Multi-party Conversation Summarization Based on Sentence Selection Using Verbal and Nonverbal Information

In this paper, we propose a method for conversation summarization. For the method, we combine two approaches, a scoring method and a machine learning technique (SVMs). First we compare important utterance extraction by the scoring method and SVMs. In the machine learning technique, we introduce verbal features, such as relations between utterances and anaphora features, and nonverbal features. Next we generate a summary from the outputs of the scoring method and SVMs. In our approach, a basic summary consists of utterances with high confidence extracted from the scoring method. Utterances from SVMs are used as supplementary information. In the experiment, we compare a combination method and a method with only SVMs. The output of our method was suitable in terms of readability and correctness as a summary of original conversation.