Query-based Instance Discrimination Network for Relational Triple Extraction

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7677 - 7690 December 7-11, 2022 ©2022 Association for Computational Linguistics Query-based Instance Discrimination Network for Relational Triple Extraction Zeqi Tan1, Yongliang Shen1, Xuming Hu2, Wenqi Zhang1, Xiaoxia Cheng1, Weiming Lu1∗, Yueting Zhuang1 1Zhejiang University, 2Tsinghua University 1{zqtan,syl,luwm,yzhuang}@zju.edu.cn 2{hxm19}@mails.tsinghua.edu.cn Abstract Joint entity and relation extraction has been a core task in the field of information extraction. Recent approaches usually consider the extrac- tion of relational triples from a stereoscopic perspective, either learning a relation-specific tagger or separate classifiers for each relation type. However, they still suffer from error prop- agation, relation redundancy and lack of high- level connections between triples. To address these issues, we propose a novel query-based approach to construct instance-level represen- tations for relational triples. By metric-based comparison between query embeddings and to- ken embeddings, we can extract all types of triples in one step, thus eliminating the error propagation problem. In addition, we learn the instance-level representation of relational triples via contrastive learning. In this way, re- lational triples can not only enclose rich class- level semantics but also access to high-order global connections. Experimental results show that our proposed method achieves the state of the art on five widely used benchmarks. 1 Introduction Extracting structured information from open do- main texts is a long-standing study topic in NLP. Joint entity and relation extraction aims to mine high-quality relational triples (subject, relation, ob- ject) from unstructured texts. For example, in the sentence “900 people cross the border between Malaysia and Singapore”, the subject entity 900 people and the object entity border have a physical location relation PHYS. The current entity and relation extraction meth- ods can be divided into two categories: pipeline methods (Chan and Roth, 2011; Lin et al., 2016) and joint methods (Miwa and Bansal, 2016; Kati- yar and Cardie, 2017). Despite the flexibility of pipeline methods, they face the error propagation and the lack of interaction problems (Katiyar and ∗ Corresponding author (b) Ours (a) Relation-specific Tagger Subject Object B-SUB B-OBJ 𝑯𝑯𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄 + 𝑹𝑹𝒇𝒇𝒇𝒇𝒇𝒇𝒄𝒄𝒄𝒄𝒇𝒇𝒄𝒄𝒇𝒇 Locate Founder Business 𝑯𝑯𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄𝒄 + Figure 1: (a) Hcontext represents sentence representa- tion, Rpotential represents potential relation type em- beddings. PRGC first filters out unlikely relations in a sentence, and then learns the relation-specific tag- ger for potential relations. (b) The gray circle below represents the initial state of query embeddings which are type-agnostic, and the star above represents rela- tion type embeddings. We can break the limitation of type-independence and learn the high-order connections between different types via contrastive learning. Cardie, 2017; Wang and Lu, 2020). Therefore, many strategies have been proposed to unify the two tasks. Zheng et al. (2017) extends tagging schema to tag both entities and relations. However, it cannot handle the triple overlapping problem. To tackle this problem, the generation-based methods (Zeng et al., 2018, 2020; Nayak and Ng, 2020; Ye et al., 2021), the table-filling methods (Wang et al., 2020b; Wang and Lu, 2020; Wang et al., 2021) and the …

Content truncated to 500 words