Initial Explorations on using CRFs for Turkish Named Entity Recognition

Proceedings of COLING 2012: Technical Papers, pages 2459–2474, COLING 2012, Mumbai, December 2012. Initial explorations on using CRFs for Turkish Named Entity Recognition Gökhan Akın ¸Seker1 Gül¸sen Eryi˘git2 (1) ITU Informatics Institute Graduate Program in Computer Science, Istanbul Technical University Istanbul, 34469, Turkey (2) Department of Computer Engineering, Istanbul Technical University Istanbul, 34469, Turkey {sekerg, gulsen.cebiroglu}@itu.edu.tr ABSTRACT This paper reports the highest results (95% in MUC and 92% in CoNLL metric) in the litera- ture for Turkish named entity recognition; more specifically for the task of detecting person, location and organization entities in general news texts. We give an in depth analysis of the previous reported results and make comparisons with them whenever possible. We use condi- tional random fields (CRFs) as our statistical model. The paper presents initial explorations on the usage of rich morphological structure of the Turkish language as features to CRFs together with the use of some basic and generative gazetteers. KEYWORDS: Named Entity Recognition , Turkish , Conditional Random Fields , ENAMEX. 2459 1 Introduction Named Entity Recognition (NER) can be basically defined as identifying and cate- gorizing certain type of data (i.e. person, location, organization names, date-time expressions). NER is an important stage for several natural language processing (NLP) tasks including machine translation, sentiment analysis and information extraction. MUC (Sundheim, 1995; Chinchor and Marsh, 1998) and CoNLL (Tjong Kim Sang, 2002; Tjong Kim Sang and De Meulder, 2003) conferences define three basic types of named enti- ties; these are 1- ENAMEX (person, location and organization names), 2- TIMEX (date and time entities) and 3- NUMEX (numerical expressions like money and percentages). But NER research is not limited to only these types; different application areas concentrate to determin- ing alternative entity types such as protein names, medicine names, book titles. The NER research was firstly started in early 1990s for English. In 1995, with the high interest of the research community, the success rates for English achieved nearly the hu- man annotation performance on news texts (Sundheim, 1995). Nadeau and Sekine (2007) gives a survey of the research for English NER between 1991 to 2006. The satisfaction on English NER task directed the field to new research areas such as multilingual NER sys- tems (Tjong Kim Sang, 2002; Tjong Kim Sang and De Meulder, 2003), NER on informal texts (LIU et al., 2011; Rüd et al., 2011; Mohit et al., 2012), transliteration (Zhang et al., 2012) and coreference (Na and Ng, 2009) of named entities . Conditional Random Fields (Lafferty et al., 2001) is a very popular method used in NLP. It is also widely used for named entity recognition task in various domains (LIU et al., 2011; Ekbal and Bandyopadhyay, 2009; Settles, 2004). Stanford NER (Finkel et al., 2005) which is a well-known NER tool also uses CRFs as its machine learning method. Morphologically rich languages pose interesting challenges for NLP tasks as it is the case for NER (Hasan et al., 2009). Turkish being one of such languages attracts attention of the NLP …

Content truncated to 500 words