An Efficient Trigram Model for Speech Act Analysis in Small Training Corpus

Abstract

Speech act analysis is essential to a dialogue understanding system because a speech act of an utterance is closely tied with the user's intention in the utterance. However, it has been difficult how to analyze a speech act of an utterance since it highly depends on the context of the utterance. For that matter, statistical approaches seem a promising direction although traditional statistical models usually require large corpus to train probability distributions. It is also not an easy job to collect dialogue corpus and annotating them with speech acts. In this paper, we propose a fuzzy trigram model as an alternative. The trigram model uses a membership function in fuzzy set theory instead of conversational probability distributions to alleviate sparse data problems. In the experiments, the trigram model performed better than a traditional statistical trigram model although the scale of training data was as small as 300 dialogues. The result showed that the fuzzy trigram model is an appropriate alternative for a traditional statistical models when training data is small.