The Effect of Answer Patterns for Supervised Named Entity Recognition in Thai

The Effect of Answer Patterns for Supervised Named Entity Recognition in Thai ∗ Nutcha Tirasaroj and Wirote Aroonmanakun Department of Linguistics, Chulalongkorn University, Bangkok, Thailand pernn39@gmail.com, awirote@chula.ac.th Abstract. In this paper, we present Thai named entity recognition (NER) systems using supervised Conditional Random Fields (CRFs) with various answer patterns to find out whether different answer patterns would affect the performance of the systems. Every system used the same set of features except the answers in the training corpus. There are 5 patterns of answer used in this study. The results show that the answer tags having more information can help improve the system performance than those having less information. Keywords: Named entity recognition, Conditional Random Fields, Supervised machine learning, Thai language processing ∗ The work reported in this paper was supported by The Thailand Research Fund (TRF) under grant no MSG53Z0008, and partially supported by Chulalongkorn University Centenary Academic Development Project. Moreover, we would like to thank NECTEC for their BEST 2009 corpus and reviewers for their suggestions and comments. Copyright 2011 by Nutcha Tirasaroj and Wirote Aroonmanakun Introduction Named entity recognition (NER) was first introduced in MUC-6 in 1990’s. According to MUC, NER is divided into 3 main groups. The first one is entity names including person, organization, and location names. The second one is temporal expressions and the last one is numerical expressions which are monetary values and percentages. Among these 3 groups, the first one is the most difficult to be recognized because the structures of the names are complicated and the contexts they occur are varying. As a result, a number of research on NER as well as this study have focused on the first group. Most research on NER using statistical models mainly focused on feature selection since features are the most important part of the models and directly affect the performance of the systems. It was assumed that the answer given to the models has no effect on the system performance. However, in Chinese, some research on NER (Mao et al., 2008; Yu et al., 2008), the work of Yu et al. (2008) proved that the number of answer tags can help improve the system performance. The use of more tags yields better results because it gives more information of entity names’ boundaries to the system for learning. For machine learning, it tends to be that the more information is provided, the better the performance of the system is. Research on Thai NER using statistical models such as SVMs (Suwanno et al., 2007), HMM (Sansing and Prayot, 2010), ME (Chanlekha and Kawtrakul, 2004), CRFs (Tepdang et al.; Tirasaroj and Aroonmanakun, 2009), etc. also focused on feature selections. None has studied the effect of answer patterns for their supervised systems. The answer pattern is the way we give the answer in the training corpus for supervised machine learning system. The answer suggests how much information of NEs provided for the systems. For example, the answer can be only the information of NE types or it …

Content truncated to 500 words