Dialog Acts Classification with Semantic and Structural Information

Dialog act, which is an essential element in dialog systems, plays an important role in understanding users' intents in the human-computer dialog. With the increasing applications of deep learning in natural language processing, more and more researchers adopt methods of text classification to identify dialog acts. However, sometimes one utterance may correspond to multiple labels, which leads to a cumbersome problem in classifying dialog acts. In this paper, we propose a multi-label classification model for solving this issue. Meanwhile, as Graph Convolutional Network can extract spatial features of topological graphs, we utilize it to capture syntactic information to enrich the words' semantic representations. In addition, due to the fact that one utterance consists of different clauses and each clause may correspond to one act label, we associate the relationship between clauses through Recurrent Neutral Network. Finally, our model combines the above two-level representations together. Compared with other classification models, experimental results can verify that our proposed model performs better.