The Construction of a Chinese Named Entity Tagged Corpus: CNEC1.0

The Construction of a Chinese Named Entity Tagged Corpus: CNEC1.0 Cheng-Wei Shih, Tzong-Han Tsai, Shih-Hung Wu, Chiu-Chen Hsieh, and Wen-Lian Hsu Institute of Information Science, Academia Sinica {dapi,thtsai,shwu,gladys,hsu}@iis.sinica.edu.tw Abstract. In order to build an automatic named entity recognition (NER) system for machine learning, a large tagged corpus is necessary. This paper describes the manual construction of a Chinese named entity tagged corpus (CNEC 1.0) that can be used to improve NER performance. In this project, we define five named entity tags: PER (person name), LOC (location name), ORG (organization name), LAO (location as organization), and OAL (organization as location) for named entity categories. In addition, we propose a special tag, DIFF (Difficulty), to annotate ambiguous cases during corpus construction. A, corpus-annotating procedure, a tagging tool, and an original corpus are also introduced. Finally, we demonstrate a part of our manual-tagged corpus. 1 Introduction Named entity recognition (NER), which includes the identification and classification of certain proper nouns in a text, is an important task in information extraction. It is useful in many natural language processing systems for document indexing and managing data with named entities [Tsai et. al 2004]. Since numerous new proper nouns are generated every day, it is not enough for an IR system to index names from Internet documents or refer to gazettes. Therefore, NER has become an important method for information processing in recent years. Machine learning (ML) is one of the most popular methods in NER, due to its easy maintenance and portability [Tsai et. al 2004]. Typical machine learning approaches applied in NER include the Hidden Markov Model (HMM) [Bikel et. al 1997], Support Vector Machine (SVM) [Asahara 2003], and Maximum Entropy (ME) [Borthwick 1998]. No matter which approach is used, a tagged named entity corpus with clear annotating criteria is needed in the training phase of building an NER system. However, constructing such a corpus is a labor-intensive task, so few researchers have focused on it. The Automatic Content Extraction program (ACE) executed by the Linguistic Data Consortium (LDC) [Http://wave.ldc.upenn.edu/] annotates seven common entities in English, simplified Chinese, and Arabic. Meanwhile, the IREX (Information Retrieval and Extraction Exercise) [Http://nlp.cs.nyu.edu/irex/] defines 8 kinds of named entities (NE) in Japanese [Sekine and Isahara 2000], and the shared task in CoNLL 2002 and 2003 [Erik 2002] [Erik et. 2002] develops the NER system using four types of NE in English, German, Dutch, and Spanish. However, as none of these methods focus on traditional Chinese, there is an urgent need for a traditional Chinese NE corpus and NE annotating standards to support an automatic Chinese NER system like Mencius [Tsai et. al 2004]. The categories of named entities defined by Message Understanding Conferences (MUC) are the names of persons, organizations, locations, temporal expressions and number expressions [Grishman and Sundheim 1996]. Since temporal and number expressions, such as “the past year” and “40 percent”, are generally used as adjectives to describe other entities, we disregard them and focus on the annotation of person names, organization names, and location …

Content truncated to 500 words