EDREL: Document-Level Relation Extraction with Evidence and Logical Rules

Document-level relation extraction (DocRE) is the task of extracting all relations between entities in a whole document. Previous research has demonstrated that sentence-level evidence information can enhance the accuracy of relation extraction. Existing joint relation and evidence extraction methods mostly rely on neural networks to learn the semantic information of local contexts, thereby capturing the interdependencies between entity pairs. However, these methods lack the ability to reason about the potential structural relations between entities, resulting in deficiencies in both relation and evidence prediction. To address this problem, we propose a Evidence-enhanced Document-level Relation Extraction with Logical rules method (EDREL), integrating semantic modeling with structural rule-based reasoning. By incorporating a logical rule mechanism as a supplement to semantic representation, this method simultaneously improves the structural modeling and reasoning capabilities for both relation and evidence extraction. We further design a simple yet effective structured evidence scoring method. Based on relation prediction, this method calculates the evidence confidence of sentences in a reverse manner through structural rules of co-occurrence and bridge entity paths. Experiments show that our method outperforms existing DocRE methods on both DocRED and Re-DocRED datasets.