AUTOMATIC TOPIC SEGMENTATION AND LABELING IN MULTIPARTY DIALOGUE

Abstract

This study concerns how to segment a scenario-driven multiparty \ndialogue and how to label these segments automatically. \nWe apply approaches that have been proposed for identifying \ntopic boundaries at a coarser level to the problem of \nidentifying agenda-based topic boundaries in scenario-based \nmeetings. We also develop conditional models to classify segments \ninto topic classes. Experiments in topic segmentation \nshow that a supervised classification approach that combines \nlexical and conversational features outperforms the unsupervised \nlexical chain-based approach, achieving 20% and 12% \nimprovement on segmentating top-level and sub-topic segments \nrespectively. Experiments in topic classification suggest \nthat it is possible to automatically categorize segments \ninto appropriate topic classes given only the transcripts. Training \nwith features selected using the Log Likelihood ratio improves \nthe results by 13.3%.