Dialogue Act Recognition for Open-Domain Based on Word-Level Sequence Annotation with CRF

In dialogue system, natural language understanding has a huge impact on human-machine interaction. Moreover, natural language contains a huge amount of noise within its content because the short length and free format. From this aspect, the demand of this paper is to design the useful feature representations and verify these feature sets by searching suitable machine learnings. Besides, the dialogue acts as labels associate with utterance can be viewed as sequence labelling problem. Considering this situation, this paper provides a word-level sequence annotation method, which annotates dialogue structural information and semantic information to each word of utterance. Meanwhile, Linear-CRF is employed in this paper because it is able to natively capture constrains of hidden state sequence and obtain optimal probability, which is perfectly suit for sequence labeling task. Compared with proposed method, this paper uses the Bag-of-word model to present utterance as baseline. According to result, all the designed features outperform the feature in baseline, which indicates that the proposed method is useful for DA recognition task.