Event

Using Unsupervised and Supervised Approaches for Extractive Meeting Summarization

Shasha Xie

ICSI

Tuesday, August 04, 2009
12:30

Extractive meeting summarization selects salient sentences from the meeting transcripts to form a summary. In this talk, we present two approaches to improve summarization performance. The first one is based on the unsupervised global optimization framework [1] that selects summary sentences by the number of unique concepts they contain. We proposed to leverage sentence importance weights in this concept-based optimization method by using them to extract more indicative concepts, prune sentence candidates, and extend directly the objective function being optimized for. Experiments on the ICSI meeting corpus showed consistent improvement over the original approach. As for the second approach, we explore integrating prosodic information in a supervised summarization framework. We propose different ways to normalize a set of widely used prosodic features according to speaker, topic, or local context, and show that using only prosodic features can outperform using non-prosodic information (lexical, structural, discourse and topic features). We also evaluate different ways to combine prosodic and non-prosodic information, and their combination at the decision level yields better performance than using a single information source.

This work is joint with Benoit Favre, Dilek Hakkani-Tur and Yang Liu.

[1] Dan Gillick, Korbinian Riedhammer, Benoit Favre and Dilek Hakkani-Tur, "A global optimization framework for meeting summarization", in Proceedings of ICASSP, 2009.