Selective and contrastive mechanism for distantly supervised relation extraction

Traditional relation extraction methods typically rely on large-scale annotated datasets. A classic solution to this dependency is the use of distant supervision learning to automatically acquire data. However, this approach inevitably introduces noisy data, which adversely affects model performance. To address this challenge effectively, a novel distantly supervised relation extraction model is proposed in this paper, integrating selective and contrastive mechanisms. Initially, semantic associations between entities are enhanced by incorporating knowledge graphs, thereby improving the model's ability to understand relations. Subsequently, text-augmented information and knowledge graph data are employed as initial positive samples for contrastive learning. Furthermore, a dynamic selection strategy is developed, which filters high-quality feature information using dynamic thresholds to serve as positive samples in contrastive learning. By leveraging these high-quality positive instances along with context-aware enhanced negative samples, richer semantic signals can be provided for sparse data scenarios during contrastive learning, thereby alleviating the adverse effects of data imbalance on relation extraction performance. Experimental results on the NYT10 and GDS datasets demonstrate that the proposed method achieves significant improvements in both accuracy and generalization capability.