An Experimental Approach For Information Extraction in Multi-Party Dialogue Discourse

International audienceIn this paper, we address the task of information extraction for transcript of meetings. Meeting documents are not usually well struc-tured and are lacking formatting and punctuations. In addition, the information are distributed over multiple sentences. We experimentally investigate the usefulness of numerical statistics and topic modeling methods on a real data set containing multi-part dialogue texts. Such information extraction can be used for different tasks, of which we consider two: contrasting thematically related but distinct meetings from each other, and contrasting meetings involving the same participants from those involving other. In addition to demonstrating the difference between counting and topic modeling results, we also evaluate our experiments with respect to the gold standards provided for the data set