Unity Is Strength: Coupling Media For Thematic Segmentation Dalila Mekhaldi, Denis Lalanne and Rolf Ingold Universit´ e de Fribourg , Chemin de mus´ ee 3, CH-1700 Fribourg {dalila.mekhaldi, denis.lalanne, rolf.ingold}@unifr.ch Abstract. In this paper we present the preliminary results and the eval- uation of a combined thematic segmentation of (a) meeting documents and (b) meeting speech transcript. Our approach is based on a clustering method applied on a 2D representation of the thematic alignment, and then the projection of the extracted clusters on each axis, corresponding to meeting documents and the speech transcript. Finally, our bi-modal thematic segmentation method is evaluated, in regards to a mono-modal segmentation method (TextTiling ). 1 Introduction In the context of multimodal applications, especially meeting recordings and lectures, research are in hand, in order to establish temporal links between the various modalities, mainly between documents and meetings dialogs [5]. Our viewpoint is that bridging temporal links between these two modalities may be attained once their thematic links, i.e. their thematic alignment, are established. The document/speech thematic alignment and the thematic segmentation are closely related. The thematic alignment is building thematic links between doc- uments and speech units, which are semantically close. While thematic seg- mentation builds thematic links between units of a unique modality (document or speech). Thematic segmentation is thus an intra-modal segmentation, while thematic alignment is an inter-modal segmentation. Since the preliminary eval- uation we have performed on state-of-the-art, thematic segmentation methods did not show good results, our assumption is that an inter-modal segmentation will be more efficient and will benefit from the various modalities information. In this article, we present briefly our bi-modal thematic segmentation method and its projection to each modality. A preliminary evaluation shows that our bi-modal segmentation is more efficient than a mono-modal segmentation. 2 Thematic alignment vs. Thematic segmentation Our document/speech alignment takes as input the speech transcript of a meet- ing and the documents related to the meeting, and generates a set of aligned pairs (document units, speech units) [5]. Currently, we are focusing on press re- views, where many speakers discuss a daily newspaper cover page.