Attention and Edge-Label Guided Graph Convolutional Networks for Named Entity Recognition

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6499 - 6510 December 7-11, 2022 ©2022 Association for Computational Linguistics Attention and Edge-Label Guided Graph Convolutional Networks for Named Entity Recognition Renjie Zhou1, Zhongyi Xie1, Jian Wan2, Jilin Zhang1, Yong Liao3 and Qiang Liu4 1Hangzhou Dianzi University, 2Zhejiang University of Science and Technology, 3Univerisity of Science and Technology of China, 4Zhejiang Police College {rjzhou, zyxie}@hdu.edu.cn Abstract It has been shown that named entity recognition (NER) could benefit from incorporating the long-distance structured information captured by dependency trees. However, dependency trees built by tools usually have a certain per- centage of errors. Under such circumstances, how to better use relevant structured informa- tion while ignoring irrelevant or wrong struc- tured information from the dependency trees to improve NER performance is still a challenging research problem. In this paper, we propose the Attention and Edge-Label guided Graph Con- volution Network (AELGCN) model. Then, we integrate it into BiLSTM-CRF to form BiLSTM-AELGCN-CRF model. We design an edge-aware node joint update module and introduce a node-aware edge update module to explore hidden structured information entirely and solve the wrong dependency label informa- tion to some extent. After two modules, we ap- ply attention-guided GCN, which automatically learns how to attend to the relevant structured information selectively. We conduct extensive experiments on several standard datasets across four languages and achieve better results than previous approaches. Through experimental analysis, it is found that our proposed model can better exploit the structured information on the dependency tree to improve the recognition of long entities. 1 Introduction Named Entity Recognition (NER) is the recogni- tion of entities with specific meanings in the text, mainly including person, organization, location, etc. NER is the fundamental tasks for many natural language processing tasks such as relation extrac- tion (Miwa and Bansal, 2016), event extraction (Chen et al., 2015; Liu et al., 2019), coreference resolution (Lee et al., 2017), question answer (Yao and Van Durme, 2014), and knowledge graph (Li et al., 2022). Previous studies have obtained useful structured information from dependency trees and Figure 1: Sentences annotated with dependency tree and named entities. have verified the effectiveness of integrating syn- tactic dependency into NER tasks (Jie et al., 2017; Jie and Lu, 2019; Aguilar and Solorio, 2019; Xu et al., 2021). A dependency tree reveals the syntactic structure of a language unit by analyzing the dependency relationships between its components, and "depen- dency" refers to the relationship between related words while specifying a dependency relationship to form a syntactic tree reflecting the syntactic rela- tionships between words in a sentence. However, how to better use relevant structured information while ignoring irrelevant or wrong structured infor- mation for NER remains a research question to be answered. For example, Figure 1 (a) shows the dependency tree and named entities of the sentence "Old Li Jingtang still tells visitors old war stories circulat- ing in the Taihong Mountain area". For the PER (person) entity "Li Jingtang" where there is an …

Content truncated to 500 words