2003 ISCA Workshop on Multilingual Spoken Document Retrieval
(MSDR2003)

Hong Kong
April 4-5, 2003

Two-Stage Story Segmentation and Detection on Broadcast News Using Genetic Algorithm

Jia-Hsin Hsieh, Chung-Hsien Wu, Kuao-Ann Fung

Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan, Taiwan

This paper proposes a two-stage story segmentation and detection approach on Mandarin broadcast news. In the two-stage paradigm, a topic classifier is first constructed to find the topic on the broadcast news within a sliding window and determine the potential story boundaries. Then, the problem for story segmentation is transformed to the determination of a chromosome (number sequence) in a search space. The genetic algorithm is then adopted to globally determine the chromosome, which represents the final story boundaries. A topic strength measure is defined as the fitness function used in the genetic algorithm. In order to evaluate our proposed approach, the word-based and syllable-based story segmentation systems were constructed. Experimental results show our proposed method achieves a better performance with 32.94% missing probability and 22.83% false alarm probability compared to the Makhoul’s method for the segmentation and detection on Mandarin broadcast news.


Full Paper

Bibliographic reference.  Hsieh, Jia-Hsin / Wu, Chung-Hsien / Fung, Kuao-Ann (2003): "Two-stage story segmentation and detection on broadcast news using genetic algorithm", In MSDR-2003, 55-60.