Global-Attn-GateNER: unified entity recognition based on global attention and dynamic gated fusion
TL;DR
This article proposes a three-stage enhancement strategy: a global self-attention mechanism to enhance contextual modeling and boundary semantic representation; a dropout reorganization strategy to mitigate over-regularization during prediction; a dynamic gated fusion network to achieve adaptive aggregation of structured features and contextual representations at the word-pair granularity.
Abstract
Unified modeling poses significant challenges for named entity recognition (NER), where effectively modeling contextual semantics and achieving efficient feature fusion remain critical challenges. The Word-Word Relation Classification for Named Entity Recognition (W 2 NER) framework, based on word-word relationship classification, offers a unified NER solution through two-dimensional word-grid modeling. However, it still faces limitations in modeling long-range dependencies and achieving flexible feature interactions. To address these limitations, this article proposes a three-stage enhancement strategy: (1) a global self-attention mechanism to enhance contextual modeling and boundary semantic representation; (2) a dropout reorganization strategy to mitigate over-regularization during prediction; (3) a dynamic gated fusion network to achieve adaptive aggregation of structured features and contextual representations at the word-pair granularity. Experiments on eight benchmark datasets covering Chinese and English flat, nested, and discontinuous NER tasks show that the proposed model achieves competitive overall performance and yields consistent gains over the reproduced W 2 NER baseline on multiple benchmarks. Further analyses indicate that the model is more effective for longer entities while introducing only mild additional inference overhead.