Evaluation of dialogue act recognition approaches

This paper deals with automatic dialogue act recognition. Dialogue acts (DAs) are utterance-level labels that represent different states of a dialogue, such as questions, statements, hesitations, etc. Information about actual DA can be seen as the first level of dialogue understanding. The main goal of this paper is to compare our dialogue act recognition approaches that model the utterance structure, and are particularly useful when the DA corpus is small, with n-gram based approaches. Our best approach is also combined successfully with prosodic models. We further show that sentence structure-based approaches significantly outperform n-gram based methods.