Story link detection based on event model with uneven SVM

Abstract

Topic Detection and Tracking refers to automatic techniques for locating topically related materials in streams of data. As a core of it, story link detection is to determine whether two stories are about the same topic. Up to now, many representation models have been used in story link detection. But few of them are specific to stories. This paper proposes an event model based on the characters of stories. This model is used for story link detection and evaluated on the TDT4 Chinese corpus. The experimental results indicate that the system using the event model achieves a better performance than that using the baseline model. Furthermore, it shows a larger improvement to the former, especially when using uneven SVM as the multi-similarity integration strategy. © 2008 Springer-Verlag Berlin Heidelberg.