Document-Level Relation Extraction Method Based on Attention Semantic Enhancement

Abstract:

Document-level relation extraction aims to extract the relations between multiple entity pairs from a document， a task characterized by high complexity. This paper proposes a method for document-level relation extraction based on attention semantic enhancement to address challenges such as handling multiple entities， capturing relationship correlations， and dealing with imbalanced relationship distributions within documents. The method proposed facilitates the inference of relationships between entity pairs. Specifically， the data encoding module enhances the encoding strategy by incorporating additional entity information， capturing semantic features of the document through the encoding network， and generating an entity pair matrix. Subsequently， a U-Net network employing an attention gating mechanism is devised to capture local information and aggregate global information from entity pair matrices， thereby achieving semantic enhancement. Finally， this paper introduces an adaptive focal loss function to mitigate imbalanced relationship distributions. The Att-DocuNet model proposed is evaluated on four publicly available document-level relation extraction datasets （DocRED， CDR， GDA， and DWIE）， yielding promising experimental results.