Document-Level Relation Extraction with Relational Reasoning and Heterogeneous Graph Neural Networks

Document-level relation extraction aims to identify the relations between the entities in an unstructured text and represents them in a structured way for downstream tasks such as knowledge graphs and question answering. In recent years, graph neural network-based methods have made significant progress in relation extraction. However, these methods usually require extracting all the entities in the document first, then a classifier is used to analyze the relations between the entities regardless of whether they have any relation. This wastes a lot of time analyzing the relations of irrelevant entity pairs and reduces the classifier’s attention to relevant entity pairs. To address this issue, this paper proposes a relation extraction module that integrates Relational Reasoning and Heterogeneous Graph neural Networks (RRHGN). The method finds a meta-path for each entity pair in a document and uses multi-hop reasoning to analyze the entities on the meta-path to determine whether there is a strong reasoning path between the entity pair. The relational reasoning module built into the method makes the classifier focus more on the relevant entity pairs in the document, thus reducing the task burden of the classifier and improving the accuracy of entity relation extraction. Experimental results on the large-scale document-level relation extraction dataset DocRED show that the proposed method achieves significant performance improvement compared with existing methods.