Complementarity of Lexical Cohesion and Speaker Role Information for Story Segmentation of French TV Broadcast News

摘要

Topic boundary detection in French TV Broadcast News is addressed in this paper with an approach based on the combination of two views: lexical cohesion and speaker role analysis. We propose an improved selection strategy from the classical lexical cohesion curve as well as an integrated supervised classification approach that jointly exploits the two views. The combination of these methods leads to significant improvements on a rich French database composed of shows from 7 different channels.