An Approach to Dialogue Act Classification Based on Utterances and Dialogue History

In the development of language technologies, dialogue acts have been broadly used in many dialogue systems. For this reason, detecting and classifying them is an important goal in order to achieve better results in this area. In this work, we present an approach for the automatic classification of dialogue acts, integrating two different classifiers with a Discourse Grammar. Our models detect dialogue acts based on both the lexical and the discourse coherence of the historical sequence. A Naive Bayes classifier and a Multilayer Perceptron have been compared and combined with a Discourse Grammar. This approach has been tested using 900 spontaneous speech dialogues for answering trains information queries (Dihana corpus). Finally, the results of this study suggest this to be a successful and promising approach to this task, achieving label error classification close to 3%.