Topic-based Structuring of a Very Large-Scale News Video Corpus.pdf

Topic-Based Structuring of a Very Large-Scale News Video Corpus Ichiro IDE Hiroshi MO Norio KATAYAMA Shin’ichi SATOH National Institute of Informatics 2–1–2 Hitotsubashi, Chiyoda-ku, Tokyo, 101–8430, Japan TEL: +81-3-4212-2585 FAX: +81-3-3556-1916 E-Mail: {ide,mo,katayama,satoh}@nii.ac.jp Abstract We introduce a topic-based inter-video structuring method that considers application to a very large-scale news video corpus as well as user interfaces that provide the users with the ability to efficiently browse through the corpus based on the topic structure. Although the proposed method is a multimedia-integrated method that refers to both text and im- age based information, this paper focuses on text-based topic segmentation and tracking / threading. First, topic segmen- tation is performed referring to inter-sentence keyword vec- tor relations within a single video. Next, topic tracking and threading is performed referring to inter-topic keyword vec- tor relations throughout the entire video corpus. Such analy- sis should reveal the underlying structure of the entire corpus which is not simply a large volume of unrelated data, but data full of rich information in the content-based relational struc- ture itself. The segmentation method evaluated by applying the proposed method showed realistic ability. The proposed method was then applied to 555 daily news video (approx- imately 270 hours) obtained from a specific Japanese news program. Although detailed evaluation is yet to be done, the user interfaces showed good browsing ability for users to re- trieve and track a topic thread of interest. Introduction Due to the recent advance in telecommunication technol- ogy, large amounts of videos have become available through various media. Such video data contain a broad range of human activities, which could be considered as valu- able cultural and social heritage of the human race. From this viewpoint, news videos contain such information most densely. Nonetheless, buildin