Application of the centering framework in spontaneous dialogues

Spontaneous speech poses problems for automatic systems. While many investigators are making progress in recognition and dialogue processing, spontaneous speech also raises interesting problems for a deeper level of discourse modeling. Here, the discourse segmentation (Grosz and Sidner, 1986) and the centering frameworks (Grosz, Joshi and Weinstein, 1995) are used to track the evolution of local and global attentional states in a spontaneous speech dialogue. Several issues characteristic of this corpus are discussed: the informality of sentence structure, the use of pronouns, and the role of prosodic cues. Empirical analysis indicates that the most useful domain for assigning discourse centers is the clause in spontaneous speech, rather than the utterance. Speakers appear to have little constraint on inserting references to discourse participants, but these assume center status only when used as active agents in the discourse. Furthermore, pitch accents can be used to promote the salience of discourse entities, allowing prominent entities to become backward-looking centers in subsequent clauses.