Story Link Detection based on Dynamic Information Extending

Story Link Detection based on Dynamic Information Extending Xiaoyan Zhang Ting Wang Huowang Chen Department of Computer Science and Technology, School of Computer, National University of Defense Technology No.137, Yanwachi Street, Changsha, Hunan 410073, P.R.China {zhangxiaoyan, tingwang, hwchen}@nudt.edu.cn Abstract Topic Detection and Tracking refers to au- tomatic techniques for locating topically re- lated materials in streams of data. As the core technology of it, story link detection is to determine whether two stories are about the same topic. To overcome the limitation of the story length and the topic dynamic evolution problem in data streams, this pa- per presents a method of applying dynamic information extending to improve the per- formance of link detection. The proposed method uses previous latest related story to extend current processing story, generates new dynamic models for computing the sim- ilarity between the current two stories. The work is evaluated on the TDT4 Chinese cor- pus, and the experimental results indicate that story link detection using this method can make much better performance on all evaluation metrics. 1 Introduction Topic Detection and Tracking (TDT) (Allan, 2002) refers to a variety of automatic techniques for dis- covering and threading together topically related material in streams of data such as newswire or broadcast news. Such automatic discovering and threading could be quite valuable in many appli- cations where people need timely and efficient ac- cess to large quantities of information. Supported by such technology, users could be alerted with new events and new information about known events. By examining one or two stories, users define the topic described in them. Then with TDT technologies they could go to a large archive, find all the stories about this topic, and learn how it evolved. Story link detection, as the core technology de- fined in TDT, is a task of determining whether two stories are about the same topic, or topically linked. In TDT, a topic is defined as ”something that hap- pens at some specific time and place” (Allan, 2002). Link detection is considered as the basis of other event-based TDT tasks, such as topic tracking, topic detection, and first story detection. Since story link detection focuses on the streams of news stories, it has its specific characteristic compared with the traditional Information Retrieval (IR) or Text Clas- sification task: new topics usually come forth fre- quently during the procedure of the task, but nothing about them is known in advance. The paper is organized as follows: Section 2 de- scribes the procedure of story link detection; Section 3 introduces the related work in story link detection; Section 4 explains a baseline method which will be compared with the proposed dynamic method in Section 5; the experiment results and analysis are given in Section 6; finally, Section 7 concludes the paper. 2 Problem Definition In the task definition of story link detection (NIST, 2003), a link detection system is given a se- quence of time-ordered news source files S ⟨S1, S2, S3, . . . , Sn⟩ where …

Content truncated to 500 words