Anaphor-Aware Document-level Entity and Relation Joint Extraction with Curriculum Learning

Document-level entity relation joint extraction aims at identifying semantic relationships between different entities within data, while assuming the entities are uncalibrated.Existing methods typically focus on enhancing the interaction between Coreference resolution (COREF) and Relation Extraction (RE) without adequately leveraging the anaphors present in the text which can provide clues for COREF and RE.Another challenge in this task is that the determination of the relationship between two entities requires logical reasoning through other entities.To alleviate these above challenges, we propose Anaphor-aware Entity and Relation Joint Extractor (AERJE), in which explicit anaphoric information is utilized in the two-stage model.We also adopt curriculum learning approach during the training process to bridge the gap between the model and external tool.Additionally, to enhance the model's logical reasoning ability at the mention level, we introduce a mention-level pairwise aggregation mechanism to allow the model to concentrate on information specific to mentions and anaphors pairs.Extensive experiments on the DocRED and Re-DocRED datasets demonstrate AERJE outperforms many former state-of-the-art methods.Our code is available now at https://github.com/PurRigiN/AERJE.