Attention Filtering Mechanism-Based Document-Level Relation Extraction

Document-level Relation Extraction (DocRE) aims to identify all possible relations between entity pairs within a document. Existing methods focus on constructing document graphs, where a document is encoded by a PLM, and node representations in the graph are iteratively updated using a GCN and a multi-head attention mechanism. Context representation of the current entity pair is generated, and relation prediction is subsequently performed by integrating these feature representations. However, due to the presence of multiple sentences within a document, context representation of the current entity pair generated by PLMs encoding inevitably contains noise, which can lead the model to be influenced by irrelevant information during relation prediction. To tackle this issue, we present an Attention Filtering Mechanism-based Document-level Relation Extraction model (AFMDRE). After obtaining the attention matrix from the PLM encoding, a dynamic threshold is applied to remove irrelevant information. The refined attention matrix is then used to construct context representation of the current entity pair, which is integrated with node representations from the document graph for the final relation prediction. Experimental results on the widely-used DocRED and Re-DocRED datasets show that AFMDRE enhances performance in DocRE tasks, validating the effectiveness of our model.