HAMGFusion:A Joint Entity and Relation Extraction Model Based on Hierarchical Attention and Feature Fusion

Joint entity and relation extraction remains a fundamental yet challenging task in natural language processing (NLP), often constrained by insufficient information fusion and severe class imbalance, especially under complex semantic structures. To address these issues, we propose HAMGFusion, a unified framework that leverages a Hierarchical Semantic Encoder to capture multi-level contextual dependencies and a Dynamic Feature Fusion Decoder to adaptively integrate fine-grained semantic features across layers. To further alleviate class imbalance, a Class-Aware Loss Fusion mechanism is introduced to dynamically calibrate loss weights and better align predictions with imbalanced label distributions. Extensive experiments on three benchmark datasets—NYT, WebNLG, and ADE—demonstrate that HAMGFusion achieves state-of-the-art F1 scores of 93.7%, 94.8%, and 86.4%, respectively. Ablation studies validate the contribution of each module, and visualizations of fusion dynamics offer insights into the model's adaptive behavior. The proposed method shows strong potential for real-world applications, such as knowledge graph construction and biomedical information extraction, and contributes to advancing learning-based approaches for joint extraction tasks.