An Encoding Strategy Based Word-Character LSTM for Chinese NER

Proceedings of NAACL-HLT 2019, pages 2379–2389 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 2379 An Encoding Strategy Based Word-Character LSTM for Chinese NER Wei Liu1,2, Tongge Xu3,1, Qinghua Xu1,2, Jiayu Song2, Yueran Zu2 1Hefei Innovation Research Institute, Beihang University 2School of Computer Science and Engineering, Beihang University 3School of Cyber Science and Technology, Beihang University {liuwei1206, xutg, xuqh buaa, heavensyc, yueranzu}@buaa.edu.cn Abstract A recently proposed lattice model has demon- strated that words in character sequence can provide rich word boundary information for character-based Chinese NER model. In this model, word information is integrated into a shortcut path between the start and the end characters of the word. However, the exis- tence of shortcut path may cause the model to degenerate into a partial word-based model, which will suffer from word segmentation er- rors. Furthermore, the lattice model can not be trained in batches due to its DAG struc- ture. In this paper, we propose a novel word- character LSTM(WC-LSTM) model to add word information into the start or the end char- acter of the word, alleviating the influence of word segmentation errors while obtaining the word boundary information. Four different strategies are explored in our model to en- code word information into a fixed-sized rep- resentation for efficient batch training. Ex- periments on benchmark datasets show that our proposed model outperforms other state- of-the-arts models. 1 Introduction Name Entity Recognition(NER) is a basic task of many NLP systems including Information Retrieval (Virga and Khudanpur, 2003), Rela- tionship Extraction (Miwa and Bansal, 2016), Question Answering (Moll´a et al., 2006). The main task of NER is to identify named enti- ties such as person, location, organization, etc. in given text. Various methods have been pro- posed to tackle this problem, including Hid- den Markov Models(HMMs) (Saito and Nagata, 2003), Maximum Entropy Models(ME) (Chieu and Ng, 2003), Support Vector Machines(SVM) (Ekbal and Bandyopadhyay, 2010) and Condi- tional Random Fields(CRF) (Feng et al., 2006). With the development of deep learning, neural net- c c c c c c � Up § Rise ƒ of 3 Water ^ River ~ Long �§ Rise ^3 River Water ~^ Yangtze River w w w O O O B-LOC M-LOC E-LOC O O O B-LOC E-LOC O (a) Original lattice model c c c c c c � Up § Rise ƒ of 3 Water ^ River ~ Long �§ Rise ^3 River Water ~^ Yangtze River w w w O O O B-LOC M-LOC E-LOC O O O B-LOC E-LOC O (b) Degraded lattice model in extreme cases Figure 1: An example of the lattice model degen- erates into a partial word-based model. Due to the shortcut path ”江(River)” → ”江 水(River Water)” → ”水(Water)”, the model incorrectly predicts that ”江(River)” and ”水(Water)” belong to the same entity. Red labels(without underline) denote predicted labels, and blue labels(with underline) denote gold labels. works (Huang et al., 2015; Lample et al., 2016; Habibi et al., 2017) have been introduced to NER task. To …

Content truncated to 500 words