Document-Level Relation Extraction Based on Distant Supervision for Spring Water Domain

Document-Level Relation Extraction (DocRE) is essential for information extraction and knowledge graph construction. However, relation extraction in the spring water domain faces challenges such as data scarcity, noisy distant supervision, and difficulty in modeling cross-sentence relations. To address these issues, we propose a DocRE method integrating evidence sentence supervision and dynamic data denoising. We construct a distant supervision dataset, apply uncertainty estimation to filter noisy pseudo-labels, and enhance long-distance dependency modeling with evidence sentence supervision. Experiments on the spring water dataset show that our method outperforms baselines by 1.14% in IgnF1 and 0.99% in F1, demonstrating its effectiveness. Ablation studies further validate the impact of our proposed modules. Our approach mitigates noise and improves relation extraction accuracy, making it well-suited for low-resource domains.