Sixth International Conference on Spoken Language Processing
(ICSLP 2000)

Beijing, China
October 16-20, 2000

Expanded Vector Space Model based on Word Space in Cross Media Retrieval of News Speech Data

Seiichi Takao, Jun Ogata, Yasuo Ariki

Department of Electronics and Informatics, Ryukoku University, Seta, Otsu-shi, Shiga, Japan

News On Demand System using speech technology usually employs automatic speech transcriptions to retrieve the news data. In the retrieval, users specify a few keywords or sentences as a query and the related news data can be retrieved using the speech transcription. However when users canít give a query clearly, a video shot of news program which users are watching will become a good query to retrieve the related news data. As one of such kinds of news data retrieval, we propose here to employ video captions as a query and to retrieve the related news data using speech transcription. We call this kind of retrieval as cross media retrieval due to its media cross over. Conventionally available method in cross media retrieval is standard cosine measure in vector space model. In this conventional method, there is a problem of impossibility of semantic level retrieval. To solve this problem, we propose here an expanded vector space model based on a word space. Experimental results found that the expanded vector space model based on the word space has superiority to the conventional vector space model.


Full Paper

Bibliographic reference.  Takao, Seiichi / Ogata, Jun / Ariki, Yasuo (2000): "Expanded vector space model based on word space in cross media retrieval of news speech data", In ICSLP-2000, vol.3, 1085-1088.