Topic Tracking Based on Linguistic Features

Topic Tracking Based on Linguistic Features Fumiyo Fukumoto and Yusuke Yamaji Interdisciplinary Graduate School of Medicine and Engineering, Univ. of Yamanashi, 4-3-11, Takeda, Kofu, 400-8511, Japan fukumoto@yamanashi.ac.jp, g03mk031@ccn.yamanashi.ac.jp Abstract. This paper explores two linguistically motivated restrictions on the set of words used for topic tracking on newspaper articles: named entities and headline words. We assume that named entities is one of the linguistic features for topic tracking, since both topic and event are related to a specific place and time in a story. The basic idea to use headline words for the tracking task is that headline is a compact representation of the original story, which helps people to quickly understand the most important information contained in a story. Head- line words are automatically generated using headline generation technique. The method was tested on the Mainichi Shimbun Newspaper in Japanese, and the re- sults of topic tracking show that the system works well even for a small number of positive training data. 1 Introduction With the exponential growth of information on the Internet, it is becoming increasingly difficult to find and organize relevant materials. Tracking task, i.e. starts from a few sample stories and finds all subsequent stories that discuss the target topic, is a new line of research to attack the problem. One of the major problems in the tracking task is how to make a clear distinction between a topic and an event in the story. Here, an event refers to the subject of a story itself, i.e. a writer wants to express, in other words, notions of who, what, where, when, why and how in the story. On the other hand, a topic is some unique thing that occurs at a specific place and time associated with some specific actions [1]. It becomes background among stories. Therefore, an event drifts, but a topic does not. For example, in the stories of ‘Kobe Japan quake’ from the TDT1 corpus, the event includes early reports of damage, location and nature of quake, rescue efforts, consequences of the quake, and on-site reports, while the topic is Kobe Japan quake. A wide range of statistical and machine learning techniques have been applied to topic tracking, including k-Nearest Neighbor classification, Decision Tree induction [3], relevance feedback method of IR [12,13], hierarchical and non-hierarchical clustering algorithms [20], and a variety of Language Modeling [15,5,10,17]. The main task of these techniques is to tune the parameters or the threshold for binary decisions to pro- duce optimal results. In the TDT context, however, parameter tuning is a tricky issue for tracking. Because only the small number of labeled positive stories is available for training. Moreover, the well-known past experience from IR that notions of who, what, where, when, why, and how may not make a great contribution to the topic tracking task [1] causes this fact, i.e. a topic and an event are different from each other. R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 10–21, 2005. c⃝ Springer-Verlag Berlin Heidelberg 2005 Topic …

Content truncated to 500 words