Document-level joint entity and relation extraction faces persistent challenges, including error propagation across sub-tasks, limited relational reasoning over long documents, and inadequate modeling of heterogeneous semantic edges. To address these issues, we propose a table-filling based multi-task explicit interaction framework. Mentions are uniformly represented as graph nodes, where coreference and relation graphs are dynamically constructed, and mention-level syntactic dependencies are incorporated as structural priors. To further explore potential correlations between sub-tasks, we design an alignment graph between coreference resolution and relation extraction, explicitly modeling structural dependencies to enable bidirectional information flow and reinforce semantic complementarity. Building on this design, an enhanced relation graph emphasizes high-confidence relational patterns to improve entity representations. In addition, a relation-guided gated graph attention mechanism is proposed to differentiate semantic edge types and adaptively regulate cross-layer message passing, improving robustness to noise and long-tail relations. Hard negative sampling and uncertainty-based weighting are employed to stabilize multi-task optimization. Experiments on two standard document-level benchmarks demonstrate that our approach achieves state-of-the-art performance, validating its effectiveness in relational reasoning and long-document scenarios.
Du et al. (Thu,) studied this question.