Abstract:
This paper discusses a statistical model for recognizing discourse intentions of utterances during dialogue interactions. We argue that this recognition process should be based on features of the current utterance as well as on discourse history, and show that taking into account utterance features such as speaker information and syntactic forms of utterances dramatically improves the system’s performance as compared with a simple trigram model of discourse acts. In addition, we propose that taking into account information about discourse structure may allow the system to construct a more accurate discourse act model and thus improve recognition results. Experiments show this proposal to be promising.