Document-level relation extraction with multi-semantic knowledge interaction

Document-level relation extraction has gained increasing attention because of its capability to discover relationship facts between entity pairs within a document. Existing studies only leverage semantic information derived from mentions, entities, and entity pairs, but overlook rich semantics embedded within relation labels that encapsulate implicit semantic knowledge capable of enhancing relation prediction. This paper proposes a multi-semantic interaction method for document-level relation extraction. First, we model relation labels and employ a template-based method to extract and incorporate their semantic features. Next, a relation label self-interaction module is introduced to capture complex semantic associations among relation labels. Then, we propose two distillation strategies with and without distantly supervised datasets. Finally, experimental results on three datasets demonstrate that our method outperforms previous methods in terms of F1 and IgnF1.