NDGR: A Noise Divide and Guided Re-labeling Framework for Distantly Supervised Relation Extraction

Distant supervision (DS) is widely used in relation extraction to reduce the cost of annotation but suffers from noisy instances. Current approaches typically involve selecting reliable instances from the DS-built dataset for model training. However, these approaches often lead to the inclusion of numerous noisy instances or the disregard of a substantial number of valuable instances. In this paper, we propose NDGR, a novel training framework for sentence-level distantly supervised relation extraction. Initially, NDGR partitions the noisy data from the DS-built dataset by employing a Gaussian Mixture Model (GMM) to model the loss distribution. Afterwards, we utilize a guided label generation strategy to generate high-quality pseudo-labels for noisy data. By iteratively executing the processes of noise division and guided label generation, NDGR helps refine the noisy DS-built dataset and enhance the overall performance. Our method has been extensively evaluated on commonly used benchmarks, and the results demonstrate its substantial improvements in both sentence-level evaluation and noise reduction.