An lnformation-Theoretic Model of Discourse for Next Utterance Type Prediction

概要

論文の詳細を見る
We propose a statistical model of dialogue that is based on an information-theoretic interpretation of discourse, to predict the illocutionary force type of the next utterance. The model consists of a second-order Markov model of utterances classified by their illocutionary force types, such as REQUEST, INFORM, etc., and it gives us a criterion for measuring whether the speech recognition candidate forms a natural local discourse in terms of the speech act sequence. By predicting the next utterance in an abstract level, we can rule out erroneous speech recognition candidates that are syntactically and semantically correct, but contextually incorrect. We show the effectiveness of the statistical dialogue model for ulterance type prediction by extensive experiments using 100 telephone dislogues containing 7,531 utterances. The model achieves 61.7% accuracy for the top candidate and 85.1% for the lop three candidales, when 50 dislogues were used for training and the other 50 dialogues were used for testing. We also show that the model can capture the basic characteristics of the local discourse structure, such as turn-taking and speech act sequencing, and dialogue-type dependent features, such as initiative, which is the allocation of the control and the manner by which the control is transferred.
一般社団法人情報処理学会の論文
1994-06-15