Medical named entity recognition via lattice-enhanced transfer learning with random attention

• A novel Lattice-based Transfer Learning (LRA) framework is proposed for Chinese medical named entity recognition (NER). • The model integrates Inter-Attention and Random Attention (RA) sequentially to enhance semantic representation and generalization. • RA introduces stochastic sparsity in attention, mitigating source-domain bias and improving cross-domain robustness. • LRA achieves F1 scores of 90.55%, 80.72%, and 83.71% on TCM-Methods, Tianchi2020, and CCKS2019 datasets, respectively. • The proposed method effectively balances character- and word-level features, outperforming existing NER models without large-scale pretraining. Medical texts are characterized by high acquisition costs, specialized terminology, imbalanced distribution, and limited annotated resources, all hindering the performance of named entity recognition (NER) tasks. Transfer learning(TL) offers a practical solution by leveraging models pre-trained on data-rich domains to improve efficiency and accuracy in low-resource scenarios. This study proposes a medical NER model that integrates a Lattice structure with a Random Attention (RA)–based domain-adaptive framework. The Lattice structure encodes both characters and potential words from the vocabulary, thereby enhancing contextual representation and incorporating prior knowledge. Inter-Attention processes contextual fragments and differentiates word variations across entity types to capture long-range semantic dependencies. RA is introduced to strengthen generalization and improve entity feature integration under diverse text structures in medical domains. Finally, a word-aware representation is fed into a conditional random field (CRF) layer for sequence labeling. Experimental results demonstrate that our model achieves F1 scores of 90.55%, 80.72%, and 83.71% on TCM-Methods, Tianchi2020, and CCKS2019, respectively, confirming its effectiveness in medical information recognition.