Document-Level Relation Extraction with Adaptive Thresholding and Localized Context Pooling

Document-level relation extraction (RE) poses new challenges compared to its sentence-level counterpart. One document commonly contains multiple entity pairs, and one pair occurs times in the associated with possible relations. In this paper, we propose two novel techniques, adaptive thresholding localized context pooling, solve multi-label multi-entity problems. The replaces global threshold for classification prior work a learnable entities-dependent threshold. pooling directly transfers attention from pre-trained language models locate relevant that is useful decide relation. We experiment on three document-level RE benchmark datasets: DocRED, recently released large-scale dataset, datasets CDRand GDA biomedical domain. Our ATLOP (Adaptive Thresholding Localized cOntext Pooling) model achieves an F1 score of 63.4, also significantly outperforms existing both CDR GDA. have our code at https://github.com/wzhouad/ATLOP.