Relation-Specific Attentions over Entity Mentions for Enhanced Document-Level Relation Extraction

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1523 - 1529 July 10-15, 2022 ©2022 Association for Computational Linguistics Relation-Specific Attentions over Entity Mentions for Enhanced Document-Level Relation Extraction Jiaxin Yu† and Deqing Yang† ∗and Shuyu Tian‡ School of Data Science, Fudan University, Shanghai 200433, China †{jiaxinyu20,yangdeqing}@fudan.edu.cn ‡sytian21@m.fudan.edu.cn Abstract Compared with traditional sentence-level re- lation extraction, document-level relation ex- traction is a more challenging task where an entity in a document may be mentioned mul- tiple times and associated with multiple rela- tions. However, most methods of document- level relation extraction do not distinguish be- tween mention-level features and entity-level features, and just apply simple pooling opera- tion for aggregating mention-level features into entity-level features. As a result, the distinct semantics between the different mentions of an entity are overlooked. To address this problem, we propose RSMAN in this paper which per- forms selective attentions over different entity mentions with respect to candidate relations. In this manner, the flexible and relation-specific representations of entities are obtained which indeed benefit relation classification. Our exten- sive experiments upon two benchmark datasets show that our RSMAN can bring significant improvements for some backbone models to achieve state-of-the-art performance, especially when an entity have multiple mentions in the document.1 1 Introduction Relation extraction (RE) is one important task of in- formation extraction, aiming to detect the relations among entities in plain texts. Recently, many schol- ars have paid more attention to document-level RE (Sahu et al., 2019; Yao et al., 2019) which aims to identify the relations of all entity pairs in a doc- ument, since it is more in demand than sentence- level RE in various real scenarios. In general, one document contains multiple entities and an entity may have multiple mentions across different sen- tences. Furthermore, one entity may be involved by multiple valid relations and different relations are ∗Corresponding author. 1Our code and trained model are publicly available at https://github.com/FDUyjx/RSMAN. Figure 1: A t-SNE visualization example from DocRED. Points of the same color and marker are different men- tions’ embeddings of an entity, which are encoded by BERT (Devlin et al., 2019). expressed by different mentions of the same entity. As a result, document-level RE is more challenging than sentence-level RE. A key step of existing document-level RE meth- ods is to aggregate the information of different mentions of an entity (mention-level features) to obtain the entity’s representation (entity-level fea- ture) at first, since relation classification is gener- ally achieved on entity level. To this end, previous RE models simply apply average pooling (Ye et al., 2020; Xu et al., 2021), max pooling (Li et al., 2021), or logsumexp pooling (Zhou et al., 2021; Zhang et al., 2021). Finally, a fixed representation is ob- tained for the given entity, which is then fed into the classifier for relation classification. However, different mentions of an entity in a document may hold distinct semantics. A simple pooling operation of generating a …

Content truncated to 500 words