill. Then HE was injured. b. John hat Bill geschlagen. ??Dann wurde der verletzt.
HE = John der =??
(13) a. Paul wollte mit Peter laufen gehen. Aber der war erkältet. b. Paul wanted to go running with Peter. But HE had a cold.
der = Peter HE =??
The question of whether the preference we found for German demonstrative pronouns to refer to object antecedents is a genuine effect of the grammatical function or a side effect triggered by some factor correlating with grammatical function cannot be settled by our data. Still, there is one observation casting doubt on the idea that the use of demonstrative pronouns is exclusively determined by the grammatical role of the antecedent, namely the clear preference of demonstrative pronouns to occur in the preverbal position of the sentence (cf. Section 3.1). There is no reason in the first place why a pronoun that prefers non-subject antecedents should preferably occur in a pre-verbal position. Word order in German is largely guided by information-structural requirements. Preference of demonstratives for a pre-verbal position suggests that information structure plays a crucial role in the use of demonstrative pronouns. This brings us back to (Zifonun et al. 1997) who propose that demonstrative pronouns are used if the antecedent is part of the rheme or is in contrast with other topics, and we might speculate that demonstrative pronouns are used for a topic change which would be supported by the findings in Winter (2003). Still, any explanation of the contrast between personal and demonstrative pronouns in terms of topicality or the theme-rheme divi-
The non-subject bias of German demonstrative pronouns
sion has to relate these notions to grammatical function to account for the non-subject bias of demonstrative pronouns.
Acknowledgements The work reported here was carried out in a study project of the Cognitive Science International Master Programme at the University of Osnabrück. Project participants were Phil Cummins, Boris Gutbrod, Kyoung-ho Park, Tom Rozario, Yufan Zhao and as staff advisers Peter Bosch, Graham Katz, Carla Umbach, and Jaqueline Griego. The authors gratefully acknowledge the contributions of all participants and advisers. The corpus data reported in Section 3 were already discussed in Bosch et al. (2003).
References Bosch, P., Rozario, T. and Zhao, Y. 2003. Demonstrative pronouns and personal pronouns. German der vs. er. In Proceedings of the EACL2003. Budapest. Workshop on the Computational Treatment of Anaphora. Comrie, B. 1997. Pragmatic binding: Demonstratives as anaphors in Dutch. In Proceedings of the Annual Meeting of the Berkeley Linguistic Society, Vol. 23, 50–61. Carminati, M.N. 2002. The Processing of Italian Subject Pronouns. PhD dissertation, University of Massachusetts, Amherst MA. Chomsky, N. 1981. Lectures on Government and Binding. Dordrecht: Foris. Duden. 1984. Grammatik der deutschen Gegenwartssprache (Ed. by G. Drosdowski). Mannheim: Bibliographisches Institut. Grosz, B. J., Joshi, A. K., and Weinstein, S. 1995. Centering: A framework for modeling the local coherence of discourse. Computational Linguistics, 21(2), 203–225. Gundel, J., Hedberg, N. and Zacharski, R. 1993. Cognitive status and the form of referring expressions in discourse. Language 69: 274–307. Hobbs, J.R. 1978. Resolving pronoun references. Lingua 44: 311–338. Kaiser, E. and Trueswell, J. (in press). The referential properties of Dutch pronouns and demonstratives: Is salience enough? In Proceedings of the Sinn und Bedeutung 8. Konstanz: Universität Konstanz, FB Sprachwissenschaft. Kaiser, E. and Trueswell, J. 2005. Investigating the interpretation of pronouns and demonstratives in Finnish: Going beyond salience. In The Processing and Acquisition of Reference, E. Gibson and N. Pearlmutter (eds). Cambridge MA: The MIT Press. Kameyama, M. 1999. Stressed and unstressed pronouns. In Focus, P. Bosch and R. van der Sandt (eds). Cambridge: CUP. Kamp, H. 1981. A theory of truth and semantic representation. In Formal Methods in the Study of Language, Vol.136, J. Groenendijk, T.M.V. Janssen and M. Stokhof (eds), 277–322. Dordrecht: Kluwer. Lappin, S. and Leass, H. 1994. An algorithm for pronominal anaphora resolution. Computational Linguistics 20(4): 535–561.
Peter Bosch, Graham Katz and Carla Umbach Mitkov, R. 1998. Robust pronoun resolution with limited knowledge. In Proceedings of COLINGACL’ 98. Montreal, Canada. Preuss, S., Schmitz, B., Hauenschild, C. and Umbach, C. 1994. Anaphora resolution in machine translation. In Text and Context in Machine Translation [Studies in Machine Translation and Natural Language Processing 6], W. Ramm (ed.), 29‑52. Dordrecht: Kluwer. Winter, Y. 2003. Parameter zur Bestimmung der Referenz von Demonstrativpronomina in Texten. Bachelor thesis, Institute of Cognitive Science, University of Osnabrück, www.cogsci. uni-osnabrueck.de/~CL Zifonun, G. Hoffmann, L. and Strecker, B. 1997. Grammatik der deutschen Sprache. Berlin: Walter de Gruyter.
Anaphoric properties of German right dislocation Maria Averintseva-Klisch I argue that the so-called ‚German right dislocation’ in fact comprises two formally and functionally distinct constructions, which I label right dislocation proper and afterthought. The main focus of the paper lies on right dislocation proper (RD) and its discourse function. I show, that RD is used as a discourse-structuring device. RD marks the referent of the right-dislocated noun phrase as the discourse topic for the following discourse segment. This explains certain anaphoric constraints on the following discourse segment. Thus, the corresponding referent has to be anaphorically resumed as sentence topic of the corresponding sentence at the level of the ‘main structure’ (i.e., the main story line).
1. Introduction German right dislocation (‘Rechtsversetzung’, cf. Altmann (1981)) is a construction like (1), which consists of an NP at the right end of the clause and a coreferent pro-form (pronoun (1a) or a full lexical NP (1b)) inside the clause: (1) a. Eri ist halt wirklich eine dämonische Figur, der Franz Josefi. Hei is particle really a demonic figure, the Franz Josephi. [Altmann (1981: 55)] b. Und dann passierte das Unglücki, (ich meine) dieser And then happened the misfortunei, (I mean) this 1. This paper is based on research financially supported by DFG within the graduate school ‘Economy and Complexity in Language’ (Humboldt University Berlin / Potsdam University). I would like to thank my supervisor Claudia Maienborn for her constant support, my fellow students at the graduate school and the audience of the DGfS workshop for helpful discussions and criticism, and Manfred Consten (Univ. of Jena) for answering my innumerable questions concerning anaphora and for helpful comments on the first draft of this paper.
Maria Averintseva-Klisch
schreckliche Autounfalli. terrible traffic-accidenti.
Since Altmann (1981) it is generally assumed that German right dislocation (GRD) is a device of spoken German, which enables the speaker to resolve or ‘repair’ a (pro)nominal reference that is unclear and thus might overcharge the hearer (see e.g. also Auer (1991), Selting (1994), Uhmann (1993, 1997), Zifonun et al. (1997)). This cannot, however, be the whole story: there are cases, as in (2), where the pronominal reference is unambiguously clear: (2) a.
(“Der Taifun!” rief Lukas dem Kapitän zu. “Da ist er!”) (“The typhoon!” Lukas called to the captain. “Here it comes!”) Ja, da war eri, der Taifuni. Yes, there was iti, the typhooni. [Ende, M. (1990): Jim Knopf und die Wilde 13: 190]
b. (Den Tag, den vergess’ ich nicht,) deri war viel zu schön, der Tagi. (That day, I’ll never forget it,) d-proni was much too wonderful, the dayi. [Altmann (1981:129)]
Here, the function of the GRD is not to disambiguate a pronominal reference, as it is not ambiguous at all, but to mark the referent of the right-peripheral NP as being especially important for the following discourse. In other words, the typhoon (2a) and, respectively, the day (2b) is set as the “subject” or “theme” (in the pretheoretical sense) for the following discourse segment. In (2a), the following segment actually offers a detailed description of the typhoon, of the dark waves, wind and so on. It has been noticed in the literature that right dislocation might have an additional function of “attracting the attention of the addressee” to the right-peripheral NP (Zifonun et al. (1997:548)). I argue that disambiguation of a pronominal reference and marking the importance of the discourse referent are not two complementary functions of one construction, but that there are in fact two constructions subsumed under the label of GRD: right dislocation proper (further right dislocation, RD) and afterthought (AT). These two constructions differ not only in their function, but also in their prosodic, syntactic and semantic features. In the following I will shortly summarize these differences (section 2). Then I will view in detail the discourse functions of RD and AT (section 3). The main focus of the paper is RD. I will show that RD marks the referent of the right-peripheral NP as the ‘discourse topic referent’ for the following discourse segment. The discourse referent in question has to be already established in the discourse model; in this sense RD is anaphoric, and not cataphoric. Besides, RD issues certain constraints on the resumption of the referent of the right-peripheral NP in the following, and supports the subdivision of the discourse segment into ‘main structure’ and ‘side
Anaphoric properties of German Right dislocation
structure’. AT is shown to be a local repair strategy, that does not have any impact on the discourse structure. A short summary follows in section 4.
2. Right dislocation vs. afterthought: prosodic, syntactic and semantic differences As stated above, there are prosodic, syntactic and semantic differences between RD and AT. These are briefly introduced in the following. Prosodically, RD is integrated into the preceding sentence (3a), whereas AT builds a prosodic unit of its own: it has a tone movement of its own and bears a clause-like main accent; it is also often divided from its host sentence by a major pause (3b): (3) a. (Den Tag, den vergess’ ich nicht,) {der war viel zu schön, (That day, I’ll never forget it,) d-pron was much too wonderful,
der Tag↓.} the day.
b.
(Lisa und Melanie haben sich gestritten.) (Lisa and Melanie quarrelled.) {Dann ist sie weggelaufen↓}, | { die Lisa.↓.} Then is she ran-away, the Lisa.
(|: pause; { }: prosodic unit; capitals: main accent)
RD
AT
Thus, RD prosodically belongs to its host sentence in a more straightforward way than AT. This is confirmed by the syntactic properties of these two constructions. The most important are the following: – RD requires morphological agreement between the clause-internal pro-form and the right-peripheral NP (4a), whereas for AT agreement is only optional, (4b,c). This is expected if RD is syntactically part of its host sentence and AT not, as morphological agreement is a sentence-bound phenomenon. (4) a. Deri war viel zu schön, der Tagi / *den Tagi D-Pronmask_nom was much too wonderful, the daymask_nom / the daymask_akk
2. It is especially easy to violate the gender agreement, as in (4b). The same is true for other kinds of anaphoric constructions: anaphoric resumption is generally possible without gender congruence, but only across sentence boundaries (see Consten (2004: 90)). Thus, gender agreement violation is strong evidence for afterthought not belonging to its host sentence syntactically.
Maria Averintseva-Klisch
/ *die Feieri. / the partyfem
b. Esi ist ein bisschen gleichförmig, | deine Melodiei. AT ItNEUTR is a bit even, your melodyFEM. [attested oral data]
c.
Der Zwiespalt [...] zerriss ihni fast: | [Fürst Georg III., AT (The dichotomy [...] tore himACC nearly: Prince Georg III, der Reformator von Anhalt-Dessau]i. the reformerNOM of Anhalt-Dessau) [Chrismon, 05.2004]
RD
– Additions (Altmann (1981): 'Floskeln') ich meine (‘I mean’), also (‘that is’) etc. are possible with AT (5a), whereas they are not with RD (5b). This is expected, as additions enhance the repairing function of afterthought: (5) a.
(Lisa und Melanie haben sich gestritten.) (Lisa and Melanie quarrelled. ) Dann ist sie weggelaufen, | (ich meine / also) Lisa. Then is she ran-away, (I mean / that-is) Lisa. AT
b.
(“Der Taifun!” rief Lukas dem Kapitän zu. “Da ist er!”) (“The typhoon!” Lukas called to the captain. “Here it comes!”) Ja, da war er , (*ich meine / *also) der Taifun. Yes, there was it, ( I mean / that-is) the typhoon.
RD
– AT is not restricted to the position at the right of the clause, but can vary its position: AT can come not only after its host sentence, but also immediately after the coreferent pro-form or even in some other position in the host sentence. RD, on the contrary, is only possible at the right periphery, cf. (6) vs. (7): (6) a. Ich habe ihn gestern nur mit Mühe wiedererkannt, I have himi yesterday only with effort recognized, | ich meine den Peter. I mean the Peteri b.
Ich habe ihn, | ich meine den Peter, | gestern nur I have him, I mean the Peter, yesterday only mit Mühe wiedererkannt. with effort recognized
c. Ich habe ihn gestern | ich meine den Peter | nur I have him yesterday, I mean the Peter, only
AT
Anaphoric properties of German Right dislocation
mit Mühe wiedererkannt. with effort recognized (I hardly recognized him yesterday, Peter.)
(7) *Der, der Tag, war viel zu schön. d-pron, the day, was much too wonderful.
RD
In (7), AT ich meine den Peter (‘I mean Peter’) comes at the very end of the sentence (7a), immediately after the coreferent pronoun ihn (‘him’) (7b) or between the temporal adverbial gestern (‘yesterday’) and the adverbial nur mit Mühe (‘with effort’) (7c). Thus, AT, but not RD, can vary its form (allowing additions) and its position in the host sentence. This makes it implausible that AT is really part of its host sentence syntactically. Prosodic and syntactic autonomy of AT is best accounted for with an analysis as an ‘orphan’ (in terms of Haegeman (1991), Shaer (2003)). An ‘orphan’ is a phrase that is syntactically independent of its host sentence and gets integrated into it only at the level of the discourse via some discourse relation. This analysis has consequences for the locating of AT with respect to its host sentence. AT is then not a right peripheral construction, as the right periphery is a sentence-bound concept (and AT is not a part of its host sentence). Actually, in spoken and even in written language, AT often comes explicitly after a sentence boundary, marked, for example, by the intervention of another speaker (8), or, in the case of written language, graphically (9): (8) a.
A: Ich weiß nicht, wann sie kommt morgen. A: I know not when she comes tomorrow. B: wer? A: na die Anna. B: who? A: interj the Anna.
[attested oral data]
b. A: [...] auf mein Bundesland und auf die Bundesrepublik A: [...] on my federal-state and on the federal-republic Schleswig-Holstein 3. This sentence is well-formed only as an afterthought, with the corresponding prosody: (a) Der, | (ich meine) der Tag, | war viel zu schön. (It, (I mean) the day, was too wonderful.) 4. Grosz / Ziv (1994) show that also in English the possibility of adding I mean as well as to vary its position in its host sentence distinguishes afterthought from right dislocation. 5. Uhmann (1993) analyses cases of reference resolution like (8a) and reference correction like (8b) as ‘alien-initiated repairs’. They are opposed to ‘self-initiated repairs’ like in (a): (a) Ich schreib da vorher der Frau der der Rodi. I write there before the woman the the Rodi. (I‘ll write first to that woman, Rodi.) [Uhmann (1993: 342)]. ‘Alien-initiated repairs’ are always afterthoughts, as their prosody shows. However, prosodic and syntactic tests suggest that both instances of RD and of AT are subsumed under ‘self-initiated repairs’.
Maria Averintseva-Klisch
Schleswig-Holstein. B: (Zuruf von der CDU/CSU): Bundesrepublik Schleswig-Holstein? B: (shout from the CDU/CSU): Federal-republic Schleswig-Holstein? A: Entschuldigung, ich meine die Bundesrepublik Deutschland. A: Sorry, I mean the federal-republic Germany. [plenary protocols of the German parliament, 1996–2003]
(9) (Der Koch war schon an Bord, der Matrose ebenfalls.) Er aß die Fliegen. The cook was already on board, the sailor too. He ate the flies. Der Koch, nicht der Matrose. The cook, not the sailor.
[Yann, Martel: Schiffbruch mit Tiger: 364]
So there is ample evidence that AT does not belong syntactically to its host sentence. RD on the contrary is, considered syntactically, a real part of its host sentence. Its exact syntactic analysis is impossible within the scope of this paper. As for the semantic differences between RD and AT: RD is much more restricted as far as the semantic status of the NP is concerned. The RD-NP can only refer to a definite specific individual (10a), whereas neither indefinite NPs (10 b) nor any kind of quantificational NPs (10 c) are possible: (10) a.
Da kommt er schon wieder, der Peter / der There comes he already again, the Peter / the blonde Mann / dieser blonde Mann. blonde man / that blonde man.
b. Da kommt er schon wieder, * so ein Typ aus dem Tanzkurs. There comes he already again, such a guy from the dancing-class. c. Alle blonden Frauen sind für ihn wunderschön. Peter liebt sie, All blonde women are for him beautiful. Peter loves them,
6. I owe this example to Konstanze Marx, p.c. 7. (10b) and (10c) would be well-formed as ATs, cf. (a) and (b): (a) Da kommt er schon wieder, | ich meine so ein Typ aus dem TANZkurs. (Here he comes again, I mean a guy from my dancing class.) (b) Peter liebt sie, | ich meine ALle blonden Frauen. (Peter loves them, I mean all blond women.)
Anaphoric properties of German Right dislocation
*alle blonden Frauen. all blonde women.
This is the reason why operator binding as in (11) is only marginally available for RD, even if considered syntactically nothing would prevent it: seine Frau (‘his wife’) here does not refer to a specific individual but to an ordered set (of women in a relation to men), and that is against the restrictions on the NP in the RD construction: (11) a.??? Jeder Manni liebt sie, seinei Frau. Every mani loves her, hisi wife.
On the other hand, AT allows nearly all possible types of NPs, both definite and indefinite, and as far as the latter are concerned, specific as well as non-specific. In (12a) the NP is indefinite specific, in (12b) indefinite generic, and in (12c) indefinite non-specific: (12) a.
Sie kommt heute zum Abendessen, | ich meine eine Frau aus She comes today to dinner, I mean a woman from meinem Tanzkurs. my dancing-class.
b. Sie sind Fleischfresser, | ich meine Löwen. They are carnivorous I mean lions. c. Hast Du eins, | ich meine ein Euro-Stück? Have you one, I mean a euro-piece? (context: standing near a locker in a library)
This difference between RD and AT correlates with the more general constraint on the RD-NP: it has to denote a discourse-old entity (in the sense of Prince (1992)), discourse-new entities are pragmatically dispreferred as referents for the NP (13a): (13) A: Und wie geht die Festvorbereitung? B: Ich weiß nicht was ich noch versuchen soll. Ich kann einfach keine JazzBand für den Abend auftreiben. 8. Note that the discourse in (10c) would be well-formed without right dislocation, cf. (a): (a) Alle blonden Frauen sind für ihn wunderschön. Peter liebt alle blonden Frauen. All blond women are for him beautiful Peter loves all blond women Grosz / Ziv (1994) state that in English, right dislocation cannot be used to refer to entities that were mentioned in the sentence immediately preceding the one with the right dislocation (Grosz / Ziv (1994: 190); see, however, objections in Ward / Birner (1996)). In German this is possible, cf. (b), so that does not account for the ill-formedness of (10c): (b) Verena ist für ihn die schönste. Peter liebt sie, die Verena. Verena is for him the prettiest Peter loves her, the Verena 9. Ward / Birner (1996) show that the same holds for right dislocation in English.
Maria Averintseva-Klisch
A: How are the festival preparations coming along? B: I don’t know what I should try next: I haven’t been able to get a jazz band for the evening. a. A: #Du könntest ihn fragen, diesen Chorleiter. A: #You could him ask, that choirmaster. (Bestimmt kennt er jemanden.) (Sure he knows somebody.)
RD
b. A: Du könntest diesen Typen fragen, | na, diesen Chorleiter. A: You could that guy ask, interj this choirmaster. (Bestimmt kennt er jemanden.) (Sure he knows somebody.)
AT
Obviously, this does not hold for AT, where the NP in (13b) can also be discoursenew. The only restriction that holds for AT is that it should be easily identifiable through the particular NP expression, so that AT can best fulfil its function of repairing an insufficient reference. So, (14a) is a better repair than (14b), even if one assumes that the information that apples are pip fruit and plums are not, and that red delicious is an apple breed, is hearer-old: (14) a.
Ich habe Äpfel und Pflaumen gekauft. Die schmecken aber I have apples and plums bought. They taste but leider nicht, | (ich meine) die Äpfel. unfortunately not, (I mean) the apples
b.
Ich habe Äpfel und Pflaumen gekauft. Die schmecken aber I have apples and plums bought. They taste but leider nicht, | (ich meine) #dieses Kernobst / #die Red Delicious. unfortunately not, (I mean) those pip-fruit / the ‘red delicious’.
So the AT-NP is semantically much less restricted than the right dislocation one. The latter has to denote a specific definite individual that is discourse-old. This means that RD is not cataphoric (as claimed e.g. in Zifonun et al. (1997)) but anaphoric: the discourse referent which the pronoun and the right-peripheral NP refer to has to be already present in the discourse representation. However, there appear to be plenty of exceptions to this latter constraint, where the sentence with RD opens the discourse, and thus the NP appears to denote a discourse-new entity, e.g. (15): (15) a. Der spinnt doch, der Typ / dieser Schröder. He is-crazy yet, the guy / this Schröder. (context: A and B are talking about linguistics. A sees a newspaper B has on his table with a picture of the German federal chancellor on the front page, and comments on it.
Anaphoric properties of German Right dislocation
b. Es gibt sie noch, die guten Nachrichten aus der deutschen Universität. It gives them still the good news from the german university. (You can still find some – good news coming from German universities. (first sentence of a lead)) (ZEIT 21, 13.5.04)
However, cases like (15) are only seeming exceptions, as such discourse-initial use is only possible under certain conditions. In spoken discourse like (15a), it requires mostly an explicit pointing gesture to the picture that identifies the referent.10 So the referent here is not discourse-new in the proper sense of the word, but rather situationally evoked (in the sense of Prince (1981)). Situationally evoked entities behave like discourse-old ones with respect to different linguistic diagnostics,11 and can thus be regarded as discourse-old. The case is a bit different with the written discourse. Here the RD implicitly embeds the beginning discourse into some larger, thematically contiguous setting that is familiar to the author and the recipient of the discourse. In (15b) such ‘meta-discourse’ is a series of articles about the German university system and its future in the weekly German periodical Die ZEIT, and the author of the text from which (15b) is quoted clearly relies on his readers to know this. In this sense, even in cases like (15), the discourse referent in question is still discourse-old.12 To sum up: there is ample evidence for the assumption that there are actually two distinct constructions subsumed under the name ‘German right dislocation’: right dislocation proper and afterthought. The former is prosodically and syntactically a part of its host sentence. The ‘dislocated’ NP refers to a specific definite NP denoting an individual. The corresponding referent is marked as the ‘theme’ for the following discourse segment. As for the afterthought, it is a so-called ‘orphan’, i.e. it is prosodically and syntactically independent of its host sentence, and gets integrated into its host sentence only at the level of the discourse. Its function is to repair an insufficient (pro)nominal reference that might otherwise remain unclear. In the following, I will focus on the discourse functions of both constructions. 10. In the case of using the name (this Schröder) instead of this guy, an explicit gesture is not necessary. That is why the right dislocation NP here is situationally evoked rather than straightforwardly deictical. 11. E.g. situationally evoked information can be preposed in inversion, which is the case with the discourse-old information, as Birner / Ward (1998) show, cf. (a): one can point to a table in the corner and utter (a) without the table being mentioned before: (a) On that table is a red pen; could you bring it to me? I owe this observation and example to Betty Birner (p.c.). 12. This requires a broad understanding of discourse involving different sorts of non-linguistic information sources. This seems to me, however, to be in line with recent discourse theories, e.g. SDRT (Asher / Lascarides (2003)).
Maria Averintseva-Klisch
3. Right dislocation and afterthought in discourse It has been shown above that RD marks some discourse referent as being the “theme” for the following discourse segment. I propose the term discourse topic referent for this case. This corresponds to the ‘entity-based’ approach to discourse topic, which is advocated for example in Oberlander (2004). According to Oberlander, the only sort of discourse topic relevant for establishing coherence is the entity the discourse segment is “about”.13 With respect to RD this means that it marks the referent of the right-peripheral NP as being the discourse topic referent for the following discourse segment. This can be best illustrated by (16). It is the very beginning of a book about Berlin and its citizens: (16) Ob sie so sein müssen, die Berliner, whether they so be must the Berliners das fragte sich das restliche Deutschland seit jenem Tag, da man ihre Stadt zur Hauptstadt des Reiches gemacht hatte. Und seit 1871 pflegten viele Deut-
sche bei dieser Frage den Kopf zu schütteln. Denn die Berliner waren – geben wir es ruhig gleich zu – nicht überall beliebt. Whether they have to be like this, these Berliners, was the question the rest of Germany had asked themselves since the day when their city had been made the capital of the state. And since 1871 many Germans have shaken their heads to this question. For Berliners were not especially loved everywhere, as we can easily concede. [Fischer-Fabian, S. (1959): Berlin-Evergreen: 5] The text is going to be about the citizens of Berlin, and thus they are the discourse topic referent of the text segment.14 This fact is explicitly marked by the RD in the first sentence. Naturally, we expect that the referent of the right-peripheral NP is 13. The existence of some kind of entity that is most salient at the given stage of the discourse and that is relevant for establishing coherence seems to be the common point which is shared by the papers in the recent issue of Theoretical Linguistics dedicated to discourse topic: cf. “recurring sentence topic” in Oberlander (2004), “protagonist” in Zeevat (2004), “local topics within discourse segments” in Kehler (2004), and “Discourse topic 1” in Stede (2004). 14. In a sense, the whole book is about the people of Berlin. But my understanding of the discourse topic referent is a more local one: it is the most salient discourse referent for a given segment. The segment is understood intuitively as several thematically connected utterances (this corresponds roughly to a paragraph in written texts). Thus, discourse topic referents change rather often as the text goes on. This understanding of the discourse topic referent corresponds more or less to the concept of «local topics within discourse segments» in Kehler (2004). Whether these local discourse topic referents can be organized into systems with discourse topics of some higher grade dominating them (similar to ‘discourse tree’ and ‘question under discussion’ with subquestions in Büring (2003)), cannot be discussed in this paper.
Anaphoric properties of German Right dislocation
going to be resumed in the following. This expectation holds: the referent is resumed with the anaphoric personal pronoun ihre (‘their’) and then once more with the full NP die Berliner (‘the citizens of Berlin’) in the following segment (marked with boldface in the text). Now, this observation is too weak as yet: not only the most salient referent in question, but also many other referents are resumed again in the following (e.g. the question and Germans). What is relevant here, is that we expect the discourse topic referent to be resumed as the sentence topic of its sentence. In other words, the discourse referent which the whole segment is about is expected to emerge as what a certain sentence of this segment is about.15 To test this point I assume that the structural position for sentence topic in German is one in the middle field directly above the base position for sentence adverbials, as proposed in Frey (2004): In the middle field of the German clause, directly above the base position of sentential adverbials (SADVs), there is a designated structural position for topics (in the aboutness-sense): all topical phrases in the middle field, and only these, are located in this position. [Frey (2004: 208)]
Frey (2004) shows that topics are generated in this position (although afterwards they may be moved into the prefield position). Sentence adverbials are those expressing “speaker’s estimation of the eventuality” (Frey (2004: 208)), like unfortunately or probably. Relevant for testing the topicality of a NP is whether this NP takes the position immediately before or after a sentence adverbial16. As one sees in (17), the referent of the right-peripheral NP has to be resumed as the sentence topic of the corresponding sentence: (17) Ob sie so sein müssen, die Berliner, das fragte sich das restliche Deutschland seit jenem Tag, da man ihre Stadt zur Hauptstadt des Reiches gemacht hatte. Whether they really have to be like this, these Berliners, was the question the rest of Germany had asked themselves since that day when their city had
been made the capital of the state.
15. I understand sentence topic, in accordance with Frey (2004), in the aboutness sense (cf. Reinhart (1981)), as the constituent the sentence is about. Although the exact relation between sentence topic and discourse topic is not quite clear, there is undoubtedly some relation leading to discourse topic referents tending to emerge as sentence topics (see also Oberlander (2004)). 16. It is important that it concerns only sentence adverbials in their neutral use, and not in the focus-inducing one, like in (a), where anscheinend (‘apparently’) is used to focus the NP dem Otto: (a) Dem Otto anscheinend hat sich Maria anvertraut. Det Otto apparently has Refl Maria confided (in) [cf. Frey (2004: 229)].
Maria Averintseva-Klisch
1. Im ganzen Land waren Berliner seltsamerweise wenig beliebt. In-the whole land were Berliners strangely little liked. 2. #Im ganzen Land waren seltsamerweise Berliner wenig beliebt. In-the whole land were strangely Berliners little liked. Strangely, in the whole state the Berlin people were not very popular.
The data from (17) confirm: as expected, the discourse topic referent of a given segment is referred to with the sentence topic expression in sentences that constitute this segment. However, this is not always the case, as the short discourse in (18) shows. This is a short report in a local paper about a certain tourist bus service called “Märkische Schweiz”. The discourse topic referent (the tourist bus) is marked by the RD in the first sentence17. In the following it is resumed through anaphoric personal pronoun ihn / er / ihm (different case forms of 3d person singular masculine), possessive pronoun seine (‘its’) and an indirect anaphor die Route (‘the route’) (resumptions are marked with boldface): (18) 1.
Viele Freunde hat er im vergangenen Sommer gefunden - der Many friends has it in last summer found - the Ausflugsbus “Märkische Schweiz”. tourist bus “Märkische Schweiz It found many friends last summer – the tourist bus “Märkische Schweiz”
2.
Deshalb haben ihn DB Regio und die Strausberger because-of-that have it DB Regio and the Strausberg Verkehrsgesellschaft auch 2004 wieder ins Ausflugsprogramm genommen. traffic-company also 2004 again in-the trip-programme taken That is why DB Regio and the traffic company of Strausberg put it once again in2004 in the trip list.
3.
Vom 15. Mai bis 12. September zieht er sonnabends, sonntags und From 15th may till 12th september pulls it saturdays, sundays, and feiertags im Zwei-Stunden-Takt seine Runden durch das on-holidays in-the two-hours-rhythm its rounds through the Minigebirge östlich Berlins. mini-mountains to-the-east-of Berlin From 15th May till 12th September it makes its rounds through the hills to the east of Berlin every two hours on Saturdays, Sundays and holidays.)
4. Wunderschöne Landschaften, Schlösser und Herrenhäuser kann man Wonderful landscapes, castles and manor-houses can one 17. However, the referent is not discourse-new here; it is introduced first by the title Bus “Märkische Schweiz” startet in die zweite Saison (Bus “MS” starts to its second season).
Anaphoric properties of German Right dislocation
mit ihm entdecken, Kunst, Kultur, Geschichte in vielen Facetten. with it discover arts, culture, history in many facets Wonderful landscapes, castles and manor houses can be discovered with it, arts, culture and history in many facets.
5.
Vom Bahnhof Müncheberg führt die Route über Buckow from-the railstation M leads the route over B durch die Teichlandschaft von Karlsdorf, ins Schinkeldorf through the pool-landscape of K, in-the S.-village Neuhardenberg, dann zum Schloss Wulkow und zurück über N., then to-the castle W and back over Bahnhof Trebnitz nach Müncheberg. railstation T to M From the Müncheberg station the route leads via Buckow through the lakes of Karlsdorf, to the Schinkel village Neuhardenberg, then to
Wulkow castle and back via Trebnitz station to Müncheberg. [Punkt 3, 13.5.04]
Now, the question is whether or not the expressions resuming the discourse topic referent are sentence topics. In (19) it is tested clause-after-clause with the help of the middle field topic position: (19) 2. a. Deshalb haben ihn logischer Weise (logically) DB Regio und die Strausberger Verkehrsgesellschaft auch 2004 wieder ins Ausflugsprogramm
genommen. b. #Deshalb haben DB Regio [...] logischer Weise ihn wieder ins Ausflugs programm genommen.18
18. To be sure that the variant in (19-2b) is really ruled out because of non-topicality being unsuitable in this discourse, I changed the position of the subject DB Regio [...]. Pronoun in a non-topical position is only possible when it is focused (contrast focus), like in (a), as Manfred Consten pointed out to me. This, however, requires that the topic position is filled with the subject NP, cf. (b): Der Bus war viel beliebter als die Straßenbahn. (The busMASK was much more popular than the streetcarFEM) (a) Deshalb hat DB Regio logischer Weise ihn, und nicht sie, wieder because-of-that has [DB Regio]TOP logically [itMASK]FOK, and not [itFEM]FOK, again ins Ausflugsprogramm genommen. in-the trip-list took. (b) *Deshalb hat [Ø]TOP logischer Weise ihn, und nicht sie, DB Regio wieder because-of-that has [Ø]TOP logically [itMASK]FOK, and not [itFEM]FOK, DB Regio again ins Ausflugsprogramm genommen. in-the trip-list took.
Maria Averintseva-Klisch
In 2., the pronoun ihn (‘itAKK’) is the sentence topic; the variant with the pronoun following the sentence adverbial is not suitable in this discourse. The case is the same in 3. and 5., while in 4. the pronoun ihm (‘itDAT’) might on the contrary be topical as well as non-topical: (19) 3. a. Vom 15.Mai bis 12. September zieht er zum Glück (luckily) sonnabends, sonntags und feiertags im Zweistunden-Takt seine Runden
durch das Minigebirge östlich Berlins. b. #Vom 15.Mai bis 12. September zieht zum Glück er [...]seine Runden durch das Minigebirge östlich Berlins.
4. a. Wunderschöne Landschaften, Schlösser und Herrenhäuser kann man mit ihm unerwarteter Weise (unexpectedly) entdecken, Kunst, Kul
tur, Geschichte in vielen Facetten. b. Wunderschöne Landschaften, Schlösser und Herrenhäuser kann man unerwarteter Weise mit ihm entdecken [...]
5. a. Vom Bahnhof Müncheberg führt die Route in einer gut durchdachten Weise (in a well-considered way) über Buckow durch die Teichlandschaft von Karlsdorf, ins Schinkeldorf Neuhardenberg, dann zum Schloss Wulkow und zurück über Bahnhof Trebnitz nach Müncheberg. b. #Vom Bahnhof Müncheberg führt in einer gut durchdachten Weise die Route über Buckow durch die Teichlandschaft von Karlsdorf [...]
It is not arbitrary whether expressions referring to the discourse topic referent are sentence topics or not. To explain this I adopt the analysis in von Stutterheim / Klein (2002). Von Stutterheim and Klein distinguish between ‘main structure’ and ‘side structure(s)’ in a discourse. The main structure is the main story line, which includes partial answers to the ‘Quaestio’, a (mostly) implicit question which the discourse as a whole is answering. The side structure consists of sentences supplying information that is not immediately relevant as a partial answer to the Quaestio.19 An example illustrating the point is (20): here, the clauses 1–3 and 6 are answering the Quaestio What was your old flat like?, whereas clauses 4 and 5 build a side structure expanding on the subject of the yard:
19. The side structure comprises such deviations from the main story line that still stay in the big thematic continuity, as opposed to real breaches with a total change of subject, which rather often occur in spoken discourse.
Anaphoric properties of German Right dislocation (20) Quaestio: What was your old flat like? 1. It had four rooms and a rather big kitchen. main structure 2. The rooms were light and airy. 3. Two faced the street and two the yard. 4. That was where the children always played.
side structure
5. That's what we miss now most. 6. Besides, all rooms were parqueted. [adapted from Stutterheim (1997: 27)]
To return to the text (18): the discourse topic referent here, or, respectively, the Quaestio, is “the tourist bus “MS” and its route”20. The main story line then is constituted through the clauses 1, 2, 3 and 5; the side structure consists of the clause 4, which does not supply information immediately relevant for the Quaestio. Clause 4 is exactly the clause where some other discourse referent, i.e. wunderschöne Landschaften etc. (‘wonderful landscapes, castles and manor houses’)21, and not the discourse topic referent of the whole segment, might be expressed as the sentence topic. This means, that the discourse topic referent is resumed as the sentence topic on the level of the main structure, whereas on the level of the side structure some other discourse referents are referred to as sentence topics22. Thus, we can conclude that RD itself belongs to the main structure of its discourse segment, and by marking a discourse referent as the discourse topic referent for the following segment, it ensures that at the level of the main structure this referent 20. For von Stutterheim (1997), Quaestio does not have to appear as a «speech act of questioning»; she defines it as the «main point» of a text, that is «theoretically reconstructable as the question that the text is bound to solve»(von Stutterheim (1997: 18) [translation mine: MA]). So, in this sense, Quaestio and the discourse topic referent do not differ much. 21. The question is whether there is a two-part sentence topic with the second part being the NP arts, culture and history in many facets at the right periphery, or whether this NP is better analyzed as an elliptical addition to the sentence, which does not really syntactically belong to the sentence. Theoretically, right-peripheral sentence topics seem to be possible in German (cf. Frascarelli / Hinterhölzl (to appear)). However, I tend to the ellipsis analysis for the NP arts, culture and history in many facets. As this issue is not immediately relevant for this paper, I do not pursue it further here. 22. For the case in (19-4a) I assume that the clause is meant by the speaker to be relevant for the Quaestio ( i.e. discourse topic referent), and thus belongs to the main structure. It might be an option to assume that side structure implies a change of the discourse topic referent, e.g. in (20) the discourse topic referent of the main structure would be the old flat, and the discourse topic referent of the side structure would be the yard. However, more evidence is required to be able either to claim or to disprove this. The introduction of a new discourse topic referent might be plausible in the case of relatively extended side structure (comprising at least several clauses).
Maria Averintseva-Klisch
will be resumed as the sentence topic of the corresponding sentence(s), and thus coherence will be maintained. This is the role RD plays in the global discourse structuring. As for AT, it does not have any consequences for the structuring of the whole discourse segment: in (21), the discourse segment is about a certain play with an actress playing the role of a nun, and the afterthought occurs in the utterance claiming that the actress was much more beautiful than the actual nun. The nun is in no way the discourse topic referent of the segment, and the afterthought does not change anything in this respect: the text continues to be about the play: (21) [...] und das ist es auch [...] was das Stück will, was man um so deutlicher sieht, als die Bethmann wirklich eine sehr hübsche Frau ist oder doch zum wenigsten viel hübscher, als sie wirklich war ich meine die Nonne was aber nichts schadet [...] as she really was I mean the nun what but nothing damages [...] and this is also [...] what the point of the play is, and one sees it even clearer, because the Bethmann is really beautiful, or at least much more beautiful than she really was, I mean the nun, but it was not so bad [...] [Newspaper Corpus of Bonn BZK: 2014916]
Thus, AT is a local strategy, which is only concerned with the previous sentence, or to be more exact, with that expression in it whose reference AT is clearing. To summarize: RD is an anaphoric construction that marks a discourse-old referent as being the discourse topic referent for the following segment (in the sense of the definite specific entity the discourse segment is about). In this sense, RD is a forward-looking strategy used to structure the following discourse segment. It is responsible for dividing the discourse segment into the main structure and side structure, and for the resumption of the discourse topic referent as sentence topic on the level of the main structure. AT is a purely local backward-looking repair strategy that resolves an unclear (pro)nominal reference in its host sentence.
4. Summary In this paper I have shown that there is ample prosodic, syntactic, semantic and discourse-functional evidence for the distinction between RD and AT in German. In short, the differences between RD and AT are the following: RD is a prosodic and syntactic part of its host sentence (presumably a right adjunct). In terms of the discourse function RD marks the referent of the rightdislocated NP as the discourse topic referent for the following discourse segment.
Anaphoric properties of German Right dislocation
This interacts with the preference for definite specific NPs with individual reference in RD constructions. The referent of the NP has to be discourse-old, i.e. already established in the discourse representation. In this sense RD is anaphoric, and not cataphoric. RD imposes certain anaphoric constraints on the following discourse segment: the discourse referent marked as the discourse topic referent has to be resumed in the following as a sentence topic of the corresponding sentence. This constraint applies to the level of the main structure of the segment. Sentences where the discourse topic referent is not resumed or is resumed as a non-topical expression are argued to construct the side structure of the segment. As for AT, it is shown to be prosodically and syntactically independent of its host sentence, and thus not a right-peripheral construction in the proper sense of the word, but an ‘orphan’, that can occur (also) after the end of the sentence. Discourse-functionally it is a local reference repair strategy.
References Altmann, H. 1981. Formen der Herausstellung im Deutschen: Rechtsversetzung, Linksversetzung, Freies Thema und verwandte Konstruktionen [Linguistische Arbeiten 106]. Tübingen: Niemeyer. Asher, N. and Lascarides, A. 2003. Logics of Conversation. Cambridge: CUP. Auer, P. 1991. Vom Ende deutscher Sätze. Zeitschrift für Germanistische Linguistik 19: 139–157. Birner, B.J. and Ward, G., 1998. Information Status and Non-canonical Word Order in English. Amsterdam: John Benjamins. Büring, D. 2003. On D-trees, beans, and B-accents. Linguistics & Philosophy 26(5): 511–545. Consten, M. 2004. Anaphorisch oder deiktisch? Zu einem integrativen Modell domänengebundener Referenz [Linguistische Arbeiten 484]. Tübingen: Niemeyer. Frascarelli, M. and Hinterhölzl, R. To appear. Types of topics in German and Italian. In Information Structure and the Architecture of Grammar: A typological perspective [Linguistik Aktuell/Linguistics Today 100], Winkler, S. and Schwabe, K. (eds). Amsterdam: John Benjamins. Frey, W. 2004. Notes on the syntax and the pragmatics of German left dislocation. In The Syntax and Semantics of the Left Peripher [Interface Explorations 9], Lohnstein, H. and Trissler, S. (eds), 203–233. Berlin: Mouton de Gruyter. Grosz, B. and Ziv, Y. 1994. Right dislocation and attentional state. In Proceedings of the 9th Annual Conference & Workshop on Discourse of the Israel Association for Theoretical Linguistics, R. Buchalla and A. Mittwoch (eds), 184–199. Jerusalem: Akademon Press. Haegeman, L. 1991. Parenthetical adverbials: The radical orphan approach. In Aspects of Modern English Linguistics: Papers presented to Masatomo Ukaji on his 60th birthday, Chiba, S. et al. (eds), 232–254. Tokyo: Kaitakusha. Kehler, A. 2004. Discourse topics, sentence topics, and coherence. Theoretical Linguistics 30: 227–240. Oberlander, J. 2004. On the reduction of discourse topic. Theoretical Linguistics 30: 213–225. Prince, E.F. 1981. Towards a taxonomy of given/new information. In Radical Pragmatics. P. Cole (ed.), 223–254. New York NY: Academic Press.
Maria Averintseva-Klisch Prince E.F. 1992. The ZPG letter: Subjects, definiteness, and information-status. In Discourse Description. Diverse linguistic analyses of a fund-raising text, W.C. Mann and S.A. Thompson (eds), 295–325. Amsterdam: John Benjamins. Reinhart, T. 1981. Pragmatics and linguistics: An analysis of sentence topics. Philosophica 27: 53–94. Selting, M. 1994. Konstruktionen am Satzrand als interaktive Ressource in natürlichen Gesprächen. In Was determiniert Wortstellungsvariation?, B. Haftka (ed.), 299–318. Opladen: Westdeutscher Verlag. Shaer, B. 2003. An ‘orphan’ analysis of long and short adjunct movement in English. In: WCCFL 22 Proceedings, G. Garding and M. Tsujimira (eds), 450–463. Sommerville MA: Cascadilla. Stede, M. 2004. Does discourse processing need discourse topics? Theoretical Linguistics 30: 241–253. Uhmann, S. 1993. Das Mittelfeld im Gespräch. In Wortstellung und Informationsstruktur, M. Reis (ed.), 313–354. Tübingen: Niemeyer. Uhmann, S. 1997. Grammatische Regeln und konversationelle Strategien. Fallstudien aus Syntax und Phonologie [LA 375]. Tübingen: Niemeyer. Von Stutterheim, C. 1997. Einige Prinzipien des Textaufbaus. Empirische Untersuchungen zur Produktion mündlicher Texte [Reihe Germanistische Linguistik 184]. Tübingen: Niemeyer. von Stutterheim, C. and Klein, W. 2002. Quaestio and L-perspectivation. In Perspective and Perspectivation in Discourse, C.F. Graumann and W. Kallmeyer (eds), 59–88. Amsterdam: John Benjamins. Ward, G. and Birner, B.J. 1996. On the discourse function of rightward movement in English. In Conceptual Structure, Discourse and Language, A. Goldberg (ed.), 463–479. Stanford CA: CSLI. Zeevat, H. 2004. Asher on discourse topic. Theoretical Linguistics 30: 203–211. Zifonun, G., Hoffmann, L. and Stecker, B. (eds). 1997. Grammatik der deutschen Sprache, [IDS 7.1, Bd.1]. Berlin: Walter de Gruyter.
Antecedents of diverse types An investigation of the syntactic and semantic relationships in a wh-relative construction Anke Holler The present paper addresses the issue of the grammatical relations in German wh-relative clause constructions. It discusses data demonstrating that the ordinary assumption that German wh-relative clauses generally relate to propositional antecedents must be revised. In particular, it shows that the syntactic antecedent of a wh-relative clause has to be separated from its semantic antecedent(s). Wh-relative clauses relate syntactically to a functionally complete sentential projection, while they relate semantically to entities of various kinds. The semantic objects that may function as antecedents of a wh-relative clause are determined by the semantic type of the anaphoric wh-expression introducing a wh-relative clause. The paper provides an account of the described empirical facts in the framework of Discourse Representation Theory.
1. Introduction In the present paper I broach the issue of the syntactic and semantic type of a possible antecedent of a wh-relative clause. Referring to complex sentences like (1) as wh-relative constructions, (1) a.
Archäologen erforschen das Umland der Polis von Milet, was archeologists explore the hinterland of the polity of Milet which sehr interessant ist. very interesting is ‘Archeologists explore the hinterland of the polity of Milet, which is very interesting.’
Anke Holler
b. In der Spätantike versandeten die Häfen von Milet, weswegen in the Late Antiquity slited the harbors of Milet why die Stadt an Bedeutung verlor. the city PREP importance lost ‘The harbors of Milet slited in Late Antiquity, that’s why the city became less important.’
I show that the ordinary assumption that a wh-relative clause always relates to a sentential or propositional antecedent is incorrect. In particular, I argue that the syntactic antecedent of a wh-relative clause has to be distinguished from its semantic antecedent(s). Whereas a wh-relative clause syntactically relates to a functionally complete sentential projection, it can have semantic antecedents of various kinds. The investigation of the syntactic and semantic relationships of a wh-relative clause is supposed to be instructive because the wh-relative construction is residing between syntax and text. Its analysis therefore helps to disclose restrictions that are effective at various grammatical interfaces, such as the syntax-semantics interface or the semantics-discourse interface. The article is organized as follows. In section 2, I introduce the phenomenon. In section 3, I outline previous analyses to the wh-relative construction. Then, in section 4, I discuss several aspects of the syntactic behavior of a wh-relative clause, and I investigate the semantic relationships in a wh-relative construction. Next, in section 5, I present an account of the described facts in the framework of Discourse Representation Theory (DRT, cf. Kamp and Reyle (1993)). The analysis is based on the assumption that the wh-relative expression introducing a wh-relative clause is anaphoric. Finally, in section 6, I summarize the article.
2. Phenomenon The class of wh-relative clauses can be identified by a (possibly complex) wh-expression introducing the respective relative clause. This wh-expression either saturates an argument of the wh-relative clause’s predicate, cf. (1a) and (2a), or acts as an adverbial modifier, cf. (1b) and (2b). In the latter case, the adverb may also contribute a causal, temporal or modal relation. (2) a. Deutschland verlor das Fußballspiel, was die Fans bedauerten. Germany lost the soccer match which the fans regretted ‘Germany lost the soccer match which the fans regretted.’ b. Er hatte uns verlassen, wonach wieder Ruhe im Haus war. he had us left whereupon again calm in the house was ‘He had left us, whereupon the calm was restored in the house.’
Antecedents of diverse types
If the wh-expression acts as a complement, it has to be theta-marked by the whrelative clause’s predicate. This is the reason of why (3) is ungrammatical. (3) *Die Stadt steht unter Wasser, was geregnet hat. the city is under water which rained has
This fact comes along with the observation that the class of verbs occuring in a wh-relative clause introduced by was (‘which’) is restricted to verbs that admit a phoric correlative es (‘it’ ) which bears a theta-role. This generalization contrasts to the assumption made by Helbig and Buscha (1986), who claim that the class of verbs occuring in a was-relative clause comprises only those verbs that require a so-called placeholder es (‘it’). If Helbig and Buscha (1986) were right, the grammatical example in (4a) would have to be unacceptable since beschließen does not tolerate a placeholder es (‘it’) as is shown in (4b). (4) a. Otto hat vor, am Wettbewerb teilzunehmen, was Max auch beschließt. Otto intends in the competition to take part which Max also decides
‘Otto intends to take part in the competition, which Max decides, too.’
b. Max beschließt (*es) auch, am Wettbewerb teilzunehmen. Max decides it also in the competition to take part ‘Max also decides (*it) to take part in the competition.’
In addition, the predicate of a was-relative clause must be subcatogarized for a DP complement which may be realized by the wh-expression was (‘which’). This is the reason of why verbs like ermahnen (‘to advice’) cannot occur in a wh-relative construction: (5) a. Anna ermahnt ihren Bruder, seine Hausaufgaben zu erledigen. Anna advices her brother his homeworks to do ‘Anna advices her brother to do his homeworks.’ b. * Anna ermahnt ihren Bruder das Erledigen seiner Hausaufgaben Anna advices her brother the doing of his homeworks c. * Ihr Bruder soll seine Hausaufgaben erledigen, was Anna ihn ermahnt. Her brother shall his homeworks do which Anna him advices 1. There is a specific variant of the wh-relative construction introduced by was (‘which’) that shall not be discussed here. In this variant, was (‘which’) substitutes a predicative argument as is illustrated in (i). Example (i) has several readings; in a predicative reading the team of France as well as the team of Germany want to win the world championship. A comprehensive analysis of this construction variant can be found in Holler (2005). (i) Die Deutschen wollen die Fußballweltmeisterschaft gewinnen, was die Franzosen auch The Germans want to the world championship in soccer win which the French also wollen. want `The Germans want to win the world championship in soccer. This want the French, too.‘
Anke Holler
Although wh-relative clauses seem to be dependent clauses as a final position of the finite verb indicates, they clearly show characteristics of root clauses as well. As can be seen by (6) and (7), they can be transformed into an independent clause, and they admit perfomative indicators, modal particles and epistemic expressions. (6) a. Der Vertrag ist abgeschlossen, was alle begrüßen. the contract is closed which everybody appreciate ‘The contract is closed, which everybody appreciates.’ b. Der Vertrag ist abgeschlossen. Das freut alle. the contract is closed this appreciates everybody ‘The contract is closed. This everybody appreciates.’ (7) a. Der Vertrag ist unterzeichnet, worüber ich Sie hiermit informiere. the contract is signed which I you herewith inform of ‘The contract is signed, which I herewith inform you of.’ b. Der Vertrag ist abgeschlossen, was glücklicherweise jeden freut. the contract is closed which fortunately anybody enjoys ‘The contract is closed. Fortunately, this enjoys anybody.’
Example (8) demonstrates that a sentence type operator introduced in the relative clause’s matrix clause cannot scope over a wh-relative clause, which is further evidence for its rootness. (8) *Spiel nicht am Radio, was ich dir schon hundertmal gesagt habe! toy not with the radio which I you already hundreds of times told have
Closely related to the rootness of wh-relative clauses is the observation that they behave like typical non-restrictive clauses as the following facts suggest: First, a wh-relative clause cannot be pronominalized together with its host clause. As can be seen in example (9), the pronominal discourse anaphor das (‘this’), also called ’complex anaphor’ by Consten, Knees and Schwarz-Friesel (this volume), either refers to the host clause or to the wh-relative clause. (9)
Die SPD erhöht die Steuern, was sie im Wahlkampf the SPD raises the taxes which she during the election campaign übrigens abgestritten hatte. by the way denied has
2. There might be a reading where das (‘this’) similarly refers to both clauses. This does not degrade the argument as long as das (‘this’) may also relate to the relative clause only.
Antecedents of diverse types
Das ist sehr ärgerlich. this is very annoying ‘The SPD wants to raise taxes which she has denied by the way during the election campaign. This is very annoying.’
Second, a wh-relative clause is always asserted and never presupposed. This directly follows from the observation that wh-relative clauses are sensitive for speech act modification, which one would not expect if they were presupposed, cf. Reis (1977). For instance, the modal adverb bedauerlicherweise (‘unfortunately’) in (10) may express the speaker’s regret. (10) Max hatte Geburtstag, was seine Frau bedauerlicherweise vergaß. Max had birthday which his wife unfortunately forgot ‘Max had birthday, which his wife forgot, unfortunately.’
Although wh-relatives are mentioned in almost every grammar book of German, to date their grammatical properties have not been studied in its entirety, the only exception being Brandt (1990), who, however, focusses on pragmatic aspects of the wh-relative construction. The next section briefly reviews pertinent previous approaches to parts of the grammar of wh-relative clauses.
3. Previous approaches There are two influential research papers that have shaped the work on wh-relative clauses: an early article by Helbig (1980) and a monography by Brandt (1990). Both authors primarily discuss the class of continuative clauses and consider the wh-relatives as an instance of this clausal class. While Helbig (1980) focusses on the syntactic and semantic description of continuative clauses, Brandt (1990) concentrates more on their pragmatic behavior. Helbig (1980) uses descriptive criteria to distinguish wh-relative clauses from ordinary non-restrictive ones. He claims that the main characteristic of a wh-relative clause is its peculiar (syntactic and semantic) relationship to the prior sentence. Thus, Helbig clearly acts on the assumption that wh-relative clauses are always sentence-related. Additionally, Helbig lists several secondary properties that he considers typical for wh-relative clauses such as the realization of an introducing wh-element, the clause final position of wh-relatives in a complex sentential construction, and the possibility to transform wh-relatives into an independent main clause. Based on Helbig’s insights, Brandt (1990) characterizes wh-relative clauses pragmatically. She argues that wh-relative clauses belong to the class of continua-
Anke Holler
tive relative clauses, which bear—according to Brandt (1990)—the same communicative weight as their host clause and, hence, express main information. For the identification of continuative relative clauses Brandt uses the so-called dennoch (‘nonetheless’) test going back to Pasch (1983). This test manipulates the right context of a relative clause by introducing a sentence that starts with dennoch (‘nonetheless’). Brandt assumes that a non-restrictive relative clause conveys main information and is thus continuative if dennoch (‘nonetheless’) may refer to this information. This applies to wh-relative clauses as is demonstrated by example (11) cited from Brandt (1990). (11)
Sie sagte die Wahrheit, was alle einsahen. Dennoch wurde sie she said the truth which everybody acknowledged nonetheless was she bestraft. punished ‘She said the truth, which everybody acknowledged. Nonetheless she was punished.’
Helbig’s and Brandt’s work had also an impact on descriptive grammar writing. In prevalent grammars such as the DUDEN grammar (Dudenredaktion, 2005) or the IDS grammar (Zifonun et al., 1997), a grammar provided by the Institut for German Language (IDS), wh-relative clauses are mainly characterized (i) by their relation to a state of affairs expressed by the preceding sentence they relate to and (ii) by their continuative function. In the generative grammar tradition, wh-relatives are mostly discussed in relation to clause linkage phenomena and/or grades of clausal subordination, see for instance Fabricius-Hansen (1981), Fabricius-Hansen (1992), Reis (1997) and Zimmermann (1993). Mostly, it has been argued that wh-relative clauses are not integrated into the sentence structure of the complex sentential construction they belong to. Summarized, previous approaches to the grammar of wh-relative clauses are either based on the assumption that a wh-relative clause is related to a sentence or rather a proposition, or they do not discuss this issue at all. In the next section, I will investigate the grammatical relationships between a wh-relative and its potential syntactic and semantic antecedents in more detail.
3.
For a critical assessment of Brandt’s approach see Holler (2005).
4. Holler (2005) gives a more detailed overview of the literature on wh-relative clauses.
Antecedents of diverse types
4. Syntactic and semantic antecedents of a wh-relative clause As presented in the previous section, it is generally claimed that a wh-relative clause expresses a state of affairs that is related to, but semantically independent of the state of affairs expressed in the matrix clause. A consequence of this claim is the prevalent assumption that the antecedent of a wh-relative clause is sentential or propositional. Contrary to this, I argue (i) that the syntactic antecedent of a whrelative clause has to be separated from its semantic antecedent, (ii) that a wh-relative clause syntactically relates to a functionally complete sentential projection, and (iii) that semantic entities of various kinds can function as an antecedent of a wh-relative clause. 4.1
For a separation of the syntactic and semantic relationships
Constructions like (12) have been taken as prototypical instances of a wh-relative construction. In both cases, the meaning of was (‘which’) arises from the meaning of a clausal entity which corresponds to an expression equivalent to a sentence in an appropriate logic formalism. This clausal entity may be identical to the matrix clause as in (12a) or to a clausal subset as in (12b). (12) a. b.
Max küsste eine hübsche Frau, was seine Ehefrau übrigens Max kissed a beautiful woman, which his wife by the way gesehen hat. seen has ‘Max kissed a beautiful woman, which his wife has seen, by the way.’ Anna erzählte uns, dass Max eine hübsche Frau küsste, was seine Anna told us that Max a beautiful woman kissed which his Ehefrau übrigens gesehen hat. wife by the way seen has ‘Anna told us that Max kissed a beautiful woman, which his wife has seen, by the way.’
In both cases, the apparent meaning of the wh-relative clause is that Max’s wife saw that Max kissed a beautiful woman, which means that the wh-relative clause relates to an entity that denotes a proposition. So far the standard assumptions are met. Example (13), however, demonstrates that a wh-relative clause may also relate to a question meaning. This is, of course, no big issue as long as one sticks to the assumption that the antecedent of was (‘which’) may express any meaning equivalent to a sentence. 5.
There might be a further reading of example (12b) where was (‘which’) relates the main clause.
Anke Holler
(13)
Anna fragte sich, wen Max geküsst hat, was seine Ehefrau Anna wondered REFL whom Max kissed has which his wife auch unbedingt wissen wollte. also necessarily know wanted to ‘Anna wondered whom Max has kissed, which his wife also was desperate to know.’
Example (14), clearly, challenges the standard approach as the wh-relative clause is related to an infinitive of the so-called second status in the sense of Bech (1957). (14)
Max versucht einen Elefanten zu dressieren, was Maria auch schon Max tries an elephant to train which Maria also already versucht hat. tried has ‘Max tries to train an elephant, which Maria also has already tried.’
Although it is still under debate whether German infinitival clauses behave like sentences, it is obvious that the proposition that Max trains an elephant cannot be the antecedent of was (‘which’). It is rather the event-type ‘to train an elephant’ was (‘which’) refers to. Similar examples made Brandt (1990) to argue that wh-relative clauses refer to syntactic units of different size. As evidence she cites the following wh-relative constructions, cf. (Brandt, 1990, p. 20). The syntactic units which she analyzes as antecedents of was (‘which’) are each identified by brackets: (15) a. Er wollte [nach Italien fahren]VP, was er schließlich auch tat. he wanted to to Italy go which he finally PART did ‘He wanted to go to Italy, which he finally did.’ b. Sie ist [immer freundlich]AP, was er wirklich nicht ist. she is always kind which he in fact not is ‘She is always kind, which he isn’t, in fact’. c. Er kann schon [schwimmen]V, was sie nicht kann. he can already swim which she not can ‘He can already swim, which she cannot, yet.’ d. Er gab mir [zehn Mark]NP, was damals viel Geld war. he gave me ten marks which at that time much money was ‘He gave me ten marks, which was much money at that time.’
Albeit Brandt is right in the observation that wh-relative clauses may relate to subsentential units, she wrongly attributes this to syntax. As can be seen by (16), was
6. In the German grammar tradition, the term ‘Satzwertigkeit’ is used to describe this aspect.
Antecedents of diverse types
(‘which’) resumes entities that are abstracted from the matrix clause even if they are not adjacent at the syntactic surface or do not function as a syntactic constituent. (16) a. b.
Sampras gewinnt meistens das Match, was ihn aber nicht Sampras wins mostly the match which he however not überheblich macht. arrogant makes ‘Sampras mostly wins the match, which, however, does not make him arrogant.’ Max hat dieses Mal seinem Chef nicht widersprochen, was Max has this time his boss not disagreed which ihm sonst sicher Ärger eingebracht hätte. him else certainly trouble earn had ‘Max did not disagree with his boss. Otherwise he certainly would have been in trouble.’ c. Karl wollte eine Maus halten, was seine Mutter ihm aber Karl wanted to a mouse keep which his mother him however nicht erlaubte. not allowed ‘Karl wanted to keep a mouse, which his mother, however, did not allow him.’ d. Becker spielt wahrscheinlich gegen Lendl, was sich Edberg auch Becker plays probably against Lendl which REFL Edberg also gewünscht hätte. wished had ‘Becker will probably play against Lendl, which Edberg has wished to do, too.’ e.
Wenn die Parteien Kandidaten ohne politische Biographie berufen, if the parties candidates without political biography appoint was wünschenswert wäre, which desirable were dann sollten sie auch ihre Meinung frei äußern können. then should they also their opinion freely give can ‘If the parties appoint candidates without any political biography—which were desirable—then they should be allowed to freely give their opinion.’ f. Peter befahl, dass alle anwesend zu sein haben, was man von ihm Peter ordered that all present to be have which one from him gar nicht kannte. PART not knew ‘Peter ordered that all have to be present, which one did not at all expect from him.’
Anke Holler
In (16a) and (16b), the adverb mostly and the negation particle not respectively are not a part of the antecedent of was (‘which’). Only parts of the verbal complex function as antecedent in (16c). Also, the meaning of the construction (16d) is not that Edberg is likely to wish to play against Lendl, but that he really wants to have such a match. Finally, the constructions in (16e) and (16f) show that a wh-relative clause may relate to entities that are not at all syntactically realized in the matrix clause. In (16e) the wh-relative clause relates to a condition, in (16f) to a speech act. There are even more arguments against Brandt’s approach. As the examples in (17) show, the predicate of a wh-relative clause may be subcategorized for an argument whose syntactic type differs from the syntactic form of the potential antecedent. For instance, behaupten (‘to claim’) in (17a) selects a dass (‘that’)-clause, whereas versuchen (‘to try’) provides an infinitival complement. Despite this mismatch in the syntactic type of the respective arguments, the wh-relative construction is acceptable without restrictions of any kind. (17) a.
Die Medien behaupten, dass die neue Regierung die Steuern erhöhen The media claim that the new government the taxes raise wird, was die alte Regierung übrigens auch mehrmals versucht hatte. will which the old government by the way also several times tried had ‘The media claim that the new government will raise taxes, which the old government had also tried several times, by the way.’
b. *Die alte Regierung hatte versucht, dass die neue Regierung die the old government had tried that the new government the Steuern erhöhen wird. taxes raise will c.
Der Minister beschloss, sein Amt niederzulegen, was das Parlament The minister decided his office to resign which the parlament vor Wochen schon verlangt hatte. ago weeks already requested had. ‘The minister decided to resign from office, which the parlament had requested weeks ago.’
d. *Das Parlament hatte schon vor Wochen verlangt sein Amt niederzulegen the parlament had already ago weeks requested his office to resign e. Kohl soll die Spender nennen, was er aber rigoros ablehnt. Kohl shall the donators name which he however adamantly refused ‘Kohl shall name the donators, which he adamantly refused, however.’ f. * Er lehnt ab die Spender zu nennen. He refused PART the donators to name
Antecedents of diverse types
g. Schröder will nicht zurücktreten, was alle überrascht. Schröder wants to not resign which everyone surprises ‘Schröder does not want to resign, which everyone surprises.’ h. * Alle überrascht Schröder will nicht zurücktreten Everyone surprises Schröder wants to not resign
Furthermore, it can be shown by examples like (1a), repeated below as (18), that the wh-relative clause may be semantically ambiguous. (18) has a propositional reading saying that it is interesting that the archaeologists explore the hinterland of the Polis of Milet. This reading seems to be the preferred one. The wh-relative clause acts on the truth value of its antecedent. However, there is a second reading that relates was (‘which’) to the event-type of the matrix clause, which is the exploration of the hinterland of the Polis of Milet. Last but not least, the wh-relative clause may refer to the singular event of exploring the hinterland of the Polis of Milet once at the time, which is expressed in the matrix clause. (18)
Archäologen erforschen das Umland der Polis von Milet, was archeologists explore the hinterland of the polity of Milet which sehr interessant ist. very interesting is ‘Archeologists explore the hinterland of the polity of Milet, which is very interesting.’
If the reference relationship between the wh-relative clause and its antecedent were purely syntactically determined, one would have to assume that it attaches at different points in the syntactic tree in order to derive these three readings from one and the same syntactic string. Such an analysis, however, could not explain why examples like (19) can be understood in at least two different ways depending on a strict or sloppy interpretation of the reflexive pronoun sich (‘himself ’). (19) a. Hans will sich rasieren, was Max auch will. Hans wants to REFL shave which Max also wants to ‘Hans wants to shave himself, which Max wants to do, too.’ b. “Max wants Hansi to shave himselfi.” c. “Maxj wants to shave himselfj.”
Taking these facts into account, it suggests itself that the antecedent of a wh-relative is semantically restricted. Further, it seems to be incorrect that a wh-relative clause only relates to the state of affairs of the matrix clause and is in this sense generally sentence-related. 7.
Actually, there is a second reading of (19) meaning that Max also wants Hans to shave himself.
Anke Holler
Before discussing the issue which semantic objects may function as an antecedent of a wh-relative clause in more detail in section 4.3, I will next give empirical evidence for an syntactic analysis that adjoins a wh-relative clause to the highest functional projection representing the prior sentence. 4.2
The syntactic relationship
Due to their root clause properties, wh-relative clauses are not canonically linked to their syntactic surrounding. As has been argued in Holler (2005), wh-relative clauses are not syntactically integrated into the complex sentence structure representing the whole construction. This is substantiated by the following facts. First of all, wh-relatives are prosodically and pragmatically independent from their matrix clause, which is indicated inter alia by an independent focus domain, cf. (20), and an autonomous illocutionary force, cf. (21). The construction in (20) for instance cannot be uttered as an answer to a sentence-wide focus inducing question such as What happened?, which indicates that the wh-relative clause is not integrated into the information structure of the host clause. (20) # [Emma kaufte einen teuren Schrank, was ÄRGERlich ist.]F Emma bought an expensive cupboard which annoying is ‘Emma bought an expensive cupboard, which is annoying.’
Similarly, (21) is ungrammatical because the wh-relative clause has been forced to be a part of the host’s speech act which is a question. In other words: By uttering a wh-relative clause an illocutionary act is performed that is independent from the illocutionary act performed by the utterance of the matrix clause. (21) *Hat Max eine schöne Frau geküsst, was seine Ehefrau maßlos ärgerte? has Max a beautiful woman kissed which his wife extremely annoyed
Moreover, wh-relatives disallow variable binding from outside as can be seen in (22). Hence, a wh-relative clause is not c-commanded by its host, which means that it does not function as a syntactic sister of the matrix clause. (22) *Niemandi gewann das Schachspiel, was ihni maßlos ärgerte. nobody won the game of chess which him extremely annoyed
In addition, wh-relative clauses are topologically restricted. They may occur only at the very end of a complex sentence as is shown by (23) and (24) illustrating that a wh-relative must follow an extraposed complement clause or an extraposed relative clause. (25) demonstrates that a wh-relative clause does not stand in the so-called 8. This is due to the general topological freedom of parentheticals.
Antecedents of diverse types
Vorfeld-position, either. The described topological behavior also suggests that a whrelative clause is not integrated into the sentence structure of the matrix clause. (23) a.
Es fiel Maria nicht auf, dass sie sich verrechnet hatte, EXPL realized Maria not PART that she REFL mistaken had weswegen sie sich jetzt ärgert. that’s why she REFL now annoyed ‘Maria didn’t realize that she made a mistake, and that’s why she is annoyed now.’ b. *Es fiel Maria nicht auf, weswegen sie sich jetzt ärgert, EXPL realized Maria not PART that’s why she REFL now annoyed dass sie sich verrechnet hatte. that she REFL mistaken had (24) a.
Anna hat einen Ring verloren, der sehr wertvoll war, weshalb sie sich Anna has a ring lost that very valuable was that’s why she REFL jetzt maßlos ärgerte. now extremely annoyed ‘Anna lost a ring that was very valuable, and that’s why she was annoyed now.‘
b.
*Anna hat einen Ring verloren, weshalb sie sich jetzt maßlos Anna has a ring lost that’s why she REFL now extremely ärgerte, der sehr wertvoll war. annnoyed that very valuable was
(25)
*Weshalb sie sich jetzt maßlos ärgerte, hat Anna einen wertvollen that’s why she REFL now extremely annoyed has Anna a valuable Ring verloren. ring lost
Furthermore, the standard constituent tests illustrated in (26) indicate that a whrelative clause is not adjoined to a V- or VP-projection: (i) A constituent consisting of the matrix verb or the matrix VP and the wh-relative clause cannot be asked for, cf. (26b), (ii) such an constituent cannot be topicalized, cf. (26c), and (iii) it cannot be coordinated either, cf. (26d). (26) a.
Anna wird am Triathlon teilnehmen, was Emma sich nicht Anna will PREP triathlon take part in which Emma REFL not traut. have the heart to do ‘Anna will take part in a triathlon, which Emma doesn’t have the heart to do‘
b. Was tut sie? *{Am Triathlon teilnehmen/ Teilnehmen}, ‘What does she do?‘ PREP triathlon take part in/ take part in
Anke Holler
was Emma sich nicht traut. which Emma REFL not have the heart to do
c.
*{Am Triathlon teilnehmen/ Teilnehmen}, was Emma sich nicht PREP triathlon take part in/ take part in which Emma REFL not traut, wird Anna. have the heart to do will Anna
d.
*Anna wird {am Triathlon teilnehmen/ teilnehmen} und was Emma Anna will PREP triathlon take part in/ take part in and which Emma sich nicht traut. REFL not have the heart to do
Thus, it has to be concluded that a wh-relative clause is not integrated into the syntactic structure of the matrix clause, but ties in with the complete matrix clause. A wh-relative clause can be considered as sentence-related in this syntactic sense only. 4.3
Potential semantic antecedents
In section 4.1, I argued that the antecedent of a wh-relative clause is semantically determined, and that a wh-relative may relate to semantic antecedents of various kinds. Now, I closer investigate which grammatical restrictions affect these potential antecedents. Note that similar aspects are discussed by Consten, Knees and Schwarz-Friesel (this volume) with respect to the complex anapher das (’this’). It is often claimed in grammars and in the philologically oriented literature, e.g. Helbig and Buscha (1986), Brandt (1990), Zifonun et al. (1997), etc., that wh-relatives generally refer to facts. Examples like (27) are usually quoted as evidence for this claim. (27)
Der Fußballspieler beleidigte den Schiedsrichter, was zu seinem. The soccer player offended the referee which PREP his Ausschluss führte disqualification resulted ‘The soccer player offended the referee, which resulted in his disqualification.’ (Helbig and Buscha, 1986)
Example (28), however, shows that this assumption is not fully correct. The whrelative clause contains a perception verb which may have an event reading. For instance, (28) can mean that Anna heard how Max sang during taking a shower.
9. Compare Mönnich (1992) for a discussion of possible readings of perception verbs.
Antecedents of diverse types
(28) Max sang unter der Dusche, was Anna mal wieder deutlich hörte. Max sang under the shower which Anna once again clearly heard ‘Max sang taking a shower, which Anna clearly heard once more.’
In addition, (29) shows that verbs like dauern (‘to last’) are admissible in a whrelative clause. This would not be expected if wh-relatives were restricted to factive antecedents since it is widely accepted that dauern (‘to last’) refers to an event. (29) Max rasierte sich, was mal wieder ewig dauerte. Max shave REFL which once again forever lasted ‘Max shaved himself, which once more took ages.’
Furthermore, a wh-relative clause may refer to an event type as (30) demonstrates. (30) means that Anna as well as Max has found a solution to a certain problem, but not necessarily the same one. (30) Anna hat das Problem gelöst, was Max auch gelungen ist. Anna has the problem solved which Max also succeeded is ‘Anna has solved the problem, and Max succeeded, too.’
Even if we ignore the aforementioned observation that wh-relative clauses can relate to non-propositional antecedents, it can be shown that the assumption that wh-relatives are restricted to facts does not hold for all propositional antecedents, either. First, factive verbs that clearly relate to activities like kritisieren (‘to critizise’) and bereuen (‘to regret’), cf. Ehrich (1991), can occur as predicates in a wh-relative clause as (31a) and (31b) illustrate. (31) a.
Anna reist mitten im Hochsommer ans Mittelmeer, Anna travels in the middle of in the midsummer to the Mediterranean was sie alsbald bereut. which she shortly regrets ‘Anna travels to the Mediterranean in the middle of midsummer, which she shortly regrets to have done.’ b. Stoiber beschimpft die ostdeutschen Wähler, was die Presse sofort Stoiber swears the East German voters which the media immediately kritisiert. criticize ‘Stoiber swears at the East German voters, which the media immediately criticize.’
Second, responce-stance verbs, cf. Hegarty (1992), which do not presuppose their complement as factive verbs do, can be found in a wh-relative clause.
Anke Holler
(32) Anna hat gekündigt, was ihr Chef akzeptieren muss. Anna has signed off which her boss accept must ‘Anna has signed off, which her boss has to accept’
Third, verbs that semantically select a question are possible in a wh-relative clause as (33) demonstrates. (33)
Der Richter will wissen, wer die Millionen ins Ausland geschafft hat, the judge wants to know who the millions PREP abroad brought has was den Staatsanwalt auch sehr interessiert. which the attorney also very interests ‘The judge wants to know who has brought the millions abroad, and this very much interests the attorney, too.’
The assumption that factivity is the decisive factor in establishing a reference relation between a wh-relative clause and its antecedent was probably motivated by unacceptable examples like (34a), which clearly contrast to examples like (34b). (34) a. *Fred hat sich den Arm gebrochen, was sein Trainer vermutete. Fred has REFL the arm broken which his trainer suspected b. Fred hat sich den Arm gebrochen, was sein Trainer bedauerte. Fred has REFL the arm broken which his trainer regretted ‘Fred has broken his arm, which his trainer regretted.’
Interestingly, however, (34a) improves a lot if the clause expressing the potential antecedent of the wh-relative clause is embedded under a predicate whose argument is restricted to the same semantic type as the argument of the verb in the wh-relative clause: (35)
Fred glaubte, dass er sich den Arm gebrochen hat, was sein Trainer ebenfalls Fred believed that he REFL the arm broken has which his trainer also vermutete. suspected ‘Fred believed that he has broken his arm, which his trainer suspected, too.’
These data suggest that the syntactic antecedent of a wh-relative clause must provide a semantic object whose type is in line with the semantic type of the wh-expression introducing the wh-relative clause. If the wh-expression is an argument of the wh-relative clause’s predicate, then the semantic type of the wh-expression is restricted by this predicate according to its selection properties; otherwise it is propositional because of the adverbial relation that is established. Thus, the construction (34a) is not grammatical since there is no content of belief was (‘which’)
Antecedents of diverse types
could refer to. This is different in (35) where the embedded clause expresses such a content of belief induced by the matrix verb glauben (‘to believe’). (36a) presents two additional examples showing that a mismatch between the semantic type of the wh-expression and its potential antecedents is responsible for the ungrammaticality of certain wh-relative constructions. (36) a. *Anna besitzt einen Bentley, was im Frühjahr geschah. Anna owns a Bentley which in spring happened b. *Max beschloss, die Ausstellung zu besuchen, was Anna fragte Max decided the exhibition to visit which Anna asked
In (36a), was (‘which’) must refer to an event; the predicate of the matrix clause, however, expresses a state. In (36b), was (‘which’) requires a question meaning as antecedent, which cannot be provided by the respective matrix clause. Summarizing the facts: Provided that the wh-expression saturates an argument of the wh-relative predicate, the selectional properties of the predicate of a wh-relative clause determine the semantic type of the left-peripheral wh-expression and, thus, the semantic type of the antecedent to which the wh-relative clause relates. In case an adverbial wh-expression modifies the respective predicate, the potential antecedent has to be propositional. This means that a wh-relative clause may refer at least to events, event types, propositions and question meanings. An antecedent of a suitable type must be available in order that a wh-relative construction is grammatical. Consequently, a wh-relative construction is ambiguous if (i) the wh-expression is ambiguous in its semantic type, and if (ii) several appropriate antecedents of these types are accessible.
5. An account of the facts This section is devoted to an analysis of the facts afore discussed. The core of the syntactic analysis advocated here is the idea that wh-relative clauses belong to a class of syntactically non-integrated clauses that attach to a functionally complete sentential projection. In principle, this can be implemented in any phrase-structural framework that is able to represent functional completeness. I will sketch a sample analysis in section 5.1. In order to account for the semantic facts it is assumend that the wh-expression is an anaphoric element introducing an incomplete anaphoric relation into the representation. This triggers the search for an suitable semantic antecedent within the clause the wh-relative is syntactically attached to. The semantic analysis as proposed in section 5.2 is couched in terms of Discourse Representation Theory
Anke Holler
(DRT, cf. Kamp and Reyle (1993), Asher (1993)) whose context-driven dynamic semantics qualifies very well to account for the presented data. 5.1
Wh-relatives are syntactically non-integrated
Given the facts discussed in section 4.2, any structural analysis of wh-relative clauses has to account for their non-integratedness. The proposals by Reis (1997) and Gärtner (2001) are pertinent for analyzing German non-integrated clauses. Both authors treat them paratactically. Reis (1997) assumes that they are adjoined at the highest sentential level, i.e. they are base-generated as a right sister of the CP representing the host clause. This analysis is depicted in figure1. As can be seen in figure 2, Gärtner (2001) stipulates a paratactic phrase π, which adjoins to the host clause and whose head is phonologically empty. The non-integrated clause acts as a complement of this empty head.
Figure 1. Paratactic structure that results from adjoining a non-integrated clause to its host according to Reis (1997)
Figure 2. Paratactic structure that results from adjoining a non-integrated clause to its host according to Gärtner (2001)
Antecedents of diverse types
Both proposals presume that there is a syntactic relation between the non-integrated clause and its host. This is in line with approaches that treat non-restrictive relative clauses as syntactically related to a constituent of the sentence which contains them, such as Emonds (1979), de Vries (2000), Arnold (2004) and many others. These approaches differ in the syntactic relation that is exploited to attach the non-restrictive relative clause to its host clause (i.e. modification, specification, theta-assignment, coordination, etc.), and in the categorial or projectional status of the constituent (N, D, NP, DP, CP, etc.), which the relative clause combines with. For a comprehensive survey of existing theories of non-restrictive relative clauses see de Vries (2002). Reis (1997) and Gärtner (2001) share the fundamentals of the approaches by Emonds (1979), Safir (1986) and McCawley (1982) to capture English non-restrictive relative clauses by adjoining them to a root clause. I follow this basic idea. Hence, wh-relative clauses are adjoined to the highest functionally complete projection representing the host clause. Figure 3 sketches this analysis.10
Figure 3. Structural analysis for a wh-relative construction
The analysis obviously explicates the fact that a wh-relative clause is syntactically sentence-related. The afore mentioned syntactic and topological properties of the wh-relative construction automatically follow from it. Separating the syntactic and the semantic relationships in the analysis of the wh-relative construction is in the spirit of Culicover and Jackendoff ’s claim that “it is plausible to view parataxis as grammatically the most primitive way to combine linguistic elements, one that leaves the semantic relations among the elements to be determined by their inherent semantic possibilities or by pragmatic considerations coming from the discourse or non-linguistic context.” (Culicover and Jackendoff, 2005, 528).
10. For a technically more elaborated analysis couched in terms of the framework of Head-driven Phrase Structure Grammar (Pollard and Sag, 1994) see Holler (2005).
Anke Holler
5.2
Wh-relatives relate semantically to abstract objects
It has been argued in section 4.3 that a wh-relative clause construction is grammatical if the matrix clause contains an object whose semantic type is consistent with the semantic type of the wh-expression introducing the relative clause, and thus may function as a semantic antecedent of the wh-relative clause. In this section, I will present an analysis accounting for this empirical fact adapting an approach developed by Asher (1993) in the framework of Discourse Representation Theory (DRT, Kamp and Reyle (1993)). The analysis proposed here is based on the assumption that the left-peripheral wh-expression may be analyzed as a wh-anaphor that introduces both into the representation, a discourse referent whose semantic type is restricted by the wh-relative clause’s predicate, and an open anaphoric condition which has to be completed by an appropriate semantic object abstracted from the clause which the wh-relative clause is syntactically attached to. Asher (1993) examining abstract entities in natural language metaphysics develops a Vendlerian-style typology of saturated abstract objects. This is given in figure 4.
Figure 4. Typology according to Asher (1993, 57)
As Asher states it, the basic types at the leaves of the tree are arranged from left to right in increasing order of abstraction along the spectrum of world immanence. The DRS construction procedure proposed by Asher (1993) distinguishes between several types of objects by assigning distinct types of discourse referents for various sorts of eventualities, concepts, propositions, concrete individuals and groups. In addition, Asher supposes that there are abstract discourse referents which may have specific types, such as fact discourse referents, pure proposition discourse referents, projective proposition discourse referents and so on. Type information is defined by means of conditions on discourse referents. Distinct letters such as f1,
Antecedents of diverse types
f2,... for fact discourse referents or p1, p2,... for pure propositional discourse referents abbreviate these extra conditions. Asher (1993) further defines a characterization relation denoted by ≈, which takes as arguments a discourse referent of some abstract type and a DRS. x ≈ K is written when the DRS K characterizes x.11 It suggests itself to apply Asher’s framework to the semantics of the wh-relative construction. The wh-anaphor introduces an abstract discourse referent which has to be characterized by an appropriate entity of the syntactic antecedent. For instance, if the wh-anaphor introduces an fact discourse referent as in example (2a), repeated as (37), it is characterized by a fact abstracted from the syntactic antecedent of the wh-relative clause as shown by the DRS in figure 5. (37) Deutschland verlor das Fußballspiel, was die Fans bedauerten. Germany los the soccer match which the fans regretted ‘Germany lost the soccer match which the fans regretted.’
Figure 5. DRS of a wh-relative clause relating to a fact
If the wh-anaphor on the other hand introduces an event discourse referent as in example (29), repeated here as (38), its antecedent has to be eventive as well. The respective event is abstracted from the syntactic antecedent of the wh-relative clause as the DR-theoretic representation in figure 6 illustrates. (38) Max rasierte sich, was mal wieder ewig dauerte. Max shave REFL which once again forever took ‘Max shaved himself which once again took ages.’ 11. For the interpretation of the characterization relations for various types of abstract discourse referents see Asher (1993).
Anke Holler
Figure 6. DRS of a wh-relative clause relating to an event
Now, it is easy to explain why an example like (36a), repeated as (39), is ungrammatical. As the DRS in figure 7 demonstrates there is no event that could be abstracted from the syntactic antecedent. (39) *Anna besitzt einen Bentley, was im Frühjahr geschah. Anna owns a Bentley which in spring happened
Figure 7. Unresolved DRS for a wh-relative clause without a suitable antecedent
Thus, the presented analysis based on Asher’s DR-theoretic approach is able to account for any anaphoric relation established between the wh-anaphor introducing a wh-relative clause and an abstract object contained in the syntactic antecedent of the wh-relative clause. There are no extra conditions that need to be defined, since everything can be derived from the properties of the wh-expression, in particular from its anaphoric status: A wh-relative construction is grammatical if the whanaphor introduces an anaphoric discourse referent that is characterized by an object abstracted from the host clause. A wh-relative construction is ungrammatical if the wh-anaphor introduces an anaphoric discourse referent that cannot be resolved at all. A wh-relative construction is ambiguous if the wh-anaphor introduces several anaphoric discourse referents that all potentially could be resolved.
Antecedents of diverse types
6. Conclusion In this article, I hope to have shown that the syntactic and semantic relationships in a wh-relative construction need to be distinguished. Syntactically, a wh-relative clause attaches to the functionally complete sentential projection representing the host clause, since wh-relative clauses behave like root clauses and are not integrated into the complex sentence structure. Semantically, a wh-relative clause may relate to abstract objects of various types, which can be attributed to the anaphoricity of the left-peripheral wh-expression.
References Arnold, D. 2004. Non-restrictive relative clauses in construction based HPSG. In Proceedings of the HPSG04 Conference, S. Müller (ed.). Stanford CA: CSLI. Asher, N. 1993. Reference to Abstract Objects in Discourse [Studies in Linguistics and Philosophy 50]. Dordrecht: Kluwer. Bech, G. 1957. Studien über das deutsche verbum infinitum. Kopenhagen: Munksgaard. (Reprint 1983, Tübingen: Niemeyer). Brandt, M. 1990. Weiterführende Nebensätze. Stockholm: Almquist & Wiksell International. Culicover, P.W. and Jackendoff, R. 2005. Simpler Syntax. Oxford: OUP. de Vries, M. 2000. Appositive relative clauses. Linguistics in the Netherland, AVT 17: 221–231. de Vries, M. 2002. The Syntax of Relativization. PhD dissertation, University of Amsterdam. Duden (eds). 2005. Duden: Die Grammatik, Vol. 4. (7th ed.). Mannheim: Dudenverlag. Ehrich, V. 1991. Nominalisierungen. In Semantik. Ein internationales Handbuch der zeitgenössischen Forschung, A. v. Stechow and D. Wunderlich (eds), 441–458. Berlin: Walter de Gruyter. Emonds, J. 1979. Appositive relatives have no properties. Linguistic Inquiry 10(2): 211–243. Fabricius-Hansen, C. 1981. Was ist nun wieder ein Korrelat? Gedanken zur Rehabilitierung eines naiven Nebensatzbegriffes. Kopenhagener Beiträge zur germanistischen Linguistik 18: 1–45. Fabricius-Hansen, C. 1992. Subordination. IdS-Jahrbuch (Tagung 1991): 458–483. Gärtner, H-M. 2001. Are there V2-relative clauses in German? Journal of Comparative Germanic Linguistics 3(2): 97–141. Hegarty, M.V. 1992. Adjunct Extraction and Chain Configuration. PhD dissertation, MIT. Helbig, G. 1980. Was sind weiterführende Nebensätze? Deutsch als Fremdsprache 17: 13–23. Helbig, G. and Buscha, J. 1986. Deutsche Grammatik. Ein Handbuch für den Ausländerunterricht.. Leipzig: VEB Enzyklopädie Verlag. Holler, A. 2005. Weiterführende Relativsätze. Empirische und theoretische Aspekte [Studia Grammatica 60]. Berlin: Akademie Verlag. Kamp, H. and Reyle, U. 1993. From Discourse to Logic. Dordrecht: Kluwer. McCawley, J.D. 1982. Parentheticals and discontinuous constituents. Linguistic Inquiry 13: 91–106. Mönnich, U. 1992. Ereignisse und Gedanken: Syntax und Semantik von Perzeptionsverbkomplementen. Ms, University of Tübingen. Pasch, R. 1983. Untersuchungen zu den Gebrauchsbedingungen der deutschen Kausalkonjunktionen da, denn und weil. Linguistische Studien Reihe A(104): 41–243.
Anke Holler Pollard, C.l and Sag, I.A. 1994. Head-Driven Phrase Structure Grammar. Chicago IL: CSLI & University of Chicago Press. Reis, M. 1977. Präsuppositionen und Syntax [Linguistische Arbeiten 51]. Tübingen: Niemeyer. Reis, M. 1997. Zum syntaktischen Status unselbständiger Verbzweit-Sätze. In Syntax im Fokus. Festschrift für Heinz Vater, C. Dürscheid, K.H. Ramers and M. Schwarz (eds). Tübingen: Niemeyer. Safir, K. 1986. Relative clauses in a theory of binding of levels. Linguistic Inquiry 17: 663–689. Zifonun, G., Hoffmann, L. and Strecker, B.. 1997. Grammatik der deutschen Sprache. Berlin: Walter de Gruyter. Zimmermann, Ilse. 1993. Zur Syntax und Semantik der Satzeinbettung. In Satz und Illokution [Linguistische Arbeiten 279], M. Reis and I. Rosengren (eds), 231–249, Tübingen: Niemeyer.
Corpus-based and machine learning approaches to anaphora resolution A critical assessment Michael Strube Anaphora resolution is an important component of natural language processing applications like information extraction, question answering, or automatic summarization. These applications have to deal with unrestricted input which is difficult to process with symbolic anaphora resolution methods. If trained on unrestricted input, machine learning based anaphora resolution methods can robustly deal with a wide variety of input documents. Those methods are mostly implemented as binary classification realizing models of local inference. While this makes the task accessible to standard machine learning techniques, it has the drawback that knowledge about the context is lost. Based on a critical assessment of the state-of-the-art, models of global inference are introduced as a possible alternative.
1. Introduction Anaphors are an important linguistic device for the speaker or writer to signal the local coherence of a discourse. Anaphoric expressions comprise mainly personal and possessive pronouns, demonstratives, proper names, and definite noun phrases when they are used to refer to something previously mentioned in the discourse. The linguistic expression realizing the previous mention is called the antecedent. In order to construct a coherent representation of the discourse, hearers or readers have to interpret, i.e. resolve anaphoric expressions. This task is difficult, because many anaphoric expressions are referentially ambiguous, i.e., they can potentially 1. The terms anaphor and antecedent may be defined differently in text linguistics and therefore in the remainder of this volume. Here, definitions taken from work in Computational Linguistics and Natural Language Processing are used.
Michael Strube
be used to refer to more than one antecedent. Their unambiguous resolution requires a considerable amount of linguistic and world knowledge. Research in the area of Natural Language Processing about anaphora resolution started out by modelling linguistic knowledge algorithmically. Early work on pronoun resolution used the syntactic structure to define constraints and search procedures for determining pronouns’ antecedents (Hobbs, 1978). Building upon refinements in linguistic theory, e.g. the development of Binding Theory (Chomsky, 1981) and Discourse Representation Theory (Heim, 1982; Kamp & Reyle, 1993) pronoun and anaphora resolution algorithms were implemented within syntactic parsers (e.g. Correa (1988)) and semantic ones (e.g. Asher (1988)). Filters provided by those theories are only one part of the knowledge needed by an anaphora resolution algorithm. Filters can only reduce the set of potential antecedents of anaphoric expressions. They cannot select the most likely antecedent. Hence it is necessary to provide knowledge about preferences for certain candidates. Preferences are provided either by linguistically and psychologically motivated models (e.g. the centering model (Grosz et al., 1995)) or by heuristics (e.g. Lappin & Leass (1994)). Most centering and heuristics-based algorithms rely on a complete syntactic analysis. Since there exist no parsers which provide that complete analysis robustly on unrestricted input, those algorithms were evaluated either only manually or automatically on a very restricted set of documents. However, in the mid-nineties the need to apply anaphora resolution algorithms to unrestricted input grew. This led to three important developments in the area of anaphora resolution. Firstly, corpora were annotated with information about anaphoric relations, in particular due to the series of Message Understanding Conferences (MUC). Secondly, heuristics were adapted such that their preprocessing demands decreased. Hence, anaphora resolution algorithms based on heuristics became more robust and could be applied to unrestricted input (e.g. Kennedy & Boguraev (1996)). Thirdly, statistical and machine learning techniques were adapted such that they could be applied to anaphora resolution (McCarthy & Lehnert, 1995; Aone & Bennett, 1995). 1.1
Definition of the task
From a linguistic point of view anaphora resolution is defined as the task of finding the antecedent for an anaphoric expression, i.e. to determine that “the sentryi” in (1a) below is the antecedent of “hei” in (1b). In the area of Natural Language Processing the task early on was considered as identifying the entity a referring 2. http://www.itl.nist.gov/iaui/894.02/related_projects/muc/ 3.
Example (1) is taken with slight simplifications from the Brown corpus cn03.
Corpus-based and machine learning approaches to anaphora resolution
expression is used to refer to, the discourse entity (Webber, 1979). If multiple referring expressions are used to refer to the same entity they are said to corefer (see Sidner (1983) for a discussion of this terminology), i.e. all referring expression indexed with with the same letter are said to corefer. Based on those theoretical assumptions, the task of anaphora resolution has been redefined as coreference resolution, i.e., to assign a referring expression the discourse entity it is used to (co)refer to. The discourse entity then may be viewed as a representation of the set of referring expressions which are used to refer to it, or as a representation of a coreference chain (all referring expressions indexed with the same letter belong to a coreference chain, e.g. the sentryi – hei – hei – himi – the mani). (1)
a. b. c. d. e.
The sentryi was not dead. Hei was, in fact, showing signs of reviving... Hei was partially uniformed in a cavalry tunicj. Mikek stripped thisj from himi and donned itj. Hek tied and gagged the mani,...
For applications like information extraction or automatic summarization, to determine the antecedent of an anaphoric expression is not relevant. Hence, coreference resolution is the task most statistical approaches attempt to solve. In the following we describe the approach due to Soon et al. (2001). Most recent machine learning approaches to coreference resolution are based on their baseline approach or compared with it. Soon et al. use the corpora which were created for the coreference task of MUC-6 and MUC-7. The input to their system consists of raw data, i.e. data stripped from all kinds of linguistic information (annotation). Hence a sequence of preprocessing components is needed to determine the extent of referring expressions and assign them values for the features shown in Table 1 (possible values of the features are given in parentheses). The preprocessing components used by Soon et al. (2001) are: tokenization and sentence segmentation, morphological processing, Part-of-speech tagging, noun phrase identification, named entity recognition, nested noun phrase extraction, and semantic class determination. In order to express coreference resolution as a machine learning task, Soon et al. (2001) transform it into a binary classification task. Annotated data (see Section 1.2) are converted into pairs of potential anaphors and potential antecedents. These data instances are realized as feature vectors which contain values of the features from Table 1 and a target concept, i.e. the value True or False, indicating whether the potential antecedent is in fact the correct antecedent of that anaphoric expression or not.
Michael Strube
Table 1. Features used by Soon et al. (2001) 1. 2. 3. 4. 5. 6. 7. 8.
Distance i-Pronoun j-Pronoun String Match Definite Noun Phrase Demonstrative NP Number Agreement Semantic Class Agreem.
9. 10 11.
Gender Agreement Both-Proper-Names Alias
12.
Appositive
Distance between ana and ante in sentences (0, 1, 2,...) is ante a pronoun? (T, F) is ana a pronoun? (T, F) are ana and ante identical? (T, F) is ana definite noun phrase? (T, F) is ana definite NP with demonstrative determiner? (T, F) do ana and ante agree with respect to number? (T, F) do ana and ante agree with respect to semantic classes which are arranged in a isa-hierarchy: female, male, person, organization, location, date, time, money, percent, object? (T, F) do ana and ante agree with respect to gender? (T, F) are both ana and ante proper names? (T, F) are ana and ante similar (substring match, abbreviations, etc.)? (T, F) is ana appositive of ante? (T, F)
Table 2 shows a few feature vectors which may have been generated using Example (1). To keep it simple, only the two referring expressions from (1e) “he” and “the man” are taken as potential anaphors. Potential antecedents are taken only from (1c-e). The target concept is the concept to be learned. Table 2. Feature vectors ante
ana
1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. target concept
hei a cavalry tunicj Mikek thisj himi itj
hek hek hek hek hek hek
2 2 1 1 1 1
T F F T T T
T T T T T T
T F F F F F
F F F F F F
F F F F F F
T T T T T T
T F T F T F
T F T F T F
F T F F F F F F F F F F F F F F F F
F F T F F F
hei a cavalry tunicj Mikek thisj himi itj hek
the mani the mani the mani the mani the mani the mani the mani
2 2 1 1 1 1 0
T F F T T T T
F F F F F F F
F F F F F F F
T T T T T T T
F F F F F F F
T T T T T T T
T F T F T F T
T F T F T F T
F T F F F F F F F F F F F F F F F F F T F
T F F F T F F
In each of the three abstract text representations in Figure 1 one referring expression is considered, the it in (1d), the he and the man in (1e). These three referring expressions marked in italics are considered potential anaphors in the following
Corpus-based and machine learning approaches to anaphora resolution
discussion. The it has five, the he six, and the man seven potential antecedents in fragment shown in Figure 1. This results in 18 data instances, i.e. pairs of potential anaphors and potential antecedents. Hence, coreference resolution formulated as binary classification results in a large number of data instances, even if we consider only eight referring expressions in three sentences.
Figure 1. Abstract text representation
Figure 2. Positive, if anaphor and antecedent are coreferent
The classification task then can be formulated as follows: Given a pair of potential anaphor and potential antecedent, classify as positive if anaphor and antecedent are in fact coreferent, i.e., both referring expressions are used to refer to the same discourse entity. Classify as negative otherwise. This is shown in Figure 2 where solid arrows indicate a positive classification, dashed arrows negative ones. This data generation strategy, which was used successfully by Aone & Dennett (1995), may lead to a problem with the distance feature (i.e., the feature which captures the distance in
Michael Strube
words or sentences between anaphor and antecedent) in the case of long coreference chains. Also, this data generation strategy leads to a highly skewed distribution of positive and negative instances. In Figure 2 there are 14 negative instances and only four positive instances. Applied to real data this strategy may yield more than 98% negative instances. Soon et al. (2001) use the following strategy for reducing the large number of negative instances: Given a pair of potential anaphor and potential antecedent, classify as positive if the antecedent is in fact the closest antecedent. Classify as negative only if the potential (wrong) antecedent is closer than the closest (correct) antecedent. Exclude all instances where the potential antecedent is farther away than the closest (correct) antecedent, i.e., exclude transitive positives and distant negatives, see Figure 3, where only three positive and six negative instances are left. However, this strategy is rather uninituitive. Intuitively, anaphor and antecedent should be rather close. Following this strategy, however, the distance between anaphor and antecedent in positive instances is always the maximum distance. Consequently, in Soon et al. (2001) the distance feature proved to be not important for solving the task.
Figure 3. Exclude transitive positives and distant negatives
1.2
Data
The most important prerequisite for statistical approaches are annotated data. There exist several corpora annotated with information about coreference, e.g. the MUC-6 and MUC-7 corpora. However, those corpora are rather small and ideo4. These corpora are available through the Linguistic Data Consortium (LDC), http://www. ldc.upenn.edu; MUC-7 sample at http://www.itl.nist.gov/iaui/894.02/related_projects/muc/ proceedings/walkthru_co_key.html).
Corpus-based and machine learning approaches to anaphora resolution
syncratic. A subset of the Penn TreeBank has been extended with information on pronouns and their antecedents by Ge et al. (1998) (available from them on request). The ACE corpus (Automatic Content Extraction, http://www.itl.nist.gov/ iad/894.01/tests/ace/) is available through the LDC. This corpus is substantially bigger than the MUC corpora, but only anaphoric relations relevant to the ACE program are marked. If a statistical approach is to be applied to a new genre, domain or language, data have to be annotated with information about anaphoric relations. This can be done manually or semi-automatically using annotation tools, e.g. the annotation tool MMAX (download at http://mmax.eml-research.de). 1.3
Evaluation
The evaluation of the anaphora resolution task is straightforward. In the key (i.e., the data representing the solution for the task) each anaphor has exactly one antecedent. The response (i.e., the output of the anaphora resolution system) also contains exactly one antecedent for an anaphor. The performance of a system then can be measured in terms of the success rate and in terms of precision, recall and Fmeasure (see Table 3 for a list of those terms and their abbreviations). The F-measure represents the harmonic mean of precision and recall. Table 3. Terms used for evaluation Na: Nc: Nh: S: P: R: F:
number of all anaphors in a document or corpus number of correctly resolved anaphors number of resolutions attempted success rate precision recall equally weighted F-measure
The success rate just compares the number of correctly resolved anaphors with the number of all anaphors in the document or corpus (Mitkov, 2002):
S=
Nc Na
An evaluation which reports precision, recall and F-measure may prove more realistic since a high recall, low precision system may still achieve a high success rate, though it may be not desirable for most applications.
Michael Strube
P=
Nc Nh
R=
Nc Na
F=
2PR P+R
Problems with those measures arise because researchers often do not report exactly which types of anaphoric expressions are relevant for the evaluation, which types are filtered out manually, and which types their systems attempt to resolve. Therefore the reported results are often not comparable (Byron, 2001). Though there exists an established evaluation metric for the coreference resolution task introduced for MUC-6 by Vilain et al. (1995), this topic is still controversial. Vilain et al.’s evaluation metric partitions the set of referring expressions into sets of coreferring expessions by computing the minimal number of links needed to create those sets. Both key and response consist of those sets. The difference between key and response is given by computing the minimal number of links to be deleted or created in order to transform the response into the key. Finally, precision, recall and F-measure are reported on the basis of those numbers. This metric abstracts from the task of finding “the” antecedent of an anaphoric expression. It reports a result which is very close to the task of coreference resolution as needed in most real world applications.
2. State-of-the-art approaches Most work in statistical coreference resolution is realized using off-the-shelf machine-learning software as provided by the Weka machine learning library (Witten & Frank, 2005), the statistical software package R, or more specialized software such as C4.5 and C5.0 (Quinlan, 1993) or TiMBL (Daelemans et al., 1999). These software packages allow to implement supervised machine learning approaches. These approaches are trained using labeled data, i.e., data annotated with the target concept. The training procedure creates a model which can be applied to a development test set, i.e., data not used for the training (a valid approach to avoid overfitting is 10-fold cross-validation). The development test set allows to tweak the model until it performs best. Finally, results should be obtained by applying the model to a holdout test data set only once during the whole process. This procedure ensures that the results are not due to overfitting but can be generalized. When the whole data set is large, it is sufficient to use between 10% and 20% of the data for the holdout test data.
5.
http://www.r-project.org.
2.1
Corpus-based and machine learning approaches to anaphora resolution
Decision trees
Decision-tree based machine learning algorithms (CART by Breiman (1984) and C4.5 and C5.0 by Quinlan (1993) and their respective reimplementations in R and Weka) are popular because they not only return a model which can be used to classify unseen test data but also a decision tree which can be inspected. That way the researcher can get an intuition which of the features are important and whether the learned model corresponds to their intuition. The state-of-the-art technique in decision-tree based coreference resolution by Soon et al. (2001) uses the MUC-6 and MUC-7 data which are already annotated with the target concept. They apply the chain of preprocessing components mentioned in Section 1.1 to extract markables (potential anaphors and antecedents) from the data and to assign the twelve features given in Table 1 to them. They generate training data following the strategy already shown in Figure 3. A standard decision-tree based classifier is trained (C4.5). The learned model is applied to testing data which are fed into the evaluation algorithm according to Vilain et al. (1995). Soon et al. report results of 62.6% F-measure for MUC-6 and 60.4% for MUC-7. The model can finally be integrated into an NLP application extending it by an anaphora resolution component. 2.2
Maximum entropy
Lappin & Leass’ (1994) heuristic algorithm for pronoun resolution proved to be popular among researchers implementing shallow approaches (e.g. Kennedy & Boguraev (1996)). However, Lappin & Leass’ algorithm has the great disadvantage that the weights assigned to features have to be determined manually. An approach, which allows to express the Lappin & Leass algorithm in statistical terms, is the maximum entropy framework (Berger et al., 1996), which learns the weights from the data automatically. That way a maximum entropy system yields a probabilistic notion of salience, which, in turn, can be used to determine the most likely antecedent of an anaphor. Applied to coreference resolution the maximum entropy framework requires to generate training and testing data also in terms of binary classification (Morton, 2000). It relies on a set of binary features which define constraints on the statistical model that is learned. Given a set of those constraints, the model with the maximum entropy has to be identified, i.e., the model with the least amount of information beyond the constraints. The maximum entropy framework has been applied to coreference resolution by Kehler (1997) and Morton (2000), to pronoun resolution by Kehler et al. (2004). Ponzetto & Strube (2006) reimplement Soon et al.’s approach
Michael Strube
using the maximum entropy framework. The reimplementation yields an improvement of 2.7% F-measure on MUC-6 and 0.7% on MUC-7 over the baseline. 2.3
Statistical modeling
Ge et al. (1998) describe a statistical approach to pronoun resolution. In contrast to the work described above, they do not use preprocessing components. Instead they derive the following features from the Penn TreeBank enriched by them by anaphoric information: distance between pronoun and potential antecedent, gender, number, and animaticity of the potential antecedent, selectional restrictions of the head of the pronoun, the number of times a discourse entity was referred to in a document. Given this information Ge et al.’s statistical approach can be described as a function from pronouns to antecedents. Then they collect the relevant statistics from training examples and run the program on test data. They compute probabilities for each potential antecedent and select the one with the highest probability. Their results cannot be compared directly with the approaches mentioned before, because Ge et al. work with already annotated data (treebank) instead of a series of preprocessing components. 2.4
Unsupervised and semi-supervised machine learning
The approaches described above are supervised, i.e., they require data annotated with the target concept as input. Since this annotation is expensive in terms of cost and labor, unsupervised and semi-supervised approaches are much sought after. Cardie & Wagstaff (1999) present an unsupervised approach to coreference resolution. They start out with the observation that all referring expressions in a coreference chain define an “equivalence class”. Therefore, it would be natural to view the coreference resolution task as clustering. Since all members of a coreference chain should be similar, their conceptual distance should be small. Given a method for measuring this distance, a clustering algorithm could group referring expressions together, such that coreferent ones end up in the same cluster. Cardie & Wagstaff ’s (1999) system uses preprocessing components to assign values to features in vectors describing referring expressions (not pairs of potential anaphors and potential antecedents). Initially each referring expression is assigned its own cluster. Then the algorithm works backwards through the document, where each referring expression is compared with all preceding ones. If the distance between two referring expressions is less than the clustering radius r, they are considered for possible merging into one cluster. Cardie & Wagstaff evaluate their approach on the MUC-6 data. They report a performance comparable to
Corpus-based and machine learning approaches to anaphora resolution
other systems presented at the MUC-6 conference. Their system outperforms the only machine learning based system presented there. Unsupervised machine learning of natural language tasks is almost too difficult. Hence, semi-supervised machine learning meta algorithms, e.g. Co-training (Blum & Mitchell, 1998), were applied. A Co-training classifier consists of two simple classifiers, which, initially, are trained separately using a small set of labeled training data. Each of them is trained on a different subset of features, i.e. on different views of the data. The views have to be disjoint or even conditionally independent. The Co-training algorithm bootstraps by gradually extending the training data with self-labeled instances. It utilizes the two classifiers by letting them in turn label the best positive and best negative instances from the set of unlabeled training data. Instances labeled by one classifier are then added to the other’s training data, and vice versa. After each turn, both classifiers are re-trained on their augmented data sets. This process is repeated either for a given number of iterations or until all the unlabeled data has been labeled. Co-training was applied to coreference resolution by Müller et al. (2002) and later by Ng & Cardie (2003). One result of this work was, that the amount of training data to achieve reasonable performance on the coreference resolution task is rather small (Müller et al., 2002).
3. Recent and future developments 3.1
Overcoming problems with binary classification
Most state-of-the-art approaches to statistical coreference resolution redefine the task as binary classification. While this helps to apply standard classification-based machine learning algorithms, it has serious drawbacks: 1. Even for small documents, the binary classification approach generates very large training and testing data sets, which may prevent researchers to conclude experiments in a reasonable amount of time. 2. Each decision is independent from all other decisions. This may lead to contradicting results (A = B, B = C, A ≠ C ). This may also lead to unnecessary competition between compatible results (A = B, B = C, A = C; there is no need to resolve A twice). 3. Training instances do not have knowledge about the context of referring expressions. They do not have knowledge about other referring expressions used to refer to the same or a different discourse entity.
Michael Strube
In order to avoid problem (1) a few modifications have been proposed. Though the testing data set must not be modified using the information about the target concept, the training data set can be minimized considerably (see Figure 3 and a variation in Ng & Cardie (2002)). In order to avoid the conflicts described in (2), different strategies can be applied. Most often, only one instance per anaphoric expression is allowed to be classified as positive. The standard approach is to select only the positive instance which contains the closest antecedent for an anaphoric expression. Ng & Cardie (2002) do not select the instance with the closest antecedent but the one with the highest confidence value returned by the classifier. The most serious problem is given in (3). Yang et al. (2003, 2004) try to address that problem. Yang et al. (2003) present a competition learning approach (similar to the one by Connolly et al. (1997)) where each anaphoric expression is paired with two antecedents, a correct one and an incorrect one. The concept to be learned is which of the potential antecedents is the correct one. During classification the algorithm is given initially one instance. It determines which of the potential antecedents is the better one. The loser is discarded. The winner is paired with some other candidate. This new instance is given to the classifier which again determines the better one. That way the overall best antecedent for an anaphoric expression can be determined. The results reported by Yang et al. (on MUC-6 and MUC-7 data) indicate that this strategy outperforms state-of-the-art systems. Yang et al. (2004) describe another method of incorporating more contextual information into a pronoun resolution algorithm. In addition to potential anaphor and potential antecedent, they extend the representation of each instance by a third referring expression: the closest antecedent of the potential antecedent (the set of features describing this referring expression is set to nil if there is no such expression). Yang et al. (2004) claim that they are able to include knowledge similar to centering approaches and report increased performance over baseline systems. Luo et al. (2004) describe a more general approach to make knowledge about the context accessible to a coreference resolution algorithm. They propose the Bell tree as representation for the search space. The leaves in the Bell tree contain all possible outcomes (which is a very large number). The task of coreference resolution then is to find the best path from the root to the leaf nodes in the Bell tree. Nodes in the tree are made up of partial solutions. They contain sets of referring expressions representing coreference chains and singletons which are not coreferent. Luo et al. use a maximum entropy framework (Berger et al., 1996) to assign a referring expression to an existing set (coreference chain) or to create a singleton set. That way they provide access to knowledge about other referring expressions in a coreference chain and to competing candidates. Luo et al. report encouraging
Corpus-based and machine learning approaches to anaphora resolution
results on ACE and MUC data. Efficient search heuristics are necessary to implement this approach. Recent work also tried to solve that problem using models of global inference instead of models of local inference (modeling coreference resolution as binary classification realizes a model of local inference). E.g., McCallum & Wellner (2005) apply conditional random fields (Lafferty et al., 2001) to the task of noun coreference. In their work conditional random fields are used to capture knowledge about the whole coreference chain and also wrong antecedents. That way this approach realizes a global optimization approach. Integer linear programming (Nemhauser & Wolsey, 1999) is another global optimization technique which has been applied to some NLP problems (e.g. Marciniak & Strube (2005)) but not yet to coreference resolution. Conditional random fields and integer linear programming are good candidates to overcome problem (3) in a general and efficient manner. 3.2
Integrating linguistic and extra-linguistic knowledge
Most of the work mentioned relies on a rather small set of linguistically simple features. There have been early attempts to include knowledge about predicate-argument structure by Dagan et al. (1995). Recently Kehler et al. (2004) evaluated the contribution of this knowledge to pronoun resolution. They found no significant improvement. Some researchers also tried to include world or domain knowledge in statistical coreference resolution. E.g., Harabagiu et al. (2001) report an improvement in performance by including knowledge from WordNet. Ponzetto & Strube (2006) integrate knowledge about semantic roles of constituents and about the semantic relatedness of nouns computed on the basis of WordNet and Wikipedia. They report significant improvements over the Soon et al. (2001) baseline system. 3.3
Covering more linguistic phenomena
Most of the research about statistical coreference resolution is focused on English language texts, either taken from MUC or from the Penn TreeBank. Aone & Bennett (1995) apply their approach to Japanese texts. Müller et al (2002) work with German texts. Only when annotated data in other languages becomes accessible, statistical methods can be applied to more languages. Existing approaches focus either on pronoun resolution in texts or on noun phrase coreference resolution in texts. Strube & Müller (2003) is the first attempt to apply machine learning methods to pronoun resolution in spoken dialogue. Müller (2006) describes an approach for detecting non-referential it in spoken multi-party dialogue. This prevents non-anaphoric pronouns from being resolved which improves the precision of a coreference resolution classifier. Nissim &
Michael Strube
Markert (2003) present a statistical approach to metonymy resolution, Modjeska et al. (2003) one to other-anaphora resolution. Due to sparse data the latter two approaches first need to identify relevant cases in large corpora. Then they apply standard machine learning techniques.
Acknowledgments This work has been funded by the Klaus Tschira Foundation, Heidelberg, Germany.
References Aone, C. and Bennett, S.W. 1995. Evaluating automated and manual acquisition of anaphora resolution strategies. In Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics, Cambridge MA, 26–30 June 1995, 122–129. Asher, N. andWada, H. 1988. A computational account of syntactic, semantic and discourse principles for anaphora resolution. Journal of Pragmatics 6:309–344. Berger, A.. Della Pietra, S.A. and Della Pietra, V.J. 1996. A maximum entropy approach to natural language processing. Computational Linguistics 22(1):39–71. Blum, A. and Mitchell, T. 1998. Combining labeled and unlabeled data with co-training. In Proceedings of the 11th Annual Conference on Learning Theory, Madison, WI, 24–26 July, 1998, 92–100. Breiman, L., Friedman, J.H., Stone, C.J. and Olshen, R.A. 1984. Classification and Regression Trees. Belmont CA: Wadsworth and Brooks/Cole. Byron, D.K. 2001. The uncommon denominator: A proposal for consistent reporting of pronoun resolution results. Computational Linguistics 27(1):569–577. Cardie, C. and Wagstaff, K. 1999. Noun phrase coreference as clustering. In Proceedings of the 1999 SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora, College Park, MD, 21–22 June 1999, 82–89. Chomsky, N. 1981. Lectures on Government and Binding. Dordrecht: Foris. Connolly, D., Burger, J.D. and Day, D.S. 1997. A machine learning approach to anaphoric reference. In New Methods in Language Processing, D. Jones and H. Somers (eds), 133–143. Oxford: OUP. Correa, N. 1988. A binding rule for government-binding parsing. In Proceedings of the 12th International Conference on Computational Linguistics, Budapest, Hungary, 22–27 August 1988, Vol. 1, 123–129. Daelemans, W., Zavrel, J., van der Sloot, K. and van den Bosch, A. 1999. TiMBL: Tilburg Memory Based Learner [Technical Report ILK 99–01]. Tilburg: ILK. Dagan, I., Justeson, J., Lappin, S., Leass, H. and Ribak, A. 1995. Syntax and lexical statistics in anaphora resolution. Applied Artificial Intelligence 9(6):633–644. Ge, N., Hale, J. and Charniak, E. 1998. A statistical approach to anaphora resolution. In Proceedings of the Sixth Workshop on Very Large Corpora, Montréal, Canada, 161–170.
Corpus-based and machine learning approaches to anaphora resolution
Grosz, B J., Joshi, A.K. and Weinstein, S. 1995. Centering: A framework for modeling the local coherence of discourse. Computational Linguistics 21(2):203–225. Harabagiu, S.M., Bunescu, R.C. and Maiorano, S.J. 2001. Text and knowledge mining for coreference resolution. In Proceedings of the 2nd Conference of the North American Chapter of the Association for Computational Linguistics, Pittsburgh PA, 2–7 June, 2001, 55–62. Heim, I. 1982. The Semantics of Definite and Indefinite Noun Phrases. PhD dissertation, University of Massachusetts (Published by Graduate Linguistics Student Organisation). Hobbs, J.R. 1978. Resolving pronominal references. Lingua 44:311–338. Kamp, H. and Reyle, U. 1993. From Discourse to Logic. Introduction to modeltheoretic semantics of natural language, formal logic and discourse representation theory. Dordrecht: Kluwer. Kehler, A. 1997. Probabilistic coreference in information extraction. In Proceedings of the 2nd Conference on Empirical Methods in Natural Language Processing, Providence, RI, 1–2 August 1997, 163–173. Kehler, A., Appelt, D., Taylor, L. and Simma, A. 2004. The (non)utility of predicate-argument frequencies for pronoun interpretation. In Proceedings of the Human Language Technolgy Conference of the North American Chapter of the Association for Computational Linguistics, Boston, MA, 2–7 May, 2004, 289–296. Kennedy, C. and Boguraev, B. 1996. Anaphora for everyone: Pronominal anaphora resolution without a parser. In Proceedings of the 16th International Conference on Computational Linguistics, Copenhagen, Denmark, 5–9 August 1996, Vol. 1, 113–118. Lafferty, J., McCallum, A. and Pereira, F. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the 18th International Conference on Machine Learning, Williamstown MA, 28 June – 1 July 2001, 282–289. Lappin, S. and Leass, H.J. 1994. An algorithm for pronominal anaphora resolution. Computational Linguistics 20(4):535–561. Luo, X., Ittycheriah, A., Jing, H., Kambhatla, N. and Roukos, S. 2004. A mention-synchronous coreference resolution algorithm based on the Bell Tree. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics, Barcelona, Spain, 21–26 July 2004, 136–143. Marciniak, T. and Strube, M. 2005. Beyond the pipeline: Discrete optimization in NLP. In Proceedings of the 9th Conference on Computational Natural Language Learning, Ann Arbor, MI, USA, 29–30 June 2005, 136–145. McCallum, A. and Wellner, B. 2005. Conditional models of identity uncertainty with application to noun coreference. In Advances in Neural Information Processing Systems 17, L.K. Saul, Y. Weiss and L. Bottou (eds), 905–912. Cambridge MA: The MIT Press. McCarthy, J.F. and Lehnert, W.G. 1995. Using decision trees for coreference resolution. In Proceedings of the 14th International Joint Conference on Artificial Intelligence, Montréal, Canada, 1995, 1050–1055. Mitkov, R. 2002. Anaphora Resolution. London: Longman. Modjeska, N.M., Markert, K. and Nissim, M. 2003. Using the web in machine learning for otheranaphora resolution. In Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing, Sapporo, Japan, 11–12 July 2003, 176–183. Morton, T.S. 2000. Coreference for nlp applications. In Proceedings of the 38th Annual Meeting of the Association for Computational Linguistics, Hong Kong, 1–8 August 2000, 173–180. Müller, C. 2006. Automatic detection of nonreferential it in spoken multi-party dialog. In Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics, Trento, Italy, 3–7 April 2006, 49–56.
Michael Strube Müller, C., Rapp, S. and Strube, M. 2002. Applying co-training to reference resolution. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, Philadelphia, PA, 7–12 July 2002, 352–359. Nemhauser, G.L. and Wolsey, L.A. 1999. Integer and combinatorial optimization. New York NY: Wiley. Ng, V. and Cardie, C. 2002. Improving machine learning approaches to coreference resolution. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, Philadelphia, PA, 7–12 July 2002, 104–111. Ng, V. and Cardie, C. 2003. Weakly supervised natural language learning without redundant views. In Proceedings of the Human Language Technolgy Conference of the North American Chapter of the Association for Computational Linguistics, Edmonton, Alberta, Canada, 27 May -1 June, 2003, 173–180. Nissim, M. and Markert, K. 2003. Syntactic features and word similarity for supervised metonymy resolution. In Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, Sapporo, Japan, 7–12 July 2003, 56–63. Ponzetto, S.P. and Strube, M. 2006. Exploiting semantic role labeling, WordNet and Wikipedia for coreference resolution. In Proceedings of the Human Language Technolgy Conference of the North American Chapter of the Association for Computational Linguistics, New York NY, 4–9 June, 2006, 192–199. Quinlan, J. R. 1993. C4.5: Programs for Machine Learning. San Mateo CA: Morgan Kaufman. Sidner, C.L. 1983. Focusing in the comprehension of definite anaphora. In, Computational Models of Discourse, M. Brady and R.C. Berwick (eds), 267–330. Cambridge MA: The MIT Press. Soon, Wee Meng., Hwee Tou Ng and Chung Yong Lim, D. 2001. A machine learning approach to coreference resolution of noun phrases. Computational Linguistics 27(4): 521–544. Strube, M. and Müller, C. 2003. A machine learning approach to pronoun resolution in spoken dialogue. In Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, Sapporo, Japan, 7–12 July 2003, 168–175. Vilain, M., Burger, J., Aberdeen, J., Connolly, D. and Hirschman, L. 1995. A model-theoretic coreference scoring scheme. In Proceedings of the 6th Message Understanding Conference (MUC-6), 45–52. San Mateo CA: Morgan Kaufmann. Webber, B.L. 1979. A Formal Approach to Discourse Anaphora. New York NY: Garland. Witten, I.H. and Frank, E. 2005. Data Mining: Pratical machine learning tools and techniques (2nd ed.). San Francisco CA: Morgan Kaufmann. Yang, X., Su, J., Zhou, G. & Tan, C.L. 2004. Improving pronoun resolution by incorporating coreferential information of candidates. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics, Barcelona, Spain, 21–26 July 2004, 128–135. Yang, X., Su, J., Zhou, G. & Tan, C.L. 2003. Coreference resolution using competition learning approach. In Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, Sapporo, Japan, 7–12 July 2003, 176–183.
section 3
Neurolinguistic Studies
Neuroimaging studies of coherence processes Evelyn C. Ferstl and Florian Th. Siebörger Text comprehension involves linking the current sentence or utterance to the previous discourse context. This process of coherence building might be aided by linguistic markers, such as cohesive ties and direct anaphora. However, when the explicitly mentioned information does not provide unambiguous and consistent cues to coherence, as for example in the case of indirect anaphora, the comprehender draws on general background knowledge for inferring plausible links. In this chapter, we summarize neuroimaging studies that shed light on the interplay between cognitive and linguistic components of text comprehension. The results indicate that the anterior temporal lobes and the left fronto-medial cortex are particularly important for coherence building.
Introduction One of the most fundamental tasks in comprehending text or discourse is to connect the current utterance or sentence with the previous context so that a coherent representation results. This coherence building process often relies on syntactic and lexical cues that provide explicit information on the type of connection intended. Particularly conjunctions and adverbial connectives provide a vast tool box to signal specific relationships between events or to modify discourse referents. However, even in the absence of such cohesive ties (Halliday & Hasan, 1976), many utterances remain easily comprehensible, as the following example shows: 1. Jimmy was invited to Sally’s birthday party. The boy went to the mall and bought the Goblet of Fire.
The reader needs a number of inferences to render the incohesive text coherent. The anaphoric reference mapping the boy to the previously mentioned Jimmy establishes continuity with respect to the discourse referents, but also induces an elaboration of the text representation previously set up. In addition to the protagonists’ gender, the recipient now adds information about their age to the mental
Evelyn C. Ferstl and Florian Th. Siebörger
model of the text. The final or causal inference connecting the two events, however, is not directly provided by lexical items, but it has to be derived based on general world knowledge. Independent of the sentence’s cohesion is the use of two pieces of cultural knowledge. First, it is necessary to know that birthday guests usually bring presents to the party, and second, listeners who know that Goblet of Fire is a book title will set up a more accurate reading of the short story compared to those who interpret the phrase in its literal sense. Schwarz (2000) has termed this type of reference indirect anaphor. Direct anaphors are definite noun phrases that refer back to a previously mentioned discourse entity, the antecedent. In contrast, indirect anaphors lack an explicitly mentioned antecedent, although in many cases their reference is easily established. The preceding expression to which the noun phrase can be linked is called the anchor of the anaphor. In our example, there are no explicit antecedents in the discourse for the noun phrases the mall and the Goblet of Fire. The listener infers that the mall might be the only one in town or the one that Jimmy usually shops in. Furthermore, she infers that the book is going to be Jimmy’s present for Sally. In this case, the noun phrase fills a conceptual role in the schema of birthday parties (cf. Ferstl & Kintsch, 1999), and the resulting coherence relation between the two sentences is causal or final. The sentences tell a short story about the reason for Jimmy’s intentional, goal-directed action. The two-step procedure Schwarz (2000) proposes for resolving indirect anaphors includes the identification of the anchor, in this case birthday party, followed by the assignment of a relationship between anchor and anaphor, in this case the conceptual role present. Alternatively, the indirect anaphor might fill a semantic role, or an inference might be needed for establishing a novel relationship not previously stored in the reader’s general world knowledge. Schwarz (2000) suggests that these different types of relationships between anaphor and anchor are checked successively and that processing times are expected to increase from semantic, over conceptual to inferred relationships (but see Hagoort, Hald, Bastiaansen, & Petersson, 2004 for evidence of similarity between semantic and general world knowledge). In psycholinguistic research, the study of inference processes during text comprehension has a long tradition and there is a vast empirical literature delineating the conditions facilitating anaphoric and pronominal reference specifically, and coherence building more generally (for reviews see relevant chapters in Gernsbacher, 1994). The discussion in the psychological literature has centered on the automaticity of these processes and the time course of knowledge use. There has been an ongoing debate about which types of inferences are drawn mandatorily, on-line, and which are optional, depending on instructions or reading goals. The minimalist hypothesis, or memory-based hypothesis claimed that only those inferences are online which are needed for bridging successive sentences, or those which are based on easily accessible knowledge (McKoon & Ratcliff, 1992). The constructionist, or ex-
Neuroimaging studies of coherence processes
planation-based theories, in contrast, claim that those inferences are fundamental which set up causal structures of texts, or which are needed for understanding the goal structure of texts (Graesser, Singer & Trabasso, 1994; van den Broek, 1994). From this more psychological point of view, there is less of a focus on the exact linguistic wording, but more on the psychological processes needed for integrating the current utterance with the previous discourse and the comprehender’s general world knowledge. In the meantime, most researchers agree that both types of inference contribute to text comprehension. The issue of interest is to describe how the text properties interact with the comprehender’s knowledge base to elicit different types of coherence building processes (Guéraud & O’Brien, 2005). In addition to refined traditional psycholinguistic methods, such as reading time measurements, priming and eye movement recording, neuroscientific methods have now become available. Evoked brain potentials (ERPs), positron emission tomography (PET) or functional magnetic resonance imgaging (fMRI), yield multidimensional information beyond a difficulty continuum. In addition to latency information, all of these methods include spatial distributions enabling qualitative dissociations between cognitive processes. For instance, syntactic and semantic processes can be separated using ERPs by their latency, polarity and scalp topography (Friederici, 2002), and fMRI shows that they engage different brain regions within left inferior prefrontal cortex (Bookheimer, 2002). Text comprehension researchers have only recently begun to take advantage of these methods (Ferstl, 2007; Mar, 2004; van Berkum, 2004). In this chapter, we will summarize two fMRI studies on inference processes and coherence building. The results of these studies help in identifying likely candidates for brain regions involved during the processing of direct and indirect anaphors. Before going on to the empirical studies, we will summarize the most important neuropsychological theories of inferencing.
1. The neuropsychology of inference processes In the neuropsychological literature, there is rather widespread agreement that comprehension processes on the text level are mediated by right hemispheric (RH) brain regions. Patients with right sided brain lesions, mostly after stroke, exhibit deficits in language comprehension, specifically with the activation of semantic associations (Beeman, 1998), the appreciation of non-literal or pragmatic interpretations, the revision of erroneous interpretations and the inference of causal relationships between sentences (Brownell & Martino, 1998). Interestingly, the description of language deficits after frontal lobe lesions or after closed-head injury is very similar (McDonald, 1993; Nicholas & Brookshire, 1995; Novoa & Ardila, 1987). Higher level language deficits concerning the text and discourse level rather
Evelyn C. Ferstl and Florian Th. Siebörger
than the level of phonology, syntax and semantics, have been termed RH syndrome (Brownell & Martino, 1998), cognitive communication impairment or more general, non-aphasic language disorder (Prigatano et al., 1986; Ferstl, Guthke & von Cramon, 2002).
Figure 1. Anatomical regions superimposed on a graphic rendition of a brain (cf. Talairach & Tournoux, 1988). The numbers refer to the zytoarchitectonic Brodmann areas, and the shading indicates regions of interest (Ferstl, 2007). The upper panel shows a left lateral view of the brain. The abbreviations are: IFG=inferior frontal gyrus, alPFC=antero-lateral prefrontal cortex, plPFC=posterior lateral prefrontal cortex, aTL=anterior temporal lobe, mTL=middle temporal lobe, pTL=posterior temporal lobe, IPL=inferior parietal lobe. Broca’s area lies within the IFG (BA44), Wernicke’s in pTL. The regions important for language processing in context are the aTL, the IPL, and the plPFC. The lower panel shows a medial view of the right hemisphere. The abbreviations are: dmPFC=dorso-medial prefrontal cortex, vmPFC=ventro-medial prefrontal cortex, ACC=anterior cingulate cortex, PCC=posterior cingulate cortex, and Prec=precuneus. Here, the most important regions are the dmPFC and the PCC/Prec.
Neuroimaging studies of coherence processes
Neuroimaging studies have also confirmed the importance of frontal and right sided brain regions for text comprehension in general and for inference processes in particular (Bookheimer, 2002; Ferstl, 2007; Gernsbacher & Kaschak, 2003; Mar, 2004). For those readers unfamiliar with the anatomical terminology, Figure 1 presents a schematic rendition of the brain including the regions important for the presented studies. Mason and Just (2004) studied the functional neuroanatomy of coherence building using sentence pairs with three different levels of relatedness. The sentence examples for the three conditions are taken from Keenan, Baillet and Brown (1984). 2. Joey’s brother punched him again and again. The next day, his body was covered with bruises. 3. Joey’s crazy mother became furiously angry with him. The next day, his body was covered with bruises. 4. Joey went to a neighbor’s house to play. The next day, his body was covered with bruises.
The lack of a definite noun phrase in the target sentence, precludes an analysis of these sentences in terms of indirect anaphors. Note, however, that a slight rewording of the second sentence in, for instance:
5. The next day, the bruises on his body hurt terribly.
introduces an indirect anaphor with the definite noun phrase “the bruises”. The reworded version contains similar information as the original one and is thus likely to elicit qualitatively equivalent inference processes. Thus, in terms of a cognitive process model, there is no need to restrict the definition of inferencing to the resolution of direct or indirect nominal anaphors. To better understand the conceptual inference requirements, we can analyse the properties of the sentences. First, all of the examples are cohesive. They are connected by a temporal marker (the next day), and the pronoun his directly refers back to Joey. With respect to the complexity of the inference, however, there are differences. In 2), the causal connection is not quite explicit, but it requires only one bridging step: hitting causes bruises. In contrast, two propositions (cf. Kintsch, 1998) or content units are needed in 3): angry people hit, hitting causes bruises. Fi1. As M. Consten noted, there is an additional pronoun reference ambiguity in examples 2 and 4. It is not clear whether «his body» refers to Joey, the brother or the neighbor. Only in example 3, the gender information disambiguates the sentence. Because many sentences were used in the experiment, the results are not likely to be influenced by this spurious linguistic feature. However, it nicely illustrates how careful psycholinguists need to be in constructing appropriate experimental materials.
Evelyn C. Ferstl and Florian Th. Siebörger
nally, in 4) there is either one more proposition needed, for example, the friend got angry, or the situation needs to be elaborated on using a different scenario to fill the gap (e.g., Joey might have climbed a tree and fallen off; he might have had a bicycle accident on his way to his friend’s house, a.s.o.). In their fMRI study Mason and Just (2004) found that activation in the dorsolateral prefrontal cortex increased bilaterally with the distance of the inference, i.e., with the difficulty of bridging two sentences. Activation in the right temporal lobe was related to the ease of later retrieval of the inferred information. The sentences used in this study were equal with respect to the lexical cohesion. However, cohesion has a clear impact on building a coherent text representation (Halliday & Hasan, 1976; Münte, Schiltz & Kutas, 1998). Definite articles, for example, are a cue to whether the information is new or has already been implied by the previous discourse (cf. Schwarz, 2000). A series of unrelated sentences is more likely to be combined into one scenario when the noun phrases all contain definite articles. Compare the examples 6) and 7):
6. The old woman sat on the bench. The dog chased the cat.
7. An old woman sat on a bench. A dog chased a cat.
Example 6) immediately elicits a situation model in which it’s the woman’s dog chasing the cat, or at least, a scenario in which the two events happen in close spatial proximity. Although the materials seem quite artificial, any resulting coherent representation must be based on the resolution of indirect anaphors. This is not the case in 7). Due to the lack of cohesive markers, readers are less likely to search for coherence, although the lexical information transporting the establishment of content based links is comparable to that in 6). Both examples use the same content words and thus elicit the same associations in general world knowledge. In an fMRI study using materials like these, Robertson, Gernsbacher et al. (2000) found more right frontal activation for the sentences with definite articles as compared to indefinite articles. In line with the RH hypothesis, the authors argued that successful coherence building requires the contribution of the right hemisphere. Although these two imaging studies on coherence building both confirmed the role of frontal and right hemisphere brain regions for inference processes, the results converged only partially. In the right prefrontal cortex, Mason and Just (2004) found less activation for more closely related sentences, as defined by the pragmatic distance, whereas Robertson et al. (2000) found more activation for more closely related sentences, as defined by the lexical cohesion.
Neuroimaging studies of coherence processes
2. Cohesion and coherence In our own experiment, we were interested in the interaction between the two factors cohesion and coherence (Ferstl & von Cramon, 2001). We constructed the following four conditions: 8. coherent / cohesive Sometimes a truck drives by the house. That‘s when the dishes start to rattle. 9. coherent / incohesive Sometimes a truck drives by the house. The dishes start to rattle. 10. incoherent / cohesive The lights have been on since last night. That‘s when the dishes start to rattle. 11. incoherent / incohesive The lights have been on since last night. The dishes start to rattle.
Word and sentence level features were controlled by using each target sentence in both a coherent and and incoherent context. The participants were simply asked to indicate after reading the second sentence, whether the sentence pair was conceptually coherent or not. The behavioral data confirmed that coherent and incoherent trials were easily distinguishable, and that cohesion facilitated the comprehension of coherent trials (Ferstl, 2007; Ferstl & von Cramon, 2001). When a cohesive tie falsely indicated a connection that was in fact absent (as in example 10), errors and reading times increased. In addition, a recent ERP experiment using a similar design once more confirmed this interaction between cohesion and coherence. The amplitude of the N400, a component indicating the ease of semantic integration, was largest in the incoherent, cohesive condition (Brehm & Saddy, 2005). In the comparison of sentence comprehension to a perceptual control condition, the fMRI data replicated a pattern described numerous times for language processing in context (e.g., Mazoyer et al., 1993; Xu et al., 2005; for reviews see Ferstl, 2007; Mar, 2004). Specifically, there was extended left-lateral activation, including the inferior frontal gyrus and regions along the entire length of the superior temporal sulcus. On the right side, homolog but smaller areas in the anterior temporal lobe and the dorsal part of the inferior frontal gyrus were active. In addition, there was left medial activation in the dorso-medial prefrontal cortex (dmPFC) and the posterior cingulate cortex (PCC/prec).
Evelyn C. Ferstl and Florian Th. Siebörger
The interaction between cohesion and coherence was a significant in left lateral PFC (BA 44/6). The activation was due to an increase in the most difficult condition 10), when a cohesive tie falsely suggested a connection. One explanation, consistent with the spatial overlap with Broca’s area (BA 44, inferior frontal gyrus), would attribute this activation to increased demands on syntactic or phonological processes (cf. Bookheimer, 2002; Stowe, Haverkort & Zwarts, 2005). However, the region close to the junction of the inferior frontal and precentral sulci (IFJ) has been shown to be important for fast and flexible behavior in socalled executive function tasks requiring inhibition, monitoring or integration of stimuli (Brass, Derrfuss, Forstmann & von Cramon, 2005). Thus, a more parsimonious account is that the IFJ contributed to the inhibition of the cohesive tie in favor of the task relevant pragmatic information. Interestingly, the arguably more difficult incoherent sentence pairs did not elicit additional activation as compared to coherent sentence pairs. In the reverse contrast, the two medial regions were the only ones sensitive to coherence. When a coherent sentence pair was read, activation in the dmPFC and the PCC regions was more extensive than when an incoherent sentence pair was read. This finding was subsequently replicated using auditory presentation and slightly different task instructions (Ferstl & von Cramon, 2002). Based on the functional neuroanatomy of these two regions, the role of the dmPFC was seen to reflect non-automatic inference processes requiring a knowledge-based integration of the current utterance with the prior context. To further investigate the issue of whether the medial activations indeed reflect goal-directed, intentional coherence building, a further experiment was conducted (Siebörger, Ferstl & von Cramon, 2003). The goal was to induce more difficult, truly non-automatic inference processes. To this end, the same sentence pairs were used, but now 75% were incoherent. The participants were instructed to search for a pragmatic connection for all sentence pairs, and to rate on a 4-point scale how easy or difficult the inference process was. Pretests had confirmed that it was easily possible to invent a plausible bridging inference for many of the incoherent trials, and that the ratings and reaction times reflected the length and complexity of the inference chain. The fMRI results confirmed that the dmPFC was now active in all conditions, not only during the processing of closely related, coherent trials as in the previous experiment. The peak latency of the hemodynamic response was correlated with the response times, indicating a lasting contribution of the dmPFC up until the bridging inference was found and the rating decision was made. The results of these experiments confirm the sensitivity of fMRI to subtle linguistic variations as provided by cohesive ties. More importantly, they established the dmPFC as a crucial region for coherence building processes, and in particular for non-automatic, non-associative inferencing.
Neuroimaging studies of coherence processes
Due to methodological differences a comparison with the aforementioned studies is difficult. Besides different materials and tasks, Mason and Just (2004) used predefined regions of interest, neglecting the medial areas. Robertson et al. (2000) used short paragraphs in which cohesion facilitated or hindered building a coherent representation, whereas in our studies, the sentences elicited different textworlds or situation models. The study presented in the next section is concerned with situation model updating in the context of coherent stories.
3. The situation model In most text comprehension theories, the global representation of the text information in its entirety is conceptualized as a mental model (Johnson-Laird, 1983) integrating the current information with the comprehender’s background knowledge. This mental model has been termed discourse model (Altmann & Steedman, 1988), textworld model (de Beaugrande & Dressler, 1981; Schwarz, 2000), or, most prominently, situation model (Ferstl, 2001; Ferstl & Kintsch, 1999; Kintsch, 1998; van Dijk & Kintsch, 1983; Zwaan & Radvansky, 1998). There is a two-way relationship between inferences and situation model building. Inferences contribute to establishing a particular situation model representation, and an elaborate, cumulative situation model greatly facilitates appropriate inferences later on during text comprehension. Thus, there is no strict separation between these two types of processes but rather a tight interaction. A number of neuroimaging experiments targeted situation model building (Ferstl, 2007). First, all studies comparing connected text to a baseline of perceptually matched stimuli, word lists, or unrelated sentences include situation model building as a subprocess. These studies consistently show a left-dominant bilateral network of brain regions involved during language comprehension in context, very similar to the one described in the previous section (Ferstl & von Cramon, 2001; 2002). We have termed this network the extended language network (Ferstl, 2007). In more specific experiments, Maguire, Morris & Frith (1999) and St. George, Kutas, Martinez and Sereno (1999) isolated situation model building by comparing the comprehension of the same texts with and without an appropriate situation model. When a picture or a title provided the necessary background knowledge, short loosely connected texts became comprehensible. Despite the smiliarities of the designs, the results once more yielded inconsistent results. While St. George et al. (1999) found right temporal regions to be sensitive to global coherence, Maguire et al. (1999) found left anterior temporal, fronto-medial and parieto-medial regions to be more active when situation model building was successful. In both studies, however, the baseline conditions consisted of incomprehensible, loosely
Evelyn C. Ferstl and Florian Th. Siebörger
structured texts. In a study conducted in our laboratory, we focussed on even more specific comparisons. For narrative texts, it has been shown that several story aspects are monitored on-line during situation model building. Among others, these include the goals, emotions and identity of the protagonists, and the time and location of events (Zwaan, Magliano & Graesser, 1995; Rinck, Hähnel & Becker, 2002). The situation model may contain non-verbal elements represented in a domain specific, appropriate format. For instance, a description of a map might elicit an analog, spatial representation, or a newspaper article about a soccer game might induce a representation of the excitement felt when watching a game. A well-established paradigm for studying situation model updating is the inconsistency paradigm which we adopted for a recent fMRI study (Ferstl, Rinck & von Cramon, 2005). Stories were used that contained globally inconsistent information, while being locally coherent. There were two story types. Emotion stories mentioned the emotional state of one of the protagonists, and chronological stories contained information about the timing or the sequence of events. Examples are given in Table 1. Table 1. Examples for the stories used to study situation model updating (Ferstl, Rinck & von Cramon, 2005). The words in bold face are the crucial target information, the words in italics provide the inconsistent information. Chronological Today, Markus and Claudia would finally meet again. Markus’ train arrived at the station 20 minutes after / before Claudia’s train. Markus was very excited when his train stopped at the station on time. He tried to think of what he should say when he met her. Many people were crowding on the platform. Claudia was already waiting for him when he got off the train with his huge bag. They were both very happy.
Emotional The semester was finally over and Sarah wanted to celebrate. A lot of her friends had shown up for her end-of-school party. It was one of these parties with everything being just perfect. Sarah’s best friend gave her a hug and told her how much fun she was having. Sarah couldn’t remember that she had ever been so happy / sad before. She put her favorite record into the CD-player and started dancing by herself.
The stories were presented over headphones while the participants were scanned using fMRI. After listening to each story they were asked to indicate whether there was a content based error. The analysis of the BOLD response immediately after hearing the target words (in Table 1 marked in bold face) showed activation in the right anterior temporal lobe for inconsistent information, independent of its type. Emotional target words elicited ventro-medial prefrontal activation consistent with findings from other studies on affective processing. In contrast, the chronological target information was accompanied by activation increases in a fronto-
Neuroimaging studies of coherence processes
parietal network related to attention and working memory processes, but also including a region that plays a role for the processing of numbers, quantity and sequence (Walsh, 2003). Analysing the activation based on a more extended window, the integration into the unfolding situation model representation was studied. Using short epochs from the target information until the end of the story, possible inference processes were studied. For the chronological stories, fronto-parietal regions once more indicated increases in memory and executive processes. For these stories, the decision on the consistency was all-or-none. Either the sequence of events matched, or it did not. It is impossible for Claudia to await Markus at the platform if she arrives later than him. Thus, processing the inconsistent information required reaccess of the relevant discourse information provided in the second sentence and a comparison with the information given in the target sentence. In contrast, integration of the emotional inconsistencies engaged the dorsomedial prefrontal cortex, i.e., the region shown to be important for inference processes in previous studies (Ferstl & von Cramon, 2001; 2002; Siebörger, et al., 2003). Inspection of the time course of the activation confirmed that only in the inconsistent emotional stories, but not in the other conditions, this brain region remained active until the end of the story. Similar to the goal-directed inferences elicited in the study described previously (Siebörger et al., 2003), a deliberate inference process might have been used for integrating the seemingly inconsistent target information into a coherent situation model. Taken together, these results clearly make two important contributions to the understanding of text comprehension. First, they show that affective and cognitive aspects of comprehension, or hot and cold aspects (cf. Goel & Dolan, 2004), can be separated using fMRI. And second, they confirm the role of the dmPFC for inference processes, but also restrict its function to specific inference types. When inconsistent temporal information had to be integrated based on discourse information only, executive functions were needed. When inconsistent emotional information had to be integrated based on general world knowledge, the inference area in the dmPFC was recruited.
4. Conclusions In this chapter we summarized recent findings on the functional neuroanatomy of coherence building and inferencing. Language processing in context was shown to elicit bilateral, but left-dominant activation in a fronto-temporal network including but not restricted to the perisylvian language areas (cf. Ferstl, Rinck & von
Evelyn C. Ferstl and Florian Th. Siebörger
Cramon, 2005; Ferstl & von Cramon, 2001, 2002; Mazoyer et al., 1993; Siebörger, et al., 2003; Xu et al., 2005). By inspecting more specific contrasts, fMRI provides the means to distuinguish qualitatively different processes. In contrast to the expectation of the right hemisphere playing a particular role during inference processes (Beeman, 1998; Mason & Just, 2004; Robertson et al., 2000; St. George et al., 2000), we found leftmedial brain regions to be consistently involved in coherence building. The dmPFC region engaged during inference processes was interpreted to be indicative of knowledge-based non-automatic inferences. Independent of the particular functional attribution, over which there has been considerable debate, the dmPFC is likely to be engaged during tasks with the following properties: there is no predefined, correct answer, individuals may establish their own response criteria, and the task requires the integration of current external stimuli with general world knowledge. Thus, the dmPFC also appears during a wide variety of tasks in which participants are challenged to “make sense” of the world: during ethical or moral decision making, evaluative judgments, inductive reasoning, mentalizing tasks, certain types of semantic categorization, and even during rest, when thoughts drift freely (Ferstl & von Cramon, 2002; Northoff & Bermpohl, 2004). The parieto-medial region (PCC/prec) is often co-activated with the dmPFC. Activation in this region has been shown to be modulated by the knowledge requirements of evaluative language tasks (Zysset et al., 2002; 2003). The second brain region specifically engaged during language comprehension in context are the anterior temporal lobes (aTL). The aTL is active when comprehensible language is interpreted, and it has been shown to be specific for language (Humphrie, Willard, Buchsbaum, & Hickok, 2001). The right aTL was activated in the study on situation model building immediately upon encountering inconsistent words. This result might be due to increased demands for the propositionalization, or the on-line integration of the information into the text representation. In contrast, the dmPFC implements knowledge-based inferences required for setting up a coherent, common scenario of the resulting propositions. Preliminary evidence for such an interpretation was recently provided by a study on syntactic processing (Wartenburger, Heekeren, Burchert, et al., 2004). Using a grammaticality judgment task, correct sentences elicited activation in the fronto- and parieto-medial regions as described here. In contrast, the incorrect sentences, in which the syntactic error rendered propositionalization more difficult, or even impossible, elicited more activation in the inferior frontal gyrus and the anterior temporal lobe. With these functional interpretations in mind, we can now return to the issue of anaphors in text. Although this type of linguistic reference has not yet been studied using neuroimaging (but see Marx et al. this volume for an ERP study), we can formulate a tentative prediction for the process sequence in indirect anaphor
Neuroimaging studies of coherence processes
resolution, as sketched by Schwarz (2000). Direct anaphors require a lexico-semantic process for linking the noun phrase to its antecedent, an anchoring process located at the text base level (Kintsch, 1998). These processes are likely to be realized in left fronto-temporal brain regions, in particular in the left anterior temporal lobe and the inferior frontal gyrus (Bookheimer, 2002; Thompson-Schill, D’Esposito & Kan, 1999). Those indirect anaphors for which a semantic or conceptual role provides the relationship of the noun phrase to the referring anchor require more extensive integration with general world knowledge – a process reflected in the engagement of the PCC/prec. And finally, the need for additional inferences is likely to be indicated by an increase of dmPFC activation. These predictions are based on a rather small number of relevant neuroimaging studies, and subtle linguistic variations have not yet been included in experiments on text comprehension. Thus, more elaborate investigations of inference processes are needed. This endeavor will greatly benefit from a close interdisciplinary interaction between linguistics, experimental psychology and neuroscience. Detailed linguistic analyses, embedded in well-established psychological process models, and paired with the available methodological sophistication are desirable. Based on a larger empirical database, neuroscientific methods will then further our understanding of the interplay between linguistic features of text and the knowledge based comprehension processes needed for inferencing in general and anaphor resolution in particular.
References Altmann, G.T.M. and Steedman, M. 1988. Interaction with context during human sentence processing. Cognition 30: 191–238. Beeman, M. 1998. Coarse semantic coding and discourse comprehension. In Right Hemisphere Language Comprehension: Perspectives from cognitive neuroscience, M. Beeman and C. Chiarello (eds), 255–284. Mahwah NJ: Lawrence Erlbaum. Bookheimer, S. 2002. Functional MRI of language: New approaches to understanding the cortical organization of semantic processing. Annual Review of Neuroscience 25: 151–188. Brass, M., Derrfuss, J., Forstmann, B. and von Cramon, D.Y. 2005. The role of the inferior frontal junction area in cognitive control. Trends in Cognitive Sciences 9: 314–316. Brehm, E. and Saddy, D. 2005. On cohesion and coherence in text processing: Two ERP-studies. Journal of Cognitive Neuroscience Suppl. 77. Brownell, H.H. and Martino, G. 1998. Deficits in inference and social cognition: The effects of right hemisphere brain damage on discourse. In Right Hemisphere Language Comprehension: Perspectives from cognitive neuroscience, M. Beeman and C. Chiarello (eds), 309–328. Mahwah NJ: Lawrence Erlbaum. de Beaugrande, R.-A. and Dressler, W.U. 1981. Einführung in die Textlinguistik. Tübingen: Niemeyer.
Evelyn C. Ferstl and Florian Th. Siebörger Ferstl, E.C. 2007. The functional neuroanatomy of text comprehension: What’s the story so far? In Higher Level Language Processes in the Brain: Inference and comprehension processes, F. Schmalhofer and C. Perfetti (eds). Mahwah NJ: Lawrence Erlbaum. Ferstl, E.C. 2001. Learning from text. In International Encyclopedia of the Social & Behavioral Sciences, N.J. Smelser and P.B. Baltes (eds),Vol. 3.13: Cognitive Psychology and Cognitive Science (Ed. W. Kintsch). Amsterdam: Elsevier. Ferstl, E.C., Guthke, T. and von Cramon, D.Y. 2002. Text comprehension after brain injury: Left prefrontal lesions affect inference processes. Neuropsychology 16: 292–308. Ferstl, E.C. and Kintsch, W. 1999. Learning from text: Structural knowledge assessment in the study of discourse comprehension. In The Construction of Mental Models during Reading, H. Oostendorp and. S. Goldman (eds), 247 – 277. Mahwah NJ: Lawrence Erlbaum. Ferstl, E.C., Rinck, M. and von Cramon, D.Y. 2005. Emotional and temporal aspects of situation model processing during text comprehension: An event-related fMRI study. Journal of Cognitive Neuroscience 17(5): 724–739 Ferstl, E.C. and von Cramon, D.Y. 2001. The role of coherence and cohesion in text comprehension: An event-related fMRI study. Cognitive Brain Research 11: 325–340. Ferstl, E.C. and von Cramon, D.Y. 2002. What does the fronto-medial cortex contribute to language processing: Coherence or theory of mind? NeuroImage 17: 1599–1612. Friederici, A.D. 2002. Towards a neural basis of auditory sentence processing. Trends in Cognitive Sciences 6: 78–84. Gernsbacher, M.A. (ed.). 1994. Handbook of Psycholinguistics. San Diego CA: Academic Press. Gernsbacher, M.A. and Kaschak, M.P. 2003. Neuroimaging studies of language production and comprehension. Annual Review of Psychology 54: 16.1–16.24. Goel, V. and Dolan, R.J. 2004. Reciprocal neural response wihtin lateral and ventral medial prefrontal cortex during hot and cold reasoning. NeuroImage 20: 2314–2321. Graesser, A.C., Singer, M. and Trabasso, T. 1994. Constructing inferences during narrative text comprehension. Psychological Review 101: 371–395. Guéraud, S. and O’Brien, E. 2005. Components of comprehension: A convergence between memory-based processes and explanation-based processes [Special Issue: Discourse Processes]. Mahwah NJ: Lawrence Erlbaum. Gusnard, D.A., Akbudak, E., Shulman, G.L. and Raichle, M.E. 2001. Medial prefrontal cortex and self-referential activity: Relation to a default mode of brain function. In Proceedings of the National Academy of Sciences, 2001. Hagoort, P., Hald, L., Bastiaansen, M. and Petersson, P.M. 2004. Integration of word meaning and world knowledge in language comprehension. Science 304: 438–441. Halliday, M.A.K. and Hasan, R. 1976. Cohesion in English. London: Longman. Humphries, C., Willard, K., Buchsbaum, B. and Hickok, G. 2001. Role of anterior temporal cortex in auditory sentence comprehension: An fMRI study. Neuroreport 12: 1749–1752. Johnson-Laird, P.N. 1983. Mental Models: Towards a cognitive science of language, inference and consicousness. Cambridge MA: Harvard University Press. Keenan, J.M, Baillet, S.D. and Brown, P. 1984. The effects of causal cohesion on comprehension and memory. Journal of Verbal Learning and Verbal Behavior 23: 115–126. Kintsch, W. 1998. Comprehension: A paradigm for cognition. Cambridge: CUP. Maguire, E. A., Frith, C. D. and Morris, R.G.M. 1999. The functional neuroanatomy of comprehension and memory: The importance of prior knowledge. Brain 122: 1839–1850. Mar, R. A. 2004. The neuropsychology of narrative: story comprehension, story production and their interrelation. Neuropsychologia 42: 1414–1434.
Neuroimaging studies of coherence processes
Mason, R.A. and Just, M.A. 2004. How the brain processes causal inferences in text: A theoretical account of generation and integration component processes utilizing both cerebral hemispheres. Psychological Science 15: 1–7. Mazoyer, B.M., Tzourio, N., Frak, V., Syrota, A., Murayama, N., Levrier, O., Salamon, G., Dehaene, S., Cohen, L. and Mehler, J. 1993. The cortical representation of speech. Journal of Cognitive Neuroscience 5: 467–479. McDonald, S. 1993. Viewing the brain sideways? Frontal versus right hemisphere explanation of non-aphasic language disorders. Aphasiology 7: 535–549. McKoon, G. and Ratcliff, R. 1992. Inference during reading. Psychological Review 99: 440–466. Münte, T.F., Schiltz, K. and Kutas, M. 1998. When temporal terms belie conceptual order. Nature 395: 71–73. Nicholas, L.E. and Brookshire, R.H. 1995. Comprehension of spoken narrative discourse by adults with aphasia, right-hemisphere brain damage, or traumatic brain injury. American Journal of Speech-Language Pathology 4: 69–81. Northoff, G. and Bermpohl, F. 2004. Cortical midline structures and the self. Trends in Cognitive Sciences 8: 102–107. Novoa, O.P. and Ardila, A. 1987. Linguistic abilities in patients with pre-frontal damage. Brain & Language 30: 206–225. Prigatano, G.P., Roueche, J.R., Fordyce, D.J. 1986. Nonaphasic language disturbances after brain injury. In Neuropsychological Rehabilitation After Brain Injury, G.P. Prigatano, D. J. Fordyce, H.K. Zeiner, J.R. Roueche, M. Pepping, B.C. Wood (eds.), 18–28. Baltimore MD: John Hopkins University Press. Rinck, M., Hähnel, A. and Becker, G. 2001. Using temporal information to construct, update, and retrieve situation models of narratives. Journal of Experimental Psychology: Learning, Memory, and Cognition 27: 67–80. Robertson, D.A., Gernsbacher, M.A., Guidotti, S.J., Robertson, R.R.W., Irwin, W., Mock, B.J. and Campana, M.E. 2000. Functional neuroanatomy of the cognitive process of mapping during discourse comprehension. Psychological Science 11 : 255–260. Schwarz, M. 2000. Indirekte Anaphern in Texten: Studien zur domänengebundenen Referenz und Kohärenz im Deutschen. Tübingen: Niemeyer. Siebörger, F.T., Ferstl, E.C. and von Cramon, D.Y. 2003. Making sense of nonsense: An fMRI study of task induced coherence processes. Journal of the International Neuropsychological Society 9: 506. St. George, M., Kutas, M., Martinez, A. and Sereno, M.I. 1999. Semantic integration in reading: Engagement of the right hemisphere during discourse processing. Brain, 122: 1317–1325. Stowe, L.A., Haverkort, M. and Zwarts, F. 2005. Rethinking the neurological basis of language. Lingua 115 : 997–1042 Talairach, J. and Tournoux, P. 1988. Coplanar Stereotaxic Atlas of the Human Brain. New York NY: Thieme. Thompson-Schill, S.L., D’Esposito, M. and Kan, I.P. 1999. Effects of repetition and competition on activity in left prefrontal cortex during word generation. Neuron 23: 513–522. van Berkum, J.J.A. 2004. Sentence comprehension in a wider discourse: Can we use ERPs to keep track of things? In The On-line Study of Sentence Comprehension: Eyetracking, ERPs and beyond, M. Carreiras and C. Clifton Jr. (eds), 229–270. New York NY: Psychology Press. van den Broek, P. 1994. Comprehension and memory of narrative texts: Inferences and coherence. In Handbook of Psycholinguistics, M.A. Gernsbacher (ed.), 536–588. San Diego CA: Academic Press.
Evelyn C. Ferstl and Florian Th. Siebörger van Dijk, T.A. and Kintsch, W. 1983. Strategies of Discourse Comprehension. New York NY: Academic Press. Walsh, V. 2003. A theory of magnitude: common cortical metrics of time, space and quantity. Trends in Cognitive Sciences 7: 483–488. Wartenburger, I., Heekeren, H. R., Burchert, F., Heinemann, S., De Bleser, R. and Villringer, A. 2004. Neural correlates of syntactic transformations. Human Brain Mapping 22: 72–81 Xu, J., Kemeny, S., Park, G., Frattali, C. and Braun, A. 2005. Language in context: emergent features of word, sentence, and narrative comprehension. Neuroimage 25 : 1002–1015. Zalla, T., Phipps, M. and Grafman, J. 2002. Story processing in patients with damage to the prefrontal cortex. Cortex 38: 215–231. Zwaan, R. A., Magliano, J. P. and Graesser, A. C. 1995. Dimensions of situation model construction in narrative comprehension. Journal of Experimental Psychology: Learning, Memory, and Cognition 21: 386–397. Zwaan, R. A. and Radvansky, G.A. 1998. Situation models in language comprehension and memory. Psychological Bulletin 123: 162–185. Zysset, S., Huber, O., Ferstl, E.C. and von Cramon, D.Y. 2002. The anterior dmPFC and evaluative judgment: An fMRI study. NeuroImage 15: 983–991. Zysset, S., Huber, O., Samson, A., Ferstl, E.C. and von Cramon, D.Y. 2003. Functional specialization wihtin the anterior medial PFC: A functional magnetic resonance imaging study with human subjects. Neuroscience Letters 335: 183–186.
Reference assignment in the absence of sufficient semantic content* Petra Burkhardt This paper investigates the establishment of referential dependencies during the interpretation of pronominal entities and presents evidence from eventrelated brain potentials (ERPs) for discrete underlying processes. Regardless of their differing internal structures, reflexives and pronouns are semantically underspecified and must enter into a dependency relation with a proper antecedent to receive referential and meaningful interpretation. This paper addresses the nature of three kinds of anaphoric entities: ordinary reflexives, ordinary pronouns, and logophors – i.e. reflexive entities whose distribution patterns with pronouns. It seeks to further substantiate the claim that logophors – like pronouns – establish dependency at the level of discourse representation and must be functionally dissociated from ordinary reflexives. This paper follows an approach of the establishment of reference which views syntactic information as a prerequisite to interpretation, and more specifically the notion of coargumenthood, whose presence or absence determines the level of representation where interpretation is carried out (i.e. syntax or discourse). Accordingly, after presenting theoretical considerations and processing background, the paper examines one particular ERP component that has recently been identified as a marker of pronoun-antecedent dependency – the N400, a negative deflection peaking around 400ms – and should only surface during the establishment of discourse-based dependency. The current research therefore focuses on the nature of reference assignment during pronominal interpretation and attends to the observation that logophors pattern with pronouns (which are uncontroversially resolved in discourse) but not with ordinary reflexives.
* The experiment presented in this article was carried out at Yale University and supported through a National Science Foundation Doctoral Dissertation Improvement Grant (BCS-0236736). I would like to thank the audience of the 2005 DGfS Workshop Anaphorik in Texten as well as the editors of this volume for useful comments and discussion of the ideas presented here.
Petra Burkhardt
1. Logophors, ordinary reflexives and pronouns The present research is concerned with three different kinds of referentially dependent elements (i.e. elements that depend on an antecedent to receive referential interpretation), which are illustrated in the following (dependency relations are italicized). 1. Lisa painted herself / *her.
2. Tom realized that Natalie painted him / *himself.
3. Sam claimed that Kristen had invited Tracy and himself / him.
Many traditional theories – most prominently Binding Theory – distinguish between reflexives (and reciprocals), which are bound in a certain domain (1), and pronouns, which are free (2) (Chomsky 1981, 1986; Lasnik 1989; inter alia). In addition, complementary distribution of reflexives and pronouns has been derived from these principles, which is empirically supported by (1) and (2), but disconfirmed by (3). This latter observation that reflexives and pronouns can appear in one and the same construction and that reflexives do not necessarily need to be bound in a minimal domain – as further evidenced by (4) – called for the development of alternative accounts.
4. There were five linguists in the room apart from myself.
One particular line of research in response to this challenge explains the distribution of referentially dependent elements on the basis of the level of representation at which the dependency is established (cf. Huang 1991; Pollard & Sag 1992; Reinhart & Reuland 1993; Avrutin 1999; Levinson 2000; Reuland 2001). Two levels of representation play a major role in these frameworks, namely syntax, which encodes phrase-structural relations, and pragmatics or discourse, which keeps track of discourse referents, event structures, and their relations. Following this line of research, the reflexive entities in (1) and (3) can be distinguished from each other, which has reflexes in sentence processing (cf. e.g. Avrutin 1999; Reuland 2001, 2003; Burkhardt 2005). In the following, I briefly summarize core aspects of Reinhart and Reuland’s (1993) Reflexivity Theory and then illustrate how one particular model – the Syntax-Discourse Model (most recently described in Burkhardt 2005) – accounts for the different operations required for the interpretation of referentially dependent elements. In Reinhart & Reuland (1993), the interpretation of a referentially dependent entity is determined by the entity’s relation to its potential antecedent and the predicate that licenses it. Consider the examples from (1) and (3). In (1), herself is an argument of the predicate painted and so is its antecedent Lisa, i.e. the reflexive
Reference assignment in the absence of sufficient semantic content
and its antecedent are coarguments of the same predicate. In contrast, in (3), himself is part of a coordination DP that is an argument of invited, but the antecedent Sam is an argument of the predicate claimed, i.e. the reflexive and its antecedent do not form a coargument relationship. This notion of coargumenthood also explains the diverging patterns of (non)complementarity of reflexives and pronouns: in the presence of coargumenthood, complementarity is preserved; in its absence, it is cancelled. In addition, conditions of reflexive-marking of the predicate must be observed to license proper coargument relations (and to rule out the availability of the pronoun in constructions like (1)), but since this has no bearing on the present discussion, I leave the details aside and refer the reader to Reinhart and Reuland (1993) or Burkhardt (2005). Crucially, the structural distinction between presence and absence of coargumenthood relations is hypothesized to have consequences for interpretation such that in the presence of coargumenthood (and proper reflexive-marking) the dependency is established at the level of syntax, while in the absence of coargumenthood the interpretation must be resolved at the level of discourse. In the following, these distinctions and the respective underlying mechanisms are spelled out within the Syntax-Discourse Model. For reasons that should be apparent from the preceding discussion, I refer to reflexive entities that form a coargument relation with their antecedent as coargument reflexives and to reflexive entities that do not share the same predicate with their antecedent as logophors. The interpretation of referentially dependent elements can generally be viewed in terms of mental model theories (e.g. Johnson-Laird 1983). Within the SyntaxDiscourse Model – which is based on Heim’s (1982) File Change Semantics and subsequent work by Avrutin (1999, 2004) and has recently been elaborated in Burkhardt (2005) – two levels of representation are of particular interest for the establishment of dependency. First, information is merged at the level of syntactic representation and phrase-structural information becomes available. This is where coargumenthood and reflexive-marking are checked. Second, information about participants, events, and so on is maintained and stored at the level of discourse representation, where the language system keeps track of information in order to make it available to future referential processes. These two levels are connected via correspondence rules that state that each DP corresponds to an individual discourse unit. In the case of referentially dependent elements, the model proposes the following operations during the establishment of dependency. First, the ele1. There is much more to be said about logophors, such as the fact that these entities convey the internal perspective or point of view of their antecedent, but in this paper I constrain myself to a purely structural definition of logophors as reflexive entities that do not hold a coargument relation with their antecedents. 2.
Similarly, each TP introduces an event unit and each PP a location unit (see Avrutin 1999, et seq.).
Petra Burkhardt
ments are merged in the syntactic representation and coargumenthood is assessed. If a coargumenthood relation is available and reflexive-marking is observed, this information suffices to form the dependency at the level of syntax. In the absence of coargumenthood, however, syntactic information cannot guide the interpretation process and an (underspecified) discourse unit corresponding to the logophor or pronoun is introduced at the level of discourse, where a search operation for a potential antecedent is carried out, resulting in the establishment of dependency at the level of discourse. In sum, the assessment of coargumenthood relations is a prerequisite to the establishment of dependency and it determines the level of representation where the dependency is formed. Of course, from the perspective of mental model theory, reference per se is never assigned in syntax – which serves to compute structural information and is independent from meaning and content information – but must be carried out within discourse or text structure. The Syntax-Discourse Model does not claim that referential interpretation of coargument reflexives is achieved by the syntactic system alone, which is considered to be blind to referential content, rather what is stated by the model is that the dependency between a reflexive and its coargument antecedent is formed in syntax (via presence of coargumenthood and proper reflexive-marking) and this linkage is carried over to the discourse level for free. In other words, the linkage is established in syntax, but irrespective of this, a discourse unit is also created for the reflexive and its antecedent (due to the correspondence rule that creates a discourse unit for every DP), and in addition, the dependency link established in syntax is projected in discourse, which renders any further interpretative processes unnecessary. Such a view is compatible with the principle of economy that states that syntax-dependent processes are more economical than discourse-dependent processes as for instance formulated in Reuland (2001) and Safir (2004). Here, this notion of economy is the critical link to connect representational considerations to processing data and to test the theoretical claims about reference assignment.
3. Independent neurolinguistic evidence for the predominance of syntactic information and the determination of coargumenthood has been reported in research with Broca’s aphasia patients, whose language deficit has been attributed to a slower-than-normal formation of syntactic structure. During the interpretation of coargument reflexives and logophors, these patients show temporally delayed performance patterns (compared to unimpaired controls), suggesting that syntactic information must be fully available to license dependency formation (cf. Piñango & Burkhardt 2005; Burkhardt, Avrutin, Piñango, and Ruigendijk 2007).
Reference assignment in the absence of sufficient semantic content
2. Referentially dependent entities and sentence processing Previous psycholinguistic studies have shown that fine-grained distinctions can be observed during reference assignment. In particular, logophors can be distinguished from coargument reflexives on the basis of processing cost. A series of cross-modal lexical decision interference studies, which generally measure processing cost associated with extra-syntactic operations (cf. Shapiro, Zurif, & Grimshaw 1987), revealed that logophor interpretation consumed more processing resources than coargument reflexive interpretation in English and Dutch (Piñango & Burkhardt 2005; Burkhardt 2005). Based on prior findings from crossmodal lexical decision interference studies, the increased processing demands observed during logophor resolution can be attributed to the level of representation at which the dependency is formed, where purely syntax-dependent interpretation (required for coargument reflexives) is resolved more easily than discourse-dependent interpretation (required for logophors). Recordings of ERPs during sentence comprehension further support the idea that distinct neural operations are involved in the resolution of semantically underspecified entities. ERPs are continuous measures of the electrical activity that occurs in the brain during sentence processing and provide information about the latency and topography of the underlying processes. In a comparison of coargument reflexives and logophors, the additional processing demands of logophor interpretation were reflected in an anterior negativity between 300–450ms after the onset of the logophor, which is interpreted as an index for the creation of a discourse unit and successive establishment of a discourse-based dependency (Burkhardt 2005). Moreover, Harris, Wexler, & Holcomb (2000) investigated agreement violations in constructions licensing coargument reflexive or logophor interpretation and report a positive deflection peaking around 600ms (labeled P600) for the coargument reflexive violation condition in comparison with the correct coargument reflexive condition, but crucially not for the logophor violation condition. Since the P600 has been reported as a marker of syntactic violations during referential processing (cf. Osterhout & Mobley 1995), these findings substantiate the syntax-discourse division for dependency relations. Finally, research on language acquisition provides further evidence for the involvement of syntax- and discourse-based operations during the formation of referential dependencies. Children’s performance errors have generally been connected to discourse processing, since preschoolers show adultlike performance on syntactic manipulations, but exhibit performance errors on discourse-related manipulations (e.g. d-linking, bridging relations, omission of determiners and tense nodes) (cf. e.g. Maratsos 1973; Avrutin 1999, 2000; Avrutin & Coopmans 2000; de Roo 2001). This split in performance has been ascribed to a limitation of processing resources (cf. Avrutin 1999, 2004) and is also reflected in
Petra Burkhardt
children’s interpretation of referentially dependent elements, where it has been attested for many languages that children perform well on coargument reflexives, but poorly on pronouns (cf. e.g. Sigurjónsdóttir 1992; Jakubowicz 1984; Grodzinsky & Reinhart 1993; Avrutin 1999; Thornton & Wexler 1999; Baauw 2002). Crucial for an investigation of logophoricity, children’s interpretation of logophors patterns with that of pronouns, i.e. they exhibit a significant number of performance errors (e.g. Avrutin & Cunningham 1997). All of these investigations have shown that logophor interpretation can be dissociated from coargument reflexive interpretation and that logophor resolution exerts additional processing demands, which have been further attributed to discourse-based operations. Most of this research has looked at the contrast of coargument reflexives and logophors. Another way of examining the discourse nature of logophor interpretation is by capitalizing on the observation that logophors share properties with ordinary pronouns, which suggests that similar operations should be carried out during the interpretation of these two entities. Thus, turning to the interpretation of pronouns, a cross-modal lexical decision interference study that compared pronoun interpretation to coargument reflexive interpretation did not yield an interpretable difference between these two conditions (Piñango & Burkhardt 2005). However, as the authors point out this might be due to the nature of pronoun interpretation, which – in contrast to coargument reflexives – allow selection of an extra-sentential antecedent. The differing internal structures (e.g. him vs. himself) might have been an additional confound and as a consequence, the reaction time task was not sensitive enough to disentangle all of the underlying processes. In contrast, a methodology that has proven to be extremely sensitive to changing processing demands across time is the measurement of ERPs, and previous studies have identified an ERP component in response to the processing of discourse-bound pronouns, which allows us to focus on the underlying mechanisms through this task. In a German ERP study that compared pronouns to repeated proper names, Streb, Rösler, and Hennighausen (1999) report an early and a late negative potential for the pronoun condition, which they interpret as reflexes of discourse processing and pronoun resolution. The late negative effect (510–600ms after pronoun onset) was replicated by another study examining pronoun integration with a latency of 360–480ms and interpreted as an N400 (i.e. a negative deflection peaking around 400ms) (Streb, Hennighausen, & Rösler 2004). In general, the N400 is best known as an index of semantic integration cost and its amplitude and/or latency increases when integration becomes more demanding (cf. e.g. Kutas & Federmeier 2000 for an overview). The wealth of N400-findings suggests that this effect marks the connection of incoming information with semantic and discourse information, and for present purposes it is generalized to a measure of depend-
Reference assignment in the absence of sufficient semantic content
ency. Additional evidence for this view is provided by a study that examined dependency formation of pronouns (5.a) and logophors (5.b) in contrast to the introduction of a discourse-new entity (e.g. Isaac in (5.c)). The results revealed an N400-like effect: a significant negativity between 400–600ms with a posterior maximum for the two referentially dependent elements (Burkhardt 2005).
5. a. Colleen explained that four judges had evaluated both her and Noah at the end of the three-hour training period.
b. Colleen explained that four judges had evaluated both herself and Noah at the end of the three-hour training period. c. Colleen explained that four judges had evaluated both Isaac and Noah at the end of the three-hour training period.
The following study seeks to extend these findings to reference assignment of logophors, pronouns, and coargument reflexives. It aims to replicate the N400-effect for discourse-based interpretation and to investigate how coargument reflexives, which are not hypothesized to form a discourse dependency, behave neurally with respect to this measure of discourse processing.
3. The present experiment In the following investigation, I capitalize on the hypothesis that logophors and pronouns enter into a discourse dependency with their antecedents, while syntactic information suffices for the establishment of a dependency between a coargument reflexive and its antecedent. Based on previous findings from pronoun interpretation (Streb et al. 1999, 2004; Burkhardt 2005), this contrast is predicted to emerge in the N400-component of the ERP signal, which is interpreted as a marker for discourse-based dependency formation. Assuming that similar interpretive processes are indexed by similar patterns of electrical activity, both pronoun and logophor resolution are expected to show an N400-effect in the comparison to a baseline of entities that do not require the formation of dependency (i.e. discoursenew entities), while coargument reflexive resolution, which is guided by syntactic information, should not exhibit an N400-effect. The N400-component is hence used as a gauge to demonstrate that coargument reflexive interpretation does not exert discourse integration cost. This study therefore takes a new approach to investigating dependency formation as a function of the level of representation that guides the interpretation process and it allows for a close, and so far missing comparison of the three dependency relations involving logophors, pronouns, and coargument reflexives.
Petra Burkhardt
3.1
Experimental manipulation
To investigate the processes underlying the formation of dependency, sentences for three dependency conditions and two corresponding no-dependency conditions were constructed, as illustrated in (6) with the critical entities underlined. (6.a) represents a case of coargument reflexive dependency, where herself and its antecedent Christine are arguments of one and the same predicate. In (6.c) and (6.d) the dependent elements and their antecedents do not form a coargument relationship based on the phrase-structural configuration, because the former are contained in a complex DP (both X and Y), which in its entirety is an argument of the predicate criticized. (6.b) and (6.e) represent the corresponding no-dependency conditions, which include a discourse-new proper name instead of a referentially dependent element. 6. a. Coargument Reflexive (dependency): Twenty viewers heard that Christine had criticized herself during the interview on public television. b. Proper Name I (no dependency): Twenty guests heard that Christine had criticized Trevor during the interview on public television. c. Logophor (dependency): The neighbors heard that Jenny had criticized both herself and Nathan for a full twenty minutes during the talkshow. d. Pronoun (dependency): Several coworkers heard that Jenny had criticized both her and Nathan for a full twenty minutes during the talkshow. e. Proper Name II (no dependency): Olivia heard that two managers had criticized both Trevor and Nathan for a full fifteen minutes at the meeting.
Although a similar construction is available for the logophor and pronoun conditions, where dependency relations are possible with the sentence-initial DP (see (5) above), the current constructions were purposely chosen to control the distance between the referentially dependent element and its antecedent across conditions and both was included to unambiguously mark the coordination structure. A potential confound resulting from this decision needs to be addressed, which is that the acceptability of the coordination construction with the pronoun (6.d) is somewhat controversial. First, native speaker intuitions diverge, such that some allow the intended coreferential reading (her = Jenny), while others reject it. A study collecting acceptability ratings was therefore conducted to obtain a measure
Reference assignment in the absence of sufficient semantic content
of native speaker judgments, which is reported below. Second, turning to the literature, Reinhart and Reuland (1993:675) point out that (7.a) does not allow for a coreferential reading between her and the queen. However, on the basis of their discussion of semantic predicates and the contrasts between collective and distributive interpretations in certain constructions, the intended coreferential reading might be recoverable under a collective interpretation as depicted in (7.c).
7. a. The queen invited both Max and her to our party.
b. distributive reading: the queen (λx (x invited Max & x invited x)) c. collective reading: the queen (λx (x invited Max and x))
Under a distributive reading, the queen is not available as a proper antecedent for the pronoun, because the pronoun and its antecedent are coarguments of the same predicate (x invited x), but the predicate is not reflexive-marked. This yields an unacceptable interpretation (of her = the queen). In contrast, the representation associated with the collective reading in (7.c) does not violate these principles, because the pronoun is contained in a coordination phrase, which in its entirety is in a coargument relationship with the antecedent. As a consequence, the predicate is not reflexive, yielding an acceptable interpretation (her = the queen). (See also Safir (2004:16) for a similar argument and a note in Reuland and Reinhart (1995:265) that further attests that for some speakers a collective interpretation is available in both X and Y constructions). The present ERP experiment took care of the possible rejection of a coreferential reading in the pronoun condition by probing interpretation via a verification sentence that proceeded every stimulus sentence and excluding participants who failed to accept the coreferential interpretation from further analyses (see below). In addition, a questionnaire study was carried out to obtain a measure of the acceptability of the coreferential reading. 3.2
Questionnaire study
3.2.1 Methods and Materials A questionnaire was created to assess the acceptability of the pronoun condition (6.d), as well as the logophor condition (6.e). Ten sentences per condition were constructed and matched with a verification sentence as illustrated directly below. 20 filler items were also created containing a dependency with the highest subject 4. The semantic representations in (7) are meant to illustrate the contrast between distributive and collective readings only and I do not intend to make any claims about the semantic analysis of distributivity and collectivity. For detailed discussion of an analysis of these phenomena, including an operation of sum-formation in the case of collective readings, see for instance Landman (1996).
Petra Burkhardt
(involving either a logophor or a pronoun), out of which 10 were matched with a correct and 10 with an incorrect verification sentence.
8. a. Pronoun:
Four journalists heard that Maureen had videotaped both her and Justin. Verification sentence: Maureen and Justin had been videotaped.
b. Logophor: Several coworkers heard that Jenny had criticized both herself and Nathan. Verification sentence: Jenny and Nathan had been criticized. c. Filler(cor): Alyssa knew that Todd had recommended both her and Sid. Verification sentence: Alyssa had been recommended. d. Filler(incor): Samantha mentioned that Darryl had visited both her and Sid. Verification sentence: Darryl had been visited.
Different verbs were used for all 40 items and two lists were created with different condition-verb pairings to eliminate any verb-related effects on the acceptability ratings. All 40 items in a list were then pseudo-randomized. Participants had to indicate whether the verification sentence was a correct or an incorrect representation of the events described in the first sentence. They were instructed to provide their first reaction and not to ponder the sentence material. 3.2.2 Participants Thirty native speakers (12 male) participated in this questionnaire study. Mean age was 21.2 years (ranging from 18 to 34). 3.2.3 Results and implications for the ERP experiment Responses were averaged over all participants and mean error rates are reported in percentages (i.e. marking (8.a-c) as ‘incorrect’ and (8.d) as ‘correct’). In the filler conditions, the average error amounted to 3% (ranging from 0 to 1 inaccurate responses per participant); in the logophor condition, the dependent reading was rejected in 6% of the responses (ranging from 0 to 6 across participants); and in the pronoun condition, the intended dependent reading was evaluated as incorrect 13% of the time (ranging from 0 to 9 incorrect responses per participant). The data indicate that the majority of the participants accepted the intended interpretation of the pronoun in (8.a) and an ANOVA comparing the pronoun to the logophor condition revealed no significant difference in error rates (F<1). The overall results from this questionnaire therefore permit the use of the pronoun construction in the following ERP study. One precaution however was taken since 5 of the 30 participants rejected the pronoun condition more than three times (i.e. > 30%), which indicates some variability across participants: the participants in the ERP
Reference assignment in the absence of sufficient semantic content
study were also screened for their acceptability ratings via a verification task and if necessary discarded from further analyses. Before proceeding to the ERP study, it should be emphasized once more that the pronoun construction was chosen to control for distance between the dependent element and its antecedent across the three contrasts presented here. However, an investigation of an uncontroversially acceptable pronoun construction has previously been reported in Burkhardt (2005) (see (5.a) above: Colleen explained that four judges had evaluated both her and Noah …) and a replication of the findings reported therein will strengthen the claim of a dependent reading. 3.3
ERP study
3.3.1 Participants Twenty-two students (8 male) at Yale University were recruited, who received either course credit or monetary compensation. Mean age was 19.6 years (ranging from 18 to 23). All participants were right-handed native-speakers of English and had normal or corrected-to-normal vision. Data from one participant were excluded prior to statistical analyses because she showed an error rate above 30% in the verification task (see below). 3.3.2 Materials All sentences used in this study had the structure DP V that DP had V (both) X and DP PP PP as illustrated in (6) above. To focus on the nature of dependency formation, the critical contrasts for the present investigation were dependency vs. no-dependency (i.e. (6.a) vs. (6.b), (6.c) vs. (6.e), and (6.d) vs. (6.e)). Stimuli were created in quintuplets and forty sentences were constructed per condition. In the referentially dependent conditions, him(self) and her(self) were distributed evenly across experimental items, and in the proper name conditions, mono- and bisyllabic proper names where chosen an equal number of times to control for the length difference between reflexives and pronouns. The final script included an additional 80 filler items and two pseudo-randomized lists were created which varied in the order of the items. Each stimulus was matched with a verification sentence. 3.3.3 Procedure Participants were instructed that their primary task was to read each sentence for comprehension and following each sentence respond to a verification sentence by pressing a response button. All 280 sentences were presented visually word-byword in the center of a computer screen. At the beginning of each stimulus, a fixation cross was presented for 500ms. Words appeared then for 450ms each and after a blank screen of 500ms, the verification sentence was presented in its entirety.
Petra Burkhardt
Participants had a maximum of 4000ms to indicate whether the verification sentence matched the preceding stimulus or not. The next stimulus appeared 1000ms after the verification decision. All sentences were assigned to 9 blocks, between which participants took short breaks. 3.3.4 EEG recordings Electroencephalograms (EEGs) were recorded and analyzed from the following 15 scalp sites by means of Ag/AgCl electrodes mounted in a standard elastic cap (QuikCap, Neuroscan, Inc.), which adhered to the position nomenclature of the international 10–20 system (Jasper 1958): FZ, CZ, PZ, F3/4, F7/8, T7/8, C3/4, P3/4, P7/8. Horizontal and vertical eye movements were monitored by two sets of electrode pairs, placed at the outer canthus of each eye and above and below the left eye to control for eye movement artifact. Recordings were referenced to the left mastoid and rereferenced offline to the average of the recorded channels. Electrode impedances were kept below 5kΩ. All channels were recorded with a band pass from DC to 60Hz and a sampling rate of 250Hz (SynAmps, Neuroscan, Inc.). ERPs were filtered offline with 12Hz low pass for the plots only. 3.3.5 Data analysis and results 3.3.5.1 Behavioral data The responses to the verification task served primarily as a criterion for data selection. Data from one participant were excluded in their entirety because the participant had a relatively high error rate in the pronoun condition, suggesting that she might not allow a coreferential reading. Correct responses of all other participants entered the final computation. Error percentages were averaged over incorrect and timed-out responses (i.e. responses that failed to be registered 4000ms after the verification sentence). Participants made 4.9% errors in the ProperName I condition, 4.4% in the Coargument Reflexive condition, 27.3% in the ProperName II condition, 10.2% in the Pronoun condition, and 9.3% in the Logophor condition. 3.3.5.2 ERP data Trials with incorrect responses in the verification task or with ocular, muscle, or amplifier saturation artifacts were excluded from the averaging procedure. This amounted to the rejection of 27.4% of trials across conditions. All ERP data were 5. The differences between the error rates might be reflecting of varying working memory requirements across the five conditions, where four distinct discourse referents and their respective roles must be maintained in the Proper Name II condition, but only two (Reflexive) or three (Pronoun, Logophor, Proper Name I) in the other conditions. However, this observation has no bearing on the present research question and overall the error rates reveal that the participants were attending to the sentences.
Reference assignment in the absence of sufficient semantic content
aligned to a 200ms baseline before the onset of the critical word (dependent entity or proper name) and average ERPs were time-locked to the onset of the critical word and computed per condition per participant. In order to investigate the discourse nature of logophor interpretation and its absence during coargument reflexive interpretation, statistical analyses were carried out for the N400-window on the basis of previous findings (Streb et al. 1999, 2004; Burkhardt 2005). As described above, the N400-effect is a negative-going waveform with a peak latency around 400ms after the onset of the critical entity and it is most pronounced over posterior electrode sites. Guided by earlier studies and after visual inspection of the data, analyses were performed on the mean amplitude values in two successive temporal windows (350–450ms and 450–550ms), because the plots revealed latency differences with respect to the N400-component. Since this study was interested in the relative contrasts between the dependency and no-dependency conditions, repeated-measures ANOVAs with the factor condition were computed for each of the three comparisons separately. Analyses for midline electrodes, which included the factor electrode (FZ/CZ/PZ), were computed separately from those for lateral electrodes. The lateral analyses included the additional topographical factor region – i.e. anterior (F3/4, F7/8, T7/8) vs. posterior (C3/4, P3/4, P7/8) – because the traditional N400 is most pronounced over posterior sites. Statistical analyses were carried out in a hierarchical manner. Figures 1 to 3 display the grand average ERPs for the relative comparisons between dependency (solid line) and no-dependency (dotted line) at a right posterior electrode site (C4).
Figure 1. Grand average ERPs (n=21) recorded to the onset of the critical word (onset at the vertical line) at C4 for Coargument Reflexives (solid line) and Proper Names I (dotted line). Negative voltage is plotted upwards.
Petra Burkhardt
Coargument reflexives vs. proper names For the contrast Coargument Reflexives vs. Proper Names I, no apparent differences emerged in the ERP waveforms, as is evident from the illustration in figure 1. The statistical analyses in the two separate time windows also yielded no significant effects (F’s<1). The results for this and the following contrasts are reported in the appendix. Pronouns vs. proper names The comparison between Pronouns and Proper Names II is illustrated in figure 2. The Pronoun condition shows a negative deflection with an onset latency around 400ms. For the 350–450ms window, an ANOVA over midline sites revealed a main effect of condition (F(1, 20)=5.242, p<.03), which was most pronounced at the posterior site PZ (F(1, 20)=6.388, p<.02). The lateral analysis in this window did not reveal a significant effect. For the 450–550ms window, there were no significant differences over midline sites, but an ANOVA over lateral sites revealed a significant negativity for the Pronoun condition (F(1, 20)=4.7, p<.04), which was resolved over posterior sites (F(1, 20)=13.17, p<.002). Logophors vs. proper names Finally, figure 3 exemplifies the contrast between Logophors and Proper Names II. The Logophor condition also shows a negativity peaking around 400ms after the onset. For the 350–450ms window, the midline analysis revealed no main effect. The ANOVA over lateral sites registered a significant difference (F(1, 20)=5.651, p<.02), which had a posterior maximum (F(1, 20)=6.637, p<.01). The midline and lateral analyses in the 450–550ms window showed no main effects.
Figure 2 & 3. Grand average ERPs (n=21) recorded to the onset of the critical word (onset at the vertical line) at C4. Figure 2 (left) illustrates the contrast between Pronouns (solid line) and Proper Names II (dotted line) and figure 3 (right) the contrast between Logophors (solid line) and Proper Names II (dotted line). Negative voltage is plotted upwards.
Reference assignment in the absence of sufficient semantic content
Table 1. Summary of effects per comparison and time window (NS = non-significant).
350-450ms 450-550ms
3.4
Coargument Reflexives vs. Proper Names I
Pronouns vs. Proper Names II
Logophors vs. Proper Names II
NS NS
Negativity Negativity
Negativity NS
Discussion
The effects observed for the three contrasts between dependent and independent entities are sketched out in table 1. This study first demonstrates that the non-complementarity of logophors and pronouns has processing correlates. The interpretation of both of these entities elicited a pronounced negativity compared to a discourse-new proper name baseline. This negativity is here interpreted as a marker of dependency formation at the level of discourse and therefore is an index for the additional processing resources required by the language system. The data thus confirm previous ERP findings from pronoun interpretation and further substantiate the claim that logophor resolution is carried out in discourse. Statistical ana lyses indicate that the negativity elicited by the logophor had an earlier onset and was only significant for the 350–450ms interval, while the negativity for the pronoun was most pronounced in the 450–550ms window. Such a latency shift possibly signals further differences between the two discourse-bound entities and additional analyses with the factor interval and the two temporal windows as levels reveal interactions of condition by interval for the contrast Pronouns vs. Proper Names in the midline (F(1, 20)=5.49, p<.03) and lateral analyses (F(1, 20)=5.66, p<.02) and a marginal interaction for the comparison of Logophors vs. Proper Names over lateral sites (F(1, 20)=3.88, p<.06), which were all further resolved over posterior electrodes. At this point, I can only speculate about the reasons for this difference. First, the more pronounced effect in the pronoun condition should not be considered a marker of unacceptability, because first, pronoun use is not prohibited and an extra-sentential antecedent is possible in this construction and second, absence of a sentential antecedent as well as morphological feature mismatches have been reported to elicit a P600-component and not an N400 (e.g. Osterhout & Mobley 1995; Burkhardt & Piñango 2003). In addition, the answers to the verification task reveal that the participants obtained the anticipated reading. The latency shift might then be due to the fact that logophors are preferred in these constructions (as indicated above) and even though the verification task probed interpretation, it does not reflect preferences. Most convincingly and supported by theory-internal considerations, however, the shift might be reflecting of the differing internal structures (him vs. himself) and related to this the observa-
Petra Burkhardt
tion that logophor resolution is more constrained than pronoun resolution, since logophors must select an antecedent whose internal perspective or state of consciousness they reflect (cf. e.g. Sells 1987), while pronouns can choose from a broader set of candidates (if available). Most importantly, the data demonstrate that logophors and pronouns register similar processing effects associated with discourse-based reference assignment. Second, this study was concerned with the claim that dependencies licensed in syntax are satisfied at no further cost. The comparison of coargument reflexives to a discourse-new proper name baseline did not elicit an N400-effect in either time interval, suggesting that no additional processing demands are exerted during these temporal windows. Taken together with the observation that coargument reflexives and logophors have the same internal structure, their differing behavior in relation to the no-dependency conditions in the N400-window must be attributed to functional distinctions. The Syntax-Discourse Model and in particular the notion of coargumenthood present an explanation according to which the emergence of a significant N400-effect for logophors is an instantiation of discoursebased operations, while its absence for coargument reflexives reflects the idea that the corresponding dependency is established in a more economical way and comes at no expense for discourse representation. The present data thus represent indirect evidence for the principle of economy and indicate that once a syntactic dependency has been formed – which is a highly automated operation – no additional resources are consumed. To conclude, the findings indicate that entities that depend on an antecedent during reference assignment are dynamically interpreted and are subject to differing dependency operations. Compared to a baseline condition, the data demonstrate that a dissociation of syntactic and discourse dependencies, as formulated in the current model, is supported by electrophysiological patterns: if available, dependencies licensed by the presence of coargumenthood (and reflexive-marking) are formed, otherwise, dependency relations must be established at the level of discourse. The findings also supplement previous psycholinguistic investigations, showing that while pronoun and logophor interpretation register additional discourse-based processing effort, coargument reflexive interpretation has already been completed.
Reference assignment in the absence of sufficient semantic content
References Avrutin, S. 1999. Development of the Syntax-Discourse Interface. Boston: Kluwer. Avrutin, S. 2000. Comprehension of discourse-linked and non-discourse-linked questions by children and Broca’s aphasics. In Language and the Brain: Representation and processing, Y. Grodzinsky, L. Shapiro and D. Swinney (eds), 295–313. San Diego: Academic Press. Avrutin, S. 2004. Beyond narrow syntax. In Variation and Universals in Biolinguistics, L. Jenkins. (ed.), 95–115. London: Elsevier. Avrutin, S. and Coopmans, P. 2000. Children who build bridges. Proceedings of the Annual Boston University Conference on Language Development 24(1): 80–91. Avrutin, S., and Cunningham, J. 1997. Children and reflexivity. Proceedings of the Annual Boston University Conference on Language Development 21(1): 13–23. Baauw, S. 2002. Grammatical Features and the Acquisition of Reference: A comparative study of Dutch and Spanish. London: Routledge. Burkhardt, P. 2005. The Syntax-Discourse Interface: Representing and interpreting dependency. Amsterdam: John Benjamins. Burkhardt, P., Avrutin, S., Piñango, M.M., and Ruigendijk, E. 2007. Slower-than-normal syntactic processing in agrammatic Broca’s aphasia: Evidence from Dutch. Journal of Neurolinguistics. Burkhardt, P. and Piñango, M.M. 2003. Implications of antecedent availability during pronominal interpretation: An ERP study, 103. New York NY: Cognitive Neuroscience Society. Chomsky, N. 1981. Lectures on Government and Binding. Dordrecht: Foris. Chomsky, N. 1986. Knowledge of Language: Its nature, origins, and use. New York NY: Praeger. de Roo, E. 2001. Root nonfinite and finite utterances in child language and agrammatic speech. Brain and Language 77(3): 398–406. Grodzinsky, Y. and Reinhart, T. 1993. The innateness of binding and coreference. Linguistic Inquiry 24(1): 69–101. Harris, T., Wexler, K. and Holcomb, P.J. 2000. An electrophysiological investigation of binding and coreference. Brain and Language 75(3): 313–346. Heim, I. 1982. The Semantics of Definite and Indefinite Noun Phrases. PhD dissertation, University of Massachusetts, Amherst. Huang, Y. 1991. A neo-Gricean pragmatic theory of anaphora. Journal of Linguistics 27(2): 301–335. Jakubowicz, C. 1984. On markedness and binding principles. In Proceedings of N(orth) E(astern) L(inguistic) S(ociety), XIV, C. Jones and P. Sells (eds), 154–182. Amherst MA: GLSA. Jasper, H.H. 1958. The ten twenty electrode system of the International Federation. Electroencephalography and Clinical Neurophysiology 10(2): 371–375. Johnson-Laird, P.N. 1983. Mental Models. Towards a cognitive science of language, inference, and consciousness. Cambridge MA: Harvard University Press. Kutas, M. and Federmeier, K.D. 2000. Electrophysiology reveals semantic memory use in language comprehension. Trends in Cognitive Sciences 4(12): 463–470. Landman, F. 1996. Plurality. In The Handbook of Contemporary Semantic Theory, S. Lappin (ed.), 425–457. Malden MA: Blackwell. Lasnik, H. 1989. Essays on Anaphora. Dordrecht: Kluwer. Levinson, S. C. 2000. Presumptive Meanings. Cambridge MA: Bradford. Maratsos, M. P. 1973. The effects of stress on the understanding of pronominal co-reference in children. Journal of Psycholinguistic Research 2(1): 1–8.
Petra Burkhardt Osterhout, L. and Mobley, L.A. 1995. Event-related potentials elicited by failure to agree. Journal of Memory and Language 34: 739–773. Piñango, M.M. and Burkhardt, P. 2005. Pronominal interpretation and the syntax-discourse interface: Real-time comprehension and neurological properties. In Anaphora Processing: Linguistic, cognitive and computational modeling, A. Branco, T. McEnery and R. Mitkov (eds.), 221–238. Amsterdam: John Benjamins. Pollard, C., and Sag, I.A. 1992. Anaphors in English and the scope of binding theory. Linguistic Inquiry 23(2): 261–303. Reinhart, T. and Reuland, E. 1993. Reflexivity. Linguistic Inquiry 24(4): 657–720. Reuland, E. 2001. Primitives of binding. Linguistic Inquiry 32: 439–492. Reuland, E. 2003. Anaphoric dependencies: A window into the architecture of the language system. Glot International 7(1/2): 3–25. Reuland, E. and Reinhart, T. (1995). Pronouns, anaphors and case. In Studies in Comparative Germanic Syntax, H. Haider, S. Olsen and S. Vikner (eds), 241–268. Dordrecht: Kluwer. Safir, K. 2004. The Syntax of Anaphora. Oxford: OUP. Sells, P. 1987. Aspects of logophoricity. Linguistic Inquiry 18(3): 445–479. Shapiro, L.P., Zurif, E. and Grimshaw, J. 1987. Sentence processing and the mental representation of verbs. Cognition 27(3) : 219–246. Sigurjónsdóttir, S. 1992. Binding in Icelandic: Evidence from language acquisition. PhD dissertation, UCLA. Streb, J., Rösler, F. and Hennighausen, E. 1999. Event-related responses to pronoun and proper name anaphors in parallel and nonparallel discourse structures. Brain and Language, 70(2): 273–286. Streb, J., Hennighausen, E. and Rösler, F. 2004. Different anaphoric expressions are investigated by event-related brain potentials. Journal of Psycholinguistic Research 33(3): 175–201. Thornton, R. and Wexler, K. 1999. Principle B, VP Ellipsis, and Interpretation in Child Grammar. Cambridge MA: The MIT Press.
Appendix Summary of statistical analyses per comparison and time window Coargument Reflexives vs. Proper Names I
Pronouns vs. Proper Names II
Logophors vs. Proper Names II
350–450ms midline ANOVA PZ lateral ANOVA posterior sites
F(1, 20)=.690, p<.41 F(1, 20)=1.262, p<.27 F(1, 20)=.437, p<.51 F(1, 20)=.071, p<.79
F(1, 20)=5.242, p<.03 F(1, 20)=6.388, p<.02 F(1, 20)=2.362, p<.14 F(1, 20)=4.214, p<.05
F(1, 20)=1.345., p<.26 F(1, 20)=.542, p<.47 F(1, 20)=5.651, p<.02 F(1, 20)=6.637, p<.01
450–550ms midline ANOVA lateral ANOVA posterior sites
F(1, 20)=1.855, p<.18 F(1, 20)=.087, p<.77 F(1, 20)=.004, p<.95
F(1, 20)=1.658, p<.21 F(1, 20)=.008, p<.93 F(1, 20)= 4.7, p<.04 F(1, 20)=.534, p<.47 F(1, 20)=13.17, p<.002 F(1, 20)=4.223, p<.05
Resolving complex anaphors Evidence from online comprehension Konstanze Marx, Ina Bornkessel-Schlesewsky and Matthias Schlesewsky This study investigates the question of whether the processing of complex anaphors require more cognitive effort than the processing of NP-anaphors. Complex anaphors refer to abstract objects which are not introduced as a noun phrase and bring about the creation of a new discourse referent. This creation is called “complexation process” (see also Consten et al. this volume). We describe ERP findings which provide converging support for the assumption that the cognitive cost of this complexation process is higher than the cognitive cost of processing NP-anaphors.
1. Introduction In this study, we distinguish two kinds of anaphors, direct noun phrase anaphors (NP-anaphors) and complex anaphors. This distinction is based on two differences. First, NP-anaphors characteristically refer back to concrete referents (Cornish 1986, Schwarz 2000) whereas complex-anaphors (see Schwarz 2000, SchwarzFriesel et al. 2004) refer to abstract objects (such as propositions, facts and events). Second, NP-anaphors refer to an entity that has already been introduced in the text by a noun phrase (NP), while complex anaphors refer to sentences or larger text segments and at the same time introduce a referent for them into the discourse representation. For instance, the discourse referent which das Auto in sen-
1. Researchers use different terms to refer to the phenomenon, e.g. “abstract object anaphora” (Asher 1993), “labelling” (Francis 1994) or “shell nouns” (Schmid 2000).
Konstanze Marx, Ina Bornkessel-Schlesewsky and Matthias Schlesewsky
tence (1) refers to is already established in the mental discourse representation (Schwarz 2001), but the first sentence in example (2) does not provide a discourse referent for the accident, its creation is due to the complex anaphor der Unfall. (1) Der BMW fuhr gegen einen Baum. Das Auto hatte einen Totalschaden. The BMW crashed into a tree. The car got totally damaged.
(2) Der BMW fuhr gegen einen Baum. Der Unfall kostete glücklicherweise keinen das Leben. The BMW crashed into a tree. Fortunately there were no casualities in this accident.
Anaphors are important coherence markers in a text since they can keep a referent activated in a mental discourse representation. Once a referent has been introduced further expressions can refer to it. This process works reasonably well as long as these referents are entities introduced by noun phrases. But how does the resolution of the anaphoric relation work if there is no such NP-referent available? We assume that abstract discourse referents are only established in a mental discourse representation once there is a complex-anaphoric expression that motivates and initiates this process. Intuitively, we would expect that this so-called complexation process (see also Consten et al. in this volume) calls for a greater cognitive effort. This greater cognitive effort can be put down to the fact that in the case of complex anaphors, a new referent must be established, whereas in the case of NP-anaphors an already existing referent is reactivated. We have conducted an event related brain potential (ERP) experiment to find out whether this prediction is borne out. Does the complexation process show up as a physiological effect? Before we introduce our study we will briefly sketch the ERP method in general and review some relevant previous ERP findings.
2. Methodological prerequisites Event-related potentials (ERPs) reflect electrical brain activity which takes place before, during or after a sensoric, motoric or psychological event in the electroencephalogram (EEG) (see Frisch 2000, Hahne 1997, Kutas et al. for further informa2. The mental discourse representation is often referred to with the term mental model, since there is no common opinion over the characteristics and we cannot review the whole debate, we used the more general term mental discourse representation which includes “all participating referents, their mutual relations as well as the events and situations in which these referents are engaged” (Schwarz 2001, 18). The mental representation generates on the basis of textual and knowledge information.
Resolving complex anaphors
tion). They make it possible to investigate the brain’s reaction to a specific linguistic stimulus. ERPs are measured over electrodes which are placed on the scalp. They have a lower amplitude extension than the basic EEG-activity and have to be extracted by averaging over many time periods which include the interesting event. The method can be used for the investigation of several steps in the text comprehension process (see among others Fischler 1990), ERPs can be recorded immediately during the visual and acoustic presentation of linguistic material and make possible a non-invasive online observation of neuronal processes during language processing. Another advantage of the ERP-method is that participants need not carry out additional tasks (see Garnsey 1993 for an overview). ERPs are relative measures, i.e., a critical condition must always be evaluated with reference to a control condition. The values for a negativity or positivity cannot be interpreted in isolation. ERPs are characterized by specific components which are language sensitive. The question of how these components should be defined is still under debate (see Coles and Rugg 1995). Based on the classical definition of Donchin et al. (1978) components are characterized by four features: – polarity Are there positive or negative peaks in the recorded waveforms? – latency How long is the time period between the presentation of a critical stimulus and the amplitude maximum? – topography At which electrode positions can a maximum deviation be observed? – sensitivity How susceptible is an effect to experimental manipulations? Important for language processing are components such as (E)LAN and P600, which have often been associated with syntactic processing and N400, which mainly correlates with lexical-semantic manipulations. (For problems with such oneto-one association between components and linguistic domains see Bornkessel et al. 2004 and Roehm et al. 2004.) As indicated by the names for these components,
3. To make sure that subjects are attending to the stimuli additional tasks are used in practice nevertheless. 4. “As a rule of thumb, differences in wave shape and/ or scalp distribution between two or more conditions are interpreted as reflecting the activity of distinct neuronal populations subserving qualitatively different processes, whereas differences in amplitude and/ or latency are interpreted as modulations in the activity of the same or related neuronal populations subserving processes that differ quantitatively” (Kutas 1993). 5. But associations between activity at a certain electrode and the directly underlying brain region are not allowed (see Streb 2000, Friederici 2003). 6. This is an abbreviation for “(early) left anterior negativity”.
Konstanze Marx, Ina Bornkessel-Schlesewsky and Matthias Schlesewsky
N400 is a negative peak latency 400 msec after its eliciting event, and P600, a positive peak latency after 600 msec.
3. Relevant previous ERP findings As just mentioned, the N400 is often correlated with semantic processing whereas the P600 often reflects syntactic processes. The following paragraph shows that both components are sensible to anaphoric processing, which might suggest that it combines syntactic and semantic processing. Streb (2000) manipulated the difficulty of integrating an anaphoric expression into a mental discourse representation by varying the detection difficulty of the antecedent: She varies the textual distance between anaphor and antecedent in a first experiment, the syntactic positions of antecedent and anaphor in a second experiment and the anaphoric expression itself in both experiments. Her test items in the first experiment consisted of three related context sentences, a target sentence including an anaphoric reference to one of these characters, and a closing yes/no-question. The anaphor referred back to sentences 1, 2, and 3 equally often (3). (3)
context: s1: Lisa schlendert über einen Basar. Lisa strolls across a bazaar. s2: Peter verkauft Edelsteine an Touristen. Peter sells gems to tourists. s3: Die Steine sind hervorragend geschliffen. The gems are cut excellently. Target: Nun wird Lisa/sie dem Händler einen Diamanten abkaufen. Then Lisa/ she will buy a diamond from the trader. Question: Kauft Lisa den Diamanten? Does Lisa buy the diamond?
Streb (2000) found that comprehension times increased with the distance between anaphor and antecedent. The ERP findings reveal a relatively larger N400 (for pronouns and proper names) for the far and the medium distance condition in comparison to the near distance condition. Streb (2000) (see also Streb et al. 2004) attributes this effect to semantic integration processes that are increasingly demanding (in relation to the increasing distance between anaphor and antecedent). The second experiment investigated whether the anaphoric processing depends on parallel (4) and non-parallel (5) sentence structures.
(4) Peter besucht Julia in der Klinik. Dort hat Peter dem Arzt eine Frage gestellt. Peter visits Julia in the hospital. There Peter asked the doctor a question. (Streb 2000, 69)
Resolving complex anaphors
(5) Peter besucht Julia in der Klinik. Dort hat die Schwester Peter das Zimmer gezeigt. Peter visits Julia in the hospital. There the nurse shows Peter the room. (Streb 2000, 69)
The first finding was that comprehension questions following non-parallel structures are answered more slowly. More important was the observed negative peak latency for non-parallel structures in a time window from 510 to 600 msec after the critical stimulus. This modulation of the N400 effect reflects the more demanding integration of a less expected antecedent in a mental discourse representation. Kulik et al. (2004) compared hyperonymic with repeated anaphoric expressions. They also varied the syntactic position of the anaphoric expression.
(6) neutral context: Peter fragt sich, was in dem Garten passiert war. Peter wonders what had happened in the garden.
(7) non-neutral context: Peter fragt sich, wer in dem Garten den Karpfen gestohlen hat. Peter wonders who had stolen the carp in the garden.
(8) Target (arg1): Dann erfuhr er, dass den Karpfen/ Fisch der Junge gestohlen hat. Then he was told that the carp/ the fish had been stolen by the boy.
(9) Target (arg2): Dann erfuhr er, dass der Junge den Karpfen/ Fisch gestohlen hat. Then he was told that the boy had stolen the carp/ the fish.
Kulik et al. (2004) found no differences between Karpfen vs. Fisch as long as the target sentences (8) and (9) were introduced in a neutral manner (6). Significant differences between hyperonymic and repeated anaphoric expressions were observed in combination with (7). Both show a reduced N400 in comparison with the same word in a neutral context, but the reduction for hyperonymic expressions sets in with a delay of 70 msec in comparison to that for repeated expressions. While the reduction for the repetition of Karpfen shows up in a time window from 300 to 500 msec, the effect for the anaphoric hyperonym Fisch was observed in a time window from 370 to 570 msec. The authors interpret their findings as evidence for the hypothesis that semantic relations facilitate the establishing of a coreference relation. The results indicate how long it might take to resolve a hyperonym-hyponym relation. The N400-effect in the above mentioned studies leads to the conclusion that this component might indicate the cognitive effort for the anaphor resolution process. The more difficult the search for an antecedent is, the higher the negative peak of the waveform (Streb 2000) seems to be. The results of Kulik et al. (2004) tie in with these findings. They observed a reduction of the N400 as soon as the anaphor processing was facilitated by lexical repetition or a semantic relation between antecedent and anaphoric expression.
Konstanze Marx, Ina Bornkessel-Schlesewsky and Matthias Schlesewsky
Burkhardt (2005) investigated how bridging relations (indirect anaphors, see Schwarz 2000) are encoded in the mental discourse representation. She compared three conditions: direct anaphors (lexical repetition) (a), indirect anaphors (b), no anaphoric relation (c), see (10). (10) context a): Regine beschreibt einen Portier aus dem Adlon. Regine describes a doorman from the Adlon. context b): Rebekka beschreibt ein Hotel in der Eifel. Rebekka describes a hotel in the Eifel. context c): Ruth schwatzt gelegentlich mit ihrer Friseurin. Ruth chats occasionally with her hairdresser. Target: Sie denkt, dass der Portier wohl überqualifiziert war. She thinks that the doorman was probably overqualified. (Burkhardt 2005, 75)
The definite noun phrase (without any anaphoric relation) in context c) elicited a N400 and a P600 effect. As expected (see among others Kulik et al. 2004) the anaphor in context a) elicited a reduced N400. The indirect anaphoric expression in context b) elicited also a reduced N400 but additionally a P600 (left posterior) as seen in context c). Indirect anaphors seem to be processed on the basis of two different patterns (Burkhardt 2005). On the one hand the resolution is comparable with the resolution of direct anaphors which refer to already known (and established) referents (reduction of the N400), on the other hand the resolution is comparable with new entities which do not yet have a slot in the mental discourse representation and need to be established in a process which demands cognitive effort (P600). These studies show that the N400 as well as the P600 component might be interpreted as indicators for cognitive effort during the anaphoric processing.
4. ERP Study 4.1
Participants
25 undergraduate students (12 female) of the University of Leipzig and one unemployed woman between 20 and 29 years of age (mean age 24.9) participated in the study. All were right-handed native speakers of German with normal or correctedto-normal visual acuity. They were paid for their participation.
4.2
Resolving complex anaphors
Stimuli
In order to find out whether the understanding of complex anaphorical relations makes a difference in comparison to noun phrase anaphorical relations four minitexts consisting of two sentences each were constructed. The conditions were Cond 1 NP-anaphorical expression (definite)(11) Cond 2 complex-anaphorical expression (definite) (12) Cond 3 critical item of cond 1 neutrally introduced (indefinite) (13) Cond 4 critical item of cond 2 neutrally introduced (indefinite) (14) The first two conditions started with identical first sentences, but the second sentences started either with a concrete (NP anaphor) or an abstract noun (complex anaphor). The experimental items were exactly these nouns, so it was not necessary to continue the sentences in the same way. In that manner it was possible to construct meaningful text passages. Nevertheless, we used the same syntactical parallel pattern for condition one and two as well as for conditions three and four. The text passages for the first two conditions start always with obwohl (although), the text passages for the last two conditions begin with als (when). (11) Obwohl die Titanic als unsinkbar galt, ging sie unter. Dieses Schiff war ein fahrendes Luxushotel. Although the Titanic was said to be unsinkable, it went down. This ship was a swimming luxury hotel. (12) Obwohl die Titanic als unsinkbar galt, ging sie unter. Dieses Unglück wurde sogar verfilmt. Although the Titanic was said to be unsinkable, it went down. This accident was even filmed.
For each such pair of conditions, two additional text passages were constructed in which the critical lexemes (in 11, ship and in 12, accident) were introduced in a neutral and indefinite manner. On the one hand, this allowed us to record the data of “plain” concrete and abstract nouns. On the other hand, these conditions were necessary in order to make sure that semantic relations between the referential and the anaphoric expression do not falsify the data (see Kulik et al. 2004). (13) Als Hanna ins Ferienhaus kam, öffnete sie die Fenster weit. Ein Schiff war am Horizont zu sehen. When Hanna entered the holiday house, she opened the windows wide. A ship could be seen on the horizon. 7. “Neutral” means that there was neither an anaphorical nor a conceptual relation between the two sentences.
Konstanze Marx, Ina Bornkessel-Schlesewsky and Matthias Schlesewsky
(14) Als Marcel aus dem Büro kam, gab es eine Menschenansammlung. Ein Unglück war passiert. When Marcel left his office, there was a crowd. An accident had happened.
60 blocks consisting of four conditions each were constructed altogether. In addition, we prepared 30 filler items which were totally senseless and without any logical context. 4.3
Norming study
Methods and Materials All items were tested via a questionnaire in order to choose the best candidates for the ERP study. There were four versions of the questionnaires in order to make sure that each person would only read one condition of a particular block. Subjects had to judge the acceptability of text passages on a seven-stage scale (one meant “absolutely acceptable” while seven meant “completely unacceptable”). Participants 48 undergraduate students (34 female) of the University of Jena between 19 and 30 years of age (mean age: 22.5) participated in the norming study. It took the students approximately 15 minutes to fill out the questionnaires. The participants did not receive any payment. Results and data selection All conditions with a mean higher than 3 were excluded. In a next step we calculated the differences between the means of the conditions that made up one block. Whenever the difference exceeded one, we discarded that specific block. In that manner 40 blocks for the final experimental sessions were selected. The filler passages could be used without exception (mean: 6.24), they also served as a model for 50 additionally constructed filler passages. In the end 40 blocks consisting of four conditions each and 80 unacceptable filler items, altogether 240 test passages, were tested. By randomizing two lists were created with different orders to avoid any processing effect that would be
8. a) Obwohl es gesund ist, isst Maria Haferflocken. Diese Frau sitzt im Schuh. Although they are healthy, Mary eats rolled oats. This woman is sitting in a shoe. b) Obwohl das Paar tanzte, bewegte es sich. Dieser Gedanke schimpfte über den Flur. Although the couple was dancing, they moved. This thought grumbled trough the hallway. c) Als die Sonne schien, war es dunkel. Ein Leopard hielt einen Vortrag. When the sun was shining, it was dark. A leopard was giving a lecture. d) Als Patrick in eine neue Stadt zog, hatte er keine neue Adresse. Ein Rat diskutierte im Plattenladen. When Patrick moved to another town he did not have a new postal address. Discussed in a record store. 9. We are grateful to Prof. Dr. Monika Schwarz-Friesel and Prof. Dr. Evyatar Friesel who allowed us to conduct this survey during their lectures.
Resolving complex anaphors
caused by the particular order of the stimuli items. The lists were subdivided into six presentation blocks of 40 trials each. 4.4
ERP Recording
EEG data were collected from 64 Ag/AgCl scalp electrodes mounted in a standard electrode cap (Elektro Cap International), a system that adheres to the standard international 10-20 system for electrode positioning of the American Electroencephalographic Society.10 Artefacts that result from ocular movement must be detected and discarded prior to data analysis. To this end, vertical and horizontal eye movements and blinks were monitored by means of two sets of additional electrode pairs, which were placed above and below the left eye (EOGV) and at the outer canthi of the left and right eye (EOGH). 4.5
Procedure
The passages were presented visually. Subjects were instructed to read the text passages carefully and to decide whether the texts made sense and were correct German texts. The texts did not have to be true (see 15). (15) Matthäus wird neuer Bundestrainer. Der frühere Nationalspieler hat einen Vertrag bis 2007 unterschrieben. Matthäus is going to be the new coach of the German national soccer team. The former national soccer player has signed a contract till 2007.
As a second task they were confronted with a single word and had to decide whether this word had been presented in the previous text or not. The minitexts (consisting of two sentences) were presented segmentally. These tasks were used to make sure that subjects were attending to the minitexts and to examine how acceptable participants would judge the passages to be under critical experimental conditions. Furthermore the subjects were asked not to move or blink during the presentation time of the texts. They were allowed to blink during the verification tasks and between the blocks, where they were given short breaks. After reading the instructions the subjects were prepared for the experimental session (application of the electrodes). The resistance of the electrodes was kept below 5 KOhm. 10. ERPs were recorded from the following positions: FP1, FPZ, FP2, AF7, AF3, AFZ, AF4, AF8, F9, F7, F5, F3, FZ, F4, F6, F8, F10, FT9, FT7, FC5, FC3, FCZ, FC4, FC6, FT8, FT10, T9, T7, C5, C3, CZ, C4, C6, T8, A2, TP9, TP7, CP5, CP3, CPZ, CP4, CP6, TP8, TP10, P9, P7, P5, P3, PZ, P4, P6, P8, P10, PO7, PO3, POZ, PO4, PO8, O1, OZ and O2. ERP data were referenced to the left mastoid. C2 served as ground.
Konstanze Marx, Ina Bornkessel-Schlesewsky and Matthias Schlesewsky
A practice session helped the subjects to get used to the modalities. It consisted of twelve minitexts, four of which were not acceptable. The text passages were presented in six blocks of 40 passages each (80 passages were not acceptable). One block lasted about ten minutes. At the beginning of each text passage, a fixation star appeared in the centre of the screen for 300msecs. Each segment was presented for 400msecs (segments longer than eight letters appeared for 500msecs). The breaks between each presented segment lasted 100msecs. There was an interval of 500msecs between the last word of the first sentence and the first word of the second sentence. 500 msecs after the last word of the second sentence three question marks appeared on the screen. They indicated that the subject should decide whether the minitext was an acceptable German utterance. 500msecs after the subject’s reaction or 2000msecs after the question marks had appeared (in case the subject did not answer) the verification word showed up for 2000msecs maximum. The next stimulus item was presented a total of 1000msecs after the verification task. The entire session lasted approximately 150 to 180 minutes, including electrode application and removal. 4.6
Results
Data of three subjects (two of them male) had to be excluded from the final analysis due to extensive EEG recording artefacts. Therefore the results were based on 23 subjects. 4.6.1 Behavioural Data Reaction time in msec and percentage of errors are presented in table 1.11
cond 1 cond 2 cond 3 cond 4 FI
RT T1
RT T2
Er T1
Er T2
363 359 411 364 352
806 789 760 802 815
.77 1.35 1.73 1.06 2.88
6.25 3.37 6.92 8.65 7.12
Figure 1. Behavioural data: Critical item by condition
11. cond 1 = NP-anaphor (introduced as definite concrete noun), cond 2 = complex-anaphor (introduced as definite abstract noun), cond 3 = indefinitely introduced concrete noun, cond 4 = indefinitely introduced abstract noun, FI = filler passages (unacceptable); RT T1 = mean reaction time probe detection task, RT T2 = reaction time lexical recognition task, Er T1 = mean errors in probe detection task, Er T2 = mean errors in lexical recognition task.
Resolving complex anaphors
There is no significant difference in the error or reaction times. The acceptability of all items was high. It took the participants approximately 50 msec longer to come up with a decision only when it came to decide whether a sentence introduced by an indefinite concrete noun (condition 3) was acceptable. Only experimental items that elicited a correct response to the probe detection task as well as to the lexical recognition task were included in the data analysis. To conclude, behavioural data showed general attention, items that were answered incorrectly were discarded from the final analysis. 4.6.2 ERP data Figure 2 displays the grand average event-related brain potentials for all conditions. The most striking effect is a large negative deflection emerging around 400msec after the critical stimulus (indefinite introduction of concrete noun) in condition 3 (dotted line).
Figure 2. Grand average ERPs of all conditions
Konstanze Marx, Ina Bornkessel-Schlesewsky and Matthias Schlesewsky
The grand average ERP waveforms for NP anaphoric expressions (condition 1) and complex anaphoric expressions (condition 2) are illustrated in figure 3. The two conditions do not appear to differ from one another.
Figure 3. Grand average ERPs elicited by the NP anaphoric expression and the complex anaphoric expression
The statistical significance of the concreteness effect was assessed by repeated measures ANOVA. The factors were concreteness (abstract vs. concrete nouns) and definiteness (definite vs. indefinite nouns). From the perspective of this factorisation, the critical complex anaphor condition was defined as abstract-definite. The statistical analysis showed a significant main effect of concreteness (F (1,22) = 18.37, p< 0.001) and a significant interaction between concreteness and definiteness (F (1,22)= 6.34, p< 0.02). This interaction indicates that definiteness modulates the processing of the distinction between concrete and abstract nouns. To examine this point further, we compared the differences between concrete and abstract nouns for each of the levels of the factor definiteness (i.e., separately for definite and indefinite nouns). These planned comparisons revealed a significant
Resolving complex anaphors
effect of concreteness for indefinite nouns (F(1,22) = 17.51, p< 0.001), but no significant difference between concrete and abstract definite nouns. 4.7
Discussion
In principle there are two results: – N400 effect for indefinite introduced (non-anaphoric) concrete nouns – no effect for complex anaphors Variable Definite Indefinite
concrete
abstract
no effect NP-anaphor significant N400
no effect complex anaphor no effect condition 4
Figure 4. Overview ERP findings
The main effect found in this experiment is a concreteness effect. Concrete nouns (introduced as indefinite NPs) show a significant N400-effect, abstract nouns (introduced as indefinite NPs) do not. These results are in line with previous studies showing increased N400 amplitudes for concrete vs. abstract nouns (Holcomb et al. 1999). But this concreteness effect interacts with definiteness in that the way in which concrete nouns are “anchored” within the discourse is crucial. More importantly, the postulated complexation process did not reflect a neurophysiological correlate. There are three potential explanations for this phenomenon: 1. There is no anaphorical complexation process (a complex discourse entity is already established by the antecedent). 2. There is an anaphorical complexation process, but it has no measurable physiological correlate. 3. There is an anaphorical complexation process and it has a physiological correlate, but the correlative cannot be found by this experiment design. The last explanation seems to be the most promising. Since previous work (Schwarz 2000, Consten et al. in this volume) has shown that complex anaphors can motivate the process of establishing a new referent, explanation 1 appears unlikely. While explanation 2 cannot be excluded unequivocally, a number of studies have shown that the ERP method is sensitive to anaphoric processing (see section 3). In particular, as argued by Burkhardt (2005), the establishment of a new referent can be associated with ERP activity. One would therefore have to conclude that the absence of a measurable electrophysiological effect for complex anaphors is a specific property of this particular anaphoric phenomenon.
Konstanze Marx, Ina Bornkessel-Schlesewsky and Matthias Schlesewsky
Turning now to explanation 3, this type of situation could have arisen on account of the following factors: 1. demonstrative noun phrases where pronouns are most likely (cond 1) 2. variability of semantic relations (cond 1) 3. subject-object variability (cond 1) 4. different degrees of resolving difficulty (cond 2) 1. One striking difficulty in constructing the test items had to do with the fact that complex anaphors usually take the form of demonstrative noun phrases. In order to create a basis for comparison, NP-anaphors had to be demonstrative noun phrases as well, which forced us to use such demonstrative NPs in the test items of condition 1 in positions where readers would expect a pronoun. But this strategy calls for a careful design of test items in order to not to bring in the so-called “disjoint reading effect” (ARIEL 2001): Full NP anaphors which refer to referents of high accessibility make readers add a referent to the mental discourse representation instead of recognizing the intended coreference. Thus we had to rule out that the test items in condition 1 might motivate the readers to activate a new referent before they realised that they only had to reactivate an already introduced referent. This would involve additional cognitive effort, since condition 1 would not be as easy to resolve as needed for the compatibility of both processes. We tried to fence off the disjoint reading effect in two ways: First, we constructed test items which started off with a subordinate clause and thus sounded more natural. Second, we used the norming study (see section 4.3) in order to ensure a high degree of acceptability. However, the results of our study might point to the fact that these attempts were not completely successful in ruling out the disjoint reading effect. 2. In condition 1 the semantic relations between antecedent and anaphoric expressions changed from time to time. While there are no indications yet that different semantic relations might affect the degree of resolution difficulty, this variability might nevertheless have made resolution more difficult. 3. We also need to take into account the syntactic role of the antecedent. Equal shares of NP anaphors in condition 1 referred to the subject and to the object of the preceding sentence. When analysing the data in a first step we did not distinguish these two groups. But there are hints in the literature which suggest differences in anaphora resolution depending on whether the anaphoric expression refers to the object or to the subject of the preceding sentence in that anaphoric expressions referring to the subject are resolved more easily (Wykes 1981, Corbett and Chang 1983, Müsseler and Rickheit 1990). Our extended analysis incorporates this idea (4.8).
Resolving complex anaphors
4. The complex anaphoric expressions in condition 2 referred to states of affairs, events, processes as well as to facts or propositions (see Consten and Marx 2006). There might be differences concerning the cognitive effort and strategies involved in resolving these different types of complex anaphors. We could not integrate these considerations in the present study. An experiment focussing on this question is in progress. 4.8
Extended ERP analysis
In the extended analysis, we investigated the question of whether a difference between the processing of NP anaphors in comparison to complex-anaphors might become obvious as soon as test items from condition 1 were separated according to the syntactic function of antecedents (see Wykes 1981, Corbett and Chang 1983). Material We separated the material into two groups depending on whether the NP-anaphor in condition 1 referred to the subject or to the object of the preceding sentence and focussed on the resulting grand averages. Results: ERP data Figure 5 shows the curves which resulted from the grand averages when reading condition 1 (antecedent = subject) and condition 2 (complex anaphor). For condition 2 a late positivity (P600) can be recognised. As figure 6 shows, there is no such effect when complex anaphoric processing is compared to NP anaphors which relate to the object of the preceding sentence. The statistical analysis showed no significant effect for one of the factors Region12, Type of Anaphor13 or Type of Antecedent14 as long as all regions of interest were integrated in the ANOVA. This might be due to reduced statistical power because of the rearrangement of the considered data. Nevertheless, results from Burkhardt (2005) for indirect anaphors motivated a closer look on possible interactions between the factors mentioned above.
12. Where does the effect take place? 13. Which kind of anaphoric expression is used; concrete anaphoric expression referring to an NP in condition 1 or abstract anaphoric condition referring to a sentence in condition 2? 14. Does the anaphoric expression in condition 1 refer to the subject or the object of the prior sentence?
Konstanze Marx, Ina Bornkessel-Schlesewsky and Matthias Schlesewsky
Figure 5. NP-anaphor vs. complex anaphor when antecedent for NP anaphor is subject
We found a marginal interaction between the factors Type of Anaphor and Type of Antecent in the right posterior region (F(1,22)=3,86, p< 0.07). Separate analyses for passages with subject vs. object antecedents showed a significant effect of concreteness in the Subject-Type (F(1,22)= 4,69, p< 0.05), which was due to an increased positivity for abstract anaphorical expressions in condition 2. There was no such difference when the anaphor referred to an object.
Resolving complex anaphors
Figure 6. NP-anaphor vs. complex anaphor when antecedent for NP anaphor is object
Discussion The reanalysis of the material corroborated the assumption that complex anaphoric processing and therefore the establishment of a new referent differentiates from NP-anaphoric processing. Complex anaphors–in contrast to NP-anaphors referring to the subject of the preceding sentence–elicit a late positivity. This component was recently interpreted as indicator for cognitive effort while establishing a new referent in the mental discourse representation. Our findings are in line with the assumption that NP-anaphors which relate to the subject of the prior sentence are easier to resolve than NP-anaphors which relate to the object of the prior sentence. We did not find a difference between complex anaphoric processing and NP-anaphor in the latter case. This might be due to the fact that the cognitive effort involved in resolving NP-anaphors which relate to object antecedents and the process of establishing a new discourse referent (complexation process) might involve similar neurophysiological processing correlates.
Konstanze Marx, Ina Bornkessel-Schlesewsky and Matthias Schlesewsky
5. Conclusion Complex anaphors motivate the integration of a new abstract referent into the mental discourse representation. The present study investigated the question of whether this postulated complexation process shows up in ERP data. In order to find out whether the referent establishing process is reflected in a prominent ERP component, we compared complex anaphors with NP-anaphors. In a first analysis we did not find a significant effect, which might be due to the variability in our first condition. An extended reanalysis of our material for complex anaphors in contrast to NP-anaphors relating to the subject of the preceding sentence showed a late positivity. This provides converging support for the assumption that the cognitive cost of processing complex anaphors is higher than the cognitive cost of processing NP-anaphors, because the former calls for the establishment of a new referent in a mental discourse representation.
Acknowledgements This experiment was conducted at the Max Planck Institute for Human Cognitive and Brain Sciences in Leipzig, Germany. Thanks to Ina Koch for her support during data acquisition.
References Ariel, M. 2001. Accessibility theory: An overview. In Text Representation. Linguistic and psycholinguistic aspects, T. Sanders, J. Schilperoord and W. Spooren (eds), 13–35. Amsterdam: John Benjamins. Asher, N. 1993. Reference to Abstract Objects in Discourse. Dordrecht: Kluwer. Bornkessel, I., McElree, B., Schlesewsky, M. and Friederici, A. 2004. Multidimensional contributions to garden path strength: Dissociating phrase structure from case marking. Journal of Memory and Language 51: 495–522. Burkhardt, P. 2005. Resolving bridging dependencies online: Evidence for given and new properties of bridged definite DPs. In Proceedings of Constraints in Discourse’, June 3-5,.2005, Dortmund. Coles, M. and Rugg, M. 1995. Event-related brain potentials: An introduction. In M. Rugg and M. Coles (eds), Electrophysiology of Mind: Event-related brain potentials and cognition, 1. New York NY: OUP. Consten, M. and Marx, K. 2006. Komplex-Anaphern–Rezeption und textuelle Funktion. In Text — Verstehen. Grammatik und darüber hinaus, H. Blühdorn, E. Breindl, and U.H. Wassner (eds), 375–379. Berlin: Walter de Gruyter. Corbett, A. and Chang, F. 1983. Pronoun disambiguating: Accessing potential antecedent. Memory & Cognition 11: 283–294.
Resolving complex anaphors
Cornish, F. 1986. Anaphoric Relations in English and French–A Discourse Perspective. London: Croom Helm. Donchin, E., Ritter, W. and McCallum, W. 1978. Cognitive psychophysiology: The endogenous components of the ERP. In Event-Related Brain Potentials in Man, E. Callaway, P. Tueting, and S. Koslow (eds), 349. New York NY: Academic Press. Fischler, I. 1990. Comprehending language with event-related potentials. In Event-Related Brain Potentials: Basic issues and applications, J. Rohrbaugh, R. Parasuraman, and Johnson, R. Jr. (eds), 165–177. New York NY: OUP. Francis, G. 1994. Labelling discourse: An aspect of nominal-group lexical cohesion. In Advances in Written Text Analysis, C. Malcom (ed.), 83–101. London: Routledge. Friederici, A. 2003. Neurobiologische Grundlagen der Sprache. In Neuropsychologie, H.-O. Karnath and P. Thier (eds), 367–377. Berlin: Springer. Frisch, S. 2000. Verb-Argument-Struktur, Kasus und thematische Interpretation beim Sprachverstehen. PhD dissertation, Universität Potsdam. Garnsey, S. M. 1993. Event-related Brain Potentials in the Study of Language: An Introduction. Language and Cognitive Processes 8: 337–356. Hahne, A. 1997. Charakteristika syntaktischer und semantischer Prozesse bei der auditiven Sprachverarbeitung. Leipzig: MPI Series. Holcomb, P.,.Kounios, J., Anderson, J. and West, W. 1999. Dual-coding, context availability, and concreteness effects in sentence comprehension: an electrophysiological investigation. Journal of Experimental Psychology: Learning Memory and Cognition 25:721–742. Kulik, S., Bornkessel, I. and Schlesewsky, M. 2004. The on-line establishment of hyperonymic anaphorical relations. Poster, 17. CUNY Conference on Human Sentence Processing. University of Maryland, College Park. Kutas, M. 1993. In the company of other words: Electrophysiological evidence for single-word and sentence context effects. Language and Cognitive Processes 8: 533–572. Kutas, M., Van Petten, C. and Kluender, R. 1994 Psycholinguistics electrified II. In Handbook of Psycholinguistics, M.A Gernsbacher (ed.). San Diego CA: Academic Press. Müsseler, J. and Rickheit, G. 1990. Komplexbildung in der Textverarbeitung. Die kognitive Auflösung pluraler Pronomen. Zeitschrift für Psychologie 198: 69–81. Roehm, D., Schlesewsky, M., Bornkessel, I., Frisch, S. and Haider, H. 2004. Fractionating language comprehension via frequency characteristics of the human EEG. NeuroReport 15: 409–412. Schmid, H.-J. 2000. English Abstract Nouns as Conceptual Shells. From corpus to cognition. Berlin: Mouton de Gruyter. Schwarz, M. 2000. Indirekte Anaphern in Texten. Studien zur domänengebundenen Referenz und Kohärenz im Deutschen. Tübingen: Niemeyer. Schwarz, M. 2001. Establishing coherence in text. Conceptual continuity and text-world models. Logos and Language 2: 15–23. Schwarz-Friesel, M., Consten, M. and Marx, K. 2004. Semantische und konzeptuelle Prozesse bei der Verarbeitung von Komplex-Anaphern. In Flexibilität und Stabilität, I. Pohl (ed.), 67–86. Frankfurt am Main: Peter Lang. Streb, J. 2000. Hirnelektrische Korrelate der Verarbeitung anaphorischer Verweise. PhD dissertation, Universität Marburg. Streb, J., Hennighausen, E. and Rösler, F. 2004. Different anaphoric expressions are investigated by event-related brain potentials. Journal of Psycholinguistic Research 33: 175–201. Wykes, T. 1981. Inference and children’s comprehension of pronouns. Journal of Experimental Child Psychology 32: 264–279.
Index A abstract object 82, 86-93, 100, 147, 202-205, 259, 260, 276 see also complex anaphor, situation accessibility 3, 6-11, 17, 22, 53-54, 86-87, 94-97, 103-105, 123-142, 226, 272 activation 3-14, 34-35, 53, 94-96, 124, 126, 129-135, 230-235 addressee 133-135, 166 see also computer addressee, human addressee afterthought 167-168, 174, 180 Altmann 165, 166, 168, 233 ambiguous complex anaphor 82, 93, 100 anaphoric chain 107, 117, 124126, 209, 216-219 anaphoric relation 43, 71, 107, 124–129, 136, 199, 204, 208, 213, 260, 264, 277 anchor 3, 8-11, 14-17, 97-99, 105, 117-118, 142, 226, 237 see also antecedent-trigger antecedentless anaphor see indirect anaphor antecedent-trigger 7-8, 22, 24, 27, 29 Ariel 6, 53, 124, 131-135, 272 Asher 82, 86, 107, 111, 113, 200, 202-203 assertion 44, 110, 117 associative anaphor 7, 14, 24, 32, 41 associative referent 27, 32-34 see also indirect anaphor attributive expression 57-58, 63-65 Averintseva-Klisch 165 B Binding Theory 145, 208, 242 Bornkessel-Schlesewsky 259 Bosch 6, 136, 145, 161 brain 227-230, 236, 245, 261
Brandt 187, 188, 190, 196 bridging anaphor see indirect anaphor Burkhardt 241, 242, 245, 264, 273 C Centering Theory 6, 130-139, 145, 161 central referent 21-22, 27-28, 32-33 Chafe 6, 124 characterization relation 203 Clark 7, 53, 57, 66 coargument reflexive 241, 243248, 254, 256 cognitive communication impairment 228 cognitive process 3, 4, 6, 8, 11, 13, 17, 18, 24, 51, 109, 112, 117, 151, 225-227, 229, 232, 235-236, 244, 245, 261-264 see also comprehension, resolution cognitive theory 14, 103, 112, 118 coherence 4, 11, 22, 53, 54, 81, 82, 96, 100, 109, 116, 118, 123, 132, 137, 142, 174, 207, 225-227, 230-233, 236-237 cohesion 226, 230-233 communicative situation 54, 56, 61, 67, 70, 73, 79, 112, 130 complementary anaphor 24, 32 complementary referent 27, 32-34 completion study 158-160 complex anaphor 13, 81-100, 105, 186, 259-260, 265, 270-276 complexation 81, 86, 87, 90, 93, 100, 259, 260, 271, 275, 276 comprehension 4, 6, 7, 8, 103, 108, 112, 153, 158, 225, 226, 227, 229, 233, 235, 236, 237, 245, 259 see also resolution computational linguistics see machine learning
computer addressee 50-56, 66 see also human addressee computer mediation 49-55 Comrie 146 conceptual structure 54, 82, 83 concreteness effect 270, 271, 277 Consten 7, 69, 81, 94, 196 Conte 72, 79, 97, 100 continuative clause 187 continuity 3-12, 82, 96-99, 115, 225 contract of communication 44 coreference 8, 97, 209, 263 coreference resolution 209-219 Cornish 8, 21, 22, 27, 94, 97 corpus 83-85, 147-151, 161, 207, 213 D Davidson 75, 86 definiteness see definite description, definite reference definites see definite description definite article see definite determiner definite description 5, 6, 7, 11, 14, 16, 33, 39, 53, 84, 123-142, 146, 171, 173, 181, 207, 270, 271 definite determiner 8, 11, 24, 39, 46, 103, 105, 107, 146, 147, 230 definite expression see definite description definite noun phrase see definite description definite reference 4, 8, 16, 69, 70, 170, 172 deixis 69-79 demonstratives 13, 15, 24, 35, 39, 46, 71, 84, 87, 97, 133, 145-163, 272 demonstrative pronoun 84, 87, 145-163 see also pronoun determiner 34, 35, 39, 46, 84
Anaphors in Text direct anaphor 10-18, 22-23, 9799, 109, 226, 264 see also indirect anaphor disambiguation 94, 166 discourse deixis 69-79, 83 discourse model 166, 233 see also text world model discourse object 82, 90, 93, 96, 99 discourse referent 39, 94, 128, 136-140, 166-181, 202-204, 275 see also discourse topic discourse representation 22, 27, 124, 128, 131, 132, 135, 172, 243, 256, 259-264, 275 Discourse Representation Theory, see DRT discourse structure 123, 124, 128, 132, 134, 135, 137, 144, 167, 258 discourse topic 161, 165, 166, 174, 175, 176, 177, 178, 179, 180, 181, 182 see also discourse referent see also topic discourse-cognitive 22 DRT 86, 94, 124, 131-139, 145, 183-184, 199, 200, 202, 208 E emotion 234 endocentric 37, 41, 42 epistemic contract 44 event 74-77, 81-99, 147, 190, 193, 196, 197, 199, 203 see also abstract object, complex anaphor, situation executive function 232, 235 exocentric 37, 41, 42 ERP see N 400 F fact 81-99, 147, 197-198, 202, 203, 273 see also abstract object, complex anaphor faithful anaphor 44 Ferstl 225-240 Fillmore 70– 72, 74 finite verb predicate 42 fMRI 227-237 focus 3, 6, 10-18, 24, 26, 34, 35, 132-142, 194
Fraurud 4, 8, 82, 127 Frey 175 functional completeness 199 G Garrod 10, 24-28 germanic languages 38-44 given / new 6-18, 33, 34, 124, 131, 134 see also discourse topic, new information Givenness see given / new grammatical function 148, 150, 161, 162 grammaticalisation 45, 46 Grosz 6, 124, 137, 145, 161, 208 Gundel 6, 21, 22, 24, 28, 34, 53, 95, 124, 131-135, 146, 161 H Harweg 73 head-noun nominal anaphor 37, 38 Hegarty 95, 197 Heim 7, 8, 208, 243 Helbig 185, 187, 196 Holler 183, 194 human addressee 49, 50, 56, 58, 66, 67 see also computer addressee I identical anaphor 38 implicit anaphor see indirect anaphor indexical form 34 indirect anaphor 3-18, 21-35, 97, 98, 105, 106, 117, 128, 141, 142, 225-230, 236, 237, 264, 273 indirect complex anaphor 97-99 indirect metaphorical anaphor 105, 106, 116-118 indirect pronominal anaphor 8, 24, 29 indirect reference see indirect anaphor inference 5, 7, 10, 11, 14, 17, 22, 23, 97, 98, 219, 225-237 infidèle 40 see also indirect anaphor information structure 150, 162, 194 see also given / new
interpretation preference 146, 154, 162 J Just 229, 230, 233, 236 K Kaiser 146, 161 Kamp 124, 136, 145, 200, 208 Katz 145 Klein 73, 77, 78, 178 Knees 8, 81, 186, 196 knowledge 5-17, 22, 26, 43, 44, 50, 51, 54, 56, 72, 97, 105, 109, 110, 112, 113, 127, 128, 130, 142, 154-162, 208, 218, 219, 225-227, 230-237 Kulik 263-265 L Lakoff 116 language types 38, 43 Lenz 69 Levinson 33, 70, 71, 242 lexical anaphor 13, 26, 33, 34, 37, 38, 39, 46, 84, 89, 165 lexicalisation 40, 41 lexically varied anaphor 38, 39 local deictics 72, 79 locative expression 63-65 logophor 241-256 Lundquist 37, 40, 41 Lyons 71, 83 M machine learning 207-220 Maes 6, 49, 54, 57 Maienborn 86, 87 main structure 165, 166, 178, 179, 180, 181 Marcelis 49, 58, 62 Marx 11, 98, 236, 259 Mason 229, 230, 233, 236 memory 5, 6, 10, 11, 13, 14, 24, 35, 95, 135, 235, 252 mental representation, see discourse representation, text world model meta-communication 69, 70, 78, 79 metaphorical anaphor 103-118 minimalist hypothesis 226 Müller 217, 219
N N400 231, 241, 246, 247, 253, 255, 256, 261-264, 271 Nass 52, 53 neuroimaging 225, 229, 233, 236, 237 neurolinguistics, see fMRI, N400, neuropsychology neuropsychology 227 see also fMRI, N400 new information 11-14, 18, 37, 43, 104, 105, 127 see also given / new non-aphasic language disorder 228 non-finite verb phrase 43, 45 non-integrated clause 199-201 non-restrictive relative clause 188, 201 non-subject antecedent 145, 162 non-subject pronoun 28, 30, 32, 34 nucleus 44-45 O Oberlander 174 overspecification 50, 57-59, 64-66 P paratactic sentence ordering 43 peripheral referent 21, 27-33, 35 personal pronoun 21-35, 84, 95, 97, 145-162, 176, 207 see also pronoun PET, 227 plausibility 11, 13, 15, 17, 18, 110, 112, 141 Poesio 127 prefrontal cortex 227, 228, 230, 231, 235 presupposition 43, 44 primary predicate 42, 45 principle of economy 244, 256 process (ontological) 74, 75, 83, 86, 88-90, 92, 273 see also abstract object, complex anaphor, situation processing see cognitive process, comprehension, processing costs, resolution
Index processing costs 28, 103, 116, 118, 245-247, 256, 259, 276 progression 81, 82, 96, 99, 100 pronominal anaphor 21, 22, 24, 29, 38, 40, 46, 84, 108, 135, 145, 166, 241 pronominal complex anaphor 84, 88 pronoun 21, 39, 108, 114, 124126, 132-134, 137, 141-142, 165, 172, 208, 241-256 see also demonstrative pronoun, personal pronoun, pronoun resolution pronoun resolution 208, 215219, 246, 256 proposition 42-44, 46, 80, 82, 83, 86, 92, 188-190, 199, 202 see also abstract object, complex anaphor psycholinguistic experiment 109, 115, 145, 152, 245 Q quaestio 178, 179 R recency 161 referential ambiguity 16, 40 referential behavior 49, 50, 54-57, 128 referentially disjoint 39 reflexive language 69-71 Reflexivity Theory 242 Reinhart 175, 182, 242, 243, 246, 249 resolution 4, 6, 8, 14, 15, 17, 18, 32, 82, 93, 94, 100, 109, 112, 115118, 136, 207-209, 229, 245-247, 263, 264, 272 see also comprehension resolution model 82, 94 see also discourse model, text world model Reuland 242, 243, 244, 249 rheme 81, 82, 96, 99, 147, 161, 162 rhetorical relation 37, 38, 44-46 right hemisphere 228, 230, 236 right peripheral NP 166, 167, 172, 174, 175 romance languages 38, 40 root clause properties 186, 194, 205
S salience 95, 123-142, 146, 161, 215 salience change 140 salience spreading 123, 124, 140-142 Sanford 10, 21, 24, 25, 26, 27, 28 satellite 45 Schlesewsky 259 Schwarz-Friesel 3, 4, 7, 8, 11, 15, 81, 94–99, 105, 109, 110, 112, 116, 117, 128, 226, 230, 233, 237, 259, 260, 264 Searle 76 self-paced reading task 25, 28, 152, 153 semantic object 124, 183, 194, 198, 202 semantic type 9, 128, 183, 198, 199, 202 semantics-pragmatics-interface 103, 109, 112, 118 semi-active 27, 33 sentence processing 242, 245 sentence topic 165, 174-181 sentence-related 82, 187, 193, 196, 201 short-term memory 6, 10, 11, 24, 135 see also memory side structure 178-181 Siebörger 225, 232, 235, 236 situation 5, 27, 29, 35, 75, 76, 77, 78, 79, 86, 154, 260, 272 see also abstract object, complex anaphor, event, process, state situation model 230, 233-236 see also discourse model, text-world-model situational anaphor 82 see also complex anaphor Skirl 103 Soon 209, 210, 212, 215, 219 specifying anaphor 12, 105, 106, 113 speech act 66, 76, 77, 187, 192, 194 state 74, 75, 77, 81, 82, 86, 88-94, 199 see also abstract object, complex anaphor, situation state of affair 188, 189, 193, 273 see also situation Strube 141, 207, 215, 219
Anaphors in Text subordinate syntactic ordering 43 Syntax-Discourse Model 242244, 256 T temporal lobe 225, 228, 230, 231, 234, 236, 237 text linguistics 112, 207 text world model 4-6, 11, 12, 1417, 81, 94, 96, 97, 99, 103, 104, 112, 113, 115-118 see also discourse model textual function 46, 81, 82, 100 thematic 17, 81, 82, 96, 99, 178 see also given / new time deixis 69, 73-75, 78 Tomasello 52
topic 35, 38, 71, 73, 78, 95, 107, 113, 126, 134, 147, 148, 161, 162, 174-181 see also discourse topic topic continuity 3, 4, 6 topic time 78 Trueswell 146, 161
W Webber 83, 124, 209 wh-anaphor 202-204 wh-relative clause 183-205 working memory 6, 11, 24, 235, 252 see also memory
U Umbach 145 underspecification 4, 54, 57, 97, 112, 241, 244, 245 unfaithful anaphor 37-46
Z Zifonun 147, 162, 166, 172, 188, 196
V Vendler 75, 76, 86, 202 Verheyen 49 von Heusinger 6, 123, 132, 140 von Stutterheim 178, 179
Studies in Language Companion Series A complete list of titles in this series can be found on the publishers’ website, www.benjamins.com 89 Stark, Elisabeth, Elisabeth Leiss and Werner Abraham (eds.): Nominal Determination. Typology, context constraints and historical emergence. Expected August 2007 88 Ramat, Paolo and Elisa Roma (eds.): Europe and the Mediterranean as Linguistic Areas. Convergencies from a historical and typological perspective. xxv, 351 pp. + index. Expected August 2007 87 Verhoeven, Elisabeth: Experiential Constructions in Yucatec Maya. A typologically based analysis of a functional domain in a Mayan language. 2007. xiv, 380 pp. 86 Schwarz-Friesel, Monika, Manfred Consten and Mareile Knees (eds.): Anaphors in Text. Cognitive, formal and applied approaches to anaphoric reference. 2007. xvi, 282 pp. 85 Butler, Christopher S., Raquel Hidalgo Downing and Julia Lavid (eds.): Functional Perspectives on Grammar and Discourse. In honour of Angela Downing. xxiv, 468 pp. + index. Expected May 2007 84 Wanner, Leo (ed.): Selected Lexical and Grammatical Issues in the Meaning–Text Theory. In honour of Igor Mel'čuk. 2007. xviii, 380 pp. 83 Hannay, Mike and Gerard J. Steen (eds.): Structural-Functional Studies in English Grammar. In honour of Lachlan Mackenzie. 2007. vi, 393 pp. 82 Ziegeler, Debra: Interfaces with English Aspect. Diachronic and empirical studies. 2006. xvi, 325 pp. 81 Peeters, Bert (ed.): Semantic Primes and Universal Grammar. Empirical evidence from the Romance languages. 2006. xvi, 374 pp. 80 Birner, Betty J. and Gregory Ward (eds.): Drawing the Boundaries of Meaning. Neo-Gricean studies in pragmatics and semantics in honor of Laurence R. Horn. 2006. xii, 350 pp. 79 Laffut, An: Three-Participant Constructions in English. A functional-cognitive approach to caused relations. 2006. ix, 268 pp. 78 Yamamoto, Mutsumi: Agency and Impersonality. Their Linguistic and Cultural Manifestations. 2006. x, 152 pp. 77 Kulikov, Leonid, Andrej Malchukov and Peter de Swart (eds.): Case, Valency and Transitivity. 2006. xx, 503 pp. 76 Nevalainen, Terttu, Juhani Klemola and Mikko Laitinen (eds.): Types of Variation. Diachronic, dialectal and typological interfaces. 2006. viii, 378 pp. 75 Hole, Daniel, André Meinunger and Werner Abraham (eds.): Datives and Other Cases. Between argument structure and event structure. 2006. viii, 385 pp. 74 Pietrandrea, Paola: Epistemic Modality. Functional properties and the Italian system. 2005. xii, 232 pp. 73 Xiao, Richard and Tony McEnery: Aspect in Mandarin Chinese. A corpus-based study. 2004. x, 305 pp. 72 Frajzyngier, Zygmunt, Adam Hodges and David S. Rood (eds.): Linguistic Diversity and Language Theories. 2005. xii, 432 pp. 71 Dahl, Östen: The Growth and Maintenance of Linguistic Complexity. 2004. x, 336 pp. 70 Lefebvre, Claire: Issues in the Study of Pidgin and Creole Languages. 2004. xvi, 358 pp. 69 Tanaka, Lidia: Gender, Language and Culture. A study of Japanese television interview discourse. 2004. xvii, 233 pp. 68 Moder, Carol Lynn and Aida Martinovic-Zic (eds.): Discourse Across Languages and Cultures. 2004. vi, 366 pp. 67 Luraghi, Silvia: On the Meaning of Prepositions and Cases. The expression of semantic roles in Ancient Greek. 2003. xii, 366 pp. 66 Nariyama, Shigeko: Ellipsis and Reference Tracking in Japanese. 2003. xvi, 400 pp. 65 Matsumoto, Kazuko: Intonation Units in Japanese Conversation. Syntactic, informational and functional structures. 2003. xviii, 215 pp. 64 Butler, Christopher S.: Structure and Function – A Guide to Three Major Structural-Functional Theories. Part 2: From clause to discourse and beyond. 2003. xiv, 579 pp. 63 Butler, Christopher S.: Structure and Function – A Guide to Three Major Structural-Functional Theories. Part 1: Approaches to the simplex clause. 2003. xx, 573 pp. 62 Field, Fredric: Linguistic Borrowing in Bilingual Contexts. With a foreword by Bernard Comrie. 2002. xviii, 255 pp.
61 Goddard, Cliff and Anna Wierzbicka (eds.): Meaning and Universal Grammar. Theory and empirical findings. Volume 2. 2002. xvi, 337 pp. 60 Goddard, Cliff and Anna Wierzbicka (eds.): Meaning and Universal Grammar. Theory and empirical findings. Volume 1. 2002. xvi, 337 pp. 59 Shi, Yuzhi: The Establishment of Modern Chinese Grammar. The formation of the resultative construction and its effects. 2002. xiv, 262 pp. 58 Maylor, B. Roger: Lexical Template Morphology. Change of state and the verbal prefixes in German. 2002. x, 273 pp. 57 Mel’čuk, Igor A.: Communicative Organization in Natural Language. The semantic-communicative structure of sentences. 2001. xii, 393 pp. 56 Faarlund, Jan Terje (ed.): Grammatical Relations in Change. 2001. viii, 326 pp. 55 Dahl, Östen and Maria Koptjevskaja-Tamm (eds.): Circum-Baltic Languages. Volume 2: Grammar and Typology. 2001. xx, 423 pp. 54 Dahl, Östen and Maria Koptjevskaja-Tamm (eds.): Circum-Baltic Languages. Volume 1: Past and Present. 2001. xx, 382 pp. 53 Fischer, Olga, Anette Rosenbach and Dieter Stein (eds.): Pathways of Change. Grammaticalization in English. 2000. x, 391 pp. 52 Torres Cacoullos, Rena: Grammaticization, Synchronic Variation, and Language Contact. A study of Spanish progressive -ndo constructions. 2000. xvi, 255 pp. 51 Ziegeler, Debra: Hypothetical Modality. Grammaticalisation in an L2 dialect. 2000. xx, 290 pp. 50 Abraham, Werner and Leonid Kulikov (eds.): Tense-Aspect, Transitivity and Causativity. Essays in honour of Vladimir Nedjalkov. 1999. xxxiv, 359 pp. 49 Bhat, D.N.S.: The Prominence of Tense, Aspect and Mood. 1999. xii, 198 pp. 48 Manney, Linda Joyce: Middle Voice in Modern Greek. Meaning and function of an inflectional category. 2000. xiii, 262 pp. 47 Brinton, Laurel J. and Minoji Akimoto (eds.): Collocational and Idiomatic Aspects of Composite Predicates in the History of English. 1999. xiv, 283 pp. 46 Yamamoto, Mutsumi: Animacy and Reference. A cognitive approach to corpus linguistics. 1999. xviii, 278 pp. 45 Collins, Peter C. and David Lee (eds.): The Clause in English. In honour of Rodney Huddleston. 1999. xv, 342 pp. 44 Hannay, Mike and A. Machtelt Bolkestein (eds.): Functional Grammar and Verbal Interaction. 1998. xii, 304 pp. 43 Olbertz, Hella, Kees Hengeveld and Jesús Sánchez García (eds.): The Structure of the Lexicon in Functional Grammar. 1998. xii, 312 pp. 42 Darnell, Michael, Edith Moravcsik, Michael Noonan, Frederick J. Newmeyer and Kathleen M. Wheatley (eds.): Functionalism and Formalism in Linguistics. Volume II: Case studies. 1999. vi, 407 pp. 41 Darnell, Michael, Edith Moravcsik, Michael Noonan, Frederick J. Newmeyer and Kathleen M. Wheatley (eds.): Functionalism and Formalism in Linguistics. Volume I: General papers. 1999. vi, 486 pp. 40 Birner, Betty J. and Gregory Ward: Information Status and Noncanonical Word Order in English. 1998. xiv, 314 pp. 39 Wanner, Leo (ed.): Recent Trends in Meaning–Text Theory. 1997. xx, 202 pp. 38 Hacking, Jane F.: Coding the Hypothetical. A comparative typology of Russian and Macedonian conditionals. 1998. vi, 156 pp. 37 Harvey, Mark and Nicholas Reid (eds.): Nominal Classification in Aboriginal Australia. 1997. x, 296 pp. 36 Kamio, Akio (ed.): Directions in Functional Linguistics. 1997. xiii, 259 pp. 35 Matsumoto, Yoshiko: Noun-Modifying Constructions in Japanese. A frame semantic approach. 1997. viii, 204 pp. 34 Hatav, Galia: The Semantics of Aspect and Modality. Evidence from English and Biblical Hebrew. 1997. x, 224 pp. 33 Velázquez-Castillo, Maura: The Grammar of Possession. Inalienability, incorporation and possessor ascension in Guaraní. 1996. xvi, 274 pp.
32 Frajzyngier, Zygmunt: Grammaticalization of the Complex Sentence. A case study in Chadic. 1996. xviii, 501 pp. 31 Wanner, Leo (ed.): Lexical Functions in Lexicography and Natural Language Processing. 1996. xx, 355 pp. 30 Huffman, Alan: The Categories of Grammar. French lui and le. 1997. xiv, 379 pp. 29 Engberg-Pedersen, Elisabeth, Michael Fortescue, Peter Harder, Lars Heltoft and Lisbeth Falster Jakobsen (eds.): Content, Expression and Structure. Studies in Danish functional grammar. 1996. xvi, 510 pp. 28 Herman, József (ed.): Linguistic Studies on Latin. Selected papers from the 6th International Colloquium on Latin Linguistics (Budapest, 23–27 March 1991). 1994. ix, 421 pp. 27 Abraham, Werner, T. Givón and Sandra A. Thompson (eds.): Discourse, Grammar and Typology. Papers in honor of John W.M. Verhaar. 1995. xx, 352 pp. 26 Lima, Susan D., Roberta L. Corrigan and Gregory K. Iverson: The Reality of Linguistic Rules. 1994. xxiii, 480 pp. 25 Goddard, Cliff and Anna Wierzbicka (eds.): Semantic and Lexical Universals. Theory and empirical findings. 1994. viii, 510 pp. 24 Bhat, D.N.S.: The Adjectival Category. Criteria for differentiation and identification. 1994. xii, 295 pp. 23 Comrie, Bernard and Maria Polinsky (eds.): Causatives and Transitivity. 1993. x, 399 pp. 22 McGregor, William B.: A Functional Grammar of Gooniyandi. 1990. xx, 618 pp. 21 Coleman, Robert (ed.): New Studies in Latin Linguistics. Proceedings of the 4th International Colloquium on Latin Linguistics, Cambridge, April 1987. 1990. x, 480 pp. 20 Verhaar, John W.M. S.J. (ed.): Melanesian Pidgin and Tok Pisin. Proceedings of the First International Conference on Pidgins and Creoles in Melanesia. 1990. xiv, 409 pp. 19 Blust, Robert A.: Austronesian Root Theory. An essay on the limits of morphology. 1988. xi, 190 pp. 18 Wierzbicka, Anna: The Semantics of Grammar. 1988. vii, 581 pp. 17 Calboli, Gualtiero (ed.): Subordination and Other Topics in Latin. Proceedings of the Third Colloquium on Latin Linguistics, Bologna, 1–5 April 1985. 1989. xxix, 691 pp. 16 Conte, Maria-Elisabeth, János Sánder Petöfi and Emel Sözer (eds.): Text and Discourse Connectedness. Proceedings of the Conference on Connexity and Coherence, Urbino, July 16–21, 1984. 1989. xxiv, 584 pp. 15 Justice, David: The Semantics of Form in Arabic. In the mirror of European languages. 1987. iv, 417 pp. 14 Benson, Morton, Evelyn Benson and Robert F. Ilson: Lexicographic Description of English. 1986. xiii, 275 pp. 13 Reesink, Ger P.: Structures and their Functions in Usan. 1987. xviii, 369 pp. 12 Pinkster, Harm (ed.): Latin Linguistics and Linguistic Theory. Proceedings of the 1st International Colloquium on Latin Linguistics, Amsterdam, April 1981. 1983. xviii, 307 pp. 11 Panhuis, Dirk G.J.: The Communicative Perspective in the Sentence. A study of Latin word order. 1982. viii, 172 pp. 10 Dressler, Wolfgang U., Willi Mayerthaler, Oswald Panagl and Wolfgang Ullrich Wurzel: Leitmotifs in Natural Morphology. 1988. ix, 168 pp. 9 Lang, Ewald and John Pheby: The Semantics of Coordination. (English transl. by John Pheby from the German orig. ed. 'Semantik der koordinativen Verknüpfung', Berlin, 1977). 1984. 300 pp. 8 Barth, E.M. and J.L. Martens (eds.): Argumentation: Approaches to Theory Formation. Containing the Contributions to the Groningen Conference on the Theory of Argumentation, October 1978. 1982. xviii, 333 pp. 7 Parret, Herman, Marina Sbisà and Jef Verschueren (eds.): Possibilities and Limitations of Pragmatics. Proceedings of the Conference on Pragmatics, Urbino, July 8–14, 1979. 1981. x, 854 pp. 6 Vago, Robert M. (ed.): Issues in Vowel Harmony. Proceedings of the CUNY Linguistics Conference on Vowel Harmony, May 14, 1977. 1980. xx, 340 pp. 5 Haiman, John: Hua: A Papuan Language of the Eastern Highlands of New Guinea. 1980. iv, 550 pp. 4 Lloyd, Albert L.: Anatomy of the Verb. The Gothic Verb as a Model for a Unified Theory of Aspect, Actional Types, and Verbal Velocity. (Part I: Theory; Part II: Application). 1979. x, 351 pp. 3 Malkiel, Yakov: From Particular to General Linguistics. Selected Essays 1965–1978. With an introduction by the author, an index rerum and an index nominum. 1983. xxii, 659 pp. 2 Anwar, Mohamed Sami: BE and Equational Sentences in Egyptian Colloquial Arabic. 1979. vi, 128 pp.
1
Abraham, Werner (ed.): Valence, Semantic Case, and Grammatical Relations. Workshop studies prepared for the 12th International Congress of Linguists, Vienna, August 29th to September 3rd, 1977. xiv, 729 pp. Expected Out of print
