MECT: Multi-Metadata Embedding based Cross-Transformer for Chinese Named Entity Recognition

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 1529–1539 August 1–6, 2021. ©2021 Association for Computational Linguistics 1529 MECT: Multi-Metadata Embedding based Cross-Transformer for Chinese Named Entity Recognition Shuang Wu1, Xiaoning Song1∗, Zhenhua Feng2,3∗ 1School of Artificial Intelligence and Computer Science, Jiangnan University, China 2Department of Computer Science, University of Surrey, UK 3Centre for Vision, Speech and Signal Processing, University of Surrey, UK shuangwu@stu.jiangnan.edu.cn x.song@jiangnan.edu.cn, z.feng@surrey.ac.uk Abstract Recently, word enhancement has become very popular for Chinese Named Entity Recogni- tion (NER), reducing segmentation errors and increasing the semantic and boundary informa- tion of Chinese words. However, these meth- ods tend to ignore the information of the Chi- nese character structure after integrating the lexical information. Chinese characters have evolved from pictographs since ancient times, and their structure often reflects more informa- tion about the characters. This paper presents a novel Multi-metadata Embedding based Cross- Transformer (MECT) to improve the perfor- mance of Chinese NER by fusing the structural information of Chinese characters. Specifi- cally, we use multi-metadata embedding in a two-stream Transformer to integrate Chinese character features with the radical-level em- bedding. With the structural characteristics of Chinese characters, MECT can better cap- ture the semantic information of Chinese char- acters for NER. The experimental results ob- tained on several well-known benchmarking datasets demonstrate the merits and superior- ity of the proposed MECT method.1 1 Introduction Named Entity Recognition (NER) plays an essen- tial role in structuring of unstructured text. is a sequence tagging task that extracts namedIt entities from unstructured text. Common cate- gories of NER include names of people, places, organizations, time, quantity, currency, and some proper nouns. NER is the basis for many Nat- ural Language Processing (NLP) tasks such as event extraction (Chen et al., 2015), question an- swering (Diefenbach et al., 2018), information re- ∗Corresponding author. 1The source code of the proposed method is publicly available at https://github.com/CoderMusou/ MECT4CNER. Character CR HT SC 题 (topic) 页 是页 日一走页 榆 (elm) 木 木俞 木人一月刂 渡 (ferry) 氵 氵度 氵广廿又 脸 (face) 月 月佥 月人一ツ一 Table 1: Structure decomposition of Chinese charac- ters: ‘CR’ denotes the Chinese radical, ‘HT’ denotes the head and tail, and ‘SC’ denotes the structural com- ponents of Chinese characters. trieval (Khalid et al., 2008), knowledge graph con- struction (Riedel et al., 2013), etc. Compared with English, there is no space be- tween Chinese characters as word delimiters. Chi- nese word segmentation is mostly distinguished by readers through the semantic information of sentences, posing many difficulties to Chinese NER (Duan and Zheng, 2011; Ma et al., 2020). Besides, the task also has many other challenges, such as complex combinations, entity nesting, and indefinite length (Dong et al., 2016). In English, different words may have the same root or affix that better represents the word’s seman- tics. For example, physiology, psychology, sociol- ogy, technology and zoology contain the same suf- fix, ‘-logy’, which helps identify the entity of a …

Content truncated to 500 words