Chinese Named Entity Recognition with Multiple Features

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 427–434, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Chinese Named Entity Recognition Based on Multiple Features Youzheng Wu, Jun Zhao, Bo Xu Hao Yu National Laboratory of Pattern Recognition Fujitsu R&D Center Co., Ltd Institute of Automation, CAS Beijing 100016, China Beijing, 100080, China yu@frdc.fujitsu.com (yzwu,jzhao,bxu)@nlpr.ia.ac.cn Abstract This paper proposes a hybrid Chinese named entity recognition model based on multiple features. It differentiates from most of the previous approaches mainly as follows. Firstly, the proposed Hybrid Model integrates coarse particle feature (POS Model) with fine particle feature (Word Model), so that it can overcome the disadvantages of each other. Secondly, in order to reduce the searching space and improve the efficiency, we introduce heu- ristic human knowledge into statistical model, which could increase the perform- ance of NER significantly. Thirdly, we use three sub-models to respectively de- scribe three kinds of transliterated person name, that is, Japanese, Russian and Euramerican person name, which can im- prove the performance of PN recognition. From the experimental results on People's Daily testing data, we can conclude that our Hybrid Model is better than the mod- els which only use one kind of features. And the experiments on MET-2 testing data also confirm the above conclusion, which show that our algorithm has consis- tence on different testing data. 1 Introduction Named Entity Recognition (NER) is one of the key techniques in the fields of Information Extraction, Question Answering, Parsing, Metadata Tagging in Semantic Web, etc. In MET-2 held in conjunction with the Seventh Message Understanding Confer- ence (MUC-7), the task of NER is defined as rec- ognizing seven sub-categories entities: person (PN), location (LN), organization (ON), time, date, cur- rency and percentage. As for Chinese NEs, we fur- ther divide PN into five sub-classes, that is, Chinese PN (CPN), Japanese PN (JPN), Russian PN (RPN), Euramerican PN (EPN) and abbrevi- ated PN (APN) like "吴先生/Mr. Wu". Similarly, LN is split into common LN (LN) like "中关村 /Zhongguancun" and abbreviated LN (ALN) such as "京/Beijing", "沪/Shanghai". The recognition of time (TM) and numbers (NM) is comparatively simpler and can be implemented via finite state automata. Therefore, our research focuses on the recognition of CPN, JPN, RPN, EPN, APN, LN, ALN and ON. Compared to English NER, Chinese NER is more difficult. We think that the main differences between Chinese NER and English NER lie in: (1) Unlike English, Chinese lacks the capitalization information which can play very important roles in identifying named entities. (2) There is no space between words in Chinese, so we have to segment the text before NER. Consequently, the errors in word segmentation will affect the result of NER. In this paper, we proposes a hybrid Chinese NER model based on multiple features which em- phasizes on (1) combining fine particle features (Word Model) with coarse particle features (POS Model); (2) integrating human knowledge into sta- tistical model; (3) and using diverse sub-models …

Content truncated to 500 words