: ·.! i .,
.. ·.� . 'f 1:;'�
JOURNAL OF SEMANTICS
AN INTERNATIONAL j OURNAL FOR THE INTERDISCIPLINARY STUDY OF THE S...
19 downloads
445 Views
22MB 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
: ·.! i .,
.. ·.� . 'f 1:;'�
JOURNAL OF SEMANTICS
AN INTERNATIONAL j OURNAL FOR THE INTERDISCIPLINARY STUDY OF THE SEMANTICS OF NATURAL LANGUAGE MANAGING EDITOR :
PETER BoscH (IBM Germany)
RE V IE W EDITOR : BART GEURTS (IBM Germany) EDITORIAL BOA RD:
PETER HoscH (IBM Germany) SIMON C. GARROD (Univ. of Glasgow) BART GEURTS (IRM Germany) PAUL HoPPER (CarnegieMellon Univ., Pittsburgh) LAURENCE R. HoRN (Yale University) STEPHEN lsARD (Univ. of Edinburgh) HANS KAMP (Univ. of Srurrgarr) LEO G.M. NoORDMANN (Univ. of Tilburg) IZoB A. VAN DER SANDT (Univ. of Nijmegcn) PIETER A.M. SEUREN (Univ. of Nijmegen)
CONSULTING EDITORS:
It
H. E. BREKI.E (Univ. of Rcgensburg)
BARTSCH (Univ. of Amsterdam)
D. S. BREE (Univ. of Manchester)
H. H. C LAR K (Stanford University)
G. BROWN (Univ. of Cambridge)
H.-J. ErKMEYER (Univ. of Bielefeld) j. HoBBS (SRI, Menlo Park)
0. DAH L (Univ. of Stockholm) G. FAuCONNIER (Univ. of California, San Diego)
D. IsRAEL (SRI, Menlo Park)
P. N.jOHNSON-LAIRD (MRC, Cambridge)
E. L. KEENAN (Univ. of California, Los Angeles)
Sra joHN LYONS (Univ. of Cambridge) J. D. McCAwLEY (Univ. of Chicago)
E. LANG (Univ. or Wupperral)
B. RICHARDS (Imperial College. London)
W. MARSLEN-WILSON (MRC, Cambridge)
H. ScHNEI. I.E ( Ruhr Univ.. Bochum)
H. REICHGELT (Univ. of Nottingham)
M. STEEDMAN (Univ. of Pennsylvania)
A. J. SANFORD (Univ. of Glasgow)
Z. Y.
VENDLER (Univ. of California, San Diego)
A. VON STECH OW (Univ. of Konsranz)
WILKS (New Mexico Stare Univ.,
D. VANDERVEKEN (Univ. of Quebec) B. L. WEBBER (Univ. of Pennyslvania)
Las Cruces)
J. VAN BENTHEM (Univ. of Amsterdam)
D. WILSON (Univ. College. London).
ED ITOR lAL AD D R ESS : Journal of Semantics, IRM Germany Scienrific Cenrer, IWHS
7000-75, Postfach 8oo88o, D-7000 Srurrgarr Xo, W. Germany. Phone: (�9-7 I 1-) 6695-559- Tclcfax: (+9-71 1) 6695-500.
BITNET: bosch@dsolilog. New Subscribers to the Journal of S�manrics should apply to the Journals Subscription Deparrmcnr, Oxford University Press, Pinkhill House. Sourhfidd IZoad. Eynsham. OXH •JJ. For further information sec the inside back cover. Volumes I-6 arc available from Foris Publications Holland, PO Box 509, 3300 Am Dordrccht, The Netherlands. Published by Oxford University Press
Copyright by NIS Foundation
ISSN
OI 67-5133
Special Issue
REFERENTIAL AND ANAPHORIC PROCESSES IN TEXT COMPREHENSION
Guest editors
CHRISTOPHER HABEL GERT RICKHEIT
JOURNAL OF SEMANTICS Volume 7 Number 4
SPECIAL ISSUE ON REFERENTIAL and ANAPHORIC PROCESSES in TEXT COMPREHENSION Guest Editors: Christopher Habel, Gert Rickheit
C O NTE NTS and U . GuNTHER Interpreting anaphoric relations during reading: inspection time evidence L. SICHELSCHMIDT
32 I
M. HIELSCHER and J. MussELER Anaphoric resolution of singular and plural pronouns: the reference to persons being introduced by different co-ordinating structures 347 A. J. SANFORD and F. LocKHART Description types and method of conjoining as factors influencing plural anaphors: a continuation study of focus
365
A. GARNHAM and J. 0AKHILL Mental models as contexts for interpreting texts: implications from studies of anaphora 379
K. fRAURUD Definiteness and the processing ofNPs in natural discourse
395
D. CARTER Control issues in anaphor resolution
435
Editor's Preface This special issue of the Journal of Semantics comprises six papers that were presented during a conference on 'referential and anaphoric processes in text comprehension and text generation'. The conference was held at ZIF, the Center for Interdisciplinary Research at the University of Bielefeld, FRG, in July 1 989. For three days, the comfortable atmosphere of the ZIF provided linguists, psycholinguists, and artificial intelligence researchers with a forum for an exchange of ideas on two issues that are important to current research. Referential and anaphoric processes are fundamental to text comprehension and generation since communication requires some common ground for information interchange. Creating a suitable cognitive basis involves the linking of verbal expressions to extralinguistic entities (reference) as well as the linking of expressions to linguistic entities (anaphora). Reference enables speakers and listeners to develop an adequate representation of-and thus. talk about-objects and events in a real or fictional world. Anaphora, on the other hand, is a means to enhance coherence of the structure and the representation of discourse. At the Bielefeld conference, empirical and theorectical approaches to reference and anaphora were presented, with emphasis on comprehension. In the Journal ofSemantics issue in hand, both empirical research and theory are represented. The first part comprises a selection of psycholinguistic experiments that focus on on-line processes in reference and anaphor resolution. Using a self-paced reading technique, Sichelschmidt and G unther investigated the comprehension of 'zero' and pronominal anaphora. The inspection time profiles observed by Sichelschmidt and Gunther suggest that anaphor resolu tion docs not operate in an exclusively retroactive fashion but includes proactive subprocesses. In a similar vein, Hielscher and Miisseler present evidence in favour of the idea that concepts co-ordinated by certain conjunctions are mentally melted into a compound concept. Hiclscher and Mi.isseler argue that co-ordination facilitates subsequent plural reference to the conceptual compound but renders singular reference to an atomic concept more difficult. The paper by Sandford and Lockhart also has plural anaphora as its subject. Based on the variation of co-ordinative and descriptive means, Sandford and Lockhart report that case role identity of the atomic concepts is crucial for the formation of a compound concept.
vi
Editor's Preface
The second part comprises a collection of papers that focus on theorectical and implementation issues. Garnham and Oakhill discuss some recent findings and some fundamental problems in anaphor resolution research from a mental models point of view. In particular, Garnham and Oakhill deal with the representation of superficial and content-based anaphora and with functional aspects of name and role filler mapping in anaphora resolution. Fraurud aims at clarifying the role of definiteness in the processing of noun phrases in natural discourse. Proceeding from an empirical study of the co referential properties of definite and indefinite noun phrases, Fraurud proposes an alternative model of the processing of first-mentioned definite noun phrases. Finally Carter points out several premises for the adequate interpretation of anaphoric expressions from the perspective of computational linguistics. Carter's plea for co-operative control structures based on numerical scoring is accompanied by examples from actual implementations. With the compilation of this special issue, we attempt to stimulate a broader discussion of referential and anaphoric processes in any of the scientific disciplines mentioned above. We hope that the observations reported and the ideas proposed here will be of use for the investigation of referential and anaphoric aspects of natural language processing as well as for the design of effective computer-based language processing systems. CHRISTOPHER HABEL
University ofHamburg Computer Science Department
GERT RICKHEIT
University ofBielefeld Faculty ofLinguistics
© N.I.S. Foundation (1990)
jo11mal ofStmamics 7: J21-J4S
Interpreting Anaphoric Relations During Reading: Inspection Time Evidence
LORENZ SICHELSCHMIDT and UDO GUNTHER University ofBielefeld Abstract
INTROD UCTION For any reader, comprehending a text means developing a mental representa tion of its form and its content. During reading, people construct a superficial representation to encode the surface structure of the verbal input and a discourse model of the state of affairs described by the text, i.e. a conceptual surrogate of the real or imaginary world portrayed in discourse Gohnson-Laird 1983; Garrod & Sanford 1988). In the simplest form; a discourse model may be pictured as an abstract token structure that comprises object concepts (nodes) to represent entities and relation concepts (links) to represent properties of or connections between these entities. The token structure is dynamic in nature as readers update their mental representations step by step along with progress in reading. At that, the surface structure of the text will determine the order of instantiation of concepts in the reader's discourse model. An easy-to-read text is structured in a way that facilitates conceptual representation. Many texts, however, contain verbal expressions that render the development of an adequate discourse model quite difficult. In the present paper we shall focus on one paradigmatic type of 'difficult' expressions-on anaphora, including ellipses and pronouns.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
On-line mechanisms in the processing of anaphora were investigated in rwo reading experiments. Shorr German texts were presented incrementally for self-paced reading, and inspection rimes were recorded for every single word. Each text contained a critical rwo-clause sentence with an elliptical gap and a personal pronoun. In the first experiment, varying the surface structure of the antecedent clause affected inspection rimes for the anaphor clause as a whole. In the second experiment, varying the semantic structure of the antecedent affected antecedent but not anaphor inspection rimes. The findings are not in agreement with the view that anaphor resolution is an exclusively retroactive process triggered on encountering the anaphor. Rather, the results suggest that anaphor resolution includes proactive subprocesses: foundations for interpreting anaphora yet to come arc laid during reading the antecedent.
322
Interpreting Anaphoric Relations During Reading
ANA PHOR PRO C ESSING
The correspondence hypothesis In a two-clause text fragment like
(I ) Rita called Frank and asked him ...
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
both the pronoun and the missing subject of the second clause are instances of anaphora. The term 'anaphor' denotes a linguistic device that takes its meaning from being tied to some antecedent in the preceding verbal or situational context. In order to find out who asked and who was asked, readers of the second clause of (I) must recourse to the mental representation they have developed so far. With respect to that recourse, linguistic theory has proposed a distinction between 'model interpretive' anaphora and 'elliptical' anaphora (Hankamer & Sag I976; s.ag & Hankamer I984). Anaphoric devices of the model interpretive type ('deep anaphora') include personal pronouns or sentential pronouns (Rita called Frank, and Ellen scoffed at it). According to theory, these need not be tied to a specific verbal antecedent but may refer to an entity or an event not explicitly mentioned. Therefore, the referent of a model interpretive anaphor is to be found in the reader's discourse model. When encountering a pronoun, for example, the reader is thought to search his or her current discourse model until an appropriate concept is come across (Frederiksen I98I; Sanford & Garrod I981). Anaphoric devices of the elliptical type ('surface anaphora') include phrasal ellipsis (Rita called Frank but Ellen didn't), sluicing (Rita called Frank, and Ellen wondered why), and gapping (Rita called Frank, and Ellen, Paul). According to theory, an elliptical anaphor always has a specific verbal antecedent that must be syntactically consistent with the 'missing' portion of the anaphor clause. This is to enable the reader to mentally substitute the 'missing' portion with a copy of the antecedent. A reader encountering a gap, for instance, is thought to search his or her surface representation for a suitable filler item (Clifton & Frazier 1986; McElree & Bever 1989). The distinction between elliptical and model interpretive anaphora proposed by linguistic theory can be summarized in the 'correspondence hypothesis' (Murphy I985; see Garnham & Oakhill I990): processing of model interpretive anaphora such as personal pronouns takes place solely on the basis of the reader's discourse model; processing of elliptical anaphora such as gaps takes place solely by recourse to the reader's superficial representation of the text. It follows from the correspondence hypothesis that surface details of the antecedent clause are crucial for understanding ellipses but not for under-
L. Sichelschmidt and U. Gunther
323
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
standing pronouns. On. the other hand, the current state of the reader's discourse model is crucial for resolving pronominal but not elliptical anaphora. Several empirical studies are pertinent to this reasoning. Two observations make a case for the corresponding hypothesis. On the one hand, pronoun processing is influenced by discourse factors. World knowledge, for instance, is a factor relevant to pronoun interpretation (Hirst & Brill I980). Foregrounding of the antecedent (Sanford & Garrod I98I), shift of focus (Anderson, Garrod & Sanford I983), thematic prominence (Carpenter & Just I977), implicit causality (Caramazza, Grober, Garvey & Yates I977), and conceptual complexity of the antecedent (Garrod & Sanford I977) likewise play a role. Also, it makes a difference whether the pronoun refers to an object or to an event (Reilly I988). On the other hand, ellipsis processing is determined by surface factors. In particular, increasing the distance between anaphor and antecedent slows processing of elliptical but not of model interpretive anaphora (Tancnhaus, Carlson & Seidenberg I985). Similarly, reading rime for verb phrase ellipses increases with the number of sentences intervening between antecedent and anaphor (Garnham I987). Murphy (I985) found that the longer an antecedent verb phrase, the longer were reading rimes for an elliptical sentence that follows. Besides, syntactic inconsistency with the antecedent slowed reading rime for a subsequent elliptical expression. So far, the results fit the linguistic views fairly well. Contrary to the correspondence hypothesis, however, the manipulation of surface form affected elliptical anaphors and model inter pretive anaphors likewise. Observations like these blur the theoretical distinction between elliptical and model interpretive anaphora. Actually, several empirical findings arc incompatible with the correspondence hypothesis. First, experiments have shown the processing of elliptical anaphora to be codetermincd by factors other than surface form. In particular, context variation has proved to take an effect. Readers arc faster in comprehending an elliptical anaphor if its antecedent is in some way set off as likely to be referenced by an expression yet to come (Malt I985). Also, subjects tend to interpret elliptical verb phrases on the basis of plausibility even if the inter pretation is at odds with the one that is syntactically correct (Garnham & Oakhill I987; Oakhill & Garnham I987). Second, experiments have shown the interpretation of pronouns to be sensitive to surface variation. A factor that has attracted much attention is the distance between pronoun and antecedent. While some studies have reported an advantage in processing for expressions containing a pronoun that refers to an entity introduced lately (Clark & Sengul I979; Dancman & Carpenter I98o; Ehrlich & Rayner I983), other studies have reported the opposite (Grober, Beardsley & Caramazza I978; Corbett & Chang I983; Miisselcr & Rickheit
324 lnrerprering Anaphoric Relations During Reading
I990). However, these findings are not necessarily contradictory. The observed differences may partly be due to confounding with a change of focus (Garnham I987) and partly to confounding with syntactic complexity (Matthews & Chodorow I988).
The backward search approach
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In face of the empirical· evidence, just how do readers resolve anaphora? Psycholinguistic accounts of the cognitive processes in comprehension generally assume that anaphor resolution involves some sort of backward search. 'Backward search' can basically be regarded as a three-step process triggered by encountering the anaphor (Frederiksen I98I; Sanford I98 s). The first step in identifying the appropriate referent is to search the surface representation for any verbal items that arc syntactically consistent with the anaphor. By comparing the anaphor and the potential antecedents as to a match of 'primitive characteristics' (Sanford & Garrod I989: 255) such as number, gender, and, in German, case, the set of potential referents can be reduced considerably from the outset (Ehrlich I98o;just & Carpenter I987). If a match is found, the item in question may be accepted as a provisional substitute for the anaphor. If, however, surface representation search fails to single out an adequate referent, the scope will have to be expanded (Marslcn-Wilson, Levy & Tyler I982; Garnham & Oakhill I988). The second step thus is to search the discourse model for any concepts that arc semantically consistent with the anaphor. Search is guided by what might be called 'elaborate characteristics' of the potential referents: agent status, forcgrounding, focus, or thematic salience (Frederiksen I98I; Sanford I985). If there is a single concept that meets the semantic constraints, it may be regarded as the likely referent. If discourse model search still fails to single out an adequate referent, supplementary information will have to be provided by knowledge-based inferencing (Sanford & Garrod I98 I; Sanford, Garrod, Lucas & Henderson I984). Readers may supplement their current discourse model accordingly and thereupon resume processing at stage two. The backward search approach is at variance with the proposals of the correspondence hypothesis in that anaphor resolution is discussed in terms of cognitive effort rather than in terms of cognitive structure. This 'quantitative' point of view becomes particularly evident with model interpretive anaphora. Unlike the correspondence hypothesis, backward search concedes that pronoun resolution may operate on both surface and 'deep' representation. At that, however, a rigid search order is postulated: discourse model search is set going only if a preceding surface representation search has failed. Strictly speaking� the backward search approach implies that anaphors are
L. Sichelschmidc and U. Giincher
325
Postponation and occupation Considering the intricate nature of anaphor comprehension, recent develop ments fall away from the rigid views of the correspondence hypothesis or the backward search approach {Bosch I988; Oakhill, Garnham & Vonk I989). At present, two proposals seem especially interesting because they make the locus of pronoun resolution their main concern. The first account shall be labelled 'postponation hypothesis'. The postpona tion hypothesis distinguishes between provisional interpretation and anaphor resolution proper. The basic idea is that final resolution of a pronoun may be postponed until readers have gathered sufficient information from subsequent verbal input (Vonk I98s; Sanford & Garrod I989). On encountering a pronoun, readers are thought to immediately develop some weak relationship between the anaphor and possible antecedents on the basis of syntactic match. Thus, a domain is marked out for semantic interpretation on the basis of subsequent input or knowledge-based inferences. However, a strong semantic relationship between the pronoun and its antecedent will not be instantiated until so required by the subsequent discourse. As relevant information may appear much later, an increase in processing time due to reference assignment may occur a long time after reading the anaphor. Empirical evidence that specifically relates the postponation idea to the immediacy claim is still sparse. One pertinent case is the occurrence of 'false
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
resolved immediately on encountering. On this premise, two conclusions suggest themselves: processing time will increase with each additional step required to resolve an anaphor, and the increase in processing time will show up at or immediately after the anaphor. The former consideration has received some empirical support (Sanford & Garrod I98I; Garnham & Oakhill I985; O'Brien, Duffy & Myers I986). For instance, a match in gender between pronoun and antecedent speeds up anaphor processing demonstrably {Kerr & Underwood I984; Vonk I984). As to the latter consideration, however, findings are inconsistent. While some on line studies have produced evidence in favour of an immediate resolution of anaphoric pronouns {Ehrlich & Rayner I983; Kerr & Underwood I984), other studies have not {Stevenson & Vitkovitch I986; Carroll & Slowiaczek I987). Apparently there are circumstances that prevent anaphora from being assigned a referent right on encountering. The argument is the more convincing since eye movement tracking has shown readers to skip some pronouns {Vonk I984; I98s). Altogether, the empirical evidence is such that researchers are well advised to treat the immediacy claim with some reservation. The locus of anaphor resolution is a question that clearly requires further empirical study.
326
Interpreting Anaphoric Relations During Reading
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
bonding' (Sanford 1985; Sanford & Garrod 1989). The immediate inter pretation of a pronoun with seemingly obvious reference sometimes may turn out to be premature so that reinterpretation is required (Ifan incendiary bomb drops near you, don't lose your head; put it in a bucket and cover it with sand). At least with such expressions, the most economical strategy will be to gather evidence from discourse until an anaphoric relationship can safely be assumed to hold. The second approach has been labelled the pronominal 'occupation hypothesis'. Here, the basic idea is that proactive subprocesses relevant to the resolution of pronominal anaphora may be initiated in advance while reading the antecedent (Mi.isseler & Terhorst 1990 ). On encountering a word that denotes an entity likely to be referenced later on, a corresponding pronoun node in the reader's discourse representation will be 'occupied' for future referential use. When reading the word Rita, for instance, a singular feminine pronoun node (she) will be activated together with the Rita concept node. This permits readers to develop co-referential relationships prior to actually encountering an anaphor. These relationships may be strengthened, weakened, or replaced in the course of further reading, depending on the subsequent text. At any rate, no retroactive search is required when eventually processing the pronoun. Proactive occupation is clearly inappropriate as a general principle of anaphor resolution, if only for reasons of economy. Much processing would be superfluous, especially if no anaphora at all occurred in the remainder of the text. In specific circumstances, however, proactive processing is an efficient strategy. Given an entity that is a likely referent of anaphora yet to come, proac tive development of an anaphoric bond will reduce subsequent processing effort. As an entity can be set off as a candidate for later referencing by way of fronting, foregrounding, co-ordination, or the like, some conjectural pre requisites for proactive occupation are often met (Sanford & Lockhart 1990). However, the conditions that are necessary and sufficient to initiate proactive subprocesses still have to be established. Some findings from recent experimentation on plural reference corroborate the pronominal occupation hypothesis (Hielscher & Mi.isseler 1990;Mi.isseler & Rickheit 1990). The experiments proceed from the assumption that single concepts co-ordinated by conjunction are mentally melted into a conceptual compound. Development of such a conceptual compound is thought to activate a plural pronoun node while deactivating any singular pronoun nodes. In consequence, processing of a plural pronoun which refers to the compound as a whole should be faster than processing of a singuar pronoun which refers to an atomic component. In fact, an advantage in processing for plural reference has been observed under the conditions described (Hielscher & Mi.isseler 1990). The pronominal occupation hypothesis and the postponation hypothesis agree in that anaphor resolution does not necessarily occur at the anaphoric
L. Sichelschmidt and U. Gunther
327
EXPE R IMENT
I
According to the correspondence hypothesis, gapping is classified as an elliptical anaphor which is interpreted in terms of the reader's superficial repre sentation. Processing of the word following a gap thus should be faster if the surface distance from the antecedent is small than if it is large, and faster if the antecedent is syntactically consistent with the gap than if it is syntactically inconsistent. On the other hand, a personal pronoun is classified as a model interpretive anaphor which is interpreted in terms of the reader's discourse model. Processing of a pronoun thus should be faster if its referent is semantically foregrounded than if it is not. According to the backward search approach, however, a discourse model search is a prerequisite for semantic variation to take an effect. Hence, processing time for a pronoun should be affected by semantic factors if the
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
device or in its immediate vicinity. However, the approaches disagree with respect to the presumptive locus of resolution. From the postponation hypothesis it follows that any delays in processing caused by anaphor resolution should show up at some point after the anaphoric device. According to the occupation hypothesis, however, any delays in processing due to reference assignment should be observable prior to the anaphor. Though tentative in nature and different in reasoning, occupation and postponation hypotheses may eventually complement each other. Occupation can be regarded as an account of anaphor processing under circumstances that facilitate reference assignment. On the other hand, postponation comes to bear under circumstances such as ambiguity or shift of focus that may render anaphor resolution quite difficult. Altogether, neither the cognitive strategies that readers employ in anaphor resolution nor the factors that govern the choice of a specific processing strategy are very well understood to date. The current state of research is characterized by the existence of a large number of particular findings but lack of an integrative theoretical framework. There are a few details, though, that enable the researcher to evaluate empirically some specific claims about the processing of anaphora in a series of on-line reading experiments. In Experiment I, predictions derived from the correspondence hypothesis shall be contrasted to predictions derived from the backward search model of anaphor resolution. In Experiment 2, the backward search approach shall be confronted with the pronominal occupation hypothesis to investigate the role of retroactive and proactive subprocesses in the comprehension of anaphora.
328 Interpreting Anaphoric Relations During Reading
antecede�t is far. In particular, processing of a pers�nal pronoun should be faster if a close antecedent matches the pronoun syntactically than if it does not, and if a far antecedent is semantically foregrounded than if it is not
Method
Subjects Thirty-six male and 24 female college students participated in the experiment in return for payment. Their mean age was 26 years; they all had German as their native tongue.
Materials Aside from several distractors, 6o experimental texts were constructed. Each comprised a headline and three sentences, the second of which was the critical one. A German example is given in (2), together with an approximate English translation: (2) Diefrohliche Studentin
Viele Passanten blickten in die Schaufenster der Fufigiingerzone. Diefrohliche Stud entin sah den modischen Rock und kaufie ihn auf der Stelle.Erschrocken stellte sie dabeiJest, daft sie ihre Scheckkarte verloren hatte. (2) The cheeiful student
Many passers-by glanced at the department store windows. The cheeiful student saw thefashionable skirt and purchased it on the spot. She was startled when she noticed she had lost her credit card. The critical sentences were made up from two clauses conjoined by und (and). The first clause always contained a subject noun phrase and an object noun phrase. The second clause always contained two anaphoric devices: a gap in place of the subject and a personal pronoun that unequivocally referred to the object of the preceding clause. Thus, the structure of the critical fragment always was as in (3):
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
A measurement method which enables the detection of effects at every single item of a text is imperative for the empirical evaluation of these predictions. We have chosen an incremental reading technique on a word-by-word basis (Gunther 1989; Rickheit, Gunther & Sichelschrnidt 1990): a text is displayed on a screen word by word, cumulatively. To begin with, only the first word of the text is displayed. Whenever the reader pushes a key, the next word is disclosed in addition to the first one until the whole text is on display.
L. Sichelschmidr and U. Giinrher J29
(3) DETr-ADjr-Nr-Vr-DET2-ADj2-N2-CONJ-V2-PRON-rrTH . . . Predictions specifically apply to the word immediately following the gap (V2) and to the personal pronoun (PRON). From each text, six variants were constructed which differed with respect to the headline and to the first clause of the second sentence. The purpose of varying the headline was to put either the subject entity, the object entity, or some entity from the introductory setting into the focus of attention. There were three 'focus' conditions for each text:
The purpose of varying the 'structure:;' of the antecedent clause was to manipulate the surface distance between the anaphoric device and its antecedent. This was accomplished by exchanging subject and object noun phrase: - SVO: subject-verb-object (DieJrohliche Studentin sah den modischen Rock . . .) - OVS: object-verb-subject (Den modischen Rock sah dieJrohliche Studentin . . .) An inversion like this is possible in German without using the passive voice; when translated into English (The cheerful student saw the fashionable skirt . . .), however, the difference is usually lost. Inversion, though grammatical, constitutes a deviation from the rule. Therefore, object fronting may lead to overall processing difficulties.
Design The independent variables, 'focus' and 'structure', were varied orthogonally within subjects. Every subject read ten texts under any of the six treatment conditions. The major dependent variable was the inspection time for each word of the critical sentence.
Procedure Texts were displayed incrementally on a monochrome computer screen (Eizo 405 r) . Subjects were instructed to press a response key 'precisely when you have read and understood a word'. A warning message was displayed prior to each text. When ready, the subject pressed the 'Enter' key to go through the incremental reading procedure. Intervals between keystrokes were auto matically recorded (Compaq Deskpro 386). After reading a text, subjects were asked to rate on a five-point scale how well the title fit the text, how original the text was, and how fluently the text
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
- HFS: headline featuring subject (The cheerful student) - HFO: headline featuring object (Thefashionable skirt) - HFI: headline featuring introduction (The department store)
330
Interpreting Anaphoric Relations During Reading
was written. The rating task had been included to make subjects attend to various aspects of the text. Besides, subjects' ratings constituted a secondary dependent measure. ·
Data
Results The average inspection time was 418 ms per word. There were no main effects of 'focus' on word inspection times. However, there were main effects of 'structure' for the second half of the critical sentence. In addition, there was an interaction of 'focus' and 'structure' at the first verb. Altogether, these findings show that processing of personal pronouns and of words after a gap is sensitive to syntactic variation. To illustrate the effects of 'structure', inspection time profiles for the critical sentence are plotted in Figure I. 0.50 r-------,
0.45 �
� r IX c "' (1)
�
0.40
-4- svo Order -a-ovs Order
J
0.35 �--�--�---L--� ... DETl ADJI
Nl
Vl
DET2 ADJ2 N2 CONJ
V2 PRON 11TH
Syntactic structure Figure
1
Inspection rime profiles by structure.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Any inspection times beyond the limit of two standard deviations above or below the cell mean were replaced with the upper or lower cell limit, respectively. Inspection time data were subjected to two analyses of variance, the first one (F.) with subjects and the second one (F m) with materials as a random factor. Degrees of freedom were calculated using the Greenhouse Geisser formula.
L. Sichelschmid r and U. Giinrher
331
The second noun of the critical sentence (N2) took 36 ms longer to read when it was the head of the subject noun phrase than when it was the head of the object noun phrase (MSe 19.40, F.{I,s9)- 5·97. P- 0.018; MSe 12.17, Fm(1,59) 10.01, P 0.003). The conjunction (CON]) took 54 ms longer to read when following an OVS clause than when following an SVO clause (MSe 64.17, F5(1,59) 4.01, P o.oso; MSe 35.71, Fm(1,59) = 7.17, P o.ow). Inspection times for the item following the gap (V2) were also affected by syntactic variation. The difference of 29 ms between the OVS and SVO versions was significant in tendency by subjects and signifi cant by materials (MSe 21.11, F5(1,59) 3.68, P o.o6o; MSe 17.20, Fm(1,59) 4.22, P 0.044). Pronoun inspection times also differed between OVS and SVO versions. Readers took 28 rns longer to read the pronoun (PRON) when its antecedent, the object, had been introduced early than when it had been introduced lately. The difference was marginally significant by subjects but significant by materials MSe- 17.12,F,(1,59) 3.96, P 0.051; MSe 16.89,Fm(1,59) 4.06,P o.049).And finally, averaging of inspection times for the word following the pronoun (11TH) yielded a significant difference of 55 ms due to syntactic variation (MSe 4·54. F.(I,s9) 12.68 P = 0.001; MSe 14.67, Fm(1.59) 4.50, P = 0.038). Incidentally, this state of findings did not change when including 'gender' as an additional predictive variable. The interaction between 'structure' and 'focus' that was observed in inspection times for the first verb (MSe 2.72, F5(2,109)- 8.62, P o.ooo; MSe = 4.61, F m(2,IIs) 5-41, P = o.oo6) is illustrated in Figure 2. Simple effects analyses showed 'focus' to be effective only for SVO versions (MSe 2.56, F5(2,115) 7.90, P 0.001; MSe 4.10, Fm(2,114) S·I7, P = o.oo8) but not for OVS versions. On the other hand, 'structure' took an effect with HFS versions (marginally significant by subjects: MSe- 3-27, F,(1,59) 3.82, P o.os6; significant by materials: MSe 2.92, Fm(I,S9) 5.03, P - 0.029) and with HFO versions (MSe- 2.03, F,(I,S9) I8.48, P- o.ooo; MSe 4·43. Fm(I,S9) 8.50, P o.oo8). There was no effect of 'structure' with headlines featuring the introductory serting (HFI), however. Finally, there were effects on subjects' judgments. Texts with an object-first sentence were rated to be less fluent (MSe- 23.66, F,(1,59)- 61.88, P- o.ooo; MSe = I 1.74, Fm(I,59)- 124.67, P- o.ooo) and less original (marginally significant by subjects: MSe- 17.20, F.(I,S9)- 3.73, P o.os8; significant by materials: MSe- 8.70, Fm(I,59)- 7·37, P 0.009) than texts with subject-first versions. In comparison to HFS or HFO versions, headlines featuring the introductory setting of a text (HFI) appeared less original (by subjects: MSe- 7.69, F.(2,118)- 7·75. P o.ooi; by materials: MSe- 15.77, Fm(2,1I8)- 3.78,P- 0.026). Moreover, they were rated to fit the text less well =
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
=
=
332
Interpreting Anaphoric Relations During Reading 0.45
� � SV 0Order � � OVSOrder '""'
�
> ...
.2 1a:
0.40
c � I!)
�
Object
Introduction
Headline featuring... Figure
2
Verb inspection time by structure and focus.
(MSe 11.01, F.(2,111)- 18.49, P o.ooo; MSe 42.09, Fm(2,II8) 4.84, P 0.010) than headlines featuring the subject or object entities of the ciritical sentence. =
=
=
=
=
Discussion At first sight, some of the evidence is in accord with the views of the correspondence hypothesis on the processing of gaps and the views of the backward search approach on pronoun processing. As predicted, the word following a gap was processed faster in the case of syntactic consistency than in the case of syntactic inconsistency. In subject-first versions, antecedent and anaphor clauses were identical in surface structure. So readers could easily substitute the 'missing' subject of the anaphor clause with the subject of the antecedent clause. In object-first versions, however, the antecedent and anaphor clauses were different in surface structure. In these versions, the gap at the beginning of the anaphor clause referred to the entity mentioned at the end of the antecedent clause. So readers could not substitute the 'missing' subject of the anaphor clause from the antecedent clause without reconsidering syntactic-semantic roles (Grober, Beardsley & Caramazza 1978). The effect can also be interpreted in terms of surface distance. Other than predicted, however, processing time did not increase but decreased with surface
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Subject
L. Sichelschmidr and U. Giinrher
333
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
distance. This observation rules out any explanations based on the bridging of physical distance at the text surface. It provides an argument against the view that backward search proceeds in a linear, right-to-left fashion starting from the anaphor. Rather, the results suggest a left-to-right search, if any (Matthews & Chodorow 1988). The failure to find an effect of semantic variation on pronoun inspection time casts some doubt on the proposals of the correspondence hypothesis on pronoun processing. It is in line with the backward search approach, though. In fact, the materials and design of the present experiment favoured a pronoun processing strategy which was restricted to surface representation search. The distance between pronoun and antecedent was short; it was varied within one clause preceding the anaphor, and the appropriate referent could be identified on the basis of gender cues alone. As this enabled readers to interpret a pronoun without recourse to their discourse representation, the zero effect of mani pulating the 'focus' variable is not too surprising. The effect of semantic variation that appeared in inspection times for the verb (V1) can be interpreted in terms of discourse model maintenance. Readers were fastest in relating a predicate to some specific entity when that entity had been introduced in the headline. By featuring a particular entity, the headline may have provided the reader with an initial concept node for the development of his or her discourse model (Gernsbacher & Hargreaves 1988). The relevance of foregrounding a particular concept is emphasized by the fact that subjects judged specific headlines to fit the text better than unspecific ones. In summary, neither the correspondence hypothesis nor the backward search approach is able to account fully for the evidence. The crucial point is that after reading an object-first antecedent clause, word inspection times increase-not just for anaphoric items but for the entire anaphor clause, even prior ro anaphora. However, there is a straightforward explanation to that. The overall increase in inspection time after an object-first clause may be a symptom of problems in the co-ordination of clauses rather than in the resolution of anaphora. The 'structure' effect at the end of the antecedent clause (N2) can be taken to hint at that. The increase in inspection time at the end of an object-first antecedent clause can be attributed to the trouble readers take with clarifying syntactic-semantic roles (Gernsbacher & Hargreaves 1988; Frazier & Flores d'Arcais 1989). This view is corroborated by the fact that OVS versions were judged to be less fluent than SVO versions. The 'structure' effect at the conjunction (CON]) may arise from trying to exploit the coordinative potential of und (and) fully. Readers may have employed some sort of 'parallel function' strategy (Grober, Beardsley & Caramazza 1978; Frazier, Taft, Roeper, Clifton & Ehrlich 1984): On en countering a conjunction, the syntactic structure of the expression prior to the
334
Interpreting Anaphoric Relations During Reading
EXPERIMENT
2
Traditional accounts of anaphor comprehension assume that on encountering an anaphor, some backward search procedure is initiated ( Frederiksen 198 1; Sanford 1985). Recently, this view has been challenged by an account that does without any retroactive processes. The 'pronominal occupation! hypothesis (Hielscher & Miisseler 1990; Miisseler & Terhorst 1990) claims that, under certain conditions, pronouns can be interpreted; on the basis of proactive subprocesses. On reading a word which denotes an entity likely to be referenced later on, a corresponding pronoun node in the reader's discourse model is 'occupied' for future referential use. Though the pronominal occupation hypothesis.is not to be interpreted as an alternative but as a complement to backward search, the approaches can be
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
conjunction may have been taken as a model for establishing some structural skeleton of the expression to follow. Parallel function can account for the remaining differences between SVO and OVS versions. According to the parallel function idea, readers should have conjecrured the expression after the conjunction to be identical in structure to the expression preceding the conjunction. In the present experiment, this was true for SVO versions but not for OVS versions. Thus, after a subject-first antecedent clause, mapping of structures was not difficult. After an object-first antecedent clause, however, there was a mismatch between the strucrure conjecrured and the strucrure encountered when reading the anaphor clause. The attempt to resolve this mismatch by reconsidering. syntactic-semantic roles could have led to a delay in processing for every verbal item in the remainder of the critical sentence ( Frazier, Taft, Roeper, Clifton and Ehrlich 1984). The parallel function idea is fundamentally different from the backward search approach to anaphor resolution. Unlike backward search, which operates in a retroactive manner, parallel function is thought to operate proactively: on reading a conjunction, the structure of a prior expression is utilized as a strucrural foundation for representing an expression yet to come. So parallel function implies that the processing of one part of a text may influence processing of a subsequent part. In this respect, the parallel function idea closely resembles· the pronominal occupation hypothesis. Both approaches argue that.proactive subprocesses play an important role in the comprehension of coordinate·expressions. At that, emphasis is on clausal conjunction which is assumed to elicit cognitive processes that affect the processing of the subsequent conjunct. Experiment 2 has been designed to inquire into the role. of proactive and retroactive subprocesses in the comprehension of anaphors.
L. Sichelschmidr and U. Giinrher
3 35
Method In summary, the backward search approach predicts processing time differences between singular and plural conditions for pronouns bur not for referents. In contrast, the pronominal occupation hypothesis predicts process ing time differences between singular and plural conditions for referents bur not for pronouns.
Subjects Sixty subjects from the same population as in Experiment 1 volunteered for this study.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
confronted empirically because they differ in their predictions as to the processing of singular and plural pronouns. The backward search approach leads to the prediction that any differences between singular and plural reference will show up in inspection times for the corresponding anaphoric expression. In fact, Murphy (1984) has demonstrated that reference to two entities resulted in longer anaphor reading times than reference to only one entity-even when the anaphoric expressions were identical. This finding can readily be explained in terms of a mental model framework: anaphor comprehension takes longer in the plural condition than in the singular condition because readers must establish referential links to several concepts in the case of plural reference but to only one concept in the case of singular reference. On the basis of the background search approach, one would therefore expect pronoun resolution times to be shorter in the singular reference condition than in the plural reference condition. The backward search approach does not, however, make any prediction as to inspection times for potential antecedents. In contrast, the pronominal occupation approach does not predict any difference in pronoun processing time between the singular and the plural condition because an appropriate pronoun node has been activated beforehand. The argument is that whenever a new entity is introduced in the text, another object node will be established and, correspondingly, another singular pronoun node be activated in the reader's discourse model. In effect, several singular pronoun nodes will be 'occupied' if several individual entities are represented in the reader's discourse model. Should the individual entities be melted into a conceptual compound, however, the individual pronoun nodes should be cleared in exchange for a single plural pronoun node. On the basis of pronominal occupation, one would expect processing of a conceptual compound to be more time-consuming than processing of a comparable singular antecedent. On the other hand, one would expect processing times for singular and plural pronouns to be in keeping.
- 3 36
Interpreting Anaphoric Relations During Reading
Materials Similar to Experiment 1, twenty three-sentence texts were constructed. A sample text is given in (4), together with an approximate English translation: (4) Irischer Friihling
Im Friihjahr wechseln Regen und Sonnenschein in Irland schnell ab. Also nahm Klaus eine leichte und wetteifesteJacke und packte sie in den Kojfer. Nun hatte erfur aile F(j{[e die richtige Kleidung . (4) Spring in Ireland
Again, the critical sentences were made up from two clauses coordinated by the conjunction (CON]2) und (and). The strucrure of the critical sentence fragment thus was as in (s): (s) ... DETr-ADJr-CON]r-(DET2)-ADJ2-N2-CON]2-V2-PRON . . . From each text, two versions were constructed that were identical except for an extra determiner (DET2) in the first clause of the critical sentence. By deleting or inserting that determiner, the 'number' of implied entities was manipulated. In the case of determiner omission (SC), a single object entity with two attributes were introduced. In the case of determiner inclusion (PC), however, two object entities with one attribute each were introduced: SC: singular condition ( . . . eine leichte und wetteifeste Jacke . . . I . . . a lightweight and waterproofjacket . . . ) PC: plural condition ( . . . eine leichte und eine wetteifesteJacke . .. I . . . a light weight and a waterproofjacket .. . ) In consequence, readers of an SC version should develop a discourse model containing one single jacket node with two attribute links.Accordingly, readers of a PC version should develop a discourse model with two separate jacket nodes. Naturally, the purported difference can be expressed in syntactic terms as well. In the SC versions, the antecedent clause contains a single noun phrase with two co-ordinated artributes. In the PC versions, the antecedent clause is made up from two co-ordinated noun phrases with one artribute each, the noun category in the first phrase being empty. In order to keep SC and PC versions comparable, the present experiment has taken advantage of a peculiariry of German. Texts make use of the ambiguous
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In spring in Ireland, sunshine may suddenly change to rain. So Carl took a lightweight and waterproofjacket and put it into his suitcase. Now he was prepared for any weather condition.
L. Sichelschmidc and U. Giincher
3 37
personal pronoun (PRON) sie which may be interpreted as a singular feminine pronoun or as a plural pronoun, depending on the context. For instance, after a sentence like (6) (6) Frank and Rita went skiing . . . German sie may refer either to Rita or to Frank and Rita (Hielscher & Miisseler I 990). In the present experiment, materials have been constructed so that sie had to be disambiguated on the basis of plausibility, guided by the preceding context. A singular reading of sie (to be translated by it) should have been induced by an SC clause and a plural reading of sie (to be translated by them ) should have been induced by a PC clause.
The independent variable was the 'number' of implied entities, varied within subjects. As in Experiment I, word inspection times for a fragment of the critical sentence constituted the major dependent variable. .
Procedure and data The procedure was identical to Experiment same way as in Experiment I .
I.
The data were analysed in the
Results The average inspection time was 3 89 ms per word, a value close to the one reported in Experiment I . In Figure 3, the inspection time profiles for the critical fragment are plotted for singular and plural conditions separately. An effect of the number of implied entities showed up in inspection times at the end of the first clause. Insertion of an extra determiner led to a 47 ms delay in processing for the following noun (MS e - I 3-44, F.{I ,59) 4.96, P - 0.030; MSe - 2.27, Fm(I , I 9 5.76, P - 0.027). The number of implied entities did not affect inspection time for any other word of the critical sentence. In particular, there was no difference in pronoun inspection times berween SC and PC versions (MSe - 4·99. F.{I ,s9) O.I 4, P - 0.709; MSe - 1 .07, Fm(I,I9) - 0.53, P 0.474). Inspection time data disclose another interesting detail. Every critical sentence contained rwo conjunctions that were lexically identical but functionally different. The first und (CONJI ) served to co-ordinate rwo adjectives or noun phrases; the second und (CON]2) served to co-ordinate rwo =
=
=
=
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Design
3 38
lnterprering Anaphoric Relations During Reading 0.50 ..-------.,
I
,..._
_....
Singular
-a-
Plural
I
0.45
� 10:: c "' "
�
0.40
L----'----''---'--.l.._---'
DETI
ADJI CONJ I DET2
ADJ2
N
CONJ2
V2
PRON
...
· Syntactic structure Figure 3 Inspection rime profiles by number.
clauses. Co-ordinative function made a difference: inspection times for und were 92 ms longer when the conjunction combined clauses than when it combined noun phrases or modifiers (MS. = 0.01, F,(1,59) = 72.,j.o, P = o.ooo; MS e 4.90, Fm(1,19) = 34.46,P - o.ooo). The difference requires further specification as it was superimposed by an interaction with the 'number' factor (marginally significant by subjects: MS. = 4.78, F,(1,59) = 3.90, P - 0.053; significant by materials: MS. = 1.22, Fm(1 ,19) = 4.52,P = 0.047). Simple effects analyses revealed a 'function' effect for SC versions (MS. == 6.1 2, F,(1,59) - 27.69, P = o.ooo; MS. = 2.74, Fm(1 ,19) 20.7 1, P = o.ooo) as well as for PC versions (MS. = 5·79, F.(1,59) 63.10, P = o.ooo; MS. = 3·39, Fm(1,19) = 34.77, P = o.ooo). On the other hand, the number of implied entities was not significant at either function. Thus the 'function' effect was somewhat more marked for the plural condition than for the singular condition. Finally, texts were rated to be more fluent under the singular condition than under the plural condition (MS. = 10.30, F,(1,59) = 6.1 3, P = o.o16; MS. = 37.24, Fm(1,19) = 5.76,P = 0.027). =
=
=
Discussion In the present experiment, inspection times for sie when referring to one entity were virtually identical to inspection times for sie when referring to two
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
0 . 35
L. Sichelschmidr and U. Giinrher
3 39
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
entities. While pronoun processing was not sensitive to variation in the number of implied entities, antecedent processing was: the noun at the end of the antecedent clause took longer to process when it served to introduce two entities than when it served to introduce only one. Or, put in syntactic terms: the same item took longer to process when it was the common head of two noun phrases than when it was the head of one complex noun phrase. These observations are incompatible with the predictions of the backward search approach. A zero effect of the number of entities on pronoun inspection time does not correspond to the assumption that readers arrive at the referents of an anaphor by way of some retroactive search process that is executed on encountering the anaphor. On the other hand, the observations arc in perfect keeping with the pre dictions of the pronominal occupation hypothesis. According to that, the incorporation of an entity into the reader's discourse model goes along with the occupation of a corresponding pronoun node. Moreover, the formation of a conceptual compound is accompanied by deactivation of individual pronoun nodes in exchange for the activation of a plural pronoun node. These processes are more time consuming than simply activating one individual pronoun node, and hence lead to a delay in processing in the case of a plural antecedent. The locus of that effect can be explained by considering the dynamics of discourse modelling. During reading, readers update their discourse model whenever so required by the incoming text. In particular, any new entities or events mentioned or implied in the text are likely to be incorporated into the discourse model as soon as possible. However, when reading at the beginning of an augmented noun phrase (a waterproof . . .), the entity to incorporate is still unknown, so that final processing of the adjective has to be postponed. Not until reading the noun (. . . jacket) does the reader come to know which object node the attribute is to be linked to (Rommetveit 1 974). In the singular condition, two attributes were to be linked to a common object node. So one object node with two one-place relation links had to be established when reading the noun. In contrast, in the plural condition an extra determiner indicated the introduction of another entity. So two separate entities of the same type but with different attributes had to be established when reading the noun. Representing two objects, however, should require more cognitive effort-and thus take longer-than representing only one. Subjects' ratings of PC versions as less fluent than SC versions may be a symptom of the increased processing effort. While the observed difference in inspection time at the noun can be attributed to the representation of either one or two entities, some supplementary assumptions are necessary to explain the failure to find an effect of the 'number' of implied entities on pronoun inspection time. Preliminary activation of pronoun nodes has been suggested to be a relevant
340
Interpreting Anaphoric Relations During Reading
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
mechanism (Miisseler & Terhorst 1 990). Pronoun processing will be indifferent to number because no backward search is required for interpretation. Irrespective of whether the pronoun refers to an individual entity or to a conceptual compound, an appropriate pronoun node should have been occupied in advance (Hielscher & Miisseler 1990; Miisseler & Rickheit 1 990). Altogether, the observations on antecedent and pronoun inspection times support an explanation in terms of proactive processing-and so do inspection times for conjunctions. In a written text, a conjunction indicates that a reader will have to co ordinate expressions. Critical sentences in the present experiment always contained two und (and) conjunctions, the first und (CONJI) combining two attributes or two noun phrases and the second und (CON]2) combining two clauses. It can be easily deduced from the parallel function hypothesis that the more complex the expressions to be combined, the more painstaking the processing should be. Therefore, co-ordination of modifiers or of noun phrases should require less cognitive effort than co-ordination of clauses.Accordingly, the first und was faster in processing than the second one; an effect which was even more marked for PC versions than for SC versions. Finally, the results of the present experiment shall be confronted with the evidence from a comparable study of anaphoric reference (Murphy 1 984). In that study, an effect of the number of referent entities manifested in reading times for anaphoric expressions: reference to a single object was faster than reference to two or more objects. In the present experiment, however, this finding did not replicate. Not the anaphor inspection times but the antecedent inspection times turned out to be sensitive to the number of entities introduced here. The divergence in findings may well ensue from dif ferences in materials and rationale between Murphy's study and the one in hand. First, while Murphy used definite noun phrases for anaphoric devices, the present study used pronouns. Since nominal and pronominal anaphors differ in processing (Sanford & Garrod 1 989), it is only natural that the empirical observations are at variance. Second, Murphys variation was not restricted to the actual 'number' of entities but touched the complexity of the described events in general, including spatiotemporal aspects. Since such aspects will eminently be represented in readers' discourse models Gohnson Laird 1 98 3), the effects reported by Murphy (1984) may have been caused by factors other than the number of entities. In the present experiment, however, the 'number' of implied entities was introduced as the only variable, without implying any difference as to time and place. Concerning materials and rationale, then, Murphy's (1 984) study and the present one are hardly comparable. In general, however, the results of Experiment 2 parallel those of Experiment 1. The findings suggest that proactive subprocesses relevant to the inter-
L. Sichelschmidr and U. Giinrher
341
pretation ofanaphoric expressions are set going during reading the antecedent. Also, the findings indicate that co-ordination is a cognitive mechanism that is highly relevant to discourse comprehension. GENE RAL D I S C U S S I O N
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
On the whole, the results of the experiments reported here do not harmonize with either the correspondence hypothesis or a backward search approach. Neither is there any evidence of differences in processing between elliptical and model interpretive anaphora, nor do the effects of syntactic and semantic variation lend themselves to a straightforward explanation by retroactive search processes. Rather, the findings suggest that people process anaphoric expressions by employing proactive strategies, not unlike the ideas put forward by the parallel function hypothesis or the pronominal occupation hypothesis. In the experi ments reported here, the most robust finding was that some cognitive ground work for anaphor resolution is laid during reading the antecedent clause, i.e. prior to actually processing the anaphor. At an early stage of processing, mapping of syntactic structures at a surface level may provide a foundation for the interpretation of anaphora. At a later stage of processing, probability-based activation of conceptual nodes on a semantic level may provide an interpretive foundation. No evidence at all has been found in favour of an independent anaphor resolution mechanism. Anaphor resolution, so to speak, is a by-product of normal text processing. In face of numerous alternatives, the explanation offered here is to be treated with some caution. The increase in inspection time for the entire anaphor clause in Experiment I may as well be caused by differences in individual response latency, by processing overflow, or by overlapping of several subpro cesses. Conceivably, phrase closure may have taken an effect at the end of the antecedent, antecedent copying at the word following the gap, and reference corroboration at the pronoun. Disentangling the intricate cognitive mechan isms that are involved in anaphor processing is an important objective for future research. Similar considerations hold for Experiment 2. If co-ordination of entities were imperative for the formation of a conceptual compound, the findings should not replicate when clauses are disjunct. Using the German equivalent of but or or instead of and should yield results similar to the singular condition (c£ Hielscher & Miisseler 1 990; Sanford & Lockhart 1990). Also, the ease of form ing a conceptual compound largely depends on the particular adjectives or nouns. While a black and a white chessm an could easily constitute a conceptual set, a black and white chessman would be absurd but a black and white photograph
342 Interpreting Anaphoric Relations During Reading
Acknowledgements This research was supported by grants from the German Science Foundation (DFG Rl-3 1 4-8). The authors appreciate the co-operation of 'KoliBri', the Bielefeld research group on coherence. Requests for reprints should be sent to either author. LORENZ SICHELSCHMIDT and UDO GUNTHER Department oJLinguistics University ofBielefeld POB 864o D - 48oo Bielefeld 1 FRG
REFERE N CE S Anderson, A., S. C. Garrod & A J. Sanford (198 3), 'The accessibility of pronominal antecedents as a function of episode shifts in narrative text', Quarterly journal of Experimental Psychology, 35A: 427-40.
Bosch, P. (1 988), 'Representing and accessing focussed referents', lAnguage and Cognitive Processes , 3: 207-3 1 . Bransford, ] . D., ] . R. Barclay & J .J . Franks (1972), 'Sentence memory: a consrrucrive
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
would not . . . Melting of individual concepts into a conceptual compound is governed by the reader's world knowledge. Conceptual combination, however, opens a new, and vital, area of research (Murphy 1 990). Eventually, it cannot be excluded that the incremental reading technique is too insensitive to uncover the subtle processing mechanisms that take place when encountering the anaphor (Corbett & Chang 1 983). Obtaining precise data on the activation of a particular concept at a given time opens up further inquiry by using other on-line methods such as priming, probe recognition, or the like. Altogether, the findings of the experiments reported above agree to a remarkable extent. The bulk of processing in the comprehension of anaphors, so it seems, has taken place when processing the antecedent. At least in the experiments in hand, the cognitive resolution of anaphoric expressions could be explained in terms of proactive processes. Besides, the relevance of co ordinative processes to developing a coherent discourse representation has been demonstrated once again. Perhaps the most important result of our studies is the failure to find any evidence in favour of a separate anaphor resolution mechanism. This justifies the conclusion that interpretation of anaphora should be regarded as a by product of normal text comprehension. This is because anaphoric devices carry neither reference nor meaning. Anaphoric devices carry functions. 'People carry meanings, and linguistic inputs merely act as cues which people can use to recreate and modify their previous knowledge of the world' (Bransford, Barclay & Franks 1 972: 207).
L. Sichelschmidt and U. Gunther 343 anaphora: rules used by readers assigning pronominal referents', Discourse Processes , 4= 323-47· Garnham, A. (1987), 'Effects of antecedent distance and intervening text structure in the interpretation ofellipses', Language and Speech , 30: 5 9-68. Garnham, A. & J. V. Oakhill (198 5), 'On-line resolution of anaphoric pronouns: effects of inference making and verb semantics', Britishjournal ofPsychology, 76, 3 8 5-93· Garnham, A. & J. V. Oakhill (1 987), 'Inter preting elliptical verb phrases', Quarterly Journal of Experimental Psychology, 39A: 6 1 1 -27. Garnham, A. & J. V. Oakhill (1988), ' "Ana photic islands" revisited', QuarterlyJournal ofExperimental Psychology, 40A: 7 1 9-35· Garnham. A. & J . V. Oakhill (1990), 'Mental models as contexts for interpreting texts: implications from studies of anaphora', Journal ofSemantics, 7: 379-393· Garrod, S. C. & A. J. Sanford ( 1 977), 'Inter preting anaphoric relations: the integra tion of semantic information while reading', journal of Verbal Learning and Verbal Behavior, 16: 77-90. Garrod, S. C. & A.J. Sanford ( 1 988), 'Discourse models as interfaces between language and the world', journal ofSeman tics, 6: 147-0o. Gernsbacher, M. A. & D. J. Hargreaves (1 988), 'Accessing sentence participants: the advantage of first mention', Journal of Memory and LAnguage, 27: 699-7 1 7. Grober, E. H., W. Beardsley & A. Caramazza (1978), 'Parallel function snategy in pro noun assignment', Cognition , 6: 1 1 7-3 3· Gunther, U. (1 989), 'Lesen im Experiment (Experimentation in reading research]', Linguistische Berichte, 122.: 283-320. Hankamer, J. & I. A. Sag ( 1 976), 'Deep and surface anaphora', Linguistic Inquiry, 7: 391-426. Hielscher, M. & J. Miisseler (1 990), 'Ana photic resolution of singular and plural pronouns: the reference to persons being innoduced by different coordinating snuctures', journal ofSemantics, 7: 347-364.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
versus interpretive approach', Cognitive Psychology, ]: 1 93-209. Caramazza, A., E. Grober, C. Garvey & J. Yates ( 1977), 'Comprehension of ana ifk' phoric pronouns',journal ojVerbal Learning � and Verbal Behavior, 16: 601-9. Carpenter, P. A. & M. A.just (1 977), 'Reading comprehension as eyes see it', in P. A. Just & M. A. Carpenter (eds), Cognitive Processes in Comprehension , Erlbaum, Hillsdale, pp. 109-39· Carroll, P. J. & M. L. Slowiaczek (1 987), 'Modes and modules: multiple paths to the language processor', in J. L. Garfield (ed), Modularity in Knowledge Representation and Natural LAnguage Understanding, Massa chusetts Institute of Technology Press, Cambridge, MA, pp. 221-47. Clark, H. H. & C.J. Sengul (1 979), 'In search of referents for nouns and pronouns', Memory and Cognition , 7: 3 5-4 1 . Clifton, C . & L . Frazier (1 986), 'The use o f syntactic information i n filling gaps',jour nal ofPsycholinguistic Research , 15, 209-24. Corbett, A. T. & F. R Chang (1983), 'Pro noun disambiguation: accessing potential antecedents', Memory and Cognition , 1 1 , 283-94· Daneman, M. & P. A. Carpenter (1 980), 'Individual differences in working mem ory and reading', journal ofVerbal Learning and Verbal Behavior, 1 9: 45o-66. Ehrlich, K. (1 980), 'Comprehension of pro nouns', Quarterly Journal of Experimental Psychology, 32: 247- 5 5. Ehrlich, K. & K. Rayner (1 983), 'Pronoun assignment and semantic integration during reading: eye movements and immediacy of processing', Journal ofVerbal Learning and Verbal Behavior, 22: 45-87. Frazier, L. & G. B. Flores d'Arcais ( 1 989), 'Filler driven parsing: A study of gap fill ing in Dutch', journal ofMemory and Lan guage, 28: 3 3 1 -44. Frazier, L., L. Taft, T. Roeper, C. Clifton & K. Ehrlich ( 1 984), 'Parallel snucture: a source of facilitation in sentence comprehension', Memory and Cognition , 12.: 42 1 -30. Frederiksen, J. R ( 1981 ), 'Understanding
344 Interpreting Anaphoric Relations During Reading Hirst, W. & G. A Brill (1 980), 'Contextual aspects of pronoun assignment', Journal of Verbal Learning and Verbal Behavior, 19: 168-75 · Johnson-Laird, P. N. (198 3), Mental models. Towards a Cognitive Science of Language, Inference, and Consciousness , Cambridge University Press, Cambridge (UK). Just, M. A. & P. A. Carpenter (1 987), The Psychology of Reading and Language Com prehension , Allyn and Bacon, Boston, MA.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Kerr, ]. S. & G. Underwood (1 984), 'Fixation rime on anaphoric pronouns decreases with congruity of reference', in A. G. Gale & F. Johnson (eds), Theoretical and Applied Aspects of Eye Movement Research , North Holland, Amsterdam, pp. 1 95-202. Malt, B. C. (1 985), 'The role of discourse structure in understanding anaphora', journal ofMemory and Language, 24: 27189. Marslen-Wilson, W. D., E. Levy & L. K. Tyler ( 1 982), 'Producing interpretable dis course: the establishment and mainten an�e of reference', in R J. Jarvella & W. Klein (eds), Speech, Place, and Action. Studies on Deixis and Related Topics, Wiley, Chichester, pp. 3 39-78. Matthews, A & M. S. Chodorow (1 988), 'Pronoun resolution in two-clause sentences: effects of ambiguity, antecedent location, and depth of embedding',Journa/ ofMemory and Language, 27: 24 5-60. McElree, B. & T. G. Bever (1 989), 'The psychological reality of linguistically defined gaps', journal of Psycholinguistic Research , 1 8: 2 1 -35. Murphy, G. L. (1 984), 'Establishing and accessing referents in discourse', Memory and Cognition , l l: 489-97. Murphy, G. L. (1 985), 'Processes of under standing anaphora', Journal ofMemory and Language, 24: 29o-303. Murphy, G. L. (1 990), 'Noun phrase inter pretation and conceptual combination', Journal of Memory and Language, 29: 25988. Miisseler,J. & G. Rickheit ( 1 990), 'Komplex-
bildung in der Textverarbeitung: Die kognirive Auflosung pluraler Pronomen [Compound concepts in text processing: the cognitive resolution of plural pro nouns]', Zeitschrififiir PsychoIogie, I 98: 6981. Miisseler, J. & E. Terhorst (1 990), 'Pro nominale Beserzung: ein alternariver Mechanismus neben der rekursiven Auf losung? [Pronominal occupation: an alternative mechanism in addition to recursive resolution), Sprache und Kogni tion , 9: 3 7-49· O'Brien, E. J.,S. A. Duffy &J. L. Myers (1 986), 'Anaphoric inference during reading', journal ofExperimental Psychology: Learning, Memory and Cognition , u: 3 46-52. Oakhill, ]. V. & A. Garnham (1 987), 'Inter preting elliptical verb phrases at different rimes of day: effects of plausibility and antecedent distance', Language and Speech , JO: 145-57· Oakhill, J. V., A. Garnham & W. Vonk (1989), 'The on-line construction of dis course models', Language and Cognitive Pro cesses, 4: 26 3-86. Reilly, R. G. (1 988), 'Sentence representations and anaphor processing', journal ofPsycho linguistic Research , 1 7: 297-3 1 5· Rickheit, G., U. Gunther & L. Sichelschmidt (1 990), 'Coherence and coordination in written text reading rime studies', in D. Stein (ed), Pragmatics and Comprehension of Written Texts , Mouton de Gruyter, Berlin (in press). Rommetveit, R. (1 974), On Message Structure. A Framework for the Study ofLanguage and Communication , Wiley, London. Sag, I. A. & J. Hankamer (1 984), 'Toward a theory of anaphoric processing', Linguistics and Philosophy , 7: 325-45· Sanford, A J. ( 1 985), 'Aspects of pronoun interpretation: Evaluation of search formulations of inference', in G. Rickheit & H. Srrohner (eds), Inferences in Text Pro cessing, North-Holland, Amsterdam, PP· 1 8 3-204. Sanford, A J. & S. C. Garrod (198 1 ), Under standing Written Language. Explorations of
L. Sichelschmidt and U. Gunther 345
·
Tanenhaus, M. K., G. N. Carlson & M. S. Seidenberg (198 5), 'Do listeners compute linguistic representations?', in D. R Dowry, L. Karrrunen & A. M. Zwicky (eds.), Natural Language Parsing. Psycho logical, Computational, and Theoretical Per spectives, Cambridge University Press, Cambridge (UK), pp. 3 5 9-408. Vonk, W. (1 984), 'Eye movements during the comprehension of pronouns', in A. G. Gale & F. Johnson (eds), Theoretical and Applied Aspects of Eye Movement Research , North-Holland, Amsterdam, pp. 203- 1 2. Vonk, W. (198 5), 'The immediacy of infer ences in the understanding of pronouns', in G. Rickheit & H. Srrohner (eds), Inftr ences in Text Processing , North-Holland, Amsterdam, pp. 205-1 8.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Comprehension Beyond the Sentence, Wiley, London. Sanford, A. J. & S. C. Garrod (1 989), 'What, when, and how? Questions of immediacy in anaphoric reference resolution', Langu age and Cognitive Processes , 4: 2 3 5-02. Sanford, A. J., S. C. Garrod, A. Lucas & R Henderson (1 984), 'Pronouns without explicit antecedents?' Journal ofSemantics, 2: 303- 1 8. Sanford, A. J. & F. Lockhart ( 1 990), 'Descrip tion rypes and method of conjoining as factors influencing plural anaphora: A continuation srudy of focus', journal of Semantics, 7: 365-78). Stevenson, R J. & M. Vitkovich (1 986), 'The comprehension of anaphoric relations', Language and Speech , 29, 3 3 5-60.
© N.I.S. foundarion (1990)
Journal ofSemantics 7: 34 7-364
Anaphoric Resolution of Singular and Plural Pronouns: The Reference to Persons Being Introduced by Different Co-ordinating Structures
MARTINA HIELSCHER and JOCHEN MUSSELER Univeristy ofBielefeld Abstract
.
.
.
INTRODUCTION The comprehension of pronouns confronts the cognitive system with the problem of identifying the concept to which the pronoun refers in the preceding or following text (for recent surveys sec, for example, Sanford & Garrod 1989; Miisseler & Terhorst 1990). In written language, pronouns arc most frequently used anaphorically, i.e. referring to concepts that were mentioned earlier in the sentence or in one of the preceding sentences. Relevant determinants for the process of anaphoric pronominal resolutions arc, for example, gender and number of the pronoun (Ehrlich 1980), as well as 'topic' or main character (Marslcn-Wilson, Levy & Tyler 1 982) and verb inflexion (Garrod & Sanford 198 5). Most of these studies refer to singular pronouns; only
8
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
For the resolution of plural pronouns referring to singularly introduced reference persons the plural antecedent has to be built up by the cognitive system itself (installing a plural complex, e.g. 'john wanted to have apicnic with Mary. They had '). For singular pronouns the antecedent is usually mentioned in the text explicitly. This contribution examined which aspects of the pre pronominal sentence structure determine the installation of a plural antecedent and at which point of rime this process is initiated. Using the German pronoun 'sie ', which is ambiguous in respect to number, it was shown in a first experiment that subjects have a preference to continue a text by referring to both singularly introduced persons, if they are combined by the conjunctions 'and', 'as well as' or neither/nor', or by the preposition 'with', if the female person is in the verb phrase. Subjects prefer to refer to the female person only after the prepositions 'without' and 'instead of', and after 'with' if the female person is in the noun phrase. The reaction rime data of the second experiment indicated that at least these conjunctions and 'with' initiate a plural reference complex before a pronoun is read. This pre-pronominal installation of a complex serves to facilitate plural reference operations executed at a later point in processing. In our view, pronominal resolution is more than a mere recursive search-and match procedure initiated by reading the pronoun; the cognitive system is better prepared for processing further referential relations. This view is discussed in the context of a 'pronominal occupation' hypothesis.
348 Anaphoric Resolution of Singular and Plural Pronouns
a few refer to plural forms (Garrod & Sanford 1 982; Maxey & Sanford 1988; Eschenbach, Habel, Hcrweg & Rehkamper 1989). This article deals with the effect of six co-ordinarj.ng structures combining several referents (i.e. conjunc tions and prepositions) on the resolution process of singular and plural pronouns. First of all, there is no reason to assume qualitatively different mechanisms with regard to the resolution processes for singular and plural pronouns. However, the resolution ofsingular pronouns may be easier than the resolution of plural pronouns. Successful search for a singular referent provides a concept which always corresponds on number and gender of the pronoun: (Mary and John wanted to have a picnic.)
( I b) Er mufite zuniichst die Zutaten einkauftn .
(He1 had to buy the picnic supplies first.)
opposed to singular pronouns, thc;rc need not be a correspondence in gender and number between plural pronouns and their referential concept. For example, if there arc two singular concepts introduced in a noun phrase co ordinated by the conjunction 'and', then the pronoun 'they' (German 'sie ') can be used to refer to the two singular objects as a plural entity, as the following example shows: As
(2a) Mary undJohn wollten picknicken . (Mary and John wanted to have a picnic.) (2b) Sie mufiten zuniichst die Zutaten einkauftn . (They had to buy the picnic supplies first.) Note that in German the meaning of the pronoun 'sie' is ambiguous. 'Sie ' can refer to just 'Mary' or to both 'Mary and John'. Only the verb inflexion determines the reference to the singular ('Mary') or plural concept ('Mary and John'). Additionally, in the German example there is at least a correspondence in number of the verb inflexions ('wollten ' and 'muBten ') used in the two sentences. This is not the case if the two singular concepts arc introduced with a preposition in the noun and verb phrase yielding different verb inflexions ('wollte ' and 'muBten '): (3a) John wollte mit Mary picknicken . Gohn wanted to have a picnic with Mary.) (3 b) Sie mufiten zundchst die Zutaten einkauftn . (They had to buy the picnic supplies first.) Obviously, for comprehension of examples (2) and (3) the reader has to combine the singular persons in such a way that one can refer to them as to a whole plural entity that fits the plural pronoun. We will refer to the process
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
( I a) Mary undJohn wollten picknicken .
M. Hielscher and J. Miisseler
349
E X PE R I M E N T
1
The first experiment was designed to investigate semantic preferences for a singular or plural interpretation of the German pronoun 'sie '. Subjects had to read sentences with a female and a male person co-ordinated by the conjunc tions 'and', 'as well as', and 'neither/nor', or the prepositions 'with', 'without', and 'instead or. The following sentence beginning with 'Sie . . .' was continued by the subjects in a sentence completion task. It can be assumed that a plural interpretation of the pronoun 'sie ' is preferred for the three conjunctions and similarly for 'with'. These four express some thing that the two persons have in common or do together; building a complex of these two persons should therefore be suggested semantically. In this case the plural should be used more frequently. 'Without' and 'instead or express a difference between the two persons: one does something without or instead of the other. For these two prepositions it is assumed that no complex should be
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
that yields such a plural entity as 'installing a complex ' (for a formal description see Eschenbach, Habel, Herweg & Rehkamper 1 989). 'Installing a complex' should not necessarily be interpreted as building a whole new entity (which may be of semantical or syntactical nature): it can be interpreted as installing a plural address with pointers to the singular objects as well. In a near sense this second interpretation does not describe an installation process as no new representational objects need to be built in which both objects melt together (eventually by losing their individual identity), but the term is adequate if it is understood to describe the referential accessibilities for a plural pronoun. The process of installing . a complex is an additional component in the text comprehension process, which should take additional time over the resolution of a singular version. The following experiments use the ambiguous meaning of the German pronoun 'sie ' to investigate differences between singular and plural anaphoric pronominal resolution. The texts we used allowed both a plural and a feminine singular interpretation. In every case a female and a male person arc introduced, combined by different co-ordinations: 'and', 'as well as', 'neither/nor', 'with', 'without', and 'instead or. Firstly we tested our intuition about the preferred pronominal interpretation following these different co-ordinations. Prefer ences for a singular or plural interpretation should lead to a more frequent usc of singular or plural verb forms respectively in a sentence completion task. However, with a sentence completion method it is impossible to determine the exact point in time at which the complex is installed during text processing. To reveal information about this point in time a second experiment was conducted using an on-line reaction time measure.
350
Anaphoric Resolution of Singular and Plural Pronouns
semantically suggested, or that at least the singular concepts will stay in focus. Therefore a singular interpretation of the pronoun will be preferred, i.e. a singular verb will be chosen more frequently to complete the sentence. In addition we looked at the sequence introducing the two persons which is seen as a possible intermediating factor (john with Mary' versus 'Mary with John'). In line with the results of Gernsbacher & Hargreaves (1988; sec also Miisseler & Rickheit 1 990a), we assume that the first position which they call 'advantage of first mention', is important. Consequently, a feminine singular interpretation of the pronoun 'sie ' should be more likely if the female person were mentioned in the first position.2
Texts Experimental texts were designed according to the schema already mentioned: setting (s)
Die Hitze war unertriiglich .
(The heat was unbearable.) co-ordination (c) John und Mary schwammen im See. Gohn and Mary were swimming in the lake.) pronoun (p) Sie . . . (She/They . . .)
o, r, or 2 setting sentences (s) at the beginning give a general description of the setting, but without mentioning either of the two persons. The following sentence (c) introduces them by their Christian names, in each case a female and a male person, combined by one of the co-ordinations mentioned above (Table r ).
Design it is possible in German to continue the sentence (c) with a plural or a singular verb form after the co-ordinations 'as well as' and 'neither/nor', these two versions were separated. This resulted in eight different co-ordinative structures that had to be examined: 'and', 'as well as (sg)', 'as well as (pl)', 'neither/nor (sg)', 'neither/nor (pl)', 'with', 'without', and 'instead of'. As a further independent variable we distinguished the sequence of introducing the female and the male person varying on two levels ('female-male', versus 'male female', e.g. 'Torsten competed with Ina ' versus 'Ina competed with Torsten '). Thus, we had two different text types with the critical variation in the co-ordinating sentence (c) for each co-ordination. As
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Method
M. Hielscher and J. Miisseler
351
Table 1 One example for each co-ordinating structure used in the sentence completion task Example
'and'
Die Hitze war unerrriiglich. Fritz und Erna schwammen im nahegelegenen See. Sie . . . (The heat was unbearable. Fritz and Ema were swimming in the nearby lake. They/She . . .) Der Motor harte plotzlich ausgesetzt. Sowohl Karl als auch Gabi basteltelbastelten an dem Motorad herum. Sie . . . (Suddenly the engine had broken down. Karl as well as Gabi tried to repair the motor-bike. They/She . . ) Gestern war die AbschluBpriifung. Weder Uwe noch Agnes war/waren zu der Priifung erschienen. Sie . . . (The terminal examination was yesterday. Neither Uwe nor Agnes came to examination. They/She . . .) Die Fahrradtour fiihrte aufs Land. Torsten machte mit Ina ein Wetrrennen. Sie . . . (The bicycle tour passed through the countryside. Torsren competed with Ina in a race. They/She . . .)
'as well as' (sg/pl)
'neither/nor' (sg/pl)
'with'
'without'
'instead ot'
Die Aussrellung war gut organisiert. Anton war ohne Perra ins Museum gegangen. Sie . . . (The exhibition was well organized. Anton came to the museum without Perra. They/She . . .) Morgen ist die Geburtstagfeier. State Werner backte Elli den Geburrsragskuchen. Sie . . . (The birthday parry is tomorrow. Instead of Werner Elli made the birhday cake. They/She . . .)
Procedure Questionnaires were constructed which included one text of each of the 16 text types and 16 distractors in a random sequence. Distractor texts were constructed following the same reasoning as the experimental texts; they did not start with 'Sie . . .' in the sentence (p), but with another pronoun or noun referring to one of the concepts mentioned in the preceding sentence. Thus, each questionnaire consisted of 3 2 small texts. Subjects were instructed to read each text thoroughly and then continue the last sentence spontaneously in written form.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Co-ordination
3 5 2 Anaphoric Resolution of Singular and Plural Pronouns
Subjects Sixty-four students from different deparrments of the University of Bielefeld took part in the experiment. Their mother tongue was German. They were paid for participation.
Results and discussion
Table 2 Frequencies of singular versus plural sentence completions depending on sequence of introducing a male and a female person, and on co-ordinating structure Co-ordination
Position'
Singular
Plural
x2(df - 1 )b
and
f-m m-f f-m m-f f-m m-f f-m m-f f-m m-f f-� m-f f-m m-f f-m m-f
6 4 17 16 6 3 6 9
54 56 43 44 54 57 54 51 55 58 I7 40
3 8·4... 4 5 · 1 ··· I 1 .3*** I 3 . 1 *** 3 8·4··· 48.6*** 3 8·4··· 29·4··· 4 1 .7··· 52·3··· I 1 .3*** 6.7** 4 1 .7*** 4 5· 1 ··· 48.6*** 6o.o•••
as well as
(sg)) (pi)
neither/nor (sg)
(pi) with without instead of
•
2 43 20 55 S6 57 6o
s
4 3 0
Sequence of introducing rhe rwo persons, female person-male person (f-m) versus male person
female person (m-�.
h Level of significance:
...
P
< 0.00 1 , .. P < o.O J .
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Only fully answered questionnaires were entered into the data analyses; that left 6o questionnaires. For each of the 16 text types frequencies of plural versus singular completions X2-tests were peformed (Table 2). The results show that a plural sentence completion is preferred after introducing the two people co-ordinated by 'and', 'as well as', and 'neither/nor', all three being independent of the sequence in which the referents were intro duced in sentence (c); 'as well as' and 'neither nor' are also independent of the verb inflexion. For the preposition 'with' results show a preference for the plural inter pretation when the female version is introduced in the second position, i.e. in
M. Hielscher and J. Miisseler
35 3
E X PE R I M E N T 2 There arc three possibilities for the resolution process of plural pronouns: Firstly, one can argue that the plural complex is not installed before it becomes unavoidable in the text reception. In a sentence like example (2a) or (3a) there is no need to build a plural complex of both persons. However, as soon as one reads the pronoun 'sie ', followed by a plural verb in the next sentence (2b) or (3b), the pronoun has to be related to both john and Mary' and at this point the complex has to be built. In this case the pronominal resolution is a recursive search-and-match procedure as described by Clark & Sengul (1 979), Hirst & Brill (1 980), Garrod & Sanford (1982), Corbett & Chang (198 3), Sanford (1985), Matthews & Chodorow (1988), and Miisscler & Terhorst (1 990). The complex is triggered by reading the pronoun, in case of ambiguity by reading the pronoun plus plural verb form. A search of the preceding sentence yields no plural objects to which the pronoun can be related. A plural reference complex has to be installed from possible singular concepts. One can assume
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
the verb phrase. But i n the case of the female person being introduced in the first position, i.e. in the noun phrase, the singular interpretation is preferred. A singular sentence completion is preferred after introducing the two persons co-ordinated by 'without' and 'instead of', again independently of sequence. To summarize, using the sentence completion task, we found a preference for a plural interpretation of the pronoun 'sie ' following the three conjunctions, and a preference for a singular interpretation following the prepositions 'without' and 'instead of'. For 'with' we found a dependency of preferred interpretation on sentence position of the female person. In line with the argumentation ofSanford & Garrod (1 989), our results indicate that readers can use quite different information from the text to determine the pronominal referent, even in formally ambiguous structures. They usc semantic informa tion of the various conjunctions as well as sentence position. Following these results, one can assume that a plural reference complex has been installed for the conjunctions, and in part for 'with'. Nevertheless, with the sentence completion technique the temporal course of building a plural reference complex is not explicitly determined. It provides only the tendency to complete a sentence referring to one or both persons after reading 'sie '. No accurate information is contributed about the point in rime of the discourse processing at which the complex is installed and is an available part of the text representa tion; it is not necessarily the point in time when the pronoun 'sie ' (Miisscler & Rickhcit 1 990a) is read. This is the central question now in the main experiment using an on-line reaction-rime measure.
354
Anaphoric Resolution of Singular and Plural Pronouns
Method Texts Four-sentence texts were constructed following the same pattern as Experiment I :
m
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
that, in contrast to the resolution of a singular pronoun, the process�g time of the plural pronoun should be increased by the additional phase of building a reference complex. Correspondingly, hypothesis I claims that the singular pronoun is processed more quickly than the plural pronoun. An alternative prediction implies that building the complex is independent of reading the pronoun. The plural complex is installed while reading the first sentence, initiated by the co-ordinating structure. In this case, there arc two possible alternatives for the pronominal resolution process: the first claims that the singular concepts stay in focus together with the plural complex. Then the singular concept fits the pronominal resolution just as well as plural one. No processing time differences should occur (hypothesis 2a); the second claims that with installing the plural complex the singular concepts are deactivated, which can mean that they stay in focus but that the plural concept is marked and more easily accessible than the singular entities, or that they arc no longer represented in focus at all; both alternatives lead to a preference for plural interpretations in the following text. In this case a re-activation or a re-focusing of the singular concept is necessary for the resolution of the singular pronoun. This additional cognitive mechanism should be expressed in a processing time disadvantage for the singular pronoun (hypothesis 2b). Experimental evidence for one of these hypotheses should depend on the preposition or conjunction used. Results from preceding experiments (Miisscler & Rickheit I 990a) suggest hypothesis (2b) to be adequate for 'and', . and partially for 'with' (in a sentence structure with the female person in the second position, i.e. the verb phrase). Under the additional assumption that a plural preference requires a formerly installed complex, the results from Experiment I suggest the installation of a plural complex to be pre-pronominal and marked for the conjunctions 'as well as' and 'neither/nor', too. This hypothesis is not necessarily adequate for the disjunctive structures 'without' or 'instead or and for 'with' in a sentence structure where the female person is in first position (i.e. the noun phrase). For these co-ordinating structures, results from Experiment I showed a preference for a singular interpretation of the pronoun 'sie ', which might indicate that the singular concepts remain marked. In order to test the procedural hypotheses as they were elaborated above for the different co-ordinating structures, an experimental on-line technique with a semantic decision task was used.
M. Hielscher and]. Miisseler
setting (s)
3 55
Die Hitze war unertriiglich .
(The heat was unbearable.) co-ordination (c) John and Mary schwammen im nahegelegenen See . Gohn and Mary were swimming in the nearby lake.) Sie hatten!hatte" in der letzten Woche Schwimmen gelernt . pronoun (p) (They/She had learnt to swim in the last week.) Besonders Mary" war noch sehr vorsichtig. end (e) (Mary was especially very cautious.)
Design The critical pronominal parts of the texts in sentence (p) differ with regard to the pronoun 'sie ' referring to a singular or a plural concept of the preceding text ('pronoun') for each of the co-ordinating structures used in Experiment I.� The second independent variation again refers to the sequence of introducing the referents in sentence (c) ('male-female' versus 'female-male'). All conditions were presented to every subject in randomized order.
Procedure Altogether, 96 texts were presented to every subject in blocks of ten. Mter each block of texts subjects had the opportunity to take a break. In order to control the reading process the texts were presented to the subjects word-by-word on a computer screen with the moving-window technique. First, masks were presented on the screen in which every letter is replaced by an 'x'; blanks and punctuation marks remained unchanged. Such masks serve to facilitate the eye fixation behaviour of the subjects. Then, with a step rate of 408 ms, each mask was replaced by the whole word until the complete sentence was shown. Next, the sentence was re-masked again for 408 ms to facilitate eye movements to the beginning of the next line; then the following sentence was presented in the new line. The text presentation halted at the critical position for the referential resolution (after 'Sie ' plus verb form), marked by " in the example. As soon as the text was stopped, a square appeared beneath the text. This was the signal for the subjects to decide whether one or two persons were involved in the current sentence. They answered by pressing the left or the right button, respectively.5 The subjects received feedback as to whether or not their answer was correct.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The critical sentence (p) with the pronominal reference was presented in a singular or a plural form after the co-ordinating sentence (c). The verb inflexions ('hatten' versus 'hatte') definitely indicated which interpretation of 'sie ' was adequate. The sentences (e) were optional depending on how many setting sentences (s) were presented.
3 56 Anaphoric Resolution of Singular and Plural Pronouns
Altogether, subjects received 96 texts. Each text stopped at the critical text section, and 96 additional stops were introduced at other text passages, serving as distractors. Two seconds elapsed between presentations of the single four sentence texts.
Subjects Thirty-three students from different departments of the University ofBiclefeld volunteered for this study. Their mother tongue was German. They were paid for their participation.
One subject was excluded from further statistical analysis because of an error rate higher than I 5%. For the remaining J2 subjects the mean error rate was 4.2%. The reaction times for these wrong decisions were replaced by the mean reaction time for each subject under the specific combination of conditions. The data were entered into two ANOVAs for each co-ordination, one with 'subjects' and one with 'texts' as the random factor.6 For the three conjunctions 'and', 'as well as', and 'neither/nor' both analyses, over subjects and over texts, showed a main effect for the 'pronoun' factor: the decision time for the pural pronoun was about 40 ms shorter than for the singular version (Figure I). For 'and' the effect was significant, with P < 0.05 in both analyses (F.(I,J I) 6.34, (MSe - I 2862.27 and Fb ,u) - 8.83, MSe - 3454.22). For 'as well as' there was the same tendency in both analyses with P < o.Io (F.(1. 3 1) - 3.17, MSe = 1 9450.30 and Ft(1,u) - 3-34, MSe - 7803.59). For 'neither/nor' again a significant main effect was found, with P < 0.05 in both analyses (F.(1.31) 7.92, MSe = 1 5928.99 and Ft(1,u) = 4.76, MSe = 9 169.00). There was no main effect for 'sequence' and no significant interactions. These results suggest that hypothesis (2b) is correct for the three conjunc tions, i.e. a complex is installed while reading the first sentence, unrelated to the need to later refer to the plural entity. Semantically the three conjunctions seem to trigger building a plural reference complex while reading the sentence (c), thereby deactivating the single concepts so that the reference to a singular concept needs 're-focusing or re-activation', which takes additional time. For the three prepositions 'with', 'without', and 'instead of' there were no main effects of factors 'pronoun' and 'sequence' but significant interactions for 'with' (F.(1 ,31) - 6.25, MSe - 9290.54, P < 0.05, and Fb ,u) - 8.83, 4-23, MSe = 3454-22), P< 0.05). and for 'instead or (F.(1,3J) MSe = 10023.19, P< 0.05, and Fr(I,I 1) = 3.14, MS. = 501 8.52, P < 0.10). As
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Results and discussion
M. Hielscher and J. Miisseler 'and'
3 57
Sequence
liWiHl I}J
Male-female Female-male
720 700 680
MMM k/:1
u 740 .., "'
s
720
.§
700
.!2 u "' .., �
680
.., c
Male-female Female-male
660 640
760
Sequence
lMMJJ 1//1
740
Male-female Female-male
700
660
640 S i ngula r
Pronoun
Plural
Figure I Mean reaction times to a singular or plural pronoun for the co-ordinations 'and', 'as
well as', and 'neither/nor', introducing the female person firstly (f-m) versus the male person firstly (m-�
Figure 2 shows, the plural interpretation is preferred if the female person is in the verb phrase, that is, the second position for 'with' (john with Mary . . .') and the first position for 'instead or {'Instead of Mary John . . .'). The singular interpretation is preferred at least for 'instead or if the female person is in the noun phrase, i.e. the second position {'Instead of John Mary . . .'). This
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Sequence
760
358
Anaphoric Resolution of Singular and Plural Pronouns 'with' 760
Sequence
740
ifillill
Male-female
720
f:?;J
Female-male
700 680 660 640
�
740
s
720
.§
700
.gu
680
u Q) V>
Q) c
"' Q)
�
bMMI [3
Male-female Female-male
660 640
760
Sequence
740
II I::::::;:::J
720
Male-female Female-male
700 680 660 640 Si n gu l a r
Plural Pronoun
Figure
l Mean reaction times to a singular or plural pronoun for the conjunctions 'with', 'instead of' and 'without', introducing the female person firstly (f-m) versus the male person firstly (m-f)
correspondence possibly indicates that the syntactic function is the more relevant factor for marking a singular person (Grober, Beardsley & Caramazza 1 978; Wykes 1981; Corbett & Chang 198 3) than the position in a sentence. Hypothesis (2b) seems to be adequate if the female person is in the verb phrase: a complex is installed and focused. If the female person is in the noun phrase,
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Sequence
760
M. Hielscher and J. Miisseler
3 59
hypothesis (2a) may be correct for 'with': a complex is built but not marked; and hypothesis ( I b) seems to be correct for 'instead of': the singular female person remains marked. For 'without', no effect was found at all. Therefore hypothesis (2a) seems to be adequate to interpret the results: a complex is installed but is not focused, unlike the singular concepts. GENERAL D I S C U S S I O N A N D C O N C L U S I O N
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
A brief review of the results from both experiments for comparison shows that the data we obtained using quite different methods are largely compatible. For the three conjunctions used, 'and', 'as well as', and 'neither/nor', results from Experiment 2 clearly showed an advantage for a plural interpretation of the pronoun 'sie ' in the critical sentence (p). The mean reaction time advantage was about 40 ms. These results supported hypothesis (2b). Comparing these results 'With the results from Experiment I under the additional assumption that a plural preference requires a formerly installed complex, the reaction time advantage was predictable. Here the results of the two experiments are compat ible, suggesting that the complex is built triggered by the conjunction before the pronoun is read. The pronoun 'sie' may be in no way ambiguous for the reader in this case (Sanford & Garrod I989) and will be resolved immediately. The re-activation or re-focusing of the singular concepts when reading a singu lar verb form afterwards requires additional processing time compared to the plural verb form. The results for the prepositions are slightly more complicated. In Experi ment 2 for 'with' and 'instead of' an interaction was found between reference to a plural or singular concept, and 'sequence'. Indeed, it seemed to be more the influence of the syntactic function than the sentence position itsel£ For 'with' a clear reaction time advantage for the plural interpretation was found if the female person was in the verb phrase (hypothesis 2b). No reaction time differ ence for the singular and plural decisions was found if the female person was in the noun phrase (hypothesis 2a). Comparing the results from both experiments, a small difference is found. Following the results from the sentence completion technique a disordinal interaction was expected, but only a hybrid interaction was found: there was no difference between singular and plural decisions if the female person was in the verb phrase. For 'instead of', the singular interpretation of the pronoun is favoured if the female person is in the noun phrase. This is not the case if the female reference person is in the verb phrase. Here the reaction time for the plural expression decreased. Following these results, one must conclude that hypothesis (2b) is correct for 'instead of MaryJohn . . .'. A complex (:John' and 'Mary') is installed and focused. Hypothesis ( I b) seems to be correct for 'instead ofJohn Mary . . .'.
360
Anaphoric Resolution of Singular and Plural Pronouns
.
.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The singular objects remain focused or, in fact, no complex is built. For the preposition 'without' no general processing time advantage was found for either the plural pronoun or the singular one. Here hypothesis (2a) seems to be adequate: the complex is built, but the singular and plural concepts have equal rights in the pronominal resolution process. One problem with the results for 'without' and 'instead of' is that they arc not in line with the results of Experiment 1. The completion task clearly suggested an advantage for the singular pronoun which was not found in Experiment 2, except for 'instead of' with the female person in the noun phrase. One reason could be the methodical problem with the text structure, always introducing two persons, which could have led to a general advantage for the plural decision (see note 4). However, as there were several interactions with advantages for the singular pronoun, this is not a very convincing argument. Instead it might be advantageous to look at the experimental methods we used in more detail. The off-line sentence completion task combines reception and production processes, while the on-line reaction-time technique reflects only a reception measure. Thus, possibly at an early stage of text processing, the singular and plural concepts are both available; consequently, no differences in reaction time measurement of reception were found. However, in the sentence completion task the subjects had to produce a coherent text by themselves. Therefore, they had to look for arguments why, for example, 'Mary withoutjohn . ' or 'instead ofjohn Mary visited an exhibition '. It may be easier to produce a coherent text by referring to 'Mary' only. So the singular pronoun was favoured, a tendency that becomes evident in the produced texts. Following this point of view, reception processes are much more flexible than production processes, at least in the context of pronominal resolution. All together, the results indicate that under certain conditions a plural complex is installed before the plural pronoun is read. This facilitates the resolution of plural pronouns. Further, one can argue that the installation of a complex serves to facilitate later reference relations. In this case pronominal resolution is more than a mere recursive search-and-match procedure triggered by reading the pronoun. Such a recursive search-and-match procedure is conceptualized in most of the current theories on pronominal resolution (e.g. Clark & Sengul 1 979; Sanford 1 985; Garrod & Sanford 1 982, 1985; van Dijk & Kintsch 1983; Marslen-Wilson, Levy & Tyler 1982;just & Carpenter 1 987). The theories differ both on content and on the underlying processes but agree on the main aspects of the pronominal resolution (Miisseler & Terhorst 1990; Miisseler & Rickheit 1 990b): their first assumption is that the resolution process starts with the encoding of the pronoun. In the simplest case, the resolution of the pronoun can be realized by the agreement of the gender and number of the pronoun with the reference object. If there is more than one possible antecedent such a
M. Hielscher and J. Miisseler 361
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
simple comparison of gender and number is insufficient. Thus, the second assumption to make is that the cognitive system has to restrict the reference domain. This process can be guided by syntactic, semantic, and textual features of the reference concept. The third assumption conceptualizes pronominal resolution as a search-and-match procedure within the reference domain which establishes the relation between the pronoun and the reference concept . (Clark & Sengul 1 979; Hirst & Brill 1 980; Garrod & Sanford 1982; Corbett & Chang 198 3; Sanford 1 985; Matthews & Chodorow 1 988). In line with Sanford & Garrod (1 98 1 : p. 141 ), our results indicate that the cognitive system is working more efficiently for processing further referential relations. As far as the pronominal resolution is concerned, this preparation could have two effects: first, possible referents are focused or marked in focus by the processing system leading to a simple identification of the antecedent, by searching and matching only a few alternatives. However, the resolution process is triggered after encoding the pronoun, but restricted more to the focused antecedents; this way the search process will be facilitated. Second, an alternative view would be that a search process is not the only possible process for pronominal resolution. In a previous paper (Miisseler & Terhorst I 990) we discussed arguments and evidence for the claim that focused or marked discourse referents activate suitable pronouns as their potential future reference forms; recursive search processes may therefore not be needed in these cases. This pronominal occupation is supposed to be the product of the normal comprehension process and does not require any specialized processes (see Bosch 1 988 for a related idea of resolution without search or selection). In terms of a connectionist view, one can say that reading a Christian name activates the pronoun node together with the person concept node; this associated activation of both (or maybe there is only one node representing both concept and pronoun) results directly from their common features, in the first place number and gender. As pronouns are activated prior to encountering the anaphor, no recursive search-and-match is required because the referential links have already been established. In the present experiments there is no direct evidence for the resolution by a facilitated search-and-match or by an occupation procedure. If the reaction times of the plural pronouns would have been increased in each condition and thus would indicate that the installation of the complex was initiated after reading the anaphor, then the recursive search-and-match resolution should be favoured for the plural pronoun. Hence the present results, which are in part opposite, fit both explanations. Other results reported by Sichelschmidt & Gunther ( 1 990) provide evidence for the occupation approach, but further research is necessary to decide which of the two procedures is used by the cogrunve system.
362 Anaphoric Resolution of Singular and Plural Pronouns
Acknowledgements ·
The research reporred here was supporred by grants from the Deutsche Forschungsgemein schaft {Ri 3 1 4/9). Thanks are due to Julie Carson and Andrea Reuther for many insighrful stylistic suggestions. MARTINA HIELSCHER and JOCHEN MOSSELER Fakultiitfiir Linguistik und Literaturwissenschajt A rbeitseinheit Psycholinguistik Universitiit Bielefeld Posifach 8640 D-48oo Bielefeld 1 FRG
1 I n this case the pronoun can be used. If both persons were of the same gender it would not be possible to usc the pronoun in this syntactical structure ('Peter andJohn wanted to have a picnic. He . . . '). 2 Note that there is a difference between sentence position and syntactic function. First versus second sentence position is nor always the same as syntactic subject versus object. This is only the case for 'with' and 'without'. In German 'instead of' is used more usually with an object-subject structure. 'And', 'as well as', and 'neither/ nor' put both people into the noun phrase. The number of analyses is statistically critical, but since the significance is always above o.oo 1 (with one exception of o.o J ) a type I error is not probable. 4 There is a methodological problem due to the text structure used: in all texts two possible reference persons were intro duced, a male and a female one. With regard to the subject's task (one or two persons?), one can argue that the decision 'two persons' is generally favoured. What is measured then would include a general reaction tendency to the global amount of possible reference persons in the texts, always two, which facilitates the resolu tion of a plural pronoun. The adequate
control condition would imply introduc ing only one possible reference person and referring to it with a singular or plural pronoun. With such texts the problem gets even worse, because in such texts no plural pronoun can be used at all. Thus, if two persons arc introduced as opposed to texts where only one person is introduced, then the plural 'sie' becomes even more prob able in texts introducing two persons. Thus, we decided to use two persons in all texts and refer to one or both via the pro noun with the same probability, so that the subjects receive no hint as to which decision might be correct, and a preference for the plural decision wouW be even hindering for the task. However, results have to be interpreted cautiously if the plural decision is facilitated in all texts independently of the variation we used. In this experiment the right burton stands for 'two persons' and the left button stands for 'one person', but results of a pilot experiment with reversed answer modus indicate no or at least negligible small reaction rime differences. 6 Exactly speaking we conducted 2 X 2-ANOVAs for rhc co-ordinations 'and', 'with', 'without' and 'instead of'. For 'as well as' and 'neither/nor' an additional
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
NOTES
M. Hielscher and J. Miisseler 36 3 faccor was inrroduced, according co the possibility of using a singular or plural verb form in sentence (c) in German. Here 2 X 2 X 2-ANOVAs were calculated. As
the additional factor yields no significant main effects or interactions the following results were presented by averaging over this factor.
REFERENCES
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Bosch, P. (1 988), 'Representing and accessing Marslen-Wilson, W. D., E. Levy & L. K. Tyler (1 982), 'Producing interpretable focused referents', Language and Cognitive Processes , 3: 207-3 1 . discourses: The establishment and main Clark, H . H. & C.]. Sengul (1 979), 'In search tenance of reference', in R J.Jarvella & W. of referents for nouns and pronouns', Klein (eds), Speech, Place, and Action , Memory and Cognition , 7: 3 5-4 I . Wiley, Chichester, pp. 3 3 9-78. Corbett, A. T. & F. R. Chang ( I 9H3), 'Pro Matthews, A. & M. S. Chodorow ( 1 988), noun disambiguation: accessing potential 'Pronoun resolution in rwo-clause sen antecedents', Memory and Cognition , n : tences: Effects of ambiguity, antecedent location, and depth of embedding',journal 28 3-94· ofMemory and Language, 27: 245-60. Ehrlich, K. (I 9Ho), 'Comprehension of pro nouns', Quarterly journal of Experimental Moxey, L. M. & A.]. Sanford ( 1 988), 'Quanti fiers and focus', journal of Semantics , s: Psychology, 32: 247-55· I 89-206. Eschenbach, C., C. Habel, M. Herweg & K. Rehkamper( 1 989), 'Remarks on plural Miisseler, J. & G. Rickheit (I 990a), 'Kom anaphora', Proceedings of the 4th Conference plexbildung in dcr Texrverarbeirung: die kognitive Auflosung pluraler Pronomcn of the Association for Computational Linguis tics, Manchester, I o- 1 2 April I 9H9. (Complex concepts in discourse process Garrod, S. C. & A.J. Sanford (I 9H2). 'The ing: the cognitive resolution of plural pro nouns)', Zcirschrift fur Psychologic , 198: mental representation of discourse in a 69-8 I . focused memory system: Implications for the interpretation of anaphoric noun Miisseler,J. & G. Rickhcit ( I 990b), 'Infercnz phrases', journal ofSemantics , I: 2 I -4 1 . und Rcferenzprozesse bei der Tcxrvcr Garrod, S. c. & A.]. Sanford ( I 9 8 5). 'On the arbeirung (Inference and reference real-rime character of interpretation dur processes in discourse processing)', in S. ing reading', Language and Cognitive Pro Felix, S. Kanngielkr & G. Rickheit (cds), cesses , I: 43-59· Sprache und Wissen , Wesrdcurschcr Verlag, Opladen, pp. 7 I -97· Gernsbacher, M. A. & D. J. Hargreaves ( I 988), 'Accessing sentence participants: Miisscler,J. & E. Terhorst ( I 990), 'Pronomin ale Beserzung: ein alrcmarivcr Mcchanis the advantage of first mention', journal of mus neben der rckursivcn Auflosung? Memory and Language, 1.7= 699-7 I 7· Grober, E. H., W. Beardsley & A. Caramazza (Pronominal occupation: an alternative (I 978), 'Parallel function strategy in pro mechanism in addition co the recursive noun assignment', Cognition , 6: 1 I 7-33. resolution?)', Sprache und Kognition , 9: 3749Hirst, W. & G. A. Brill (1 980), 'Contextual aspects of pronoun assignment', Journal of Sanford, A.J. ( 1 98 5), 'Aspects of pronoun Verbal Learning and Verbal Behavior, I9: interpretation: Evaluation of search for 168-75· mulations of inference', in G. Rickheit & Just, M. A. & P. A. Carpenter (I 987), 'The H. Srrohner (eds.), Inferences in Text Process Psychology ofReading and Language Compre North-Holland, Amsterdam, ing, pp. I 8 3-204. hension , Allyn and Bacon, Boston, MA.
364 Anaphoric Resolution of Singular and Plural Pronouns Sanford, A J. & S. C. Garrod (1981), Under standing Written Language, Wiley, Chi chester. Sanford, A J. & S. C. Garrod (1 989), 'What, When, and How?: questions of immediacy in anaphoric reference resolution', Langu age and Cognitive Processes, 4: 2 3 5-62. Sichelschmidt, L. & U. Gunther (1990), 'Interpreting anaphoric relations during
reading: Inspection time evidence', journal ofSemantics, 7: 321-45). van Dijk, T. A & W. Kintsch (1 983), Strategies ofdiscourse comprehension , Academic Press, London. Wykes, T. (19!! 1), 'Inference and children's comprehension of pronouns', Journal of Experimental Child Psychology, JZ: 264-79.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
© N.I.S. Foundation (1990)
journal ofSrnrantia 7: 365-378
Description Types and Method of Conjoining as Factors Influencing Plural Anaphora: a Continuation Study of Focus
A . J . SANFORD and F. LOCKHART University ofGlasgow Abstract
I NTRODUCTION Establishing the conditions under which plural reference is possible or even preferred has received relatively little attention until recently. The present paper is concerned with the conditions under which individuals introduced separately become available for reference by means of plural pronouns. Plural anaphors may be used to refer to collections or sets ofthings which arc introduced in a variety of different ways. One major subgroup comprises collections or sets introduced through a single description, as in the class , or through a plural description, as in horses or hot days . In these cases, pronominal reference is narurally plural, as with they , them , these , and so on. Generics and conceprual anaphora (Gernsbacher 1986), and subsets resulting from quanti fication (Maxey & Sanford 1 987) may be included in this group, although they present their own problems for anaphoric interpretations. A second major subgroup for which plural anaphora may be appropriate is collections or sets introduced through descriptions which, in fact, identify the individual elements out of which the sets arc formed. The most obvious of
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
An experiment is reported which investigates the impact of two variables on the likelihood of obtaining plural pronoun anaphors in a continuation task. The first variable is syntactic: the use of and versus with as a means of relating two singular characters. Use of and enhances the likelihood of obtaining a plural anaphor in continuations, but the incidence of plural is never as high as 6o%. The second variable is description type: whether the characters arc introduced through proper names (e.g. Harry), or through a simple definition noun phrase (e.g. the doctor). When two descriptions are of the same type, plural reference is enhanced, regardless of syntactic condition, and regardless ofwhether the two descriptions arc names or noun phrases. This effect is traced in part to the special status afforded to characters which arc introduced through proper names. A discussion is presented within the framework of an account of plural anaphora based on singular individuals mapping into common role-slots in background knowledge (scenario).
366 Description Types and Method of Conjoining as Factors lnfluecing Plural Anaphora
these is the use of connective and , which offers direct syntactic support for grouping, as in ( I ): (I) Jack and Jill went to the pictures. They . . . Sometimes the basis for grouping may be somewhat more tenuous, although plural reference remains entirely possible, as in (2) and (3):
(2) Donald went to the village with Margaret. They . . . (3) Donald chased after Margaret. They . . .
Accounts ofplural reftrence and the context ofthe present experiment The present study was conceived ofin the context of a theory that what licenses plural reference is the possibility of mapping atomic individuals into a common role-slot in some interpretative structure (work in progress). We are assuming that when a sentence is understood, it is understood by mapping its constituents into background knowledge structures (a central notion in the framework
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Our concern in the present paper is with the conditions which license or favour plural reference when individuals arc introduced in such ways. With (I )-(3), it is clear that reference with he , she, or they is possible. What variables determine a preference for one form of reference over another? How arc such preferences to be understood in terms of underlying processes? Our questions arc thus more concerned with preferences than with boundary conditions in which plural references may become impossible. Two variables have been mooted as possible determinants of success in making plural reference. One variable, ontological sameness, holds that plural pronominal reference becomes possible when two mentioned individuals (henceforth 'atomic individuals') are of the same ontological type (Herweg I 988). On this line of reasoning, two humans mentioned separately might be good for plural reference, while a human and an animal will be less good (although plural reference is by no means ruled out in many cases). This is closely related to the more general idea of a common association base (CAB): the argument is that some sort of conceptual link has to be formed between individuals which have been introduced if they arc to be referred to by a plural pronoun. A CAB is viewed as necessary for the licensing of plural reference (Eschenbach, Habel, Herweg & Rehkampcr, I 989). A second variable is the syntactic relation which holds between individual atomic descriptions, so conjoining atomic individuals through the usc of and, for instance, is taken as a very strong influence in bringing about felicitous plural reference.
A J. Sanford and P. Lockharr 367
devised by Sanford & Garrod 1 98 1 , and some other theories). One example of such a structure might be a main verb-schema (e.g. Rumelhart & Ortony I 977), with the mappings representing cases. For example: (4) John asked Mary out.
( s ) John showed Mary how to mend a car. The two characters play different roles in the conception suggested by the sentence, and it is likely (though not inevitable), that they will continue to play different roles in any future development of the discourse. In conrrast, we would expect an example like (6) to produce a relatively high rate of plural continuations, since it is likely that the two meet in order to carry out joint activities. Thus in an elaborated set of mappings, the two atomic individuals will play a common role in some of them: (6) John met Mary at the station. Some verbs, like 'meet', arc likely to be highly associated with events where people play a common role in some developing story. Full discussion of this model is ourwith the scope of the present paper, and we arc leaving vague the discussion of rypcs of role-mapping which might be important: for the present purpose, by role-mappings of atomic discourse elements, we mean any mapping between a description and an underlying knowledge srrucrurc which has been invoked during the understanding process (and, of course, there may be more than one). This line of reasoning is intended to provide a means for felicitous plural and
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
might be represented by mappings in which John -+ Agent and Mary -+ patient in the verb schema. In a number of theories of text comprehension, such mappings are assumed to be made (potentially) at a number of levels. For instance, if someone knew that John and Mary always talk a lot when they go out, then that inference might be made in the face of example (4), leading to a further set of mappings in which both John and Mary map into the same role slot (both arc agents in a talking schema). Within the framework which we arc currently developing (an extension of the framework of Sanford & Garrod ( 1 98 1; Garrod & Sanford 1 982), a plural reference is possible if two (or more) atomic clements map into a common role slot. If they do not map in this way, then plural reference with the atomic individuals is not possible. However, the mapping may be in terms of repre sentations which arc close to the utterance in question, such as the verb-schema rypc of representation, or to srructurcs based on more elaborate inferences, such as the one described above. With this conception, it is possible to explain why, in a continuation task, some examples, like ( s ), produce a very low incidence of plural continuations:
368 Description Types and Method of Conjoining as Factors Influecing Plural Anaphora
Constraints on plural reference Although usc will be made of the concepts discussed above, the experiment reported below does not in any way test various theories of plural reference resolution. Rather, it is an examination of constraints which militate for or against plural reference. The first type of constraint examined here is the usc of and or with as a means oflinking two individuals, as in (7) and (8): (7) Mary andJohn went to the shops. (8) Mary went to the shops with John. It is generally assumed that the usc of and will enhance plural reference in this example. The second constraint is the form of description of the atomic individuals. Two important ways of referring to individuals (characters) is by means of proper names such as Henk , or by means of fuller, definite noun phrases such as the plumber. This latter type of description specifies a typical role which a character plays in a scene, but the term role will be reserved for the role-filler
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
singular pronominal reference under appropriate circumstances. Provided atomic individuals are mapped to the same role-slot in one representation, then plural reference is possible, and provided they map to different role-slots in another (simultaneously), then singular references are possible. However, one set of mappings may predominate over others, thus producing a bias in the availability or accessibility of plural or singular referents. Approaching the same problem from a different perspective, Eschenbach, Habel, Herweg & Rehkamper (1 989) introduce the notion of a complex referential object (complex ref-0). They suppose that the presence of a suitable CAB allows the formation of a complex ref-0 from the atomic objects related through the CAB. The complex ref-0 is that which enables plural reference to take place. Much of the work of these researchers is concerned with the conditions which either favour or result in the formation of a complex ref-0 (hence enabling plural reference), or preclude complex ref-0 formation. There is at least one similarity between the two approaches: there has to be some mechanism by which something corresponding to a complex ref-0 is repre sented. In an account based on the Sanford-Garrod framework, we take it that plural reference is possible if more than one atomic entity is mapped into a single role-slot in an elicited representation. in a sense, two or more pointers from atomic descriptions to a common role-slot constitute a condition for plural reference, like a complex rcf-0, while pointers from atomic objects to different role-slots constitute the condition for singular reference, like simple rcf-Os. If multiple mappings arc allowed, then the equivalent of both simple and complex objects can co-exist.
A J. Sanford and P. Lockhart 369
notion discussed already, to avoid confusion. We will call these descriptions noun, and the first type name. Consider the following pair of examples: {9) Mary and John ran into the house. (w) Mary and the plumber ran into the house.
{I I) Mary and the plumber talked. we may expect continuations to have to do with plumbing, in which the
plumber is offering advice to Mary, and so they arc playing different roles. Another case of interest is that of two characters introduced through noun phrases: (12) The doctor and the plumber talked. Here there are many opportunities for the two individuals to be mapped into different roles. The plumber is ill, the doctor's pipes have burst, and so on. The authors' intuitions arc to the effect that in (12), a plural continuation would be preferred, however. If experimentally it is shown that there is no difference in the number of plural continuations which occur with two names and with two roles, then we would conclude that common description type (rather than description content) is a major force in controlling the likelihood of plural reference. Whether any of these intuitions arc valid depends on the data yet to
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Intuitively, it seems to the authors that (9) would lead more readily to continuations in which Mary and John are referred to by they or them than would be the case with (w). The validity of this intuition awaits the report of the experiment, but if this were the case, what foundation could it have? There arc several possibilities. First, proper name and noun arc different description types, serving different functions. Indeed, their different uses have been investigated by Sanford, Moar & Garrod {I 988; Garrod & Sanford I 990), who showed that pronominal references in continuations arc more likely to be made to proper-named characters, and that characters introduced by proper name arc more easily accessed for anaphoric reference, as indexed by self-paced reading time. The argument is that proper name is an important signal to the processor to treat the referent as a main character, which tends to separate that character from others in terms of the roles they play in interpretative scenarios (sec also Garrod & Sanford 1988 for a fuller discussion of the concept of main character). So, one reason for getting less plural continuations with (w) might be that the proper name serves as a strong cue for treating the named character as special and separate from other individuals. Another possibility is that the noun descrip tions suggest a particular set ofactivities which become dominant in addressing schcmas for interpretation. For example, plumber might suggest the set of activities such as getting pipes mended , etc. So, given a sentence like
3 70
Description Types and Method of Conjoining as Factors Influecing Plural Anaphora
E X PE R I M E N T
Method Design and materials Forty one-line materials were devised depicting two characters in easily imagined situtions. The characters could either be introduced by using proper names (name) for one or both, or by a definiteNP (noun ) for one or both. Two types of sentence structure were used. In the first (the '&' condition), the two characters were conjoined using and , followed by a verb-phrase. In the other ('with' condition), the characters were separated by a verb phrase and conjoined by with . Thus, eight versions were possible for a given material. A typical material, showing the eight versions, is: ( 1 3) Aileen/The girl and Steve/the boy ran into the cinema. (14) Aileen/The girl ran into the cinema with Steve/the'boy. Following the discussion made earlier, in making up the examples, care was taken to eliminate any action where there was a clear role-asymmetry between the characters. Thus the interactions were all things which did not clearly put one of the protagonists in an agent position and the other in a patient position. This was done by examining the name cases. if there was any obvious possibility of separate role-mappings predominating at the level of sentence-based case, the material was not used. This was done to maximize the possibility of detecting contrasts over conditions, since an asymmetry would be expected to lead to high proportions of singular continuations which might make differences between conditions hard to detect.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
be reported. It is possible to analyse possible interactions between the syntactic cue and description type also, but these arc best understood on the basis of the data. There arc several possible methods to usc to address the problem of whether plural or singular reference preferences exist. For example, reading time and related on-line measures can be used to determine the relative speed of reference resolution (c£ Garrod & Sanford 1 982 for an application to plural anaphora). Alternatively, subjects can be invited to write continuation sentences after a sentence of interest, and the incidence of plural or singular anaphoric pronouns in the continuations can be used as an index of preference. These two methods arc not equivalent, since the time-scales over which reading time integration and generation of a continuation take place arc different. In the experiment below, we usc the continuation method in order to see which variables lead to preferences for singular or plural reference.
A. J.
Sanford and P. Lockhart 3 7 1
Subjects Thirty-two subjects were tested. They were students at the University of Glasgow, and had no knowledge of the point of the experiment.
Procedure Each subject was given a booklet consisting of seven pages. The first contained written instructions. The remainder each had ten sentences with a ··line underneath where a continuation was to be written. Subjects were told to add a continuation sentence to each of the sentences presented, so as to continue the 'narrative' in a sensible way. The need to produce complete sentences was emphasized.
Results and discussion Simple plurals The continuations were classified as to whether or not a plural description was used which referred to the target individuals, or whether singular references were used. In the case of singular references, the target individual referred to was noted. The proportion of occasions on which a plural reference (they, them, both, etc.) was made is shown in Figure I. Two effects arc striking. First, syntactic grouping through the usc of and increases the probability of a plural reference being used in continuations, as expected. Second, similarity of description-type raises the probability: whether both individuals arc described by role descriptions or proper names, the enhancement is much the same. Plural references arc less likely to be elicited by mixed description-types. Analyses of variance by subject and by materials support these observations.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
A given subject saw a sentence of 40 materials, each material appearing in only one of its eight possible versions. So that each material could appear in every version in the complete design, there were eight separate sequences. If material I appeared in its condition I form in sequence I , then it appeared in its condition 2 form in sequence 2, and so on to sequence 8. In a given sequence, five of the materials appeared in condition I form, five in condition 2 form, and so on, so that a given subject, receiving one sequence, saw 40 materials, five in each ofthe eight conditions. However, a subject did not sec the same material in more than one of its forms. Since there were 32 subj ects, four subjects saw a given sequence. The order of the sentences was randomized within a sequence, and a further 20 sentences, unrelated to the present design, were introduced as fillers.
A. J.
Sanford and P. Lockharr
3 73
but the difference is not substantial or reliable. Thus, while conjoining with and serves as a powerful cue to put the atomic individuals together, it can be offset by mixing the description rypes of the atomic descriptions.
Influence of names in mixed conditions
Noun-with-name Name-and-noun • Noun-and-name l!:l Name -with- noun III
0
3
"'
·= E
0 u ... 0 ... G) .C)
2
E
::l c c "' G)
::E
Name
Noun Continuation to
Figure 2 Mean number of singular continuations co-referential with either the named or the
role-described character in the mixed description conditions.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Turning first to the tendency to produce relatively less plural references under the mixed condition, it is possible to analyse the effect in terms of the different effects which name and noun descriptions might have. One argument given earlier was that the processor might treat a proper name as signalling the status of main character, and so bring about a relatively high proportion of singular continuations made to this character. The other argument was that different description rypes might merely reduce the tendency to group the atomic elements together without any prediction regarding which singular continua tion would predominate. Figure 2 shows the distribution of reference to name and noun descriptions in the case of singular pronominal anaphors to the name-noun and noun-name conditions. It is clear that there is a strong
374 Description Types and Method of Conjoining as Factors Influecing Plural Anaphora
Occurrence of 'both' In some cases, plural continuations were complex in that constructions included the term both , as in Both . . ., They both . . ., and Both ofthem . Although the incidence was not high, an attempt was made to discover when such constructions might be likely to occur. One possibility relates to the many cases where plural reference is made under conditions which our results show to be relatively unfavourable. It seems possible that the occurrence of both in they both and related expressions might serve to reinforce the plural grouping which is required by the presence of the pronoun alone. Thus, in conditions where to continue with a plural may produce some kind of strain, both may serve to ameliorate the strain. If this conjecture is at all sound, then the distribution of occurrences of both in conjunction with a pronoun should be more prevalent in just those cases where plural references arc in the minority, i.e. in the 'with' condition, and in the mixed description condition. The distribution of occurrence of both is shown in Table I. The numbers arc collapsed by same description types, since the absolute incidence of both in plural conditions is not that great. The only reliable difference in Table I is that between the pooled combining conditions, and versus with : by a chi-square test, X 2 = I 1.9 for df = I , P < o.o r . Thus, using a plural in a continuation based on the verb + with conditions, where we might expect some sort of strain, docs indeed produce a reliably higher number of continuations using both . However, it might be supposed that the poorest conditions of all for producing plural continuations (mixed conditions, especially in the 'with' condition) would show the highest proportion of use of both . This is not the case, as can be seen from Table I , even in terms of simple numerical trend. It is difficult to interpret this without further purpose-designed experimentation, but it is still possible to speculate about possible reasons for the observed pattern. One possibility is that the
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
tendency to continue with the named individual, the only exception being the noun-with-name condition. (The discrepancy could perhaps be explained if it is assumed that there is also a weak preference to continue with the first-named character.) Statistical analyses using Tukey's method by subjects and by materials showed that the incidence of singular continuations to the name noun contrasts shown in Figure 2 were reliable at at least the o.os level. It is therefore arguable that the tendency towards singular reference with themixed-dcscriptions is at least in part due to a strong tendency for named characters to be used as thematic subjects or main characters in continuations, although there is a small first-mention effect. Thus, part of the story must be that the name/noun contrast puts the two individuals in different roles with respect to situation, with the asymmetry resulting from the thematic subject component.
A.J. Sanford and P. Lockhart 375 Table 1 Occurrence of 'borh' in plural references as a function of condition Description
Combining condition A and B
Pooled same Pooled mixed Overall
0. 1 2 0.08 0. 103
A verb wich B [ 1 9] [9] [28]
0.29 0. 1 9 0.2 1
[22] [I I) [33]
The first numbers refer to the percent o f occasions o n which both i s included, conditional on a plural reference being made anyway. The numbers in square
not bear a simple correspondence to numbers. Because of the relatively small number of cases, when rwo description rypcs arc the same, they arc poo led under the Same group. The rwo orders ofmixed descriptions arc likewise pooled under the Mixed group. the only statistically reliable effect is on the Overall difference between combining conditions.
tendency to usc both is restricted to reducing strain induced by unfavourable syntactic constructions (i.e. verb-with as against and). An alternative is that under conditions which arc maximally unfavourable for plural reference, when a plural reference is made, it is made only because no strain exists in those (rare) cases. CONCLUSIONS The results of this study reveal a number of factors which vary the likelihood of plural references being made in a continuation task. In particular, homogeneity of description-type and direct conjunction through the connective and produce the most favourable environments for plural production. First consider the overall incidence ofplural references. The experiment was designed using situations and events which do not clearly put individuals in different major roles. For example, ifJohn ran into the cinema with Mary, then it is quite possible to read this as them both taking some action together. Other types ofinteraction can be much more marked in cueing individuals into major roles. The situations chosen were so chosen in order to achieve a reasonable level of plural continuations. In other studies where two individuals have been introduced, plural anaphora in continuations have typically been far below 1 0%. For instance, in the study of the effect of using proper names (Sanford, Moar & Garrod 1 988), individuals were introduced through such interactions as: (1 s) Mr. Bloggs was dictating a letter. Mary was taking shorthand. Now there is nothing to prevent plural continuation here (e.g. They had been working hardfor hours), but the incidence of such continuations was less than 5%,
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
brackets arc the absolute numbers, given to illusrratc the incidence of both in rhc dara. Since base rarcs of plural description vary over condition, proportions do
3 76 Description Types and Method of Conjoining as Factors Influecing Plural Anaphora
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
and continuations mostly were confined to singular references with individuals being in relatively stereotypical roles with respect to each other. Even though the actions in the present experiment were specifically designed to eliminate any overt suggestion of role differentiation, singulars still accounted for nearly half of the continuations under the best condition for plurals (same description type, use of and). While it has been widely recognized that both plural and singular references are possible under many circumstances (Eschenbach, Habel, Herweg & Rehkamper 1 989; Garrod & Sanford 1 982), the present results reveal just how prevalent singulars can be, even given a grammatical grouping cue like and. In terms of the role-mapping model of plural reference, we would suppose that the use of and imposes a constraint which specifics the two individuals as playing the same grammatical role. However, in terms of formulating what happens next (as required by a continuation study), the constraint is soft in that it only raises both the probability of the two characters playing the same role in an elaborated representation. The second major observation is that mixed description types result in a low incidence of plural continuations. Two possible reasons were identified for this. First, different description types (noun/name) may suggest that the individuals depicted in the descriptions concerned arc to be contrasted. A possible basis for the. contrast was suggested: proper names single out principal protagonists in narratives, and so, even in the simplest of materials, will tend to put the named character in a role which is likely to be different from that played by a character introduced through a noun. Indeed, it is plausible that in looking for a way to relate a proper-named individual to one introduced by a noun, the processor would be dominated by stereotypical associations. Thus, ifJack and the girlgo to the cinema , the descriptions and setting may suggest thatJack is wooing the girl, for instance. This possibility may be less prominent if it wereJack andjill who went to the cinema. A second possibility is that noun descriptions arc always contrasted, and so will lead to the atomic individuals being mapped into different roles in possible scenarios. So, if an actress and a doctor arc having coffee, say, the processor may use the different stereotypical activities associated with these sorts of character to specify a contrast between them. In the present study, the noun-noun combination does not produce any more reliable plural continuations than does the name-name combination, so such a contrast mechanism is not operating in the present study. However, since, in the mixed condition, the high incidence of singular references is associated with continuations based on the named-character, support is gained for the idea of contrast over mixed description types, influenced or driven by a tendency to isolate a named character through its elevation to status of 'principal character' (c£ Garrod & Sanford 1988; Grimaud 1 988). The observation of expressions using both is intriguing. It is possible that this
A.]. Sanford and P. Lockhart 3 77
'
(16) Jack and Pauline met Richard and Liz at the station. (16 ) They (all) went our for dinner. '
Finally, the sentence production task used in the present srudy is an easy way of determining preferences for plural or singular reference patterns. However, it is recognized that other tasks may well tap other aspects of plural reference. For example, while the present data show how soft constraints affect preference (and thus the state of the comprehension system), they cannot address the question of how the preferences develop over rime on reading a stimulus sentence. Self-paced reading rime, for instance, may show that different con straints have different rime courses. For example, and may have an effect as soon as the noun phrase it appears in is encountered, and the effect may be counteracted by what is suggested in a later part of the sentence. It remains an empirical issue to track the rime-course of the constraints uncovered in the present experiment. Acknowledgements This research was supported by grant number Rooo23 1 5 592 from the Economic and Social Research Council to the first author. Special thanks are due to Linda Moxey, Simon Garrod, Joy Aked, and Kari Fraurud for useful discussions. Thanks also to an anonymous reviewer for extensive comments on an earlier draft. A]. SANFORD and F. LOCKHART Department ofPsychology University ofGlasgow Glasgow G t z 8RT UK
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
expression is sometimes used to reinfo�ce a plural reference pattern in circum stances where such a reference pattern may be overshadowed by the potential for singular reference. In the present srudy, the expression was used more frequently in the pooled with conditions than in the pooled and conditions, and thus corresponded to conditions in which plurals were less frequent. We tentatively suggest that these observations support our working hypothesis, although there is no monotonic inverse relationship between incidence of plurals and incidence of use of both, as one might suspect. While it is possible to produce ad hoc arguments for non-monotonicity, it is probably wiser to design a srudy aimed at investigating the functions of both specifically. In much the same way, there are siruarions where all seems to be a candidate for reinforcing certain patterns of plural reference. For instance, given (16), the inclusion of all may make (16 ) sound better. The usc of such additional reinforcing terms deserves attention:
378 Description Types· and Method of Conjoining as Factors I nfluecing Plural Anaphora
RE FERE N CE S
'
anaphora', in Proceedings of the Eighth Annual Conference of the Cognitive Science Society . Lawrence Erlbaum, Hillsdale, NJ. Grimaud, M. ( 1 988), 'Discourse anaphora and the functioning of proper names in narrative', in C. Martindale (ed.), Psycho logical Approaches to Literary Narratives. Helmut Buske, Hamburg. Herweg, M. (1 988), Ansatze zu einer seman tischen und pragmatischen Theorie der Interpretation pluraler Anaphem, GAP AP 2, Universitat Hamburg. Moxey, L. M., & A J. Sanford (1 987), 'Quantifiers and focus', Journal of Seman tics, s: 198-206. Sanford, A J. & S. C. Garrod (198 1), Under standing Written Language, Wiley, Chi chester. Sanford, A J., K. Moar & S. C. Garrod, ( 1988), 'Proper names as controllers of discourse focus', Language and Speech , 3 1 : 43-56.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Eschenbach, C., C. Habel, M. Herweg & K. Rehkiimper (1 989), 'Remarks on plural anaphora', in Proceedings of the Fourth . Conference of the European Chapter of the Association for Computational Linguistics, pp. 1 6 1 -67, Manchester. Garrod, S. C. & A J. Sanford ( 1982), 'The mental representation of discourse in a focused memory system: implications for the interpretation of anaphoric noun phrases',journal ofSemantics, 1: 2 1 -4 1 . Garrod, S. C. & A. J. Sanford (1988), 'Thematic subjecthood and cognitive constraints on discourse structure , journal ofPragmatics, 1 2, 5 1 9-34. Garrod, S. C. & A J. Sanford (1 990), Refer ential processes in reading: focusing on roles and individuals, in D. A Balota, G. B. Flores D'Arcais & K. Rayner (eds), Compre hension Processes in Reading, Lawrence Erlbaum, Hillsdale, NJ. Gernsbacher, M. A ( 1 986), 'Conceptual
]o11rnal ofSemantics 7: 379-393
© N.l.S. Foundation (1990)
Mental Models as Contexts for Interpreting Texts: Implications from Studies of Anaphora A L A N G A R N H A M and J A N E O A K H I L L
University o fSussex
Abstract
I NTRODUCTION The 'mental models' approach to language processing, and thinking and reasoning (see especially Johnson-Laird I 98 3) has two principal aspects. It provides a general framework within which theories of cognitive abilities can be formulated and it suggests the form that specific theories of those abilities should take. The best example of a specific theory developed witin the frame work, Johnson-Laird & Bara's (I 984) account of syllogistic reasoning, clearly illustrates these two aspects of the approach. In its application to language processing Oohnson-Laird I 98 3; Garnham I 987), the mental models approach derives in part from the earlier ideas of Bransford (e.g. Bransford & Franks I 97 I ; Bransford, Barclay & Franks I 972). In particular, it accepts his claim that comprehension is both an integrative and a constructive process. However, it is also informed by more rigorous approaches to language-those of generative linguistics and formal semantics-and this greater leaning towards formalism can be seen both in its treatment of integration and construction, and in the attention it pays to aspects oflanguage processing such as parsing.
c
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
One of the major tenets of the mental models theory of text comprehension is that the model of the text so far provides (part of) the context for understanding the-current sentence. Using two sets of findings on the comprehension of anaphoric expressions, we attempt to provide a more specific interpretation for this statement. We first consider the linguistic distinction between deep and surface anaphors, and the proposal that they are interpreted with reference to mental models and to representations of surface form, respectively. Although the linguistic distinction is reflected fairly directly in considered judgements, in on-line processing both aspects of representation are implicated in the interpretation of both kinds of anaphora. The second set of findings shows char the interpretation of texts containing pronouns can be incomplete-only part of the information in the model is used to interpret the anaphor. Readers may effect mappings between role fillers in different clauses of a text or they may effect mappings between names and role fillers. We discuss evidence that these two rypes of mapping can be carried out separately and that, in certain circumstances, role-to-name mapping in particular may not take place at all.
380 Mental Models as Contexts for Interpreting Texts
Bransford's ideas about integration were imprecise. However, the Bransford & Franks (1971) experiment illustrates one point very clearly. The integration of information from different sentences into a single representation, for example one corresponding to: The ants in the kitchen ate the sweet jelly which was on the table. relies on taking different occurrences of an expression such as the table in The ants were on the table. the kitchen.
AND The table was in
I walked over to the table and put the book on it.
or tee-shirt and one in: Jackie is wearing a red tee-shirt and I am wearing a blue one . A major problem in integrating information from different parts of a text is, therefore, identifying anaphoric expressions, such as it and one , and assigning meanings to them. In the context of the mental models theory of compre hension a specific account of anaphor interpretation must be developed. Over the past few years we have been working on such an account, and in this paper we focus on the implications of this account for a central tenet of the mental models approach: that the mental model of the text so far provides (part o� the context for the interpretation of the current sentence or clause. In particular, we consider two sets of findings. The first relates to the idea, taken from linguistic theory, that there are two main types of anaphoric expression: those that are interpreted with respect to the surfaceform of a text and those that are inter preted with respect to its meaning. Since a mental model forms only part of the context for interpreting the current sentence, a representation of surface form might well comprise another part of the context. However, we will show that the mental model plays a role in the interpretation of anaphors that ought, according to the linguistic theory, to be interpreted with reference to a superficial representation only, and vice versa. The second set of findings suggests that, when the information in a clause containing an anaphoric expression is integrated into the existing mental model, the interpretation of that clause is often incomplete, in a sense that we will define below. The model may be the context for interpreting the clause, but the interpretation need not be a full one. Neither set of findings shows that this tenet of the mental models theory is incorrect, but both show that, in a fully
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
to be co-referential. However, in ordinary texts, expressions with the same meaning (either same reference or same sense) typically have very different forms, for example the table and it in:
A.
Garnham and J. Oakhill 3 8 1
specified theory of text interpretation, it must receive a more specific formulation. RE PRESE N T AT I O N S U S E D TO I N TE RPRET A N A P H O R S
Someone else has been invited. I wonder who. Sag & Hankamer rename them ellipses and argue that they are interpreted by 'copying over' a piece of superficial representation. For technical reasons, they argue that this representation is Chomskyan logical form. Deep anaphors, on the other hand, take their meaning directly from an element of the content based representation, or mental model, and arc renamed model-interpretive
anaphora .
A straightforward, attractive, and testable interpretation of Sag & Hankamer's suggestion is that ellipses arc interpreted with reference to only a representation of the superficial features of a text, while model interpretive anaphors are interpreted with reference to only a mental model. Much of our recent research has been directed to testing this idea, which we have shown to be incorrect. Although Sag & Hankamer's theory provides the basis of an excellent account of the considered judgement of literate people, it docs not fare so well in explaining their ordinary language comprehension. By saying that the account explains considered judgements we mean the following. If a literate person is asked whether the use of a surface anaphor is acceptable, they will tend to say 'yes' if it can be interpreted by Sag & Hankamer's copying in process and 'no' if it cannot. For example, The children asked to be squirted with the hose, so they were. is judged to be acceptable, because it can be 'completed' by squirted with the hose , and that exact phrase occurs in the first .part of the sentence. The continuation
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
An account o f anaphoric processing based o n the idea that there are two broad classes of anaphoric expression was suggested by Sag and Hankamer (1984}, whose ideas were already known to, and tentatively endorsed by,Johnson-Laird ( 1 98 3) in his most complete account of the mental models framework. Hankamer & Sag (1976) had previously proposed a division of anaphoric . expressions into deep and surface anaphors. In their later paper they suggest that these two types of anaphor arc interpreted with respect to two aspects of text representation that are well-established within psycholinguistics: a superficial or surface-based aspect and a content-based one. Surface anaphors typically involve the omission of a second occurrence of material that has recently occurred in a text, for example:
3 82 Mental Models as Contexts for Interpreting Texts
. . . so we did is less acceptable because the elided material (squirt them with the hose) is not literally present. However, the continuation . . . so we did it, which substitutes a deep anaphor, do it, for a verb phrase ellipsis, is usuallyjudged to be
In theatres all over the rest of the country The Tart and the Vicar's Wife packs them in, and has been ever since it opened in Preston in february I 986. (The Sunday Times Magazine , 22 January 1989) Alan Yentob, controller ofBBC2, said last night: 'It is fantastically important that these issues are debated. If one cannot-and that is happening in some cases-it only means that . . .' (The Sunday Times, 3 March 1 989) Car hire: don't. The traffic jams are appalling. (Article on Tokyo in The Sunday Times, 9 April 1989) This modulator allows the 1 040-STfM to be plugged directly into any domestic TV set, and comes complete with a lead to allow you to do so. (Advertisement for a computer in Practical Computing, May 1 989)
The vast majority of our examples do not seem to cause comprehension problems for the ordinary reader. Indeed, we ourselves often have difficulty in finding new examples for our corpus, unless we are specifically looking for them. Furthermore, when we do not notice an example, we often begin with a
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
good. If the deep anaphor can be assigned a suitable interpretation from a content-based representation it does not matter, at least as far as a considered judgement of acceptability is concerned, if the previous expression of that content does not parallel the one that would have to be substituted for the. anaphor to make a full second clause. Indeed, model-interpretive anaphors, unlike ellipses, do not require linguistic antecedents at all. This phenomenon is explained by the fact that elements can be introduced into mental models without being explicitly mentioned. So it is possible to refer to a contextually prominent woman as she , the first time she is mentioned in a conversation. Or a child can say to its mother, angry at finding a valuable vase broken, the dog did it . However, a contextually prominent action cannot justify the use of a surface anaphor, such as I did, yesterday. Despite. the ability of Sag & Hankamer's ideas to account for considered judgements, several lines of evidence suggest that they do not provide any simple basis for a theory of comprehension. Some of this evidence comes from a corpus of anaphoric expressions from published material that we have been collecting for several years (for a preliminary report see Garnham & Oakhill 1 989). In this corpus we find that it is very rare for an ellipsis to occur without a linguistic antecedent, and in those cases where one docs, it is usually for special effect. For example, The Independent , a British daily newspaper, recently ran an advertising campaign with posters that read It is, are you? The interpretation of these posters, at least on first reading, required a certain amount of problem solving activity. However, we have many examples of ellipses with antecedents that do not parallel the elided material, for example:
A.
Garnham andJ. Oakhill 383
It is surprising how many people in this capital arc critical of the regime, though anyone who did so openly would be promptly advised to book their plot in the cemetary.
Here, do so means 'criticized the regime', but its antecedent-containing clause expresses this content in a different way. Do so cannot be replaced by a 'copy' of part of the first clause. Of I 3 subjects, only one claimed that the passage was difficult to understand. However, I I stated that it was poorly written, and all but one of those I I rewrote the passage so as to make it conform to Hankamer & Sag's parallelism condition-some changed the antecedent from are critical of to criticize, others changed the form of the ellipses (e.g. . . . anyone who is . . .) or eliminated it (e.g. . . . anyone who was openly critical . . .). So, at least according to off-line judgements, the message conveyed by this passage was not difficult to extract, but it was expressed in a way that was not fully acceptable to focused critical judgement. The conclusion that we have drawn from these observations, which is supported by independent experimental evidence to be discussed below, is that although people expect ellipses to take their meanings from the preceding text, they do not derive that meaning solely from a representation of the superficial features of the preceding text. Indeed, in many cases it appears that the ellipsis is interpreted with reference to a content-based representation. Provided that this interpretation seems reasonable in the context and provided that it is not in conflict with other possible interpretations suggested either by the mental model or the superficial representation, it is accepted without difficulty. Experimental evidence supporting this conclusion was reported by Garnham & Oakhill (I 987). In this study subjects read passages such as: The new houses had been left in an untidy state. The plumbers had been blamed by the contractors (for leaving such a mess). The electricians/estate-agents had too.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
vague feeling that the text is poorly written and have to work quite hard to say just what is wrong with it. Nevertheless, in line with Sag & Hankamer's suggestions, once the attention of a literate person is directed to the examples in our corpus, they often appear less than fully acceptable. This intuition was confirmed in a study in which we asked graduate students and other members of staff in our department to comment on short passages derived from examples in our corpus. The subjects were asked to make rwo judgements about each passage: Was it hard to understand? Was it badly written? If they decided it was badly written they were asked to rewrite it. There was no time pressure, so the data should reflect considered judgements, rather than on-line processing. The following is an example of a passage from the study, based on an extract from a newspaper report. It contains a surface anaphor that docs not conform to Hankamer & Sag's criteria.
384 Mental Models as Contexts for Interpreting Texts
Reading was self-paced and the primary focus of attention was on reading times for the elliptical clause and on answers to following yes/no questions such as: Did the contractors blame the electricians/estate-agents? Did the electricians/estate-agents blame the plumbers?
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The choice of electricians or estate-agents in the elliptical clause affected the plausibility of its interpretation in the following way. With electricians the linguistically correct interpretation of the ellipsis_:..the one that satisfies Hankamer & Sag's parallelism criterion-is a plausible one. The elliptical clause means the electricians had been blamed by the contractors . With estate-agents not only is the linguistically correct interpretation less plausible, there is a more plausible scenario suggested by the context, one that might be described by saying the estate-agents had blamed the plumbers . Indeed, if the preceding part of the passage had been worded differently (The contractors had blamed the plumbers), the linguistically correct interpretation of the elliptical clause would be that the estate-agents had blamed the plumbers. Our results showed that this inter pretation, suggested by a content-based representation, but at odds with the one licensed by linguistic criteria, did affect the interpretation of ellipses. First, the number of 'incorrect' answers to the yes/no questions was a very high is% following the short versions of the passages and an even higher 39% when the additional phrase (for leaving such a mess) was included. These figures compared with 8 and I I% for the corresponding conditions in which the linguistically correct interpretation was the only possible one. Second, the reading times for the elliptical clauses were greatly increased when the two interpretations were at odds (24 3 3 versus I 8oo ms for the short version and 3299 versus I 99 I ms for the longer versions). When the linguistically correct interpretation of the elliptical clauses in the passages in this experiment was not in conflict with any other more plausible interpretation those clauses were understood comparatively easily. However, when there was a conflict it appears that both interpretations were considered, and that it took a considerable amount of time to resolve the conflict. Furthermore, in these cases the final interpretation was sometimes derived from a representation of the superficial aspects of the preceding text and sometimes from a content-based representation. We would, of course, expect people to judge that the passages in which there was a conflict were badly written. However, it is not true that only a superficial representation deter mines the interpretation of a surface anaphor, or that the final interpretation is always the one that the superficial representation licenses. This finding that the interpretation of surface anaphors is not determined solely by a representation of the superficial aspects of the preceding text prompts the corresponding question: arc deep anaphors interpreted only with reference to a mental model? In other words: is 'model-interpretive anaphora'
A. Garnham and J. Oakhill
385
really a good name for such expressions? It is quite hard to obtain direct evidence against this claim in English, for reasons that will become apparent below. However, we have obtained some indirect evidence for the involvement of a superficial representation in the interpretation of deep anaphors in a study that addressed a quite different issue (Garnham & Oakhill 1 988). In this study we investigated whether a definite pronoun, they , was easier to interpret when its referent was introduced directly by a plural noun phrase than when that referent was only implied by the use of a cognate verb, together with an adverbial phrase suggesting a repeated event. For example, we compared Little Billy always tells lies to his mother, bur they are never convincing. Little Billy always lies to his mother, bur they are never convincing. The subject's task was to judge as quickly as possible whether the second clause was a sensible continuation of the first. The main findings were that the continuations were judged less acceptable when the pronoun's antecedent was introduced by the verb (54 versus 92%) and that even when the continuations were judged to be acceptable the judgements were made more slowly (22 8 1 versus I 8 7 1 ms). The finding relevant to the present argument is one that. arises from a comparison of the results of the experiment reported above with those from a minor variation of this study, in which the passages were changed to the past tense. Little Billy always told lies to his mother, bur they were never convincing. Little Billy always lied to his mother, but they were never convincing. The important point in the present context is that the change from present to past tense destroys the identity of the plural noun (in this example lies) and the third person singular verb (again lies ). In this study the acceptability of the continuations following the introduction of the antecedent by a verb fell dramatically (29 versus 93% following the introduction of the antecedent by a noun phrase). However, when positive judgements were made, the difference in the time to make such a judgement between the two conditions remained at just over 400 rns. It therefore appears that in the first experiment a match between parts of the superficial representation of the passage was influencing the way that a deep anaphor, the pronoun they, was processed. The reason why it is difficult in English to obtain direct evidence for the involvement of a superficial representation in the interpretation of definite
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
with
3 86 Mental Models as Contexts for Interpreting Texts
Er isc zu gross It (masc.) is too big. Thus, the choice of a pronoun in language production (and also a compre hender's judgement ofwhether the pronoun is appropriate) depends not simply on a representation of the object in a mental model, hue on how that object would be linguistically labelled. Examples such as the one above suggest that the linguistic label for an object is pare of the representation of a scene containing that object, even if the scene has not been described. However, if the scene had been described, and the noun phrase der Schrank used, it would be natural co regard the encoding of the face chat it had been used, as part of a representation of the text's superficial characteristics. If the face chat the label is encoded there helps in the interpretation of a pronoun referring co that object, . the other half of Sag & Hankamer's hypothesis-chat deep anaphors are inter preted with respect co only a mental model-will be shown to be false. The example just discussed does not suggest an obvious experiment but other, related, phenomena do. For example, in languages with non-semantic gender it may happen that the syntactic gender of a noun is different from the natural gender of the object it refers co. In German das Miidchen (neuter, the girl) is the most common example. It is a neuter because it has the diminutive ending -chen , though in modern German it has replaced the non-diminutive (feminine) form die Magd . Indeed, any German diminutive chat refers to a person will produce a gender clash. A slightly different type of case is illustrated by the French feminine noun Ia sentinelle (the guard), which usually refers to a man, and the masculine noun le ministre (minister in the government), which may refer to a woman .
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
pronouns (the most common kind of deep anaphor) is that the form of a pronoun does not usually depend on how its antecedent was (or might be) described. Although the form of English pronouns reflects their number and gender, it almost always depends on the semantic number and gender of what they refer to, rather than to any linguistic property of the names of their referents. However, the situation is completely different in languages with non semantic gender (many European languages, for example), and the relevance of this fact for theories of anaphora has been brought to the fore in a recent debate between Bosch ( r987; see also 1988) and Tasmowski & Verluyten ( r985). One pertinent observation is that a pronominal reference to an inanimate object will have the gender of the noun that would be used to describe the object in that context. For many contexts this noun will name the basic level category, in the sense of Eleanor Rosch (see Rosch, Mervis, Gray, Johnson & Boyes-Braem 1 976), under which the object falls. For example, if some men are struggling co gee a large cupboard (German der Schrank , masculine) into a small van, a bystander might comment
A. Garnham and]. Oakhill 387
Bosch and Tasmowski & Verluyten have shown, pronouns following such words may agree with the syntactic gender of the noun or with the natural gender of the person referred to. Very roughly, the closer the pronoun is to the noun the more likely the pronoun is to agree with the syntactic gender. This phenomenon again suggests that the interpretation of a pronoun is influenced by a representation of superficial aspects of the preceding text. It also lends itself to experimental study. If a match between the gender of the pronoun and the formal gender of the noun introducing its antecedent facilitates the inter pretation of the pronoun, at least under some conditions-as we suspect it will our hypothesis about the involvement of a superficial representation in the interpretation of deep anaphors will be supported. We arc conducting an empirical investigation of this issue on a NATO supported project in collabora tion with Daniele Dubois and Marie-France Ehrlich in Paris, Uli Glowalla in Marburg, and Manuel Carrciras in Tcncrifc, and preliminary results suggest that information about non-syntactic gender can speed pronoun resolution. Within the mental models framework, we therefore propose a more complex account of anaphor interpretation than that suggested by Sag & Hankamer. On this account, several types of information influence the on-line interpretation of anaphors, though for any particular kind of anaphor one type of information predominates in determining considered judgements. In particular, although Sag & Hankamer's theory suggests that a mental model might form part of the context for interpreting model-interpretive anaphors, but not for interpreting ellipses, it actually forms part of the context for interpreting both. A representation of the superficial aspects of the immediately preceding text is also used in the interpretation of both kinds of anaphors. The first part of our proposal, therefore, is that information in the mental model is used rather indiscriminately as part of the context for the interpretation of the current clause-even when linguistic considerations suggest that discrimination is in order. In the next section we will argue that there arc circumstances under which people selectively pay attention to certain kinds of information in a mental model, even though other information in the model might be used to interpret the current sentence. Our observations arc related to an issue that we believe has been unduly neglected-the question of how completely we interpret texts. As
There arc a few well-rehearsed cases in which all the information provided by a text is not used to interpret it. For example, people often fail to see what is wrong with asking of an air crash on a national frontier 'where were the survivors buried?' or they fail to sec why saying that a book 'fills a much-needed
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
I N C O M P LETE I NTERPRETAT I O N S O F A N A P H O R S
3 8 8 Mental Models as Contexts for Interpreting Texts
Shirley questioned Diane because she had seen what had happened. This sentence mentions two events, one of questioning and one of seeing. Questioning involves two roles: a questioner and a person questioned. Likewise seeing involves a seer and a person or thing seen. Part of understanding the sentence is to decide that the person questioned in the questioning event is the same person as the seer in the seeing event. To make this connection is to perform role (or role-to-role) mapping. Another part of understanding the sentence is to decide that the person doing the seeing is called Diane. In this particular sentence there is no basis for (rolc-to-)namc mapping other than the role(-to-role) mapping, but in a similar sentence with a gender cue: Larry questioned Diane because she had seen what had happened. it is clear that the name and role mapping can take place independently. In understanding isolated sentences (for example, the ones above, presented in a psycholinguistic experiment) role mapping is more important than name mapping. The names of the people arc arbitrary labels, but the point of the sentence cannot be grasped unless it is realized that the reason for the question ing was to find out what had happened from someone who had seen it. Where there is a gender cue, role mapping can be indirectly affected via name mapping (she is Diane, so the seer was the person questioned), but when there is no such cue role mapping must be brought about in some other way. Very generally, role mapping requires the use of world knowledge in such cases. More specifically, in sentences such as those above, in which an event
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
gap' is an insult to its author. In such cases two principal factors arc at work. First, context suggests the sort of thing that might be said (the crash was on a border, so the question of which side of the border something will happen on is appropriate). Second, the sentence itself provides semantic clements that can be assembled into a plausible message, for example there is a gap (something that usually needs filling), the gap is being filled and there is something that is much-needed (the gap-filler?). However, if on hearing or reading these sentences people computed their structures, which conform to standard rules of English syntax, and used their knowledge of the semantics of English to interpret these structures, they would not be misled by them. In a clear sense, people apparently do not usc all the information available to them to interpret these sentences, at least in certain contexts. Together with Wictskc Vonk (Oakhill, Gamham & Vonk 1989) we have suggested that the phenomenon of incomplete interpretation is a much more common one than these contrived examples suggest. In particular we discuss two aspects of the interpretation of definite pronouns, which we call role mapping and name mapping, and suggest that either one of them may not occur in the interpretation of any particular sentence. Consider a sentence such as:
A.
Garnham and J. Oakhill 3 89
Shirley questioned Diane because she . . . they typically make she refer to Shirley. The ending we gave above, . . . because she had seen what had happened , makes she refer to Diane. Therefore this sentence, although it is perfectly grammatical, docs not conform to the 'bias' associated with the verb in its main clause. Experimental evidence (e.g. Caramazza, Grober, Garvey & Yates 1977; Vonk 1 98s) suggests that such sentences arc harder to understand than those with verbs that do conform to their biases, for example: Shirley questioned Diane because she wanted to find out what had happened. As
we have seen, congruence with 'verb bias' can facilitate role mapping directly, but it can only facilitate name mapping indirectly. A gender cue, on the other hand, facilitates name mapping directly. Experimental evidence suggests that, under certain circumstances, role mapping but not name mapping occurs. In these circumstances, the mental model of a text is, in a clearly defined sense, incomplete, since links between names and participants in actions arc not included in the representation. Indeed, the names themselves may not be included in the representation. An experiment that produced role mapping, bur not name mapping, was conducted by Vonk ( 1 98 s). who presented (Dutch) sentences similar to the ones above, followed by a related sentence for verification. She included difficult filler items without pronouns to divert the subjects' attention from the
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
mentioned in the first clause is explained by a reason or cause stated in a following because clause, that world knowledge includes information about what Garvey & Caramazza ( 1974) call the implicit causality of the verb in the first clause. Use of the verb question , for example, implies that the cause of the action {of questioning) is to be located in the questioner. Although Garvey and Caramazza describe implicit causality as a property of verbs, it is, as we have already implied, best regarded as an aspect of world knowledge and, as such, its effects can readily be accommodated within the mental models framework, particularly if one assumes with Ladusaw & Dowry ( 1988) that generalizations about roles associated with verbs lie in the domain of world knowledge rather than linguistic theory. A verb such as question denotes a kind of event that our knowledge of the world tells us is typically prompted by something about the questioner (usually their wanting to find something out). Therefore, if a clause giving a reason for such an action has a pronominal subject, we expect it to refer back to the questioner. This intuition is supported by sentence completion data. If subjects arc asked to complete a sentence such as:
390 Mental Models as Contexts for Interpreting Texts
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
anaphoric expressions. In this experiment congruity with bias, but not gender cue, affected reading time, though both factors affected verification time. It appears, therefore, that role mapping, but not name mapping, occurred as the original sentence was read. Using similar sentences, Garnham & Oakhill (1 985) found an effect of gender cue on reading times, but under conditions in which it was much clearer that name mapping was crucial-every sentence contained a pronoun and each one was followed by a question that required role-to-name mapping for its answer. We also found, surprisingly, that the effect of congruity with verb bias was largely restricted to the condition in which there was a gender cue. Recent work in our laboratory by Hannah Cruttenden has confirmed that, for English at least, both gender cue and congruity with bias have effects on reading time, but that the effect of gender cue, although usually larger in absolute size, is usually statistically less reliable. Cruttenden's data also rend to confirm that congruiry has a greater effect when it is paired with gender cue, which is the condition in which it would seem, a priori , to be less useful. Again, our conclusion is that name mapping does not always occur during reading and that the mental representation of texts is, in that sense, often incomplete. We believe that it is the failure to map from roles to names that accounts for the difficulty in keeping track of who is doing what to whom in certain complicated texts with many characters-some Russian novels, for example. It is not that we cannot remember characters' names, but that minor characters arc often not important enough to bother about. Of course, a literary scholar may not be satisfied with such an incomplete understanding of a text, but the ordinary reader often will be. In interpreting the studies by Vonk and by Garnham & Oakhill we have taken an effect of gender cue to indicate name mapping and one of congruity with bias to indicate role mapping. However, it would be wrong to assume any simple connection between these effects and types of mapping. In particular, either type of mapping can be established using (other kinds o� world know ledge, and in some cases it must be. For example, although the reason for an act of questioning usually rests with the questioner, it need not, as some of our examples have illustrated. If you arc told that someone 'had seen what has happened', that person is more likely to be having questions addressed to them than to be posing questions themselves. In some recent studies we have been investigating how factors that have a simple theoretical relation with name and role mapping, gender cue and congruity with verb bias, interact with the difficulty of other world-knowledge based inferences needed to establish the referent of a pronoun. For example, the 'questioned' sentences above require relatively complex inferences compared with
A.
Garnham and J. Oakhill
391
Shaun questioned Elaine/Guy because he was Inquisitive. (congruent with the bias of the verb) Alan/Gwen questioned Brian because he was a suspect. (not congruent with the bias of the verb}
CONCLUSIONS I n this paper we have outlined two lines of research that address the question of exactly how the mental model of the text so far provides (part ofj the context for interpreting the current sentence. In the first we considered the linguistically motivated division of anaphoric expressions into two broad classes and its possible relation to a theory of how anaphoric expressions arc interpreted. The division is reflected rather straightforwardly in considered judgements about
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
We presented both types of sentence in two parts and asked people to judge as quickly as possible whether the because clause was a sensible continuation of the main clause. The because clauses that required the more complex inferences were, in general, longer than those that required simple inferences, so the combined reading and judgement times were longer overall for these sentences. More interestingly, the effects of gender cue and congruity were reduced when more complex inferences were required. In one experiment the effect of congruity was found primarily in the number ofpositivejudgements made. For example, for the simple inference sentences congruity increased the proportion of positive judgements from o.6 s to 0.89, whereas for complex inference sentences the respective figures were 0.72 and o.8 s. In another experiment, with a different set of materials that were less complex overall and, therefore, more clearly sensible in all the conditions, we found effects on judgement rimes. For example, congruity with the bias decreased the time to make positive judgements-that the because clause was a sensible continuation of the sentence-by over 300 ms for the simple inference materials (1667 versus 1 998 ms), but had no effect for the complex inference materials (2 1 17 versus 2 105 ms). Our explanation of these findings is as follows. Although a gender cue gives an unambiguous cue to name mapping and the bias of the verb in the main clause suggests a plausible role mapping in such sentences, a proper understanding of the sentences requires a full processing of the content of the subordinate clause. This processing, which is relatively complicated in the case of the complex inference sentences, can also produce role- and name-mapping. Presumably, this relatively protracted usc of world knowledge masks effects of simple factors such as gender cue and congruity with bias on the interpretation of the pronouns.
392
Mental Models as Contexts for Interpreting Texts
Acknowledgements Our research is supported by ESRC grant Coo 23 of Anaphora).
ALAN GARNHAM and JANE OAKHILL Laboratory ofExperimental Psychology University ofSussex Brighton BN1 9QG UK
2 4 3 9 {Mental
Models and the Interpretation
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
the acceptability of sentences containing anaphoric expressions, but not so straightforwardly in on-line processing. The theory suggests that one type of anaphor, ellipsis, is interpreted with reference to a superficial representation of a text, while the other, model-interpretive anaphora, is interpretated using a content-based representation. However, we have shown that mental models have a role in the interpretation of ellipses, and that a surface representation is used to interpret pronouns. The second line of research relates to the completeness of the interpretation of anaphoric expressions. In the case of definite pronouns, we have shown that mapping from roles to roles and from roles to names can be influenced by manipulating the experimental task and, presumably subjects' perception of it. Our research on this topic is much less complete. For example, we have yet to explain why congruity with a verb's bias has a bigger effect when it is paired with gender cue, since the availability of a gender cue would appear to make congruity a redundant cue. More generally, if we arc to incorporate our results into a theory of ordinary language comprehension, we must explicate the relation between experimental task and the 'tasks' that people set themselves in ordinary reading and listening. This last observation allows us to speculate on an underlying principle that might link our two sets of findings. Readers do not just read, they read for a purpose. Their purpose in reading a text may, therefore, determine how they use a mental model as part of the context for interpreting the current sentence. Generally, it will make sense to usc whatever information is available to develop an interpretation-usc both mental model and representation of surface form to work out the meaning of both deep and surface anaphors. However, if a person does not want to remember names of minor characters in a novel, or does not have to remember arbitrary names to perform an experi mental task, a representation of content from which that information is omitted-an incomplete mental model-will suffice.
A Garnham and]. Oakhill 393
RE FE R E N C E S
,
journal of Experimental Psychology ,
39A:
6 1 1 -27. Garnham, A. & J. V. Oakhill ( 1 988 ), ' "Ana phoric islands" revisited', Quarterlyjournal ofExperimental Psychology, 40A: 7 1 9-35. Garnham, A. & J. V. Oakhill ( 1 989), The everyday use of anaphoric expressions: implications for the 'mental models' theory of text comprehension', in N. E. Sharkey (ed.), Models of Cognition: An Annual Review of Cognitive Science, Vol. 1 , pp. 78- 1 1 2, Ablex, Norwood, NJ.
Garvey, C. & A Caramazza ( 1 974) , 'Implicit causality in verbs', Linguistic Inquiry, s: 459-04· Hankamer, J. & I. A. Sag ( 1 976), 'Deep and surface anaphora', Linguistic Inquiry, 7, 391-428. Johnson-Laird, P. N. ( 1983 ), Mental Models:
Towards a Cognitive Science of Language, Inference, and Consciousness, Cambridge
University Press, Cambridge. Johnson-Laird, P. N.& B. Bara ( 1 984), 'Syllo gistic inference', Cognition , 16: 1 -6 1 . Ladusaw, W. A & D . R. Dowty ( 1988), 'Toward a nongrammatical account of thematic roles', in W. Wilkins (ed.), Syntax and Semantics 2 1 : Thematic Roles , pp. 6 1 -73, Academic Press, New York. Oakhill, J. V., A. Garnham & W. Vonk ( 1 989), 'The on-line construction of discourse models', Language and Cognitive Processes, 4: 26 3-86. Rosch, E., C. B. Mervis, W. D. Gray, D. M. Johnson & P. Boyes-Braem ( 1 976), 'Basic objects in natural categories', Cognitive Psychology, 8: 3 82-439. Sag, I. A. & J. Hankamer ( 1 984). 'Toward a theory of anaphoric processing', Linguistics and Philosophy, 7: 32 5-45· Tasmowski, L. & P. Verluyten ( 1985 ), 'Control mechanisms of anaphora' Journal ofSemantics , 4: 341-70. Vonk, W. ( 1 98s). 'On the purpose of reading and the immediacy of processing pro nouns' in R. Groner, G. W. McConkie & C. Menz (eds), Eye Movements and Human Information Processing, pp. 207- 1 5, North Holland, Amsterdam.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
BransfordJ. D. & J . J . Franks ( 1 97 1 ), The absrraction of linguistic ideas'. Cognitive Psychology, 1: 3 3 1-50. Bransford, J. D., J. R. Barclay & J. J. Franks ( 1 972), 'Sentence memory: a constructive versus interpretive approach', Cognitive Psychology , 3: 1 9 3-209. Bosch, P. ( 1 987), 'Pronouns under control: a reply to Liliane Tasmowski and Paul Verluyten', journal ofSemantics , s: 65-78. Bosch, P. ( 1 988), 'Representing and accessing focused referents', Language and Cognitive Processes , 3: 207-3 I . Caramazza, A., E. Grober, C. Garvey & J. Yates ( 1 977), 'Comprehension of ana phoric pronouns' journal oJVerbal Learning and Verbal Behavior, 16: 60 1 --9. Gamham, A. ( 1987), Mental Models as Repre sentations of Discourse and Text , Ellis Horwood, Chichester. Gamham, A. & J. V. Oakhill ( 1985 ), 'On-line resolution of anaphoric pronouns: effects of inference making and verb semantics', BritishJournal oJPsyclzology, 76: 3 8 5--9 3· Garnham, A. & J. V. Oakhill ( 1 987), 'Inter preting elliptical verb phrases', Quarterly
© N.I.S. Foundation (1 990)
jo11mal oJSemamics 7: 39S-·B 3
Definiteness and the Processing of Noun Phrases in Natural Discourse
K ARl FRAURUD Universitiit Hamburg and Stockho lms Universitet
Abstract
I INTRO D U C T I O N In this paper I want to question some assumptions that have commonly been made in recent work on the interpretation of noun phrases (NPs) in discourse; more specifically, with regard to the role of definiteness in text comprehension. According to one popular view, text comprehension consists of the construction of a discourse model, containing among other things a set of discourse referents (also called for example 'discourse entities') which represent the things we are talking about. An NP may either introduce a new discourse referent, or else 'refer back' to one already established. Most treatments of noun phrase interpretation atrribute a decisive role to definite ness here: it is assumed that indefinite noun phrases, e.g. a man in (I) below, trigger the establishment of a new discourse referent in the discourse model, while definite NPs, e.g. the man in (I), trigger the search for, or the retrieval of, a suitable prior discourse referent: (Ia) A man entered a pub. ( I b) The man ordered a beer.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Definiteness is commonly seen as the watershed between those .noun· phrases (NPs) that introduce new referents and those that refer to referents already familiar. Furthermore, for definite NPs, the anaphoric use is taken to be the paradigm case, while. other, so-called first mention uses are regarded as secondary. The aim of the present paper is to challenge this view, and to argue for a more complex picrure of the role of definiteness in the processing of NPs. The paper consists of two parts. The first part presents a corpus-based srudy of the co-refer ential properties of definite and indefinite NPs in narural, unrestricted texts. The data bring into light several issues with regard to co-referentiality in unrestricted discourse and the pos sible referential functions of indefinite and definite NPs. Particular attention is drawn to the fact that the most common function of definite NPs is not anaphoric but different types of first-mention uses. This is the point of deparrure for the second part of the paper, in which three different approaches to first-mention definites are discussed, and some preliminaries to an alternative model of the processing of first-mention definite NPs are presented.
396 Definiteness and the Processing of Noun Phrases in Natural Discourse
Although it is often acknowledged that there are definite NPs that refer to entities that have not been introduced by another NP in the preceding text, such first-mention cases tend to be treated as in some sense secondary relative to the anaphoric use of definite NPs, or even as deviations from the norm. In line with this, it is often implicitly or explicitly assumed that the understanding of first-mention NPs involves additional or more difficult processing than anaphoric definites.1 As an illustration of this view, we may take Heim's idea of 'file card semantics'. In Heim (I 982: 276), the processes triggered by indefinite and definite NPs are described metaphorically as follows: The cards here correspond to discourse referents in the discourse model, referred to as 'the file'. If for some definite NP no old card can be found, an accommodation process is started which is described as an adjustment of the file that is triggered by a violation of a felicity condition and consists of adding to the file enough information to remedy the infelicity (ibid .: 371 £, my italics)
This accommodation process is assumed to apply uniformly to all first mention definites. In psycholinguistic literature (e.g. Clark & Haviland I977; c£ note I), the processing of first-mention definite NPs has sometimes been described as a 'bridging inference', potentially time-consuming and giving rise to processing difficulty. In an analogous way, computer algorithms for definite NP resolution (e.g. Sidner I 979; Carter I 987; but c£ Bosch & Guerts I 9892) always contain a search for an antecedent match as their first step. Other alternatives are considered only after the search has failed. Although it is not clear what such a rule ordering should be taken to imply in terms of processing assumptions, the idea that first-mention NPs are more difficult to process is discernible. This is illustrated in the following quotation from Sidner's discussion of the judge ments involved in the interpretation of such NPs: Perhaps because there is additional processing time associated with these judgements, it is not possible to extend the judgements to the focus stack (ibid.: 1 1 2)
Commenting on the text in (I) above, Bosch & Geurts (I989) ask: 'Why should there be a problem of finding referents for definite NPs and not for indefinite NPs?' The question is legitimate, but I think it carries a more fundamental implication than Bosch and Geurts seem to be aware of, judging from their answer that this is part of the understanding of the definite NP: instead of taking for granted that NPs are processed in the way they mention, we should, in my opinion, question the assumptions that (i) definiteness is the primary deter-
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
For every indefinite, start a new card; for every definite, update a suitable old card.
K. Fraurud 397
rninant in the NP interpretation process, and (ii) first-mention uses of definite NPs are secondary relative to anaphoric uses. In the following section, I will argue for this by providing some data obtained in a corpus-based study of noun phrases in unrestricted texts. I will subsequently deal with the theoretical issues involved.
I
A S T U D Y O F N P s I N U N RE S TR I C T E D T E X T
2. 1 2.1.1
Sample and classifications
The corpus
The corpus consists of eleven texts arbitrarily chosen from a larger corpus of professional, non-fiction, written Swedish prose.3 The texts are taken from four different sources: brochures (informative), newspapers (articles), textbooks, and debate books. The total number of words is I 0,3 5 5 , evenly distributed between the four text rypes. My main reason for choosing to work with this corpus is that it is the only larger corpus of Swedish natural text which is tagged with lexical, morphological, and syntactic information. Tagged corpora are a necessary pre requisite for doing more extensive quantitative text studies on phenomena that are not restricted to word occurrences. Manual analysis is slow and ineffi cient, and this restricts the amount of data that can be sampled within a reason able amount of time. And, perhaps more importantly, the human analyser will always overlook some occurrences of the categories searched for, which decreases the reliability of the results. Furthermore, if we want to make explicit exactly what kinds of information (morphological, syntactic, semantic, pragmatic) have to be used in order to identify a certain phenomenon, it seems a good idea to let the computer do as much as possible on the (well specified) information given in the corpus, and then do the rest manually. In this study I first let the computer do the main part of the work ofpicking out and classifying all occurrences of NPs by means of an algorithm based on the morphological
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The empirical oasis for theories of definiteness is commonly confined to illustrative examples from relatively short and, often, constructed texts. The aim of the study to be presented here was thus to provide a richer empirical background for analyses of definite and indefinite NP interpretation. In particular, the investigation focused on (i) the distribution of definite and indefinite NPs with regard to their different possible uses, and (ii) the properties of those uses of definite and indefinite NPs that diverge from the alleged paradigm.
398 Definiteness and the Processing of Noun Phrases in Narural Discourse
and syntactic properties of the Swedish noun phrase. An analysis of the co referential relations between those NPs was then conducted manually. 2.1 .2
Formal classes of NPs
the man's car). Thus, in the following, the term definite/indefinite noun should be taken to refer to the formal property of having an enclitic definite article or not. A definite NP, on the other hand, need not necessarily have a definite head noun, but may be rendered definite by its determiners or consist of a definite pronoun. Since this study will focus on the clear cases of full definite NPs, namely those with a definite head noun, I will not dwell here upon the controversial issue of which particular determiners make an NP with an indefinite head noun definite. In the few cases where I need to refer to the whole class of definite NPs (e.g. in Table 6), I follow Teleman's (I969) four cri teria for distinguishing between definite and indefinite determiners, one being that only indefinite NPs can be the proper subject of a sentence with a formal subject (c£ there-insertion in English)! The classes in Table I are defined as follows (the figures within parentheses are the number of NP occurrences belonging to each sub-class of the major class): indefinite NPs with an indefinite head noun consist of all full (lexical) NPs that are syntactically indefinite, with or without an indefinite article. Definite NPs with a definite head noun constitute a sub-class of full definite NPs. lt contains those definite NPs that contain a head in the form
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The total sample contains 3,877 occurrences of constituents which were identified as NPs or, sometimes, only 'potential' NPs (see below) on the basis of morphological and syntactic criteria. These were classified as, for example, definite or indefinite NPs of different sub-types, with regard to the internal properties of the NPs (type of head, determiners, and modifiers). The principle behind the classification was to try to cover all possible structural sub-types of Swedish NPs that could be supposed to differ from other structural types with regard to definiteness and/or referential properties. In all there were I 8 such formally defined sub-classes of NPs, which can be fused into different major classes for different purposes of study. Here it will suffice to mention a few major classes. The distribution of NPs over these classes is shown in Table 1 . For the reader who i s not familiar with Swedish, i t should be mentioned that this language has an enclitic definite article in the form of a definite suffix, -(e )n I-(e )t , on the head noun of the NP, e.g. bilen � car). In addition, there is a preposed definite article, den Idet , which is used mainly when there is an adjectival modifier, and which occurs together with the definite head, e.g. den riMa bilen � red car). A noun determined by a demonstrative or possessive pronoun or a genitive noun is non-inflected, e.g. denna/min/mannens bil (thisi!!J)/
K. Fraurud 399 Table I Major classes ofNPs in che total sample Count
%
Indefinite NP with an indefinite head noun Definite NP with a definite head noun Other full NP Pronominal NP NP with an elliptical head Potential NP Coordinate NP
1 224 745 465 876 56 277 234
3 1 .6 1 9.2 1 2.0 22.6 1 .4 7- 1 6.0
Total
3 877
1 00.0
(2) En riskgrupp ar gam/a . De . . . /One risk group is old (people) (lit.: old + PLURAL). They . .
.I
it is not possible to categorically exclude constituents with these properties as non-NPs. When manually analysed, 24 of these were found to be actual NPs and 2 5 3 to be adjective phrases and other non-NPs. Coordinate NPs, finally, are those containing more than one head, i.e. NPs composed of two or more co ordinated NPs.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
of a definite noun. Other full NPs consist of those with an indefinite head noun determined by a genitive/possessive determiner (1 7 1), a demonstrative pronoun (57), a totality pronoun (all, every) (57), a preposed definite article (in Swedish grammar called 'determinative' pronoun) ( 30), or a pronoun-like definite determiner (1 3 ); NPs with an adjectival head and a preposed definite article ( 1 3 ); proper noun NPs (104); and a rest class (2o). Depending on the preferred definition of'definite NP', some or all of these sub-classes ofNPs can be analysed as definite or, at least, non-indefinite (c£ the remark above and note 4). Pronominal NPs are those with a pronoun head. The largest sub-classes were reflexive and other syntactically bound pronouns (281), definite third person pronouns (1 72), and a diverse group of other pronouns consisting mainly of first and second person pronouns (28 1). Elliptical NPs are those with indefinite or definite determiners and/or modifiers but no head (of which 21 were definite and 3 s were indefinite). Potential NPs consist of constitu ents that could not be identified as NPs or non-NPs on the basis of morpholog ical and syntactic criteria alone, but where the decision depends on semantic properties. These are constituents which (i) have a non-nominal head (adjective, participle or count word) and no determiners or modifiers, and (ii) occur in certain syntactic positions such as predicate complement, postposed nominal modifier or adverbial. As shown by examples such as:
400 Definiteness and the Processing of Noun Phrases in Natural Discourse
2.1.3
Classification according to co-referential properties
In order to get an estimate of the distribution of defNI>s and indefNI>s with different discourse functions, the NP occurrences were classified with respect to whether they were preceded or followed by any co-referential NPs. Since one of the points of the present paper is that the role of co-referentiality in natural discourse has been somewhat over-estimated, a fairly generous concept of co-referentiality has been adopted in order not to 'strengthen' the argument on the basis of what could be considered doubtful evidence. Thus, anything that could possibly count as a co-reference relation between two NPs was included in the count. 'Co-referent NPs' is used as a cover term for NPs that either refer to the same referent, co-extensional NPs, or to the same concept, co-intensional NPs. In the study of co-referential properties of indefNI>s and defNI>s, the total sample of NPs had to be taken into consideration. This was necessary because, for example, a defNp may have an antecedent which does not belong to the sub-samples of defNps and indefNp s. Thus, all NPs were (manually) supplied with information regarding their participation in co-referential chains. This was done in the following, rather 'mechanical' way:
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The present study focuses on the first two classes in Table 1, in the following called inde£NPs and de£NPs. This sub-sample, constituting about half of the .total sample, thus contains most of the full NPs, with some exceptions that will be explained here. My motivation for not including other full NPs in a larger class of 'definite NPs' was that I wanted to direc_t my attention to the well-defined sub-class of definite NPs with a definite head whose co-referential properties are least predictabe. Among the excluded classes of full (definite) NPs, those with a demonstrative determiner are almost exclusively anaphoric, while the other classes are all good candidates for being first-mention uses (c£ 2.2.2.1 below). Coordinate NPs were excluded partly because they pose difficult problems for a computational analysis and partly on the basis of semantic considerations. First of all, co-ordinate NPs can be structurally ambiguous as regards the scope of determiners and modifiers. Secondly, the semantic and referential properties of co-ordinate NPs sometimes differ from their non-co ordinate counterparts.5 The complexities involved in the processing of co ordinate NP fall outside the scope of the present study. Thus, the co-ordinate NPs, as well as the single NPs being co-ordinated ( s 8 3), have been excluded from the sub-samples of defNps of indefNI>s, and have only been taken into consideration in the manual analysis of the co-referential properties of these sub-samples.
K. Fraurud 40 1
(i) if there is one or more NPs in the preceding text with which the NP is co referential, the NP is indexed by a reference to the last preceding co referential NP and classified as a subsequent mention; (ii) or, if there is one or more NPs in the following text with which the NP is co-referential, it is marked as being the first in a co-referential chain and classified as initial mention; (iii) or, if the NP is a potential subsequent or initial mention but does not have any co-referential links to other NPs in the text, it is classified as an isolated mention. ·
2.2
Results
The distributions of defNFs, indefNFs, and other NPs into initial, isolated, and subsequent mentions are presented in Table 2, and will be commented on fur ther in the following sub-sections. The examples from the corpus are given in English translation with additional informacion on literal meaning when necessary. 2.2. 1
IndefNFs
2.2. 1.1 First-mention indefNI>s For 1 r 5 , or about 9.4%, of the indefNFs the subsequent text contained at least one co-referential NP (c£ Table 2). It is interesting to note that these initial mention indefNFs constitute no more than 34.8% of all initial-mention NPs,
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
This classification leaves a rest class of other occurrences (n = 692) including non-NPs and clear cases ofNPs that could be neither initial- nor subsequent mentions.6 The NPs classified as initial mention and isolated mention NPs (ii iii) will be collectively referred to as first mention NPs. The methodological assumption behind this classification was that the properties of an NP with respect to its co-referentiality with other NPs in the text can be seen as an approximate measure of the particular aspect of its referential function that has been associated with 'typical' indefinite and definite NPs, namely whether it introduces or refers back to a discourse refer ent, respectively. Thus, (i) the number of first mentions represents the upper bound ofNP occurrences that could be taken to introduce a discourse referent, and (ii) the number of subsequent-mentions represents the upper bound of anaphoric NP occurrences.7 This approximativity does not constitute a major problem in the present context, since the primary concern in this paper are the occurrences of first-mention defNFs, which are clearly identifiable by lacking an antecedent.
402 Definiteness and the Processing of Noun Phrases in Narural Discourse Table 2 Distribution of indefNPs, defNFs, and other classes of NPs over initial, isolated, and subsequent mention Count Row percentage Column percentage Total percentage Initial-mention
Subsequent-mention
Other
Column total
DefNFs
Other
Row total
115 34·8 9·4 2.9
89 n.o I 1 .9 2.3
1 26 3 8.2 6.6 3 ·3
3 30 1 00.0
929 49·4 75·9 24.0
365 1 9·4 49·0 9·4
585 3 I. I 30·7 1 5.0
! 879 1 00.0
101 1 0.3 8.3 2.6
269 27.8 36. 1 6.9
6o6 62.1 3 1 .8 1 5.6
976 1 00.0
79 I 1 .0 6.5 2.0
22 J.2 3·0 0.6
591 8q 3 1 .0 I 5.2
692 1 00.0
8.5
5 1 .8
22.8
1 7.8
1 224
74 5
1 908
3 877
100.0 3 1 .6
1 00.0 19.2
1 00.0 49·2
1 00.0 1 00.0
while as much as 27.0% of the latter were defNI>s. In addition, it should be mentioned that if the other full definite NPs are also taken into account (c£ section 2.1 .2), 41% of all initial-mention NPs are full definite NP. Furthermore, the starus of the discourse referents established by initial mention NPs may differ considerably. On the one hand, there are introduc tions of discourse referents that will play a significant role in the discourse. On the other, there are introductions of what we, with an extension of a term from Kartrunen ( I 976), might call 'short-term' referents. These differences are partly reflected in the number of subsequent-mentions and the scope of the initial mention NP, i.e. the number of co-referential NPs in the subsequent text (cf Table 3) and the number of sentences from the first to the last co-referent NP (c£ Table 4). For the sake of comparison, the corresponding figures for defNI>s are also presented.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Isolated-mention
IndefNFs
K. Fraurud 403 Table 3 Distribution of indefNFs and defNFs according to number of subsequent mentions IndefNFs
Number of subsequent mentions
DefNFs
Count
%
Cum. %
Count
%
Cum. %
5 5.6 I 9.I 8.7 1 .7 I 4.8
5 5.6 ]4.8 83.5 85.2 I OO.O
56 I4 I2 3 6
6 1 .5 I 5·4 I 3.2 J.3 6.6
6 1 .5 76·9 90. I 9J.4 I OO.O
Total
II5
I OO.O
I OO.O
9I
I OO.O
I OO.O
Table 4 Distribution of indefNPs and defNFs according to scope (where the last subsequent mention occurs in: o - the same clause, I the same sentence, 2 - the following sentence, 3 - the 2nd following sentence, etc.) Scope
DefNFs
InddNPs Count
%
Cum. %
Count
%
Cum. %
4 5-10 1 1 -66
4 13 34 13 7 23 21
3·5 I 1.3 29.6 I 1.3 6. 1 20.0 1 8.3
3·5 1 4.8 44·3 5 5·7 6 1 .7 8q 100.0
6 9 24 6 3 21 22
6.6 9·9 26.4 6.6 3·3 23.1 24.2
6.6 1 6.s 42·9 49·5 52·7 75·8 1 00.0
Total
115
1 00.0
100.0
91
1 00.0
1 00.0
0 2
Given these measures, we see that most initial-mention indefNFs only introduce short-term referents. For example, more than half of the initial mention indefNFs have only one subsequent mention, and about three-quar ters of them have no more than rwo. And the scope of more than half of the intial-mention indefNFs does not exceed the second following sentence. Since it was not deemed possible to formulate any absolute syntactic and/or lexical criteria for excluding indefNFs that could not even potentially be antecedents of a following anaphor (c£ note 6), such an analysis had to be done manually by looking at each NP. The clearest cases of 'impossible-antecedent' indefNFs in the present sample are those occurring in certain idiomatic expressions. In a handful ofother cases the borderline was hard to draw, but it is nevertheless clear that the overwhelming majority of indefNFs must be
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
4 5-27
64 22 IO 2 I7
2
404 Definiteness and the Processing of Noun Phrases in Natural Discourse
Subsequent-mention inde£NPs A quite substantial number of indefNFs ( I O r , or 8.3%), referred to something that had already been mentioned in the preceding text. A smaller number of these were similar to some cases already familiar from discussions of (in)definiteness. Here only a brief mention of some of the relevant literature will be given. Ushie (1986) provides an interesting analysis of subsequent-mention in definite NPs, which 'serves to present an already identified referent in a new light and from a different perspective', a use which is quite common especially in narratives. Wald (1983) presents some examples of co-referential indefinite NPs from oral discourse, where 'the inanimate specific referents are reintro duced as if they were new', a function of indefinite NPs which he shows to be highly dependent on the discourse structure. I n her attempt at establishing a functional taxonomy of givenness, Prince (198 1 ) also touches upon a closely connected problem, although in her study it turns up with definite NPs. She discusses subsequent-mentions of referents that have been 'introduced' by NPs in 'capsule statements' or 'abstracts'. These NPs, she says, are 'ambiguous as to whether they represent Evoked or Inferrable entities', which, in our terms, is synonymous with whether they should be interpreted in relation to the first mention or independently of that. In the present study of written texts, indefNFs, as well as defNFs, presenting the same problems as those discussed by Wald and Prince, turn up as subsequent-mentions of referents that have already been mentioned in headlines or preambles of, for example, newspaper articles. Most of the subsequent-mention indefNFs in the present study were generic NPs. Thus the relatively large number of subsequent-mention indefNFs is 2.2. 1 .2
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
considered potential introductions in the sense that they could be referred back to by an anaphoric pronoun or definite NP. As seen in Table 2, most (75.9%) of the indefNFs were isolated-mentions, i.e. what can be described as potential but not actual introductions. Thus, from the point of view of modelling NP processing, one major problem with indefinites is the discrepancy between, on the one hand, the vast number of entities that the text makes available for anaphoric reference and, on the other, the small number of entities actually referred back to. In connection with anaphor resolution, it is often suggested that the syntactic position of an NP determines its ranking among possible antecedents. The most popular candidate for signalling the introduction of a new discourse referent is maybe the proper subject NP of 'there-insertion' sentences (e.g. Sidner 1 979). The present sample of indefNFs contains 27 proper subjects in the Swedish equivalent to there-insertion sentences. Only one of these NPs was actually referred back to later in the text, however.
K. Fraurud 405
mainly due to the general high frequency of generic NPs in the texts, a feature which is quite common in many types of non-fiction prose, for example: (3) Psychiatry is the science of psychical disturbances . . . On the whole, the human mind, with its complexity and sensitivity, is exposed to many kinds of strains, and psychical pains are also very common. The science which is concerned with these is psychiatry . . . As has earlier been emphasized, psy chical disturbances can emerge from external strains . . . In some cases psychical disorders may emerge without there being any external cause to be found.
2.2.2
DefNI>s
First-mention detNPs Of the 745 defNI>s in the present sample, as many as 454 {89 + 36 5), or 60.9%, turned out to be initial or isolated mentions, i.e. first-mention uses. This large number of detNPs constitutes a syntactically and semantically diverse class of detNPs, which have up till now merely been defined in the negative sense of not allowing anaphoric reference due to the 'lack' of an antecedent. In order to get a picture of the frequencies of different types, it would thus be desirable to make a functional sub-classification of the first mention detNPs in the sample, according to, for example, Hawkins' {1 978) criteria for distinguishing between different uses of definites. However, such a classification raises a number of methodological and theoretical problems. A qualitative analysis of the present first-mention data was instead taken as a 2.2.2.1
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
As far as I know, the problem of how to treat 'chains' of co-referential generic NPs has not been discussed in the literature on anaphora. One might of course choose to regard these syntactically indefinite NPs as semantically definite, due to their genericity. But this does not provide an immediate solution to the problems of(i) how to recognize that a particular indefNI> is gen eric and thus potentially co-referent with a preceding one,8 and (ii) how to model the interpretation of such instances of indefNI>s. It should also be noted that the latter problem also turns up with definite subsequent-mention generic NPs. For the moment, it cannot be excluded that certain occurrences of indefinite NPs are anaphoric in the sense that the interpretation of the NP involves the identification of an already established discourse referent. Consider the follow ing case of generic indefinites. One of the texts in the corpus is from the year when the zip code was introduced in Sweden. The first mention of the zip code could thus be seen as an instruction to the reader to establish a new, generic discourse referent, which is then subsequently referred to by definite as well as indefinite generic NPs.
406 Definiteness and the Processing of Noun Phrases in Narural Discourse
Subsequent-mention defNFs A little more than a third of all defNI>s (269, or 36.1%) were considered to be co referent with a NP in the preceding text. Of these 2 1 7, or 80.7%, were simple defNFs (c£ Table 5 ). That a NP is co-referent with another"NP in the preceding text does not 2.2.2.2
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
point of departure for some considerations regarding the characterization of the differences among first-mention definites, which I will return to in section 3· Here we will instead take a look at the syntactic complexity of the defNI>s, which at least gives a hint about one difference in the ways defNI>s can be interpreted. Common to many first-mention definites is that the definite NP is interpreted by means of relating the referent to, for example, another referent. One difference among first-mention defNI>s that might be assumed to have processual implications is whether the defNI> contains a mention of such a referent, in other words, whether the relation is explicit in the defNI>. The typical way of explicitly signalling the relation is by using a genitive/possessive construction of the form 'the X's Y' or 'the Y of X' (c£ Brodda 1 975; Fraurud I 986). As already mentioned, this was also one reason for excluding from the main study the sub-class of NPs with a preposed genitive/possessive modifier, which accordingly were predicted to occur more often as first-mentions, a prediction which turned out to be correct; 8 5% of the genitive NPs were first mentions. But a relation to another referent may also be explicitly signalled by other types of modifiers, such as postposed prepositional phrases or (less commonly) restrictive adjectival modifiers. If most of the first-mentioned defNFs were of this 'self-contained' type, this would be reflected in the overall syntactic complexity of defNI>s. In order to get an approximate estimate of the frequency of defNI>s where this possibility is at hand, the number of defNI>s consisting of a single definite noun (a noun with the enclitic definite article) was computed, as was the number of those containing any kind of modifiers, here referred to as simple and complex defNps, respectively (c£ Table s). If we consider the possible role that certain modifiers may play in the inter pretation of defNI>s, it might be expected that complex defNI>s are more often used as first- than as subsequent-mentions. This also turns out to be the case. Looking at the column percentages, we see that 75.1% of the complex defNI>s are first-mentions. In general, however, simple defNI>s are more common than complex ones. And, as shown by the row percentages for first-mention defNI>s, a considerable number of them ( 58.8%) are simple. It is interesting to note that for a large part ( 36%) of all defNI> occurrences, the interpretation appeared to involve a relation to contextual elements outside the defNI> itsel£
K. Fraurud 407 Table 5 Distribution of simple and complex defNI>s over first- and subsequent-mention defNFs Count Row percentage Column percentage
Complex
Row total
First-mention
267 58.8 53.8
1 87 4 1 .2 75.1
454 1 00.0
Subsequent-mention
217 80.7 46·3
52 1 9·3 20.9
269 1 00.0
12 54·4 2.4
10 45·5 4·0
22 1 00.0
496 66.6
249 3 3-4
745 1 00.0
Other
Column total
mean that it is necessarily anaphoric (cf note 7). For the lack of a strict definition of anaphoricity, let us tentatively characterize an anaphoric NP as one that in some sense relies on a previous NP. In other words, the inter pretation of an anaphor would necessarily involve an identification of a discourse referent in the discourse model which has been introduced by a pre vious mention. Obviously, this loose definition of anaphora does not provide us with any operational criteria for classifying all occurrences of subsequent mention defNFs as anaphors and non-anaphors. But in order to get some idea of the proportion of defNFs that seem to 'rely' on a previous mention, we may look at the types of initial-mention NPs of the co-referential chains in which the subsequent-mention defNFs occur (cf Table 6). The best candidates for being 'true' anaphoric defNFs are those occurring in a referential chain initiated by an indefinite NP. Of the subsequent-mention defNI>s, I I 4, or 42.4% were of this type. A slightly higher percentage of the subsequent-mention defNFs, I 5 5, or 5 7.6%, have a definite initial-mention NP, and quite often the defNF is even identical to this initial-mention. Since the initial-mention definite NPs in these cases must be interpretable without a previous mention, it might be assumed that the subsequent-mention defNI>s in principle are, too. This would imply that they do not rely on the previous mention, in the narrow sense of being 'uninterpretable' without this initial mention. But if we take reliance of previous mention in a broader sense, the situation
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Simple
408
Definiteness and the Processing of Noun Phrases in Natural Discourse Table 6 Distribution according to types of initial-mention subsequent-mention defNI>s Type of initial-mention NP
Count
%
Indefinite NP Definite NP Identical defNP Non-identical simple detNP Non-identical complex detNP Other definite NP
I I4
42·4
I SS
57·6
Total
269
NP
Count
for
%
53
I 9·7
27
IO.O
35
I J .O
40
I 4·9
I OO.O
Table 7 Distribution of subsequent-mention detNPs according to distance from last mention (where the last preceding mention occurs in: o the same clause, I the same sentence, 2 - the preceding sentence, 3 - the next preceding sentence, etc.) -
Distance 0
Count
%
Cum. % 0. 7
2
0. 7
20
7·4
8.2
2
IOJ
3 8· 3
46. s
36
1 3 ·4
59·9
4
21
7· 8
6 7-7
s
20
7·4
7 S· I
6
9
J.4
7 8·4
7
12
4· 5
82.9
8
1 .9
84.8
9
4
r .s
86.2 96 · 3
1 0- 1 9
26
9·7
2D-29
7
2.6
98·9
J o-6 S
4
r .s
1 00.0
Total
269
1 00.0
1 00.0
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
will be somewhat different. One factor that seems to play a role in our intuitive judgements about the relatedness of two co-referent NPs is the distance between them. Table 7 shows the distance, measured in sentences, between the subsequent-mention defNI> and the last preceding mention. We see that most of the subsequent-mention defNI>s occur at a short distance from their last preceding co-referent NP, which might be taken as a
K. Fraurud 409
sign of relatedness in the broader sense. But a quite substantial number of the subsequent-mention defNI>s can be found further away. If two co-referent NPs occur at a considerable distance from each other in the text, and if the establishment of a link between the two NPs is essential to the understanding of the discourse, the second occurrence can be regarded as a {proper) re-introduction of the same discourse referent. Re-introductions are sometimes signalled explicitly by a non-restrictive relative clause or adjectival modifier; c£ the co-referential chain in (4). (4) Sentence no.: 257:
324:
This example was, however, the only instance of explicit re-introduction found among the 'long-distance' subsequent-mention defNI>s in the present sample. A more conclusive discussion of co-referentiality and anaphoricity has to be postponed to a later occasion, one important reason being that our empirical knowledge of co-referentiality in natural discourse is still very limited. For the moment, it will suffice to conclude that the number of defNI>s that can be regarded as anaphoric is probably even less than was indicated by the 36.r % subsequent-mention defNI>s. 2.2. 3
Summary
The statistical data in the present study can be summarized as follows. First, only about one-third of all defNI>s are subsequent-mentions, and probably even fewer should be considered actual anaphors. In addition, about one-tenth of the indejNPs are subsequent-mentions, some of which may possibly be regarded as anaphoric. Second, only one-tenth of all inde£NPs are actual introductions in the sense of introducing a discourse referent which is subsequently referred to, and most of these have a very short scope. Further more, these initial-mention indefNPs represent no more than about one-third of all initial-mention NPs. Finally, as much as one-third of the indefNPs and half of the defNI>s were isolated mentions. Taken together, I think that these statistics presents us with a complex picture of the role of definiteness and co referentiality in natural text, although the methodological and theoretical issues raised in the evaluation of this kind of data deserves further attention. With regard to indefinite NPs and subsequent-mention definite NPs, I have here had to confine myself to mentioning some observations and problems
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
2 59:
NP: . . . the, subduing, sedatives . . . ( r sentence) . . . them . . . (64 sentences) . . . the earlier mentioned sedatives . . .
4 1 0 Definiteness and the Processing of Noun Phrases in Narural Discourse
which may be taken as a point of departure for further empirical investigations and theoretical discussion. In the case of definite NPs, I think the least one can say is that the claim that the anaphoric use of definite NPs is primary and the first-mention uses secondary is not substantiated by the data in this study. If one would still like to maintain this claim, one has to account for why the alleged secondary function is the most common one.
3
On the basis of the statistical data on first-mention defNFs presented above, and a further examination of the sub-sample of first-mention defNFs, two main conclusions were drawn. First, the fact that the most common uses of definite NPs are first-mention uses motivates a search for an alternative to the traditional anaphora based accounts of the role of definiteness in the processing ofNPs. Second, a closer look at the first-mention definites shows that, besides being very frequent, they differ in a number of ways which cannot be described adequately in terms of discrete types, but call for a more flexible description of the information and processes involved. I start with a discussion of three contributions within the field ofdefiniteness in which an ambition to take the first-mention definites more seriously can be traced. The discussion is confined to some features of these theories that are relevant in the present context. Then I attempt to develop some tools for describing the differences among first-mention definites with regard to the information and processes possibly involved in their interpretation. 3.1
Uibner'sfunctional theory ofdefiniteness
important contribution to the search for an alternative to the anaphora based theories of definiteness is presented by Lohner (1985), who suggests a radically different approach. Instead of taking anaphoric definites as the paradigm case and extending the analysis to other cases, he takes certain (pos sible) first-mention uses as the point of departure for a general theory of defi niteness. Lohner's paradigm cases are those where the definite article, as it were, is necessary for semantic reasons. Nouns like fother, he argues, in a sense require the definite article, e.g. the fother (of Mary), and can be used with a possessive determiner, e.g. herfother, whereas other nouns, like man , are more often used with an indefinite article and seldom occur with a possessive, e.g. ? her man . An
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
THE P R O CE S S I N G O F F I R ST :. M E N T I O N DE F I N I TE NPs
K. Fraurud 4 1 1
D
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Lohner distinguishes three semantic subclasses of nouns or concepts. Sortal nouns, e.g. man , only classifies objects. Relational nouns, e.g. son , describes objects as standing in a certain relation to other objects. Functional nouns, e.g. husband , are a sub-class of relational nouns which relates objects unambigu ously to others. Lohner claims that the meaning of the definite article is that the head noun of the definite NP is to be interpreted as a functional concept The relation that defines the reference of a functional concept is a (partial) function, which relates the object unambiguously to another object. In short, a definite NP is interpreted as a function with arguments determining the reference of the NP. Furthermore, Lohner distinguishes between semantic and pragtnatic definites. Semantic definites are functional concepts which are interpreted 'independently of the immediate situation or context of utterance'. Pragmatic definites include anaphoric and deictic definite NPs. Lohner claims that his functional theory can be extended so as to also cover the pragmatic definites. However, there would seem to be no straightforward solution to the problem of how a proper treatment of anaphora could be integrated into a functional theory of definiteness (c£ below, section 3.4). Here I will focus on some aspects of Lohner's analysis of the so-called semantic definites and, in particular, on the distinction between sortal and (functional or non-functional) relational nouns. It should be noted that a functional interpretation is not confined to occurrences of functional nouns, e.g. the clutch (of a car), but also applies to cases ofplural definite NPs containing a non-functional relational noun, e.g. the tyres (of a car). In some cases, 'being relational' can correctly be described as an inherent semantic property of the noun . For example, mother belongs to a class of nouns that have traditionally been called relational and that take obligatory arguments as part of their definition; a mother is always the mother of someone . This is particularly clear when we have a sortal : relational pair of nouns which can denote the same person or object, e.g. woman : mother. There are, however, few objects in the world for which such alternative perspectives are encoded in language, that is, for which there is a choice of noun carrying either a relational or a sortal reference to an entity. Other nouns that often have an inherent relational meaning are verbal abstracts (norninalizations), which carry over the arguments of the verb. When considering a number of other nouns, however, it seems more problematic to assume that all nouns can be semantically classified as being either sortal, relational, or functional. Lohner also recognizes that many nouns are 'ambivalent' in that they can be used either as sortal or relational concepts. He illustrates this by an example where the definite NP my table is used to refer to an orange box used as a table, i.e. to an object which is not classified by the noun but whose function is defined by the noun. The possibility of a relational interpretation of 'sortal' nouns is, however, quite
412
Definiteness and the Processing of Noun Phrases in Natural Discourse
(s) FC1: the weather (in England)
FC2: the Prime Minister (ofBritain . . .) FC3: the distance (between A and B . . .)
But how are we, in each particular case of a functional concept, to distinguish between situational and object arguments and determine the
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
common and not restricted to cases where the referent of a definite NP, the X, is a non-typical X functioning as an X Most NPs that receive a relational interpretation do so because of properties of the objects they are commonly used to refer to. A door, for example, is usually the door of a house or the like. But, in the (less common) context of a carpentry shop selling doors, it is possible to talk about a door as an unconnected object. Thus, while we can analyse occurrences of nouns as having either a relational or sortal interpretation, for most nouns in the lexicon these terms do not describe a discrete distinction. In terms oflexico-encyclopaedic knowledge connected to nouns, I think we should view it as a tendency among language users to (think and) talk about objects in the world as being more or less closely tied to other objects in the world. That is, the distinction reflects our structur ing of the world. Most of the concepts that we intuitively conceive of as relational are so by pragmatic reasons rather than semantic. And these are relational only by default, i.e. if noting else is indicated by the context. They denote objects that usually belong to or in some other way are related to other objects or situations. Let us now turn to the issue of the arguments of such nouns, which for the sake of discussion, I will talk about as inherently relational. Inherently relational nouns may take one or more arguments. What are the characteristics of these arguments? Lohner considers two qualitatively different types of arguments: situations and objects. The situational argument corresponds to what in possible world semantics would be called a possible world index. It is an obligatory (often implicit) argument of all functional concepts. Lohner's motivation for distinguishing situational arguments from object arguments is that almost all functional concepts relate to at least some components of the situation. Exceptions are, for example, mathematical concepts like the product of two andfour. But since it is often difficult to cate gorically determine just which elements of the situation are involved in a cer tain functional concept, Lohner chooses to assume a situational argument in all functional concepts 'in case it may be needed'. In addition to the situational argument, a functional concept can have one or (less commonly) two object arguments which may be explicit or implicit. According to the number of arguments, Lohner differentiates between functional concepts with only a situational argument, FC1 s, and those that in addition also have object argu ments, FC2s or FC3s, for example:
K. Fraurud 4 1 3
number and types of arguments? One possibility is to study whether a difference between situation and object arguments is reflected in language. Lohner suggests that an argument for treating the weather as an FC1, in contrast to the FC2 the Prime Minister, would be the way their arguments are specified when explicitly provided, c£ (6a) the Prime Minister ofBritain the Prime Minister ??in Britain (6b) the weather ??ofEngland the weather in England
(6a ') Britanya'nin ba§bakaru /Britain + GENITIVE Prime Minister + POSSESSIVE/ ??Britanya'daki ba§bakan /Britain + LOCATIVE + 'ki' Prime Minister/ (6b ') in_giltere'nin havasi /England + GENITIVE weather + POSSESSIVE/ ??Ingiltere'deki hava /England + LOCATIVE + 'ki' weather/ In other cases, the locative construction would be preferred, just like in English, c£ (7a) ??ingiltere'nin geli§mcleri /England + GENITIVE developments + POSSESSIVE: 'the developments of England'/ (7b) ingiltere'deki geli§meler /England + LOCATIVE + 'ki' development: 'the developments in England'/ Thus, one and the same argument can be treated differently by different languages. As also suggested in 2.2. 1.1, the genitive construction may be seen as the paradigm case for explicitly signalling a relation between two objects. But it cannot be taken as an absolute criterion for distinguishing between different types of arguments. Rather than a binary distinction between object and situation arguments, there seems to be a scale along which different languages draw different borderlines. At the one end of the scale there are arguments specifying a 'possessor' of the referent, and at the other, arguments that specify the location of the referent. Thus, there does not seem to be any principled way
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
A possessive construction is preferred for the object argument Britain , denoting a state, whereas a locative construction is chos�n for the situation argument England, denoting a location. But the evaluation of such structural preferences is complicated by the fact that they are partly language-specific. In Turkish, for example, the 'situational' argument of the weather would be provided by a genitive rather than by a locative attribute (which in Turkish is a preposed noun with a locative sufHx plus the sufHx -ki signalling that the noun is a modifier). Consider the Turkish equivalents of(6a-b):
414
Definiteness and the Processing of Noun Phrases in Narural Discourse
to distinguish between arguments that are themselves objects and those that are components of a situation. Furthermore, the number of relevant arguments is not determined only on the basis of the definite NP itself, but sometimes also depends on contextual factors. A definite NP such as the dream to become rich , which Lohner regards as a FC1 , may take additional specific arguments in certain contexts. Consider the Swedish example: (8) Drommen om att bli rik fick John att arbeta dag och natt. /The dream to become rich kept John working day and night./
3 .2
Hawkins' notion ofshared sets
In a rich and detailed analysis of the necessary and sufficient conditions for the use of the definite article, Hawkins (1978) provides an interesting contribution
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In order to understand the causality expressed in this sentence, the definite NP must be interpreted as 'john 's dream to become rich' and not only as a generic unrelated concept. This interaction between properties of the NP and the context in determining the number and types of arguments is also seen in the way a noun like door, as mentioned above, in some contexts may be interpreted as a sortal concept and in others as a relational concept. A special case of this is seen in the generic use of inherently sortal nouns. In a generic context a definite NP, such as the woman , may very well take restricting arguments of time/place, e.g. the woman o[Sweden today. In conclusion, Lohner's theory is an interesting contribution to the discus sion of first-mention definites in two important respects. First, in the per spective of modelling the processing of definite NPs, I think that the main insight to be gained from his functional approach is that we should direct our attention to the role of relationality in the interpretation process. In this connection, there are two senses of relational that should not be confused. A noun, on the one hand, can be more or less inherently relational, depending on the lexico-encyclopaedic knowledge connected with the noun. An NP occurrence in a specific context, on the other hand, can be described as h�ving a relational or non-relational interpretation. Second, Lohner's theory is an attempt to do away with the traditional view of anaphora as the paradigm case for definiteness. However, I do not think that our recognition of the non-secondary role of first-mention definites necessarily suggests a reverse view, in which these uses are taken to be more basic than the anaphoric use. In section 3 ·4 I will suggest an approach that treats neither anaphora nor first-mention as primary, but in which the choice of inter pretation strategy, in interaction with the definiteness of the NP, is influenced by the lexica-encyclopaedic information connected with its head noun.
K. Fraurud . p 5
to the discussion of different first-mention definites. The basic notion in Hawkins' theory of definiteness are (i) pragmatic set recognition: the referent of a definite NP is located in a 'shared set' of objects which is identifi able by the receiver by means of the context; (ii) totality (or inclusiveness): the definite NPs refers to 'the totality of the objects of mass' within this 'shared set' which satisfy the referring expression. I will here focus on the notion of shared sets and the question of how these sets are identified by the receiver. A shared set is a set of objects of which the speaker and hearer have shared knowledge on the basis of the previous discourse or the situation of utterance:
Common to most first-mention definite NPs is that the shared set is deter mined by a trigger such as the NP a book , which is said to trigger a set of associated objects like author, pages, title, etc. For reasons that I will return to in the following, I will introduce, instead of 'triggers', the term anchors for talking about these entities or elements in relation to which first-mention definites may be interpreted. (Note that the term 'anchors' is also to some extent similar to Lohner's 'arguments'.) Let us examine the taxonomy of first-mention uses suggested by Hawkins. His criterion for differentiating between larger-situation and associative uses is whether the 'shared set' is rriggered by the larger situation or by an NP in the preceding discourse. In addition to these first-mention uses, Hawkins discusses cases like thefront page ofthe Guardian . These so-called 'unfamiliar' uses with explanatory modifier are distinguished from the other uses by two properties: (i) the 'rrigger', or anchor, is contained in the definite NP itself, and (ii) the referent of the definite NP is not 'familiar' to the receiver. If these criteria are taken to define a taxonomy of uses, which is often done, several difficulties arise. These become particularly evident if one rries to operationalize such criteria for classificational purposes and apply them to NP occurrences in natural texts. In the study presented earlier in this paper, this difficulty first turned up as a methodological problem. But, as I will try to show, a further analysis suggests that it should rather be taken as a theoretical problem. While some instances fit well with Hawkins' types, there are others which call for a more flexible description of the information and processes involved. Some illustrations of this will be given in the following. Hawkins' account implies that the anchor is either the larger situation or an NP in the preceding discourse or a modifier of the definite NP. However, it is often the case that the interpretation of a first-mention definite NP involves more than one anchor, which can come from different sources. Consider (9) uttered at a ticket office of the central station in Stockholm.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
An important assumption which I am making is, therefore, that the objects available to speaker and hearer to which linguistic expressions with the definite article can refer, are arranged in these discrete mental or physical sets defined by shared knowledge and the shared siruation of utterance.' (Hawkins, ibid., p. 1 30)
4 1 6 Definiteness and the Processing of Noun Phrases in Natural Discourse (9)
I am going to Gothenburg. When does the next train leave?
In interpreting the definite NP the next train as the next train from Stockholm to Gothenburg , the clerk at the ticket office has to make use of both discourse externally and discourse internally provided anchors. Consider also the follow ing sentence from a text about EEC, where the link referred to is the link between EEC and NATO.
( I o) Through De Gaulle the link to NATO was broken.
( I I a) The engine ofmy car has much more power than yours. ( I I b) The engine has much more power in my car than in yours. In both these cases, the receiver either may or may not have previous knowledge of the referent of the definite NPs. The difference lies only in the way the anchor my car is provided. Furthermore, by using 'unfamiliarity' in the way he does, Hawkins fails to make two necessary distinctions, namely, between the identification of anchors and the identification of referents, and between the knowledge involved in the first process and the knowledge involved in the second one. This also becomes evident in his discussion of what he calls associative and larger-situation uses involving specific knowledge. As an example of the latter, he mentions the definite NP the Little Mermaid (referring to a statue in Copenhagen). However, if the receiver has previous specific knowledge of the referent, there is no need
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Here, the modifier to NATO provides one anchor needed to interpret the definite NP, and the second anchor is found in the discourse model which contains a representation ofEEC due to previous mention. In cases like (9) and ( w), it seems inadequate to talk about the interpretation of the definite NPs in terms of discrete sets determined either by modifiers of the definite NP or by other NPs in the text, or by the larger situation. In this connection, I want to mention a problem of a more methodological nature. From the point of view of the analyst it is often difficult to judge whether a particular anchor could be said to originate from the discourse or from the larger situation. For example, when the reader of a Swedish newspaper article encounters the definite NP the government, it is of little consequence whether Sweden has been mentioned explicitly in the preceding discourse or not. Thus, there may be several potential sources for one and the same anchor, and the question of which one is actually 'used' in the interpretation process cannot be answered categorically. As regards Hawkins' 'unfamiliar' uses with explanatory modifier, it should also be pointed out that 'unfamiliarity' of the referent is neither a necessary property of definite NPs with 'explanatory modifiers', nor is it a property that is confined to this structural type, c£
K. Fraurud 4 1 7
for any anchor Copenhagen to be contextually present (for example, in a discourse about famous statues). As a matter of fact, this description is very similar to a proper noun, something which is also reflected in the capitalization of the words. As an example of a definite NP involving (some amount o� specific knowledge in the identification of an anchor, we may instead consider the definite NP in ( 1 2), uttered in Oslo:
( I 2) I am going on a cruise to see thej}ords .
a narrow arm of the sea berween cliffs or steep slopes, esp. in Norway (Longman Dictionary of
Contemporary English 1 978)
But there arc other countries with fjords, e.g. Iceland. So, in order to know which fjords the speaker is going to sec, the anchor 'Norway' has to be available in the context. However, the interpretation docs not necessarily involve any specific knowledge of the referents, the fjords of Norway. In conclusion, what makes Hawkins' work interesting in the present context is that it attempts to account for a wide range of examples of first-mention uses of definite NPs. Furthermore, his theory makes explicit certain assumptions regarding the interpretation of first-mention definites to which more or less direct parallels are commonly also found in more process-oriented theories. There would, however, appear to be two main problems in Hawkins' treatment of first-mention definitcs. First, a taxonomic view based on the notion of discrete shared sets overlooks the necessary distinction between four independent aspects of the interpretation of first-mention definite NPs: (i) Is the referent of the definite NP 'familiar' or not, that is, does the inter pretation involve an identification of a previously known referent? (ii) Does the interpretation of the definite NP involve (one or more) anchors or not? (iii) Are the anchor(s), if any, given in the text or by the discourse situation, that is, what are the discourse internal and external sources of the anchors? (iv) Is the knowledge involved in establishing the relation between the referent and the anchors specific or generic? Secondly, the point of departure for Hawkins' theory are cases where the referent is familiar to the listener. The analysis of these cases is then extended to the so-called 'unfamiliar' cases. In this sense, it could be argued that Hawkins' theory is still essentially a familiarity theory of definiteness, although he himself
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The fjords referred to are the fjords in Norway. The relation between the fjords and the relevant anchor 'Norway' cannot be identified on the basis of general knowledge about countries, but only by means of some degree of specific knowledge about the relation between fjords and Norway, c£ the dictionary definition offJord:
4 1 8 Definiteness and the Processing of Noun Phrases in Natural Discourse
argues against such traditional approaches. This is also implicit in his choice of the term 'trigger', which I find unfortunate, since it seems to imply that the sets are always evoked by an NP or the global situation, i.e. that all first-mention definites except those with an 'explanatory modifier' deP-end upon their referents being 'already there' in the discourse model.9 When studying first mention defNFs in natural text, one is struck by the vast amount of objects that appear to be accessible to this kind of reference at each point of the discourse. That all these potential discourse referents would be included in the discourse model does not seem very plausible, since the overwhelming majority will never play a role in the discourse. What has to be there though, in some way or another, are the anchors.
Bosch & Geurts' model ofdefinite NP processing
As already mentioned,
problems analogous to those pointed out in connection with Hawkins' 'distinct set' account of first-mention definites also turn up in computational models of definite NP resolution. In contrast to Hawkins, however, such processual approaches also attempt to capture the dynamic aspects of what is accessible to definite reference. For the sake of illustration, I will here discuss a recent proposal for a computational model of definite NP processing, presented by Bosch & Geurts (r989) (henceforth B&G), which is representative for current models of the processing of definite NPs in both these important respects. In what I interpret as a reaction against former treatments of first-mention definites as secondary, B&G suggest that the term anaphora should be under stood in a wider sense than has previously been the case. In their use, anaphora is 'the resumption of referents that are already represented in the current discourse model', which is said to contain the following sets of objects: (i) (ii) (iii) (iv)
objects in explicit focus; objects that can be accommodated in relation to explicit focus; unfocused objects in the discourse model; objects in the global context.
Given this definition of discourse model, anaphora would also cover the processing of many first-mention definite NPs. I do not, however, believe that such a widening of the definition of anaphora contributes to a better treatment of first-mention definites or to a more adequate general theory of definiteness. First, ifwe broaden the definition ofanaphora to such an extent, the term seems to lose its raison d'etre , since it becomes almost synonymous to definiteness. A more restrictive definition of anaphora is needed for other purposes. Second, as B&G themselves note, even with this wide definition, there are still cases of
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
3· 3
K. Fraurud 4 1 9
first-mention uses which will be difficult to integrate into an anaphora-based account, e.g.
( I 3) The product ofthree andfour is twelve.
The range of objects to which a referent can be linked via focal accommodation is the same as the range of objects from which an anaphoric pronoun can select its referent, i.e. the discourse referents that are currently in explicit focus.
It is obvious that there are unfocused objects in the discourse model to which accommodation is excluded, as is also shown by B&G. But their proposed constraint is clearly too strong, as can be seen in the following passage from a novel.10
( I 4) . . . She waited for them to realize that the car was no longer in motion.
( 8 graphic sentences without either explicit or implicit reference to the car) He bent his head over Maisie's fair gleaming curls as they walked into the
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Such examples are taken by B&G as an argument for LOhner's functional approach to definiteness, which they claim to essentially agree with. It is not, however, clear to me how any implications of this view are integrated into their sketch of a computational model of definite NP processing. & with Hawkins' theory, the basis for their model still seems to be the familiarity of the referent rather than the relational property of the description. This is reflected in the formulation that definite NPs refer to entities that are already represented 'somewhere in some form or other' (B&G). The definite NP resolver searches through the four sets of objects in the discourse model until a referent is found. Thus, a 'distinct set' approach here turns up as a 'compartmentalization' of search spaces. In B&G's model, first-mention definite NPs are resolved by a process which they, like many other authors, call accommodation. Simply put, this is a way of linking the referent of the antecedent-less definite NP to elements of the discourse model. The two (empirical) questions that arise then are: 'What can a first-mention definite NP be accommodated in relation to?' and 'What does the process of accommodation look like?' B&G's answer to the first question is that first-mention definite NPs are resolved by reference to either (i) objects that-can be accommodated in relation to objects in explicit focus or (ii) objects in the global context. The two processes are referred to as focal accommodation and global accom modation, respectively. The concept of(explicit) focus (Grosz 1 977; Sidner 1 979; Sanford & Garrod 1981) plays a central role in most current theories of anaphora. It is commonly thought of as a limited set of prominent discourse referents that are accessible to pronominal reference at a particular point of the discourse. Echoing Sidner, B&G suggest that:
420 Definiteness and the Processing of Noun Phrases in Narural Discourse
house where she had her room. Martha watched them going inside, cheeks laid together, dancing a half-mocking half-dreamy sliding step. She wished that her principles would allow her to cry. But this would not do; she efficiently let out the clutch , and drove herself back to the flat . . .
The range of objects to which a referent belongs that is introduced via global accommodation is the global context of the discourse.
Recall the discussion of example (9) above, where the definite NP the next train is interpreted in relation to elements provided by both the global context
and the discourse. It does not seem reasonable to assume that the referent, the next train from Stockholm to Gothenburg, is included together with a whole set of other trains in the discourse model, as is implied by Hawkins' 'shared set' approach. And this is probably not what B&G have in mind either. In their discussion about global context, they make the following important remark with regard to constraints on global accommodation. What is contained in the discourse model, they say, is a partial representation of the global context, determined by the local context, i.e. the preceding discourse and the immediate situation of utterance, etc. Applying this to our example, we may get something like this: discourse-initially the discourse model contains representations of all trains for which one could possibly buy tickets. This set of possible referents is then stepwise restricted, first by the context ofbeing in Stockholm (the default ifnothing else is said about some other place of departure), then by the mention ofgoing to Gothenburg, and finally by the modifier next ofthe definite NP. But even with these restrictions on a particular kind of objects in the discourse model, namely trains, there is a vast amount of other kinds of objects that could be referred to by first-mention definite NPs in the same context, such as the
wholefamily, the post office, the weather.
And, more importantly, there are numerous first-mention definites which, in my view, involve identification of anchors provided by the global context, rather than the identification of a referent in the global context, such as the economy or the traditions (of this country). In this sense, B&G's definition of
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
The definite NP the clutch must be interpreted in relation to the car, which is no longer in focus at that point of the discourse. Thus, the range of discourse referents that a referent can be 'accommodated in relation to' is wider than the set of objects that are accessible to pronoun reference. One conclusion of this, I think, is that we have to question whether the concept of focus in its present definition is applicable in modelling the interpretation of first-mention definites. In the case of global accommodation, B&G's model includes the constraint that the referent cannot be accommodated in relation to objects in the global context, but only by direct reference to such objects, c£
K. Fraurud 42 1
3 ·4
Preliminaries to an alternative model ofdefinite NP processing
In order to sum up some of the conclusions from the empirical investigation and the theoretical discussion above, I will here try to sketch some preliminaries to a more adequate and flexible way of modeling the processing of first mention definite NPs. Before I go on, I would like to point out that I do not want to claim that the procedures proposed here necessarily correspond to the cognitive processes involved in the interpretation of first-mention definites. In our present state of ignorance, it is only possible to speculate about the actual processes going on in the mind of the receiver. By analysing natural or constructed examples of definite NP uses, however, we can try to provide partial answers to questions such as: what information has to be available in order to interpret this or that definite NP? How may this information be provided by the discourse and the global context? On the basis of such considerations, we may generate hypotheses about what people do when they hear or read a first-mention definite NP. In view of the vast amount of first-mention definites in natural text, a model where the processing of first-mention definites always involves a failing search for an already established discourse referent as a first step seems less attractive. A reverse ordering of the procedures is, quite obviously, no solution to this problem, whereas a simultaneous processing, as proposed by B&G (c£ note 2), might be. Here I want to suggest another possibility, which to my knowledge has not been considered in earlier discussions on this topic. Imagine the following situation. If someone, after an accident at the daycare
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
global accommodation is also too narrow, since it would appear to cover only those cases of first-mention definites which are direct references to a referent in the global context. As regards the question of what the process of accommodation looks like, B&G's model is less explicit. If they were correct in their claims about what the definite NP can be accommodated in relation to, focal accommodation could be envisaged as a sequential search through the limited set of discourse referents in focus, in order to see if the representation of any of these has a link to an entity that would fit the definite description. But if, as I believe, the set of possible anchors is wider and more differentiated, it would be desirable to find some clues in the definite NP that facilitates this search, or to put it simply: the interpreter, when encountering the NP, would know at that point what kind of anchor(s) to look for. Furthermore, we would obtain a less crowded discourse model if we assume that it is the anchors of the referents, rather than the referents themselves, that have to be represented in some way or other in the discourse model.
422
Definiteness and the Processing of Noun Phrases in Natural Discourse
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
center, says Call the mother!, 1 1 a reasonable question from the newly arrived assistant might be · OJ whom? or Whose mother?, and he would probably be satisfied by the answer OJMary, or Mary's . But if the first person, on the other hand, had said Call the woman!, it is more likely that the assistant would ask Which woman? and expect to get the answer Mrs Smith or the like. What I would like to illustrate with this example is the possibility that, in addition to (in)definiteness of an NP, other properties of the NP may also guide the selection of an appropriate interpretation strategy. Partial support for this assumption was found in an earlier study of defNI>s in natural discourse (Fraurud 1 989). Dictionary definitions of nouns occurring as heads in subsequent-mention and first-mention defNI>s were compared. It turned out that, whereas the former commonly had the form of indefinite descriptions, the latter were more often relational/functional descriptions such as 'the X of a/the Y', where X was a hyponym of the noun and Y described a type of anchor. A useful heuristic principle for the reader,.or computer system, might thus be to expect definite NPs with a head noun which has a classifica tional, or sortal, definition to be anaphoric, and a definite NP with a head noun which is defined in relation to other objects to be a first-mention definite. With regard to interpretation procedures, this could be formulated in the following hypothesis: the order in which different interpretation procedures are applied is determined by lexico-encyclopaedic knowledge associated with the head noun of the NP. That is, if the noun occurrence is judged as relational, a non-anaphoric procedure is applied first, and if it is judged as sortal, an anaphoric procedure is given priority. The assumption that this would apply uniformly to all first-mention definites is, however, probably too strong. Thus, when a definite NP is encountered, a non-anaphoric interpretation procedure may be triggered (i) directly, by lexico-encyclopaedic knowledge connected with the head noun, suggesting a relational interpretation,or (ii) indirectly, by a failure to find an already esablished discourse referent. In the preceding discussion, I preferred to use the word 'anchors' instead of Hawkins' 'trigger', Lohner's 'arguments' or 'what first-mention definites can be accommodated in relation to' in B&G's model, all concepts which to some extent are similar in meaning. As I have tentatively used the term in the discus sion above, anchors would be j ust any kind of elements in relation to which first-mention definite NPs are interpreted, including discourse referents and elements provided by the global context. The term discourse referents is here used for representations in the discourse model of entities explicitly mentioned in the discourse, as distinct from what may be called background referents, which are representations of entities that have so far not been mentioned in the discourse but of which the receiver has previous knowledge. The discourse referents that are available as anchors are presumably a restricted number of all discourse referents in the model, but, as we have seen
K. Fraurud 423
(PI ) Establish a new discourse referent, D (P2) Identify one or more anchors to which D can be linked by suitable relations (P2:I ) Determine the relevant number and types of anchors (P2:2) Select anchors (P3) Identify a background referent to which D can be linked by an identity relation where (i) no general temporal ordering is assumed (ii) either (P2) or (P3), or both (P2) and (P3) are applied The relations referred to in (P2), which for the lack of a better cover term is called 'suitable' relations, range from part-of and belong-to relations to spatia temporal and 'situational' relations. One way to think of the interpretation of a definite NP such as, for example, the king occurring in a Swedish newspaper article from I 989, could then be as follows: lexica-encyclopaedic knowledge of the noun suggests a relational interpretation, and thus a new discourse referent is established (PI). It also provides the information that a king is a man that, within a certain PERIOD, has a unique role in relation to a COUNTRY (P2:I). The contextually provided
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
above, these cannot be equated with those in focus. The anchors stemming from the global context range from time/place/circumstance 'co-ordinates' to objects and persons in the discourse situation (not to be equated with back ground referents). A more precise characterization of the restrictions on avail able anchors has to await the results of further empirical investigations. Informally, anchors can be thought of as elements of a contextual frame work or setting built up as the result of an interaction between the global context and the discourse. The anchors can either be established prior to, or be 'contained in', the definite NP. At the beginning of a discourse, the only anchors available are those determined by the global context. As the discourse evolves, this set of anchors may be altered and/or extended by means of explicit or implicit signals in the discourse. In the same way, more local anchors are continuously specified and altered. Common to most first-mention definite NPs is that they are interpreted in relation to one or more contextually provided anchors. I will refer to this aspect of their interpretation as anchoring. In addition, or alternatively, the inter pretation may involve the identification of a background referent. The interpretation of a first-mention definite NP can be thought of as the construction of a new discourse referent with pointers, or links, to one or more anchors and/or a background referent. Analytically, this can be described as involving some or all of the following procedures.
424
Definiteness and the Processing of Noun Phrases in Natural Discourse
3 .4. 1
(P2): the identification of anchors
I will first consider those first-mention definites that appear to involve an identification of one or more anchors, (P2). In these cases, the requirements for a felicitous interpretation are: (i) the anchors have to be accessible at the point of the discourse where the definite NP appears and (ii) the receiver has to know which of these accessible anchors are relevant to the interpretation of the definite NP. That is, he has to be able to recognize the possible relation between the referent and the anchor(s). How are the anchors made accessible? As we saw above in the discussion of Hawkins' theory, the anchors may be provided (i) discourse internally: by modifiers of the NP, or by constituents of the same sentence or elsewhere in the discourse; (ii) discourse externally: by the global context, including the participants of the discourse; or (iii) by a combination of (i) and (ii). In our example with the king above, the anchor Sweden may alternatively be given by a preposed adjectival modifier the Swedish king , a postponed prepositional phrase modifier, the king ofSweden , another constituent in the text, In Sweden, the king . . ., or by the global situation of being in Sweden. In addition, it should be pointed out that anchors may also be implicitly evoked by the discourse. If you start talking about sitting behind the steering wheel this may evoke a vehicle ' anchor, after which a reference to the windscreen would be unambiguously interpretable. Or consider the possible situation when, on the basis of several different details in a story, you gradually understand that the event takes place in a certain country during a certain period, and then use these anchors when encountering definite NPs like the capital and the king. Let us now consider the kind o f knowledge required for recognizing the
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
anchors PERIOD = 1 989 and COUNTRY = Sweden are selected {P2:2). Given these anchors, the background referent Carl XVI Gustafis identified {PJ). The resulting representation is a discourse referent with pointers to the background representation of Carl XVI Gustaf and to the anchors 1989 and Sweden. Now an important conclusion to be drawn from the discussion in the preceding sub-sections is that such a sequential procedure, involving all of those steps, cannot be assumed to apply uniformly to all first-mention uses. This is what I try to capture by the remarks (i) and (ii) above. First of all, the procedures in {P2-P3) are sometimes complementary, and I find it likely that sometimes a certain procedure does not apply although the relevant anchors and/or background referents are available. Secondly, there seems to be no reason to assume a general temporal ordering of the procedures. I will discuss these issues in the following two sections, starting from some considerations regarding the information that must be assumed to be available in order to account for the interpretation of first-mention definite NPs.
K. Fraurud
425
(1 s) I can't do anything about it right now. The carburettor is out of order and I have to get a new one. Even if the car owner has never heard the word carburettor before, he would probably interpret the definite NP as 'something in the car which is called a
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
relation between the referent and a particular anchor. It is generally assumed that the interpretation of a definite NP such as the author in the context of talking about a book involves generic knowledge about a relation between authors and books. Regardless of whether we think of the activation of this knowledge in terms of a search in a lexicon or in a network of nodes connected by such relations, this knowledge can be looked at from two different per spectives. On the one hand, it can be viewed as a generic knowledge of books, e.g. that books have authors, pages, covers, contents etc., and on the other, as a generic knowledge of authors, e.g. that an author is always the author of a book, an article or the like. The implication of the first view is that the mentioning of a book, activating generic knowledge about books, would trigger a set of associates (an author, pages, etc.) which is then 'already there' to be searched for when the definite NP the author is encountered. The opposite view implies that the occurrence of the definite NP the author, activating generic knowledge of authors, would trigger and guide a search for a suitable anchor, for example, a book, which is in this case found in the discourse model due to previous mention. In different instances of first-mention definites, one or the other of these processing hypotheses may be more plausible. But neither of them can be assumed to apply uniformly to all first-mention definites. As a general theory of first-mention definites, the first hypothesis can be questioned on the basis of the observation that there, at any particular point of a discourse, appears to be an innumerable amount of entities which could be referred to by a first mention definite NP. Whereas there are restrictions regarding what anchors the referent can be linked to, there seem to be hardly any limitations on the number of entities that can be linked to these anchors. This does not exclude the possibility that the first processing hypothesis applies to some cases of first-mention definites. In the tentative description of the interpretation of the definite NP the king above, all the procedures (P2:1 ), (P2:2), and (P3) were assumed to apply. But in a general model of the processing of first-mention definites, these procedures must be q-eated as complementary. Here a few illustrations of the optionality of different'procedures will be given. Firstly, (P2:2) may be applied without guidance from (P2:1). This is the case where there is no lexico-encyclopaedic knowledge at hand to guide the selection of anchors. Imagine a person taking his car to the mechanic, who examines it and says:
426
Definiteness and the Processing of Noun Phrases in Natural Discourse
carburettor'. That is, his identification of the anchor does not involve any generic knowledge of either carburettors or cars, but is just a matter of making sense of the utterance by picking out the most salient anchor at hand. Notice, however, that this strategy might have failed, if the mechanic instead had referred to the only welding set in his little garage, and ifwe assume that the car owner is as ignorant about welding sets as he is of carburettors, c£ (16) I can't do anything about it right now. The welding set is out of order and I have to get a new one.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Secondly, many first-mention definites involve (P2) but not (P3): cases when interpretation depends entirely upon the identification of anchors and no back ground referent is identified. The reason may be either that the particular receiver lacks previous knowledge of the specific referent or that there is no such referent whatsoever, as in ( 1 3) above. Some further aspects of the optionality of(P3), the identification of referent(s), will be discussed in the next section. Thirdly, there are cases where only (P3) applies: cases where the background referent is directly identified without relations to any anchors. The most obvious cases arc definite NPs which arc similar to proper names such as the Little Mermaid, which I, contrary to Hawkins, do not think require a previous presence of any anchor. (In fact, it can very well be the other way around; an anchor 'Copenhagen' could be implicitly evoked by the definite NP.) Intuitively, even some other kinds of first-mention definite NPs seem to function more or less like proper names, providing direct access to an entity in the global context without necessarily involving any identification of anchors. That is, unless nothing else is stated (if no non-default anchors are evoked by the discourse), a default referent 'named' by the definite NP is merely picked out. In our example above, then, the king would simply and directly be identified as 'our' present king, Carl Gustaf (unless it is indicated that we talk about, for example, Norway), in the same way as the proper name Chomsky when uttered within the linguistic community would pick out the syntactician Noam Chomsky (unless someone else with the same family name had been introduced in the discourse). An even more persuasive example is perhaps the definite NP the moon . For someone who is ignorant of the fact that there are also other planets than the earth which have moons, it is obvious that 'the moon' simply is the moon, rather than the moon of the earth. But it seems likely that he would interpret the moon in the same way even after having acquired knowledge of the planetary system (for a discussion see Dahl in press). The identification of anchors could possibly be more adequately described as a simultaneous processing of different kinds of available information. That is, the procedures in determining the relevant number and types of arguments (P2:1) and selecting the actual anchors (P2:2) are not necessarily temporarily
·
K. Fraurud
42 7
ordered, and both of them need not always be applied. Rather there would appear to be an interaction between lexica-encyclopaedic knowledge associated with the head noun of the definite NP and the general knowledge associated with currently present anchors.
3.4.2 (P3 ): the identification ofbackground referents already mentioned, the identification of background referen�s), (P3), is not an obligatory procedure in the interpretation offirst-mention definite NPs, but is applicable only in those cases where the receiver has a previous representation of the referent and is able to identify the referent. The difference between cases where a specific referent is identified or not has often been discussed in terms of Donnellan's (1 966) binary distinction between a referential and an attribu tive use of definite descriptions. In its referential use, a definite description serves to identify a particular individual. In the attributive use, the description the so-and-so is taken to mean 'whoever or whatever is the so-and-so'. The referential : attributive distinction has been very influential in theories of reference as well as in process models of definite NP resolution. However, as pointed out by Partee ( 1972), this view is probably too simple. Referring to Kaplan (1968-9), she criticizes Donnellan's treatment of definite descriptions as being 'ambiguous', and remarks that the typical examples of referential and attributive definites may more properly be regarded as ·'two extremes on a continuum of "vividness" '. The receiver's knowledge of a referent is not just a matter of having a previous representation of the referent or not, but may more appropriately be described as a matter of degree. That is, in the cases where the reader is able to identify a background referent, his previous representation of the referent can be, as it were, more or less 'loaded' with information. In fact, it seems to be quite common for definite NPs to be underdetermined with respect to 'degree of identification', especially in texts written with the intention that they should be interpretable by readers with different amounts of background knowledge, i.e. without confusing the ignorant or boring the acquainted. It is obvious that the label 'ambiguity' is not adequate for those uses. It should also be pointed out that the sender may have a certain referent in mind, without necessarily expecting or intending the receiver to identify it. In fact all 'combinations' as regards the specific knowledge and intentions of the sender and receiver seem possible: both the sender and the receiver or any one of them, or neither of them may have specific knowledge of the referent. However, the degree of identification is not only a matter of the amount of knowledge that the receiver actually possesses. Consider the following example from an article about the composer Schubert in an encyclopaedia on music history. As
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
E
428
Definiteness and the Processing of Noun Phrases in Natural Discourse
( 17) . . . when he got the opportunity to play with the orchestra and to get to know the classics . . .
4
CONCLUSI O N
In the first part of the paper, we saw that the traditional view, according to which indefinite NPs introduce new discourse referents, and definite NPs refer hack to already introduced ones, gives a distorted picture of the facts. Whereas subsequent-mention indefinites are arguably somewhat marginal, the pro portion of first-mention definites is much too large to warrant the current treatment of them as secondary to the anaphoric cases. Furthermore, frequency data on co-referentiality relations suggest that the role of co-referentiality has been over-estimated in earlier treatments. Most indefinite NPs and half of the definite NPs in natural texts do not partake in any co-referentiality chains. The predominant type of referential relations are instead relations that hold
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
In the given context, the classics might be interpreted as something like 'the first rank composers (in Europe) preceding Schubert'. To what degree will the actual referents of this definite description he identified? On the one hand, this depends on the reader's amount of background knowledge. A reader ignorant about music history may be content with the description itself, while the lover of classical music may think of Haydn, Mozart and perhaps Beethoven and some others. On the other hand, there seems to be reason to assume that not even the latter will necessarily 'bother' to identify all or even anyone of the particular referents. A plausible assumption is that the degree of identification will depend on the reader'sjudgement of what is relevant for the understanding of the whole discourse. This, in turn, may, among other things, depend upon the readers' predictions regarding what role the discourse referent(s) will play in the following discourse. In conclusion, the differences in interpretation with regard to the identifica tion of referents cannot be described in terms of a referential : attributive ambiguity of the NP. Instead we may talk about degrees of identification , which can be assumed to depend on such factors as: (i) the interpreter's amount of previous knowledge of the referent; (ii) the interpreter's judgement regarding which parts of this previous knowledge are relevant in the current context; and (iii) the ontological status of the referent (specificity, genericity, concreteness, animacy, individuation). This implies that the procedure (P3), identification ofbackground referents, does not have to apply even if it is possible in principle due to previous knowledge of the referent. Furthermore, when the procedure is applied, it is presumably only a relevant part of the full background representation of that referent that is activated.
K. Fraurud .429
( I 8) John had three children and Bill two. John loved his children and spent most of his spare time with them. For Bill, however, the children were merely a nuisance.
In order to interpret the children not as John's three children but as Bill's two children, introduced in the first sentence, the receiver first has to identify the anchor Bill. Thus, the procedure of identifying anchors, (P2), may also be relevant in an account of some anaphoric NPs. Furthermore, the procedure of constructing a new dicourse referent, (PI ), has a parallel in the processing of a certain type of anaphor, namely plural (or singular) anaphors, with what is commonly described as split antecedents, cf (I9) A man and a woman sat at a cafe. The couple . . . The interpretation of the couple can be analysed as involving a construction of a new complex discourse referent related to two previously introduced referents. The relation in question is not a relation of simple identity or ofanchoring, but could be described as a summation relation. Without being able to provide a further analysis ofanaphors with split antecedents here, I would like to take the opporrunity to emphasize the necessity of taking these cases into consideration
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
between discourse referents, or between discourse referents and the global context. The second part of the paper concentrated on the role of these relations in the processing of first-mention definite NPs, and some preliminaries for modelling the anchoring of discourse referents were developed. The present version of the model obviously needs further elaboration. The most important remaining problems concern the dynamic aspects of the representation of anchors, i.e. the problems of how and when anchors are evoked and how their 'scope' is determined. These questions can only be answered on the basis of further empirical studies of natural discourse. For the moment, I can only conclude that the problem of the representation of anchors is not solved by the currently existing focusing mechanisms, which has been developed in order to account for the interpretation of anaphoric NPs. Moreover, several issues raised in connection with anaphoric and indefinite NPs remain to be investigated. In my view, however, some of the insights gained in the analysis of first-mention definite NPs could also be carried over to the analysis of anaphoric and indefinite NPs. I would like to conclude this paper by considering briefly the possible relevance of some distinctions developed in the present paper for modeling the processing of anaphoric and indefinite NPs. The interpretation of an anaphoric NP can be described as a process of identifying a previously introduced discourse referent. Sometimes, this identification appears also to necessarily involve an identification of anchors, c£
430 Definiteness and the Processing of Noun Phrases in Natural Discourse
at an early stage in the development of a general model of definite NP processing. So far, I have assumed the identification of anchors to be a procedure involved in the interpretation of first-mention and, sometimes, anaphoric definite NPs. One might ask whether the interpretation of indefinite NPs also may involve anchoring. My present judgement is that the interpretation of certain occurrences of indefinite NPs appear to require the identification of anchors (with the obvious difference from definite NPs that the referent is not un-ambiguously related to the anchor(s) ) but that these occurrences are not very frequent. If this is the case, it could be taken as a further argument for a more differentiated treatment of both indefinite and definite NPs, in which we would not necessarily assume a separate module for processing definite NPs. In conclusion, instead ofletting definiteness be the main determining factor in NP interpretation, we need a more flexible mechanism, where the way in which an NP in discourse is processed depends on several factors, including in addition to the definiteness of the NP, among other things, the lexico encyclopaedic knowledge associated with the head noun and set of currently available anchors. ,
Part of the work reported in this paper was done within the DFG project 'Gruppierungs-und Abgrenzungsprozesse' (Processes of Complex Formation and Delimitation) in Hamburg. I wish to thank Christopher Habel, Carola Eschenbach, Osten Dahl, Christina Hellmann, Gunnel Kallgren, Staffan Hellberg, Karen Ebert and two anonymous referees of this journal for valuable comments on earlier drafts of this paper. KARl
FRAURUD
InstitutionenfOr lingvistik Stockolms universitet S-1 o6 9 1 Stockholm Sweden Fachbereich Informatik Universitiit Hamburg Bodenstedstrafie 1 6 D-zooo Hamburg 5 0 FRG
N O TE S I This theory has been influential for a long rime in computational and linguis tic theories of definiteness. In more recent psycholinguisric research, it has
been shown that first-mention definites do not necessarily involve additional pro cessing rime (e.g. Sanford & Garrod 1 98 I ; see also note 9).
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Acknowledgements
K. Fraurud 4:3 1 I
(a) Svenssons har renoverar sin lagenher.!the Svensson's have renovated their apparrment/ (b) De har moderniserar koket och badrummer./they have modern ized the kitchen and the bath room/ (c) De har moderniserar kok och badrum./they have modernized kitchen and bathroom/ (d) De har moderniserat koket./they have modernized the kitchen/ (e) •De har moderniserat kok./they have modernized kitchen/ The difference berween (b) and (c) is only stylistic; the articleless use is more
·
common in formal languge. In the pre sent sample, half of the co-ordinate NPs were of this articleless type. 6 As is well known, not all indefinite NPs introduce discourse referents, c£ John is a teacher'. On the basis of such examples, ir has been suggested that, for example, NPs in the position of predicate comple ment could not be introductory (Kart tunen 1 976). Counter-examples can, however, be found to each of the pro posed syntactic and lexical constraints (c£ Fraurud 1 986), e.g. 'The flies arc a great problem. It has to be solved.' When it comes to acrual occurrences ofNPs, some of these can be judged as clearly non introductory, but on the basis of semantic rather than syntactic considerations. Because of the lack of any absolute formal criteria, such an analysis was here made manually. 7 In text-books on semantics, anaphora is usually illustrated by examples where an indefinite NP, the 'antecedent', is 'referred back to' by an anaphoric definite NP or a pronoun, e.g. a book: tlze book/it . When considering cases when the first NP is nor an indefinite bur a first mention definite NP, the problem arises whether subsequent co-referential defin ite NPs should be regarded as anaphors, and in that case, under what conditions. Imagine a situation where you, at the first page of a novel, read about a dark night when the moon was hidden by heavy clouds, and, several pages later, rhar an American space ship was going to the moon . The rwo occurrences of the moon arc obviously co-referential (it is rhe same moon). But inruitively it seems to make little sense to speak about the second occurrence as anaphorically referring to a discourse referent evoked by rhe first one. This is a somewhat extreme case, but the principal problem is the same in all analogous cases. 8 An examination of Swedish texts indic ates that most generic NPs in narural texts defy identification by means of morpho-
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
Bosch and Guerts explicitly refrain from claiming psychological realism of the subsequentiality in their algorithm, sug gesting that parallel search might be a likely alternative. The large corpus is 'Skrivsyntax: Profes sionell prosa' (The syntax ofwriting: Pro fessional prose') from Lund University, containing about 8s,ooo words. 4 An analysis which divides all sub-classes ofNPs into definite and indefinite can of " course be questioned. In addition to the possible objections to this dichotomy on principle, it should be mentioned that there are cases (among articleless NPs) when Teleman's criteria conflict and have to be weighted against each other. As an example of the strucrural ambiguity of co-ordinate NPs, consider the NP 'Carola's wonderful parties and sub sequent headaches'. Whether the scope of the genitive determiner Carola's is re stricted to the first noun or exceeds over the whole co-ordinate NP is not syntacti cally signalled. One example of the semantic peculiarities of co-ordinate NPs is the following interesting property of Swedish articleless NPs (which in Swedish are used in a wider variety of functions than in English). When co-ordinated, articleless NPs can have the same referen tial functions as definite NPs. 2
432 Definiteness and the Processing of Noun Phrases in Natural Discourse theory of certain first-mention definites. Since it is not a theory of definiteness in particular, its emphasis is somewhat dif ferent from that of this paper. At present, I am not able to go further into the important bur complicated issue of how linguistic data on the wide variety of first-mention definites in unrestricted natural texts should be related to psycho logical data on the processing of rela tively short experimental texts. I o Doris Lessing: A Proper Marriage. I 1 Maybe the use of a definite article rather than a possessive pronoun sounds slightly less natural in English than in Swedish, but that is of no importance here.
RE FERE NCE S Bosch, P. & B. Geurts (I 989), 'Processing definite NPs', IWBS Report 78,july 1989 . Brodda, B. (I975). 'Har genitiven nagon grundbetydelse?', in Nionde sammankom stenfor svenskans beskrivning, Goteborg I 8I9 0crober I 975· Carter D. (I987), Interpreting Anaphors in Natural Language Texts, Ellis Horwood series in articifical intelligence, Ellis Hor wood, Chichester. Clark, H. H. & S. E. Haviland (I 977), 'Com prehension and the given-new contract', in R. 0. Freedle (ed.), pp. I-40, Discourse Production and Comprehension , Ablex, Nor wood, NJ. Dahl, 0. (in press). 'Contextualization and de-contextualization', in Proceedings from
tuionen for lingvistik, Stockholms univer sitet, Stockholm. Fraurud, K. (I 989), 'Towards a non-uniform treatment of definite noun phrases', in 0. Dahl & K. Fraurud (eds), Papers from the
First Nordic Conference on Text Comprehen sion in Man and Machine, Sigtuna, Sweden, October 27-28, 1988 , pp. 75-87, Institu tionen for lingvistik, Stockholms univer sitet, Stockholm. Grosz, B. J. (I 977), 'The representation and use of focus in a system for understanding dialogues', in Proceedingsfrom the Fifth Inter
nationaljoint Conference on Artificial Intellig MA, William ence, Cambridge,
Donellan, K. (I 966), 'Reference and Definite Descriptions', Philosophical Review, LXXV: 28I-304. Fraurud, K. (I 986), 'The introduction and maintenance of discourse referents', in 0. Dahl (ed.), Papers from the Ninth Scandin
Kaufmann Inc., Los Alstos. Reprinted in B. J. Grosz, K. Sparck Jones & B. L. Webber (eds) (I 986), Readings in Natural Language Understanding , pp. 3 5 3-63, Mor gan Kaufmann Publishers, Los Altos. Heim, I. R (I 982), The semantics of definite and indefinite NPs. Ph.D. dissertation, Department of Linguistics, University of Massachusetts, Amherst, MA. Hawkins, ]. A (I978), Definiteness and Indefi
avian Conference on Linguistics, Stockholm, January 9-1 1, 1986, pp. I I I -22, Insti-
niteness-A Study in Reference and Grammati cality Prediction . Croom Helm, London.
the IBM Symposium 'Natural Language and Logic', Hamburg, May I989.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
logical and syntactic criteria. For exam ple, even bare plurals in subject position are not always generic, c£ 'Severe earth quakes have devastated the area'. 9· What Hawkins seems to have in mind, although he says nothing explicitly about the processual implications of his theory, is something similar to the psychological concept of priming, a process that would have to be highly constrained in some way or other in order not to result in an overcrowded discourse model. The theory of implicit focus (Sanford & Gar rod I98I) provides a partial solution to this problem, suggesting an attractive alternative to the bridging-inference
K. fraurud 433 Kaplan, D. (I 968-9), 'Quantifying in', Syn these 19: I I 8-29. Karttunen, L. ( 1 976), 'Discourse referents', in J. D. McCawley (eeL), Notes from the Lin
guistic Underground. Syntax and Semantics, Vol. 7, pp. 363-8 5, Academic Press, Lon
Wiley, Chichester. Sidner, C. L. (I 979), 'Towards a computa tional theory of definite anaphora com prehension in English discourse', TR-5 37, MIT, MA Teleman, U. (I 969), Deflnita och inde.flnita attribut i nusvenskan , Srudenditterarur, Lund. Ushie, Y. (I 986), 'Corepresentation-a text ualfunction of the indefinite expression', Text 6: 427-46. Wald, B. ( 1 983), 'Referents and Topic within and across Discourse Units: Observations from Current Vernacular English', in f. Klein-Andrew (ed.) Discourse Perspectives on Syntax, pp. 9 1 - 1 1 6, Academic Press, Lon don.
Downloaded from jos.oxfordjournals.org by guest on January 1, 2011
don. LOhner, S. (I98 5), 'Definites', Journal of Semantics 4: 279-326. Partee, B. ( 1 972), 'Opacity, Coreference and Pronouns', in D. Davidson & G. Harman �eds), Semantics of Natural Language, pp. 4 I 5-4I, D. Reidel Publishing Com pany, Dordrecht. Prince, E. f. ( I981 ), 'Toward a Taxonomy of given-new information', in P. Cole (ed.), Radical Pragmatics, pp. 223-5 5, Academic Press, New York. Sanford, A J. & S. C. Garrod (198I), Under-
standing Written Language-Explorations of Comprehension Beyond the Sentence . J.
Journal oJSemamics 7: 435-454
© N.l.S. Foundation (1990)
Control Issues in Anaphor Resolution
DAVID M. CARTER SRI International Cambridge Research Centre Abstract
I
I N TRO D U C T I O N
To derive appropriate interpretations for its input, a natural language process ing (NLP) system must take account not