Learning conditional random field with hierarchical representations for dialogue act recognition

The analysis of dialogue act is important for computers to understand natural-language dialogues because the dialogue act of an utterance characterizes the speaker’s intention. In this paper, we create a new model that adapts Conditional Random Field (CRF) with efficient hierarchical representations of the raw inputs to solve the dialogue act recognition problem. The proposed model has two advantages. First, CRF can model the statistical dependencies between the utterances which carry important information to determine the dialogue act label of an utterance. Second, the hierarchical representations potentially contain some more abstract concepts with greater predictive power. To verify the effectiveness of our model, we compare it with several baseline methods on a dialogue act classification task. The results of the experiments demonstrate that our model performs much better than all the baseline methods.