ASP-HR: An Adaptive Spatial Perception and Hierarchical Reasoning mechanism for document-level biomedical relation extraction.

Abstract

Document-level biomedical relation extraction is crucial for mining entity relationships from literature, aiding disease diagnosis, drug development, and knowledge graph construction. Existing methods face three key challenges: first, they adopt a static entity representation strategy, ignoring the specificity of entity mentions and distance dependency, resulting in the loss of key semantic information. Second, they insufficiently mine document structural features, causing the attenuation of relation specificity information. It is easy to cause excessive smoothing phenomenon, which reduces the ability of the model to distinguish different entity relationships. Third, they lack a global consistency verification mechanism, resulting in limited robustness in reasoning under noise interference and cross-sentence scenarios. To address these issues, we propose the Adaptive Spatial Perception and Hierarchical Reasoning (ASP-HR) model with a three-level progressive architecture. Initially, the Dynamic Distance-Aware Entity Interaction Enhancement (DDA-EIE) module dynamically weights entity mentions by fusing semantic and sentence distance information, enhancing contextual adaptability. Secondly, the Multi-Relation Graph Attention Reasoning Network (MR-GARN) establishes independent feature propagation channels for different relation types via a dynamic relation-aware mechanism, deeply mining global semantic dependencies between distant entities. Finally, the Dual-Stage Inference Enhancement (DSIE) module corrects initial prediction biases through an "initial inference - enhanced verification" mechanism, combining local and global features to optimize prediction robustness. Evaluated on the CDR dataset, ASP-HR demonstrates superior performance, notably achieving an Overall-F1 of 88.9%, Intra-F1 of 90.7%, and Inter-F1 of 84.3%. It represents an improvement over the current SOTA of 1.8% in Overall-F1, 1.3% in Intra-F1, and 2.6% in Inter-F1. On the GDA dataset, ASP-HR achieves an Overall-F1 of 84.9%, Intra-F1 of 88.6%, and Inter-F1 of 69.2%. On the BioRED dataset, ASP-HR also achieves competitive performance, obtaining an Overall-F1 score of 61.8%, which surpasses the previous state-of-the-art model by 0.6%. The results confirm the efficiency of our adaptive spatial perception and hierarchical reasoning mechanisms for document-level relation extraction.