Multi-Encoder with Entity-Aware Embedding Framework for Distantly Supervised Relation Extraction

Distantly supervised relation extraction (DSRE) inherently faces challenges arising from labeling noise due to the distant supervision assumption. Most existing studies employ Piece-wise CNN (PCNN) to extract semantic features from sentences, often neglecting to comprehensively capture contextual features. As a consequence, valuable information is lost, leading to a decline in relation prediction performance. In this paper, we propose a Multi-Encoder with Entity-Aware Embedding Framework for Distantly Supervised Relation Extraction (MEEA), designed to enhance the prediction of entity relations by effectively capturing comprehensive contextual features. Specifically, MEEA employs a novel entity-aware word embedding method that employs an attention fusion mechanism to integrate relative position information and information of two entities, which can emphasize the importance of entity pair in RE. Meanwhile, we adopt a multi-encoder framework that utilizes PCNN and two distinct sentence encoders to extract diverse features. Subsequently, these features are effectively fused through an attention fusion mechanism to capture global contextual dependencies comprehensively. Experiments demonstrate that our proposed MEEA framework exhibits significant enhancements over previous methods when applied to NYT10, GDS, and Wiki-20m datasets.