Key Mention Pairs Guided Document-Level Relation Extraction

Proceedings of the 29th International Conference on Computational Linguistics, pages 1904–1914 October 12–17, 2022. 1904 Key Mention Pairs Guided Document-Level Relation Extraction Feng Jiang1, Jianwei Niu1∗, Shasha Mo2∗, Shengda Fan2 1 State Key Laboratory of Virtual Reality Technology and Systems, School of Computer Science and Engineering, Beihang University, Beijing 100191, China 2 School of Cyber Science and Technology, Beihang University, Beijing 100191, China {jfeng, niujianwei, moshasha, fanshengda}@buaa.edu.cn Abstract Document-level Relation Extraction (DocRE) aims at extracting relations between entities in a given document. Since different mention pairs may express different relations or even no relation, it is crucial to identify key mention pairs responsible for the entity-level relation la- bels. However, most recent studies treat differ- ent mentions equally while predicting the rela- tions between entities, leading to sub-optimal performance. To this end, we propose a novel DocRE model called Key Mention pairs Guided Relation Extractor (KMGRE) to di- rectly model mention-level relations, contain- ing two modules: a mention-level relation ex- tractor and a key instance classifier. These two modules could be iteratively optimized with an EM-based algorithm to enhance each other. We also propose a new method to solve the multi-label problem in optimizing the mention- level relation extractor. Experimental results on two public DocRE datasets demonstrate that the proposed model is effective and out- performs previous state-of-the-art models. 1 Introduction Relation Extraction (RE), which aims to identify the relations between entities in a given text, has been explored at the sentence level for decades (Culotta and Sorensen, 2004; Zeng et al., 2014, 2015). However, according to Yao et al. (2019), a large amount of relations can only be identified across multiple sentences in the real-world scenar- ios. Therefore, researchers have recently turned to extracting relations directly in documents (Zeng et al., 2020; Zhou et al., 2021; Huang et al., 2021; Ru et al., 2021). Document-level Relation Extraction (DocRE) encounters many new challenges compared to its sentence-level counterpart. A document may in- clude numerous entities, and the same entity may appear multiple times in different sentences. It ∗Corresponding author. requires the DocRE models to recognize and fo- cus on the part of the document that has relevant context for a particular entity pair. Many previ- ous works solve the above problems by obtaining stronger context-aware entity pair representations. There are two main ways to achieve this: the graph- based methods (Guo et al., 2019; Nan et al., 2020; Zeng et al., 2020) and the sequence-based methods (Yao et al., 2019; Zhou et al., 2021). The graph- based methods construct a document graph and then use Graph Neural Networks (GNNs) to aggre- gate information across nodes. Besides, as Trans- former (Vaswani et al., 2017) could be regarded as a fully connected GNN, the sequence-based meth- ods attempt to directly use Transformer-based Pre- trained Language Models (PLMs) for DocRE with- out graph structure. The sequence-based methods generally use strong PLMs (e.g., BERT (Devlin et al., 2019) and RoBERTa (Liu et al., 2019)) to model the input text and use different strategies to …

Content truncated to 500 words