Mencius: A Chinese Named Entity Recognizer Using Hybrid Model

Mencius: A Chinese Named Entity Recognizer Using Hybrid Model Tzong-Han Tsai *�, Shih-Hung Wu�, and Wen-Lian Hsu� *Department of Computer Science and Information Engineering National Taiwan University Taipei, Taiwan, R.O.C. d90013@csie.ntu.edu.tw �Institute of Information Science, Academia Sinica. Taipei, Taiwan, R.O.C. {thtsai, shwu, hsu}@iis.sinica.edu.tw Abstract This paper presents a maximum entropy based Chinese named entity recognizer (NER): Mencius. It aims to address Chinese NER problems by combining the advantages of rule-based and machine learning (ML) based NER systems. Rule-based NER systems can explicitly encode human comprehension and can be tuned conveniently, while ML-based systems are robust, portable and inexpensive to develop. Our hybrid system incorporates a rule-based knowledge representation and template-matching tool, InfoMap [1], into a maximum entropy (ME) framework. Named entities are represented in InfoMap as templates, which serve as ME features in Mencius. These features are edited manually and their weights are estimated by the ME framework according to the training data. To avoid the errors caused by word segmentation, we model the NER problem as a character-based tagging problem. In our experiments, Mencius outperforms both pure rule-based and pure ME-based NER systems. The F-Measures of person names (PER), location names (LOC) and organization names (ORG) in the experiment are respectively 92.4%, 73.7% and 75.3%. 1 Introduction Information Extraction (IE) is the task of extracting information of interest from unconstrained text. IE involves two main tasks: the recognition of named entities, and the recognition of the relationships among these named entities. Named Entity Recognition (NER) involves the identification of proper names in text and their classification into different types of named entities (e.g., persons, organizations, locations). NER is not only important in IE [3] but also in lexical acquisition for the development of robust NLP systems [4]. Moreover, NER has proven fruitful for tasks such as documents indexing, and maintenance of databases containing identified named entities. During the last decade, NER has drawn much attention at Message Understanding Conferences (MUC) [5] [6]. Both rule-based and machine learning NER systems have had some success. Previous rule-based approaches have used manually constructed finite state patterns, which match text against a sequence of words. Such system (like University of Edinburgh's LTG [7]) do not need too much training data and can encode expert human knowledge. However, rule-based approaches lack robustness and portability. Each new source of text requires a significant tweaking of the rules to maintain optimal performance; the maintenance costs can be quite steep. Another popular approach in NER is machine-learning (ML). ML is more attractive in that it is more portable and less expensive to maintain. The representative ML approaches used in NER are HMM (BBN's IdentiFinder in [8, 9] and Maximum Entropy (ME) (New York Univ.'s MEME in [10] [11]). Although ML systems are relatively inexpensive to develop, the outputs of these systems are difficult to interpret. As well, it is difficult to improve the system performance through error analysis. The performance of a ML system can be very poor when training data is insufficient. Furthermore, the performance of …

Content truncated to 500 words