Dialog Act Segmentation and Classification in Vietnamese

Natural Language Understanding (NLU) is a critical component in building a conversational system. So far, most systems have processed the user inputs at the utterance-level and assumed single dialog act (DA) per utterance. In fact, one utterance might contain more than one DA which are denoted by different continuous text spans inside it (a.k.a functional segments). As a step towards achieving natural and flexible interaction between human and machine especially in poor-resource languages, this paper presents a work for dialog segmentation (DS) and DA classification in Vietnamese. We first introduce the corpus and then systematically investigate different pipeline and joint learning approaches to deal the two tasks. Experimental results show that the joint learning approach is superior in boosting the performance of both tasks. It outperforms the conventional pipeline approach which looked at the two tasks separately. Moreover, to further enhance the final performance, this paper proposes a technique to enrich the models with useful DA knowledge. Compared to the standard models which don't use DA knowledge, we achieve considerably better results for two tasks. Specifically, we achieved an F1 score of 86% in segmenting dialogues, and an Fl-micro score of 74.75% in classifying DAs. This provides a strong foundation for future research on this interesting field.