Description of the NTU System used for MET-2

DESCRIPTION OF THE NTU SYSTEM USED FOR MET2 Hsin-Hsi Chen, Yung-Wei Ding, Shih-Chung Tsai and Guo-Wei Bian Natural Language Processing Laboratory Department of Computer Science and Information Engineering National Taiwan University Taipei, TAIWAN E-mail: hh_chen@csie.ntu.edu.tw Fax: +886-2-23628167 ABSTRACT Named entities form the major components in a document. When we catch the fundamental entities, we can understand a document to some degree. This paper employs different types of information from different levels of text to extract named entities, including character conditions, statistic information, titles, punctuation marks, organization and location keywords, speech-act and locative verbs, cache and n-gram model. In the formal run of MET-2, the F-measures P&R, 2P&R and P&2R are 79.61%, 77.88% and 81.42%, respectively. INTRODUCTION People, affairs, time, places and things are five basic entities in a document. When we catch the fundamental entities, we can understand a document to some degree. Natural Language Processing Laboratory (NLPL) in Department of Computer Science and Information Engineering (CSIE), National Taiwan University (NTU) starts to study named entity extraction problem in 1993. At first, we focus on the extraction of Chinese person names, transliterated person names [1] and organization names [2]. The training data and the testing data in these experiments are selected from three Taiwan newspaper corpora (China Times, Liberty Times News and United Daily News). Chen and Lee [3] reported the precision rate and the recall rate for the extraction of Chinese person names, transliterated person names and organization names are (88.04%, 92.56%), (50.62%, 71.93%) and (61.79%, 54.50%), respectively in the 16th International Conference on Computational Linguistics. We employ these results to several applications. Chen and Wu [4] considered person names as one of clues in sentence alignment. Chen and Lee [3] show its application to anaphora resolution. Chen and Bian [5] proposed a method to construct white pages for Internet/Intranet users automatically. We extract information from World Wide Web documents, including proper nouns, E-mail addresses and home page URLs, and find the relationship among these data. Chen, Ding and Tsai [6,7] dealt with proper noun extraction for information retrieval. In MUC-7 and MET-2, we attend named entity extraction tasks for both English and Chinese. We extend our previous work on this problem to cover more named entity types such as locations, date/time expressions and monetary and percentage expressions. Several issues have to be addressed during extension. One of the major differences between Chinese and English language processing is that segmentation is required for Chinese. That is, we have to identify word boundary in Chinese sentences beforehand. That makes Chinese named entity extraction tasks more changeable. Besides, the vocabulary set and the Chinese coding set used in Taiwan and in China are not the same. The documents adopted in MET-2 are selected from newspapers in China, thus we have to transform simplified Chinese characters in GB coding set to traditional Chinese characters in Big-5 coding set before testing. A word that is known may become unknown due to transformation. For example, the character "�" in "��" (early morning) is used …

Content truncated to 500 words