Bayesian Networks Model for Dialogue Act Recognition Based on Intra-Utterance Features and Inter-Utterances Context

Abstract

Previous dialogue act recognition models assume inter-utterances independency, in which each utterance is independent of the preceding utterance given the preceding utterance dialogue act. Accordingly, in these models, the recognition of the dialogue act of an utterance depends on the linguistic features extracted from the utterance itself and the dialogue act of the preceding utterance. This paper presents a Bayesian Networks model for dialogue act recognition in a dialogue system. In addition to the linguistic features of the user utterance and the previous utterance dialogue act, the presented model employs inter-utterances context which results from relaxing inter-utterance independency assumption. To design the model two sets of linguistic features have been identified, intra-utterance features extracted from the user utterance and context features extracted from the previous utterance. Bayesian networks machine learning has been used to induce the networks from a task oriented dialogue corpus. A series of experimental cases have been conducted to evaluate the Bayesian Networks model. In each case, different features have been used. The results show that the inter-utterance context is an effective factor in the recognition of dialogue act and the model which is based on intra-utterance features and inter-utterances context has the highest recognition accuracy.