Document-Level Relation Extraction via Pair-Aware and Entity-Enhanced Representation Learning

Proceedings of the 29th International Conference on Computational Linguistics, pages 2418–2428 October 12–17, 2022. 2418 Document-Level Relation Extraction via Pair-Aware and Entity-Enhanced Representation Learning Xiusheng Huang1,2, Hang Yang1,2, Yubo Chen1,2, Jun Zhao1,2, Kang Liu1,2,4, Weijian Sun3 and Zuyu Zhao3 1School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China 2National Laboratory of Pattern Recognition, Institute of Automation, CAS, Beijing, China 3Huawei Technologies Co., Ltd, Shenzhen, China 4Beijing Academy of Artificial Intelligence, Beijing, China huangxiusheng2020@ia.ac.cn, {hang.yang,yubo.chen,jzhao,kliu}@nlpr.ia.ac.cn, {sunweijian,zhaozuyu1}@huawei.com Abstract Document-level relation extraction aims to rec- ognize relations among multiple entity pairs from a whole piece of article. Recent meth- ods achieve considerable performance but still suffer from two challenges: a) the relational entity pairs are sparse, b) the representation of entity pairs is insufficient. In this paper, we pro- pose Pair-Aware and Entity-Enhanced(PAEE) model to solve the aforementioned two chal- lenges. For the first challenge, we design a Pair- Aware Representation module to predict poten- tial relational entity pairs, which constrains the relation extraction to the predicted entity pairs subset rather than all pairs; For the second, we introduce a Entity-Enhanced Representa- tion module to assemble directional entity pairs and obtain a holistic understanding of the en- tire document. Experimental results show that our approach can obtain state-of-the-art perfor- mance on four benchmark datasets DocRED, DWIE, CDR and GDA. 1 Introduction Relation extraction (RE) is a primary task in the field of information extraction, which aims to iden- tify the relationships between two entities in a doc- ument. Previous works mainly focus on sentence- level relation extraction, i.e, recognizing the rela- tionships between entities in a sentence. However, large amounts of relationships are expressed over multiple sentences in real-world applications. Ac- cording to DocRED (Yao et al., 2019), above 40.7% of the relational facts can only be extracted from multiple sentences. Therefore, it requires the model to capture complex interactions among entities in the whole document. Previous work commonly referred to this problem as document-level relation extraction which has attracted much attention re- cently (Nan et al., 2020; Zhou et al., 2021; Zhang et al., 2021). Although the considerable perfor- mance of these methods, there are still two critical challenges in document-level RE to be addressed. Fig. 1: An example with entity pairs and relations from DocRED. Entity mentions only involved in these rela- tion instances are colored, other entities in the document are highlighted in grey. The first challenge is how to identify relational entity pairs that are sparse in a document. Specifi- cally, given a document with n entities, there will be n(n − 1) combinations of entities to classify. However, only a few entity pairs have predefined relationships. For example, as shown in Figure 1, this document contains 21 entities with 420 poten- tial entity pairs. However, the number of relational entity pairs is only 11, accounting for 2.62 % of the total entity pairs. According to statistics, for DocRED (Yao et al., 2019) dataset, the proportion are 3.18% and 3.11% in the train set and dev …

Content truncated to 500 words