Fuzzy trigram model for speech act analysis of utterances in dialogues

A speech act means a linguistic action intended by a user. In many cases, the speech act of an utterance varies depending on the context of the utterance. Since it is difficult to represent such contextual information in hand-crafted rules, statistical approaches suggest very promising directions. Traditional statistical models, however, need large training corpus to train probability distributions. We propose a new trigram model, named fuzzy trigram model. We use a membership function in fuzzy set theory instead of conversational probability distributions to alleviate sparse data problems and to achieve high performance even with small training data. In the experiments, the model performs better than a statistical trigram model when the size of training data is small, less than about 300 dialogues. This result shows that the fuzzy trigram model is suitable for the applications such as dialogue analysis where large training corpus is difficult to be obtained.