UniRel: Unified Representation and Interaction for Joint Relational Triple Extraction

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7087 - 7099 December 7-11, 2022 ©2022 Association for Computational Linguistics UniRel: Unified Representation and Interaction for Joint Relational Triple Extraction Wei Tang 1,2, Benfeng Xu 1,2, Yuyue Zhao 1,2, Zhendong Mao 1,2, Yifeng Liu 3 Yong Liao 1,2 ∗, Haiyong Xie 1,2 1 University of Science and Technology of China, Anhui, China 2 CCCD Key Lab of Ministry of Culture and Tourism, Anhui, China 3 National Engineering Laboratory for Risk Perception and Prevention (RPP), Beijing, China {weitang, benfeng, yyzha0}@mail.ustc.edu.cn, zdmao@ustc.edu.cn, liuyifeng3@cetc.com.cn {yliao, hxie}@ustc.edu.cn Abstract Relational triple extraction is challenging for its difficulty in capturing rich correlations be- tween entities and relations. Existing works suffer from 1) heterogeneous representations of entities and relations, and 2) heteroge- neous modeling of entity-entity interactions and entity-relation interactions. Therefore, the rich correlations are not fully exploited by existing works. In this paper, we propose UniRel to address these challenges. Specif- ically, we unify the representations of enti- ties and relations by jointly encoding them within a concatenated natural language se- quence, and unify the modeling of interactions with a proposed Interaction Map, which is built upon the off-the-shelf self-attention mech- anism within any Transformer block. With comprehensive experiments on two popular re- lational triple extraction datasets, we demon- strate that UniRel is more effective and compu- tationally efficient. The source code is available at https://github.com/wtangdev/UniRel. 1 Introduction Relational Triple Extraction (RTE) aims to iden- tify entities and their semantic relations jointly. It extracts structured triples in the form of <subject- relation-object> from raw texts in an end-to-end manner, and is a crucial task towards automatically constructing large-scale knowledge bases (Nayak et al., 2021). Early works try to solve RTE tasks in a pipelined fashion involving two sub-tasks (Zelenko et al., 2003; Chan and Roth, 2011), where entities are firstly recognized, and relations are then assigned for each extracted entity pair. Therefore, such meth- ods fail to capture the implicit correlation between these two isolated sub-tasks and are thus prone to propagated errors (Li and Ji, 2014). Instead, many researchers (Miwa and Bansal, 2016; Zheng et al., 2017; Zeng et al., 2018) seek ∗ Corresponding author. ··· Cities ··· is capital of ··· Countries ··· Semantic information ··· Cities ··· is capital of ··· Countries ··· London is capital of UK London is capital of UK ··· Cities ··· is capital of ··· Countries ··· Corpus Interaction Discrimination Entity-Entity Entity-Relation Entity-Relation Figure 1: We leverage semantic information to unify the representation of entities and relations. Relational triples are extracted by modeling the entity-entity inter- action (blue dashed line) and entity-relation interaction (red solid line) in a unified way. to jointly extract <s-r-o> triples in an end-to-end manner. For example, Wei et al. (2020) propose a cascaded network that identifies subjects first and then recognizes corresponding objects for relations. Zheng et al. (2021a) decompose RTE into three sub- tasks. Wang et al. (2020) and Shang et al. (2022) extract relational triples in one …

Content truncated to 500 words