Transcript Segmentation Using Utterance Cosine Similarity Measure

One of the problems addressed by the Tracker project is the extraction of the key issues discussed at meetings through the analysis of transcripts. Whilst the task of topic extraction is an easy task for humans it has proven difficult task to automate given the unstructured nature of our transcripts. This paper proposes a new approach to transcript segmentation based on the Utterance Cosine Similarity (UCS) method. Our segmentation approach is based on the notion of semantic similarity of utterances within the transcripts that measures the content similarity, semantic relationships, and use distance to differentiate same topics that appear in different context. The method is illustrated using one of the 17 transcripts in our study.