EmRel: Joint Representation of Entities and Embedded Relations for Multi-triple Extraction

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 659 - 665 July 10-15, 2022 ©2022 Association for Computational Linguistics EmRel: Joint Representation of Entities and Embedded Relations for Multi-triple Extraction Benfeng Xu1∗, Quan Wang2, Yajuan Lyu3, Yabing Shi3 Yong Zhu3, Jie Gao1 and Zhendong Mao1† 1University of Science and Technology of China, Hefei, China 2Beijing University of Posts and Telecommunications, Beijing, China 3Baidu Inc., Beijing, China benfeng@mail.ustc.edu.cn, zdmao@ustc.edu.cn Abstract Multi-triple extraction is a challenging task due to the existence of informative inter-triple cor- relations, and consequently rich interactions across the constituent entities and relations. While existing works only explore entity rep- resentations, we propose to explicitly intro- duce relation representation, jointly represent it with entities, and novelly align them to identify valid triples. We perform comprehensive exper- iments1 on document-level relation extraction and joint entity and relation extraction along with ablations to demonstrate the advantage of the proposed method. 1 Introduction Relation extraction aims at discovering struc- tured knowledge in the form of <subject-relation- object> triples from plain text. It is an essential task towards constructing knowledge bases. Al- though a lot of efforts have been made in build- ing advanced relation extraction systems, it is still a challenging problem under certain practi- cal scenarios where multiple entities and relations are involved, e.g., document-level relation extrac- tion (Yao et al., 2019) and joint entity and rela- tion extraction (Riedel et al., 2010; Gardent et al., 2017). Existing works mostly take the entity perspec- tive that focuses on exploring cross-entity interac- tions (Xu et al., 2021; Zeng et al., 2020). They either treat relations as atomic labels specified in a final classifier (Xu et al., 2021; Zeng et al., 2020; Wang et al., 2020), or simply search subjects and objects for each individual relation(Wei et al., 2020; Zheng et al., 2021). However, as an essential com- ponent, relations also interact with entities and con- text, which jointly exhibit informative inter-triple Inc.∗Work done while the first author was an intern at Baidu †Corresponding author. 1The code will be available at https://github.com/ BenfengXu/emrel Figure 1: Different formulations for multi-triple ex- traction. 1) entity perspective constructs only entity representation and feed them into a relation-specific classifier. 2) joint triple perspective constructs both en- tity representation and relation representation to model comprehensive correlations across all components. correlations. e.g., the two relations capital of and located at often co-occur between the same pair of entities but with different probabilities conditioned on specific contextual clues. As a consequence, the capability to model and make full use of rich interactions across relations, entities, and context is crucial for the task. In this paper, we advocate a novel joint triple perspective for relation extraction (see Figure 1 for illustration). Different from previous works that only seek to represent entities, we propose EmRel that creates, refines and leverages the Embedded representations of Relations. Specifically, we first explicitly create relation representations as embed- ded vectors; then refine these relation …

Content truncated to 500 words