NN-Denoising: A Low-Noise Distantly Supervised Document-Level Relation Extraction Scheme Using Natural Language Inference and Negative Sampling

The task of document-level relation extraction (DocRE) is crucial in the field of natural language processing, as it aims to extract semantic relations between entities in a given document to facilitate deeper comprehension. Previous methods have primarily focused on fully supervised learning for DocRE, which requires a large amount of human-annotated training data, making it a tedious and laborious task. Recently, more and more attention has been paid to the incomplete labeling problem in human-annotated data, and it is believed to be the bottleneck of model performance. To address this limitation and mitigate annotation costs, we propose a low-noise distant supervision scheme for DocRE, called NN-Denoising, that combines natural language inference (NLI) models and negative sampling to filter out noise in the training data. The NLI model serves as a pre-filter for denoising the distant supervision (DS) labels, while negative sampling is employed to overcome the false negative problem in the filtered data. Our experimental results on a large-scale DocRE benchmark demonstrate the superiority of the proposed approach over existing baselines in distant supervision learning. Specifically, NN-Denoising achieves an improvement of 15.83 F1 points and 10.34 F1 points compared to the ATLOP and SSR-PU models, respectively.