Semantic Frame-based Statistical Approach for Topic Detection

PACLIC 28 !75 Semantic Frame-based Statistical Approach for Topic Detection Yung-Chun Chang1,2 Yu-Lun Hsieh2,3 Cen-Chieh Chen2,3 Chad Liu2 Chun-Hung Lu4 Wen-Lian Hsu2 1Department of Information Management, National Taiwan University, Taiwan 2Institute of Information Science, Academia Sinica, Taiwan 3Department of Computer Science, National Chengchi University, Taiwan 4Innovative Digitech-Enabled Applications & Services Institute, III, Taiwan {changyc,morphe,can,hsu}@iis.sinica.edu.tw, 4enricoghlu@iii.org.tw Abstract We propose a statistical frame-based approach (FBA) for natural language processing, and demonstrate its advantage over traditional ma- chine learning methods by using topic detec- tion as a case study. FBA perceives and iden- tifies semantic knowledge in a more general manner by collecting important linguistic pat- terns within documents through a unique flex- ible matching scheme that allows word inser- tion, deletion and substitution (IDS) to cap- ture linguistic structures within the text. In ad- dition, FBA can also overcome major issues of the rule-based approach by reducing hu- man effort through its highly automated pat- tern generation and summarization. Using Ya- hoo! Chinese news corpus containing about 140,000 news articles, we provide a compre- hensive performance evaluation that demon- strates the effectiveness of FBA in detecting the topic of a document by exploiting the semantic association and the context within the text. Moreover, it outperforms common topic models like Na¨ıve Bayes, Vector Space Model, and LDA-SVM. 1 Introduction Due to recent technological advances, we are over- whelmed by the sheer number of documents. While keyword search systems nowadays can efficiently retrieve documents, users still have difficulty assimi- lating knowledge of interest from them. To promote research on this subject, the Defense Advanced Re- search Projects Agency (DARPA) initiated the Topic Detection and Tracking (TDT) project, with a goal of automatically detecting topics and tracking re- lated documents from document streams such as on- line news feeds. In essence, a topic is associated with specific times, places, and persons (Nallapati et al., 2004). Thus, detecting the topic of a doc- ument can help readers construct the background of the topic and facilitate document comprehension, which is an active research area in information re- trieval (IR). Linguistic information provides useful features to many natural language processing (NLP) tasks, in- cluding topic detection (Nallapati, 2003). Such in- formation is usually represented as rules or tem- plates. The main advantages of the rule-based ap- proach are its high precision as well as the capabil- ity of knowledge accumulation. When confronting a new domain, they can be adapted by adding rules that exploit the missing knowledge. However, only a limited number of cases can be captured by a single rule, and increasing the number of rules could create undesired conflicts. Thus, the inflexibility of rule- based systems has put their competence for NLP tasks in doubt. On the other hand, there are several machine learning-based approaches. For instance, Nallap- ati et al. (2004) attempted to find characteristics of topics by clustering keywords using statistical sim- ilarity. The clusters are then connected chronologi- cally to form a time-line of the topic. Furthermore, many previous methods treated topic detection …

Content truncated to 500 words