This paper describes our on-going work toward the improvement of Broadcast News story segmentation module. We have tried to improve our baseline algorithm by further exploring the typical structure of a broadcast news show, first by training a CART and then by integrating it in a 2-stage algorithm that is able to deal with shows with double anchors. In order to deal with shows with a thematic anchor, a more complex approach is adopted including a topic classification stage. The automatic segmentation is currently being compared with the manual segmentation done by a professional media watch company. The results are very promising so far, specially taking into account that no video information is used.
展开▼