Combining Linguistic Features with Weighted Bayesian Classifier for Temporal Reference Processing

Combining Linguistic Features with Weighted Bayesian Classifier for Temporal Reference Processing Guihong Cao Department of Computing The Hong Kong Polytechnic University, Hong Kong csghcao@comp.polyu.edu.hk Wenjie Li Department of Computing The Hong Kong Polytechnic University, Hong Kong cswjli@comp.polyu.edu.hk Kam-Fai Wong Department of Systems Engineering and Engineering Management The Chinese University of Hong Kong, Hong Kong kfwong@se.cuhk.edu.hk Chunfa Yuan Department of Computer Science and Technology Tsinghua University, Beijing, China. cfyuan@tsinghua.edu.cn Abstract Temporal reference is an issue of determining how events relate to one another. Determining temporal relations relies on the combination of the information, which is explicit or implicit in a language. This paper reports a computational model for determining temporal relations in Chinese. The model takes into account the ef- fects of linguistic features, such as tense/aspect, temporal connectives, and discourse structures, and makes use of the fact that events are repre- sented in different temporal structures. A ma- chine learning approach, Weighted Bayesian Classifier, is developed to map their combined effects to the corresponding relations. An em- pirical study is conducted to investigate differ- ent combination methods, including lexical- based, grammatical-based, and role-based methods. When used in combination, the weights of the features may not be equal. Incor- porating with an optimization algorithm, the weights are fine tuned and the improvement is remarkable. 1 Introduction Temporal information describes changes and time of the changes. In a language, the time of an event may be specified explicitly, for example “他们在 1997 年解决了该市的交通问题 (They solved the traf- fic problem of the city in 1997)”; or it may be related to the time of another event, for example “修成立交 桥以后, 他们解决了该市的交通问题 (They solved the traffic problem of the city after the street bridge had been built”. Temporal reference describes how events relate to one another, which is essential to natural language processing (NLP). Its major appli- cations cover syntactic structural disambiguation (Brent, 1990), information extraction and question answering (Li, 2002), language generation and ma- chine translation (Dorr, 2002). Many researchers have attempted to characterize the nature of temporal reference in a discourse. Iden- tifying temporal relations1 between two events de- 1 The relations under examined include both intra-sentence and inter- pends on a combination of information resources. This information is provided by explicit tense and aspect markers, implicit event classes or discourse structures. It has been used to explain semantics of temporal expressions (Moens, 1988; Webber, 1988), to constrain possible temporal interpretations (Hitzeman, 1995; Sing, 1997), or to generate appro- priate temporally conjoined clauses (Dorr, 2002). The purpose of our work is to develop a computa- tional model, which automatically determines tempo- ral relations in Chinese. While temporal reference interpretation in English has been well studied, Chi- nese has been rarely discussed. In our study, thirteen related features are identified from linguistic per- spective. How to combine these features and how to map their combined effects to the corresponding rela- tions are the critical issues to be addressed in this paper. Previous work was limited in that they just con- structed constraint or …

Content truncated to 500 words