A Novel Semantic Inference Model With a Hierarchical Act Labels Embedded for Dialogue Act Recognition

As an important component in dialogue system, Dialogue Act Recognition (DAR) has attracted much attention of researchers. Most of the existing approaches only focus on the information of utterances, but relatively few investigate the roles of DA labels in different layers of a conversation and also some semantic inference in irregular conversations, such as questions and answers skipping back and forth, has not been probed properly so that some dialogue acts are not always recognized correctly. To overcome the aforementioned deficiencies, we propose a hierarchical project method, in which DA labels are embedded in three levels (word, utterance, and DA) to calculate the importance of each word, provide contextual and pragmatic information, and guide the model to generate the category of utterance, respectively. Both multiple attention mechanism and semantic update mechanism are applied to infer the more accurate representation of utterance for DA recognition in a general case. Experimental results show that our model can achieve comparable performance to the state-of-the-art models with less extra information.