Topic-Features for Dialogue Summarization

Texts such as news reports and academic papers come from one single speaker and are well-structured. However, dialogues often come from two or more speakers exchanging information. In this case, the topic or intention may change in a dialogue, and the key information is often scattered in utterances of different speakers, which brings challenges to the abstractive dialogue summarization. It is difficult to apply the traditional topic modeling approaches because of too much noise and the inherent characteristics of dialogue. In order to effectively model the entire dialogue and capture various topic information, this paper proposes a topic-feature approach based on neural topic model, including word-level embedding and dialogue-level representation. Experimental results on the largest dialogue summarization corpus SAMSum show that the proposed approach can significantly improve over competitive baselines. In addition, we also conduct experiments on other datasets from different domains to verify the effectiveness and generality of our proposed approach.