Segmentation thématique : apport de la vectorisation

This paper deals with topic segmentation of TV broadcasts using their transcription obtained by automatic speech recognition. Topic segmentation has been studied for several years, and most often the techniques proposed rely on information retrieval techniques to compute similarities between segments. In this paper, we propose a new segmentation approach inspired by mathematical morphology studies developed in the framework of image segmentation. We also show that using the similarity technique called vectorization and first developed for IR (Claveau et al., 2010) can be efficiently used in this context. This new topic segmentation technique is evaluated on two corpora of TV broadcasts on which it outperforms other existing approaches. MOTS-CLÉS : Segmentation thématique, vectorisation, ligne de partage des eaux, calcul de similarité, flux TV, plongement