Fault-Tolerant Learning for Term Extraction

Fault-Tolerant Learning for Term Extraction ∗ Yuhang Yang, Hao Yu, Yao Meng, Yingliang Lu and Yingju Xia Fujitsu Research & Development Center Co., LTD. 15/F, Tower A, Ocean International Center, No.56 Dong Si Huan Zhong Rd, Chaoyang District, Beijing, 100025, P.R. China {yyh, yu, mengyao, luyl, yjxia}@ cn.fujitsu.com Abstract. This paper presents the Fault-Tolerant Learning approach for term extraction. The approach extracts terms using automatically generated seeds instead of prior domain knowledge or annotated corpora. Thus it is applicable to any domain specific corpus and it is especially useful for resource-limited domains. Two classifiers are separately trained for prediction and verification to ensure the performance of the proposed approach. Evaluations conducted on two different domains for Chinese term extraction show significant improvements over existing techniques and also verify the efficiency and relative domain independent nature of the approach. Keywords: Fault-Tolerant Learning, term extraction, machine learning. ∗ Copyright 2010 by Yuhang Yang, Hao Yu, Yao Meng, Yingliang Lu and Yingju Xia 1 Introduction Terms are the lexical units to represent the most fundamental knowledge of a domain. Term extraction aims to extract meaningful words or phrases representing domain specific meaning or concepts. Thus two issues are considered in term extraction. The first issue is to identify boundaries of meaningful words and phrases. The second issue is to verify terms by calculating domain specificity (Kageura and Umino, 1996). Existing term extraction techniques can be divided into four main categories including statistics based measures, trigger words (or characters) based algorithms, domain knowledge based methods and supervised methods. The first category is statistics based measures which identify terms by their statistical significance. The most widely used statistical measurement is TF-IDF (Salton and McGill, 1983; Frank, 1999), which is based on the hypothesis that “if a candidate occurs frequently in a few documents of a domain, it is likely a term”. The co-occurrences between the target string and its components or context, referred to as Internal association (e.g. Schone and Jurafsky, 2001) and context dependency (e.g. Sornlertlamvanich et al., 2000), are used for term extraction. There are also studies evaluating the distribution of a term within a domain or across domains through different metrics, such as term representativeness (Hisamitsu and Niwa, 2002), Inter-Domain Entropy (Chang, 2005) and the Lexicon Set Algorithm (Chen et al., 2006). Statistics based techniques can extract common used terms with statistical significance. However, the techniques are very sensitive to term frequency, and thus terms with low frequencies cannot be extracted. The second category is based on trigger words or characters. According to (Feng et al., 2004) and (Yang et al., 2008), characters and words immediately before and after these terms are proven to be useful for term extraction. Accessor Variety Criteria proposed in (Feng et al., 2004) considers the characters that are directly before or after a string as important factors for determining the independence of the string. TEDel (delimiter based term extraction) proposed in PACLIC 24 Proceedings 321 (Yang et al., 2008) identifies terms by finding their predecessors and successors …

Content truncated to 500 words