Assessing the effectiveness of conversational features for dialogue segmentation in medical team meetings and in the AMI corpus

This paper presents a comparison of two similar dialogue analysis tasks: segmenting real-life medical team meetings into patient case discussions, and segmenting scenario-based meetings into topics. In contrast to other methods which use transcribed content and prosodic features (such as pitch, loudness etc), the method used in this comparison employs only the duration of the prosodic units themselves as the basis for dialogue representation. A concept of Vocalisation Horizon (VH) allows us to treat segmentation as a classification task where each instance to be classified is represented by the duration of a talk spurt, pause or speech overlap event in the dialogue. We report on the results this method yielded in segmentation of medical meetings, and on the implications of the results of further experiments on a larger corpus, the Augmented Multiparty Meeting corpus, to our ongoing efforts to support data collection and information retrieval in medical team meetings.