Semantic Features for Dialogue Act Recognition
3rd International Conference on Statistical Language and Speech Processing (SLSP 2015) (2015)
Dialogue act recognition commonly relies on lexical, syntactic, prosodic and/or dialogue history based features. However, few approaches exploit semantic information. The main goal of this paper is thus to propose semantic
features and integrate them into a dialogue act recognition task to improve the recognition score. Three different feature computation approaches are proposed, evaluated and compared: Latent Dirichlet Allocation and the HAL and COALS semantic spaces. An interesting contribution is that all the features are created without any supervision. These approaches are evaluated on a Czech dialogue corpus. We experimentally show that all proposed approaches significantly improve the recognition accuracy.