In this paper, I explore the interaction among several factors argued to constrain discourse anaphoric expressions: the local centering context consisting of successive pairs of utterances; the informational content of a discourse anaphoric expression relative to the set of entities that need to be discriminated; the choice among different kinds of discriminating information; and the global attentional state that derives from the structure of an entire discourse. Specifications for the model are based on analysis of such factors in a corpus of spontaneous oral narratives. Then I use a reserved set of narratives for evaluating a simplified version of the generation model. The model integrates informational constraints, based on Dale's (1992) model, with centering. I show that by integrating centering constraints, the integrated model can account for two types of violation of informational constraints: informationally inadequate pronouns where the pronoun realizes the current backward-looking center, and certain full noun phrases in utterances where the backward-looking center is not overtly mentioned, and use of a pronoun would produce garden-path effects. I argue that for generation, centering constraints take precedence over informational constraints. The evaluation tests three conditions: informational constraints (1) with no focus structure imposed on the discourse, (2) with an empirically derived global focus structure, and (3) with global focus plus centering. Results improve dramatically under each subsequent condition.
All Science Journal Classification (ASJC) codes
- Language and Linguistics
- Sociology and Political Science
- Linguistics and Language
- Speech and Hearing