PRGC: Potential Relation and Global Correspondence Based Joint Relational Triple Extraction

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 6225–6235 August 1–6, 2021. ©2021 Association for Computational Linguistics 6225 PRGC: Potential Relation and Global Correspondence Based Joint Relational Triple Extraction Hengyi Zheng1,2,3, Rui Wen3, Xi Chen3,4∗, Yifan Yang3, Yunyan Zhang3 Ziheng Zhang3, Ningyu Zhang5, Bin Qin2∗, Ming Xu2∗, Yefeng Zheng3 1College of Electronics and Information Engineering, Shenzhen University 2Information Technology Center, Shenzhen University 3Tencent Jarvis Lab, Shenzhen, China 4Platform and Content Group, Tencent 5Zhejiang University zhenghengyi2019@email.szu.edu.cn, {qinbin,xuming}@szu.edu.cn {ruiwen,jasonxchen,tobyfyang,yunyanzhang,zihengzhang,yefengzheng}@tencent.com Abstract Joint extraction of entities and relations from unstructured texts is a crucial task in infor- mation extraction. Recent methods achieve considerable performance but still suffer from some inherent limitations, such as redundancy of relation prediction, poor generalization of span-based extraction and inefficiency. In this paper, we decompose this task into three sub- tasks, Relation Judgement, Entity Extraction and Subject-object Alignment from a novel per- spective and then propose a joint relational triple extraction framework based on Potential Relation and Global Correspondence (PRGC). Specifically, we design a component to pre- dict potential relations, which constrains the following entity extraction to the predicted re- lation subset rather than all relations; then a relation-specific sequence tagging compo- nent is applied to handle the overlapping prob- lem between subjects and objects; finally, a global correspondence component is designed to align the subject and object into a triple with low-complexity. Extensive experiments show that PRGC achieves state-of-the-art per- formance on public benchmarks with higher efficiency and delivers consistent performance gain on complex scenarios of overlapping triples.1 1 Introduction Identifying entity mentions and their relations which are in the form of a triple (subject, rela- tion, object) from unstructured texts is an impor- tant task in information extraction. Some previous works proposed to address the task with pipelined approaches which include two steps: named en- tity recognition (Tjong Kim Sang and De Meulder, 2003; Ratinov and Roth, 2009) and relation predic- tion (Zelenko et al., 2002; Bunescu and Mooney, *Corresponding author. 1The source code and data are released at https://github.com/hy-struggle/PRGC. Subtask Model Component Relation Judgement CasRel None (Take all relations) TPLinker None (Take all relations) PRGC Potential Relation Prediction Entity Extraction CasRel Span-based TPLinker Span-based PRGC Rel-Spec Sequence Tagging Subject-object Alignment CasRel Cascade Scheme TPLinker Token-pair Matrix PRGC Global Correspondence Table 1: Comparison of the proposed PRGC and previ- ous methods in the respect of our new perspective with three subtasks. 2005; Pawar et al., 2017; Wang et al., 2020b). Re- cent end-to-end methods, which are based on either multi-task learning (Wei et al., 2020) or single- stage framework (Wang et al., 2020a), achieved promising performance and proved their effective- ness, but lacked in-depth study of the task. To better comprehend the task and advance the state of the art, we propose a novel perspective to decompose the task into three subtasks: i) Rela- tion Judgement which aims to identify relations in a sentence, ii) Entity Extraction which aims to extract all subjects …

Content truncated to 500 words