The representation and processing of coreference in discourse

Abstract

The phenomenon of coreference-where two linguistic expressions refer to the same thing-is a central topic in attempts to understand the meaning and structure of language. This is true both for disciplines that attempt to model knowledge of language and for disciplines that attempt to model language processing. Despite their often divergent goals, the disciplines of generative linguistics, psycholinguistics, computational linguistics and formal semantics have all expended considerable energy on aspects of coreference, particularly on the distinction between coreference with full referring expressions, such as names or descriptions, and with reduced referring expressions, such as pronouns. We believe that progress in these different disciplines has provided important ideas that converge in the Direct all correspondence to: Peter C. Gordon, Department of Psychology, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599-3270; E-Mail: pcg@email.unc.edu; Randall Hendrick, Department of Linguistics, University of North Carolina at Chapel Hill, Chapel Hill, NC 27599-3155; E-Mail: hendrick@email.unc.edu. 389 390GORDON AND HENDRICK domain of local coreference, that is coreference within sentences and short strings of sentences that are semantically coherent. In this paper we attempt to integrate that progress in such a way as to provide a unified account that addresses some of the central concerns of each of these disciplines. This account is based on the following principles: (1) The primary function of pronouns (and other reduced expressions) is to refer to things that have already been mentioned in a discourse and which are mentally represented in a discourse model; therefore pronouns are the natural vehicle for coreference. In contrast, the primary function of names (and of other full expressions) is to introduce entities into the discourse model; therefore coreference with repeated names requires additional mental processes. (2) The syntactic and sequential structure of language strongly influence mental representation in the discourse model which in turn influences the interpretation of referring expressions in linguistic input. (3) The discourse model is constructed incrementally with each utterance directly adding conditions that further specify the meaning embodied in the model; exceptions to this direct incremental construction occur in well-marked cases where linguistic forms indicate that a phrase or clause serves to modify the meanings that follow it rather than the meanings that precede it.The paper is organized as follows: First, we briefly review how reference and coreference are approached by different disciplines within cognitive science, with particular attention on how the work contributes to the present effort. Next, we sketch out some basics of the approach taken towards coreference in Kamp and Reyle's (1993) Discourse Representation Theory; our work employs the mechanisms developed in that approach to formal semantics. The main body of our work then shows how the three principles outlin...

The interplay between semantic and referential aspects of anaphoric noun phrase resolution: Evidence from ERPs

“…. .The bird"; Garrod & Sanford, 1977; see also Garnham et al, 1997;Gordon & Hendrick, 1998). Furthermore, concepts central to real-world schemata are immediately available as referents for anaphoric noun phrases, even when they have not been explicitly mentioned (e.g., ''Mary dressed the baby.…”

An Incremental Discourse Parser Architecture

“…A possible rating is done in correlation with the resolution of the pronoun he and the definite noun the money. On CDT 5 0 vein(5) = 1 3 4 5, on CDT 5 1 vein(5) = 1 3 5 and on CDT 5 2 vein(5) = 1 5. CDT 5 0 where he in 5 is [Bill] is a rather strange interpretation where the money are those Bill intended to give for buying John's bike, and because they were not enough he decided to use them for going in a trip.…”

“…The resulting trees are CDT 4 0 that gives vein(4) = 1 2 3 4 and CDT 4 1 with vein(4) = 1 3 4. The variant when he in unit 3 is bound to [Bill] would result in the case of CDT 4 0 to the extremely low probable reference (see [5]) from the proper noun Bill in unit 4 to the entity [Bill] realised by a pronoun in unit 3. This makes that at this moment, of the two possible variants of CDT 4 0 , to be retained the one that indicated he in 3 being bound to [John] and Bill in 4 referring back in unit 2 to [Bill].…”

“…CDT 5 1 gives an interpretation where the decision of John to use the money for going in a trip came as a result of John obtaining a good price for his bike, and is inferable that the money are the money John obtained for the bike. CDT 5 2 evokes the interpretation in which the decision to go in a trip is only a result of John selling the bike. The money are again those obtained for the bike.…”

Coherence-driven resolution of referential ambiguity: A computational model

“…This formal theory was extended by Gordon and Hendrick (1998) to incorporate the control of referent accessibility (i.e., bias) by structural factors. In their conception, referent accessibility drives the initial interpretation of the ambiguity before the plausibility of the ambiguity's readings is taken into account to select a final interpretation.…”