An Empirical Study on Fine-Grained Named Entity Recognition

Proceedings of the 27th International Conference on Computational Linguistics, pages 711–722 Santa Fe, New Mexico, USA, August 20-26, 2018. 711 An Empirical Study on Fine-Grained Named Entity Recognition Khai Mai1,∗ Thai-Hoang Pham1,∗ Nguyen Minh Trung1 Nguyen Tuan Duc1 Danushka Bolegala2 Ryohei Sasano3 Satoshi Sekine4 1) Alt Inc 2) University of Liverpool 3) Nagoya University 4) Riken AIP {mai.tien.khai, pham.thai.hoang, nguyen.minh.trung, nguyen.tuan.duc}@alt.ai, danushka.bollegala@liverpool.ac.uk, sasano@i.nagoya-u.ac.jp, satoshi.sekine@riken.jp Abstract Named entity recognition (NER) has attracted a substantial amount of research. Recently, sev- eral neural network-based models have been proposed and achieved high performance. However, there is little research on fine-grained NER (FG-NER), in which hundreds of named entity cat- egories must be recognized, especially for non-English languages. It is still an open question whether there is a model that is robust across various settings or the proper model varies depend- ing on the language, the number of named entity categories, and the size of training datasets. This paper first presents an empirical comparison of FG-NER models for English and Japanese and demonstrates that LSTM+CNN+CRF (Ma and Hovy, 2016), one of the state-of-the-art methods for English NER, also works well for English FG-NER but does not work well for Japanese, a language that has a large number of character types. To tackle this problem, we propose a method to improve the neural network-based Japanese FG-NER performance by removing the CNN layer and utilizing dictionary and category embeddings. Experimental results show that the proposed method improves Japanese FG-NER F-score from 66.76% to 75.18%. 1 Introduction Named entity recognition (NER) is a well studied topic in natural language processing. There have been many methods proposed for NER, including the conventional methods based on Conditional Random Fields (CRF) (McCallum and Li, 2003), Support Vector Machines (SVM) (Yamada et al., 2002; Takeuchi and Collier, 2002) and Hidden Markov Model (HMM) (Zhou and Su, 2002). Recently, neural net- work based methods, such as LSTM+CNN+CRF (Ma and Hovy, 2016) or BiLSTM/LSTM-CRF/Stack LSTMs (Lample et al., 2016; Misawa et al., 2017), have achieved state-of-the-art performance. How- ever, while most existing studies mainly focus on recognizing a relatively small number of named entity (NE) categories (e.g., ten or twelve categories) such as Person, Organization, Location, Artifact, etc., modern NLP applications often require domain-specific and fine-grained (FG) NER with hundreds of NE categories. For example, a movie recommendation system might require the recognition of movie names, but does not need to recognize Locations. Similarly, a chatbot software might require not only the recognition of Organization, but also the fine-grained classification to recognize a music band name to answer the question “Which band was Paul in”, from the information shown in Figure 1. A fine-grained named entity recognition (FG-NER) model refers to a NER model that can recog- nize and classify a large number of entity categories (e.g., hundreds of NE categories). In classical coarse-grained named entity (NE) definition, often less than ten named entity categories are defined. For example, in the CoNLL-2003 Named Entity Recognition task, there are four NE categories: Person, Location, Organization …

Content truncated to 500 words