Performance Analysis of Dynamic Entity Representation and Dual-Perspective Query Mechanisms for Document-Level Relation Extraction

Document-level relation extraction must identify relations when entity mentions and supporting evidence are distributed across multiple sentences. Existing methods often use entity representations that remain fixed after encoding and provide limited support for progressively retrieving and combining dispersed evidence. This study improves complex cross-sentence relation reasoning by allowing entity representations to evolve during inference and by retrieving evidence from both entities in a target pair. The proposed multilayer iterative framework combines the Dynamic Entity Representation Mechanism (DERM) and Dual-Perspective Query Mechanism (DPQM). DERM propagates and updates entity states across reasoning layers using document context. DPQM retrieves evidence separately from head- and tail-entity perspectives and adaptively fuses the resulting representations. The framework was evaluated on Re-DocRED using BERT-base with Adaptive Thresholding and Localized cOntext Pooling (ATLOP) and DocuNet as backbones. Evaluations included overall F1, Ign F1, inter-sentence F1, evidence-based recall, ablation studies, and computational cost analysis. With ATLOP, the proposed method achieved 76.42% F1 and 75.22% Ign F1, exceeding the baseline by 2.72 and 2.46 percentage points, respectively. Inter-sentence F1 increased by 3.74 percentage points. For relations supported by three or more evidence sentences, recall improved from 54.65% to 66.40%. In ablation results, DERM and DPQM provided complementary improvements, with DERM making the larger contribution in complex cases. Explicit entity-state propagation and evidence retrieval from both entity perspectives improved dispersed evidence integration in long documents. The largest gains occurred in cross-sentence relations and instances requiring multiple supporting evidence sentences; however, stronger evidence constraints and calibration are required to reduce unsupported predictions.