A Named Entity Recognition Method using Rules Acquired from Unlabeled Data

Proceedings of Recent Advances in Natural Language Processing, pages 170–177, Hissar, Bulgaria, 12-14 September 2011. A Named Entity Recognition Method using Rules Acquired from Unlabeled Data Tomoya Iwakura Fujitsu Laboratories Ltd. 1-1, Kamikodanaka 4-chome, Nakahara-ku, Kawasaki 211-8588, Japan iwakura.tomoya@jp.fujitsu.com Abstract We propose a Named Entity (NE) recogni- tion method using rules acquired from un- labeled data. Rules are acquired from au- tomatically labeled data with an NE rec- ognizer. These rules are used to identify NEs, the beginning of NEs, or the end of NEs. The application results of rules are used as features for machine learning based NE recognizers. In addition, we use word information acquired from un- labeled data as in a previous work. The word information includes the candidate NE classes of each word, the candidate NE classes of co-occurring words of each word, and so on. We evaluate our method with IREX data set for Japanese NE recog- nition and unlabeled data consisting of more than one billion words. The exper- imental results show that our method us- ing rules and word information achieves the best accuracy on the GENERAL and ARREST tasks of IREX. 1 Introduction Named Entity (NE) recognition aims to recognize proper nouns and numerical expressions in text, such as names of people, locations, organizations, dates, times, and so on. NE recognition is one of the basic technologies used in text processing such as Information Extraction and Question Answer- ing. To implement NE recognizers, semi- supervised-based methods have recently been widely applied. These methods use several differ- ent types of information obtained from unlabeled data, such as word clusters (Freitag, 2004; Miller et al., 2004), the clusters of multi-word nouns (Kazama and Torisawa, 2008), phrase clusters (Lin and Wu, 2009), hyponymy relations extracted from WikiPedia (Kazama and Torisawa, 2008), NE-related word information (Iwakura, 2010), and the outputs of classifiers or parsers created from unlabeled data (Ando and Zhang, 2005). These previous works have shown that features acquired from large sets of unlabeled data can contribute to improved accuracy. From the results of these previous works, we see that several types of features augmented with unlabeled data contribute to improved accuracy. Therefore, if we can incorporate new features augmented with unlabeled data, we expect more improved accuracy. We propose a Named Entity recognition method using rules acquired from unlabeled data. Our method uses rules identifying not only whole NEs, but also the beginning of NEs or the end of NEs. Rules are acquired from automatically la- beled data with an NE recognizer. The appli- cation results of rules are used as features for machine-learning based NE recognitions. Com- pared with previous works using rules identifying NEs acquired from manually labeled data (Isozaki, 2001), or lists of NEs acquired from unlabeled data (Talukdar et al., 2006), our method uses new features such as identification results of the be- ginning of NEs and the end of NEs. In addi- tion, we use word information (Iwakura, 2010). The word information includes the candidate NE classes of each …

Content truncated to 500 words