Segmentation of Transcribed Free Conversation Taking Utterance Dynamics into Account

Informal communication plays an important role in many conversations such as for creating ideas and managing an organization smoothly. For this purpose, we have been studying methods to support informal communication. Currently, typical data for structuring conversations in the literature has come from goal-oriented conversations or general texts, i.e. scientific papers, articles of newspapers or magazines. In this paper, we propose a segmentation method for free conversation based on topics. We obtain free conversation in text form from transcribed round-table discussions. In addition to segmenting using standard cohesion and cue-phrases, our segmentation method also uses global-cohesion. Global-cohesion is a measure of regularity of keywords and topic transitions over an entire conversation. Among other factors, we use time transition information to obtain our global-cohesion measurement. In our experiments, we obtained a 73.2% recall factor and a 66.7% precision factor indicating that our segmentation scheme using a mixture of cohesion, cue-phrases and global-cohesion is useful for structuring conversations.