Towards Better Document-level Relation Extraction via Iterative Inference

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8306 - 8317 December 7-11, 2022 ©2022 Association for Computational Linguistics Towards Better Document-level Relation Extraction via Iterative Inference Liang Zhang1,2∗ Jinsong Su1,2∗ Yidong Chen1,2† Zhongjian Miao1,2 Zijun Min1,2 Qingguo Hu1,2 Xiaodong Shi1,2 1School of Informatics, Xiamen University, China 2Key Laboratory of Digital Protection and Intelligent Processing of Intangible Cultural Heritage of Fujian and Taiwan (Xiamen University), Ministry of Culture and Tourism, China lzhang@stu.xmu.edu.cn {jssu,ydchen}@xmu.edu.cn Abstract Document-level relation extraction (RE) aims to extract the relations between entities from the input document that usually containing many difficultly-predicted entity pairs whose relations can only be predicted through rela- tional inference. Existing methods usually di- rectly predict the relations of all entity pairs of input document in a one-pass manner, ignoring the fact that predictions of some entity pairs heavily depend on the predicted results of other pairs. To deal with this issue, in this paper, we propose a novel document-level RE model with iterative inference. Our model is mainly composed of two modules: 1) a base module expected to provide preliminary relation pre- dictions on entity pairs; 2) an inference mod- ule introduced to refine these preliminary pre- dictions by iteratively dealing with difficultly- predicted entity pairs depending on other pairs in an easy-to-hard manner. Unlike previous methods which only consider feature informa- tion of entity pairs, our inference module is equipped with two Extended Cross Attention units, allowing it to exploit both feature infor- mation and previous predictions of entity pairs during relational inference. Furthermore, we adopt a two-stage strategy to train our model. At the first stage, we only train our base module. During the second stage, we train the whole model, where contrastive learning is introduced to enhance the training of inference module. Experimental results on three commonly-used datasets show that our model consistently out- performs other competitive baselines. Our source code is available at https://github. com/DeepLearnXMU/DocRE-II. 1 Introduction Relation extraction (RE) aims to identify the rela- tion between two entities from raw texts. Due to its wide applications in many subsequent natural ∗Equal Contribution. †Corresponding Author. Figure 1: An example comes from the DocRED dataset. (a) is an input document, where different colors repre- sent different entities. (b) lists some easily-predicted entity pairs whose predictions do not require reference to the predictions of other pairs. (c) shows a difficultly- predicted entity pair, of which prediction depends on the predicted results of its overlapping pairs. The arrows between (b) and (c) indicate the dependencies among entity pairs, which can be exploited to benefit the rela- tion prediction of the difficultly-predicted one. language processing (NLP) tasks, such as large- scale knowledge graph construction (Zeng et al., 2020) and question answering (Yu et al., 2017), RE has attracted increasing attention and become a fundamental NLP task. Most of the existing works focus on sentence-level RE, where both consid- ered entities come from a single sentence (Zhang et al., 2018; Baldini Soares et al., 2019). How- ever, large amounts …

Content truncated to 500 words