CAN-NER: Convolutional Attention Network for Chinese Named Entity Recognition

Proceedings of NAACL-HLT 2019, pages 3384–3393 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3384 CAN-NER: Convolutional Attention Network for Chinese Named Entity Recognition Yuying Zhu∗ Nankai University Tianjin, China yuyzhu@mail.nankai.edu.cn Guoxin Wang Microsoft Research Asia Beijing, China guow@microsoft.com Abstract Named entity recognition (NER) is a common task in Natural Language Processing (NLP), but it remains more challenging in Chinese be- cause of its lack of natural delimiters. There- fore, Chinese Word Segmentation (CWS) is usually necessary as the first step for Chi- nese NER. However, models based on word- level embeddings and lexicon features often suffer from segmentation errors and out-of- vocabulary (OOV) problems. In this paper, we investigate a Convolutional Attention Network (CAN) for Chinese NER, which consists of a character-based convolutional neural network (CNN) with local-attention layer and a gated recurrent unit (GRU) with global self-attention layer to capture the information from adjacent characters and sentence contexts. Moreover, differently from other approaches, CAN-NER does not depend on any external resources like lexicons and employing small-size char em- beddings makes CAN-NER more practical for real systems scenarios. Extensive experimen- tal results show that our approach outperforms state-of-the-art methods without word embed- ding and external lexicon resources on differ- ent domains datasets. 1 Introduction Named Entity Recognition (NER) aims at identi- fying text spans which are associated with a spe- cific semantic entity type such as person (PER), organization (ORG), location (LOC), and geopo- litical entity (GPE). NER has received constant research attention as it is the first step in a wide range of downstream Natural Language Pro- cessing (NLP) tasks, e.g., entity linking (Gupta et al., 2017), relation extraction (Miwa and Bansal, 2016), event extraction (Chen et al., 2015), and co- reference resolution (Fragkou, 2017). The stan- dard approach in existing state-of-the-art models ∗ This work was performed when the first author was an intern at Microsoft Research Asia. for English NER treats the problem as a word-by- word sequence labeling task and makes full use of the Recurrent Neural Network (RNN) and Condi- tional Random Field (CRF) to capture context in- formation at the word level (Lample et al., 2016; Ma and Hovy, 2016; Chiu and Nichols, 2016; Liu et al., 2018). These models for English NER pre- Sentence: 南京市长江大桥 Segmentation 1: 南京市 | 长江大桥 Nanjing City, Yangtze River Bridge Location, Location Segmentation 2: 南京 | 市长 | 江大桥 Nanjing, Mayor, Jiang Daqiao Location, Title, Person Figure 1: Entity Ambiguity with Word Segmentation. dict a tag for each word assuming that words can be separated clearly by explicit word separators, e.g., blank spaces. As the Chinese language has no natural delimiters, it would be intuitive to ap- ply Chinese Word Segmentation (CWS) first to get word boundaries and then use a word-level se- quence labeling model similar to the English NER models. However, word boundaries can be am- biguous in Chinese, which leads to the possibility that entity boundaries do not match word bound- aries. For example, the term “西藏自治区 (Ti- bet …

Content truncated to 500 words