Analysis of Spoken Dialogues Based on Local Discourse Structures

摘要

One of important abilities that an intelligent agent must have in order to appropriately communicate with a human in natural language is to recognize a dialogue act conveyed with a human utterance. There are basically two methods to recognize dialogue acts: plan recognition method and linguistic-based method. Especially as dialogue corpora increase nowadays, linguistic-based methods such as n-gram model become popular. However, the n-gram model is difficult to treat subdialogue structures. In order to perform natural dialogues, recognition of subdialogue structures is needed. This paper presents a method for recognizing dialogue acts of utterances by analyzing local subdialogue structures. We define a local structure based on information state of dialogue participants. Based on the analysis of local structures in the corpus, we defined rewriting rules which capture patterns of local discourse structures and whose terminals are dialogue acts. Using such rules, a system can analyze local structures and accordingly recognize dialogue acts of utterances. Our preliminary experiment shows the effectiveness of the method.