Detection of topic and speech act type on utterance-by-utterance basis for conversational interfaces

Abstract

SUMMARYA conversational interface must extract information from texts transcribed by a speech recognizer and utilize the information for various kinds of applications. We propose a method for detecting topic and speech act type of an utterance as a fundamental technique for such a conversational interface. First, in the training phase of the proposed method, the score is determined on the basis of training data with topic and speech act tags, representing relevance scores for words relative to topic or speech acts. In the execution phase, for the words contained in the input utterance, the sum of the scores determined in the training phase is calculated for each topic or speech act, and the topic or speech act of the input utterance is inferred from the order of the sum. The score in this study is determined from the mutual information on the occurrence pattern of the words and the topics or speech acts in the training data, as well as the entropy of the occurrence of the word. The proposed method considers only the one-to-one relevance between the word and the topic or speech act, and the co-occurrence of multiple words is not considered. Consequently, the method works effectively for short sentences, such as often occur in conversation. In experiments using a conversational expression corpus with short mean length of utterances on diversified topics, a correct topic estimation rate of approximately 83% and a correct speech act estimation