Intra-Speaker Topic Modeling for Improved Multi-Party Meeting Summarization with Integrated Random Walk

This paper proposes an improved approach of summarization for spoken multi-party interaction, in which integrated random walk is performed on a graph constructed with topical/lexical relations. Each utterance is represented as a node of the graph and the edge between two nodes is weighted by the similarity between the two utterances, where we use two types of edges, one from topical similarity evaluated by probabilistic latent semantic analysis (PLSA) and another from word overlap. We model intra-speaker topics by partially sharing the topics from the same speaker in the graph. We did experiments for ASR and manual transcripts. For ASR transcripts, experiments showed intra-speaker topic sharing and integrating topical/lexical relations can help include the important utterances.