COMPUTING DISCOURSE ANAPHORA FROM GRAMMATICAL REPRESENTATION

The paper presents an implemented algorithm to compute anaphora resolution and coreference, relatively to pronominal and nominal expressions in discourse. The algorithm works on the output of other modules: the parser and its associated module for pronominal binding within the sentence(see Delmonte & Bianchi, 1991); the module for scope assignment to quantified expressions(see Bianchi & Delmonte, 1989a; Delmonte, 1990); the module for inferential processes based on KL- Prolog(Adorni et al.,1987). Our model is strongly syntactically based: in particular, access to the modules is guided by the grammatical basis of the computation, the two remaining modules being triggered by specific items: the presence of quantified indefinite expressions for the module of scope assignment; the failure of some referring expression to be picked up as Main Topic, thus conflicting with evaluation procedures independently set up by the algorithm, as the need to set up an adequate plural antecedent for a plural pronoun or noun to corefer. As to inferential processes, these are only activated at a given state whenever the scoring procedures independently set up by the algorithm require a nominal expression to corefer in the text. They can also be triggered in a given state whenever the system is in an ambiguous context: one or two pronouns to be coreferred and no MAIN or SECONDARY TOPIC.