Addressing Span Imbalance and Semantic Complexity in Nested Medical Named Entity Recognition

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

As a fundamental task in biomedical natural language processing, Medical Named Entity Recognition (MNER) aims to identify and classify medical entities from unstructured medical texts. A major challenge in this task is the prevalence of nested entities, which arise from the syntactic complexity and domain-specific characteristics of medical language. Recently, span-based models have been proposed to handle nested entities by reformulating the task as a multi-label classification problem over all possible spans. However, when directly applied to medical texts, these models encounter two problems. First, existing span-based methods typically rely on general-domain pre-trained language models to learn textual semantics, which hinders their ability to capture contextual dependencies in complex medical texts, thereby limiting their generalization across diverse medical data. Second, these methods treat all candidate spans equally during training, which results in an underestimation of the gradient contributions from entity spans due to their relatively small proportion, thereby degrading recognition performance. To address these issues, we propose AGPNer, a novel method for recognizing nested medical named entities. The proposed AGPNer integrates two key components: (1) a heterogeneous dependency fusion encoder, which reconstructs masked entities to enhance token representations and fine-tunes a hybrid dependency modeling block to learn domain-specific patterns in medical texts; (2) an imbalance-adaptive span decoder, which decouples entity and non-entity spans and adaptively assigns them different exponential decay factors to regulate their contributions during training. Experimental results on four public benchmarks demonstrate that AGPNer achieves absolute improvements in F1-score of 0.90, 1.02, 0.47 and 0.75 percentage points on CMeEE-V1, CMeEE-V2, GENIA, and CLUENER, respectively, showing consistent and competitive performance among the compared baselines.