BOAT: Boundary and Order Aware Transformer for Biomedical Named Entity Recognition
OBJECTIVES: Biomedical named entity recognition (BioNER) is essential for clinical information extraction and knowledge discovery. This study aims to address the persistent challenges of acronym ambiguity and incomplete extraction of long-span entities, which often impair model accuracy in real-world biomedical informatics applications. METHODS: We propose the Boundary and Order Aware Transformer (BOAT), a novel model specifically optimized for biomedical text. The methodology incorporates two key methodological designs: (1) optimizing the encoder by adopting relative-position self-attention and simplifying the position-wise feed-forward network, thereby enhancing contextual order modeling and reducing redundant parameters, and (2) introducing an auxiliary task of Span Entity Classification (SEC) based on pointer annotation. This multi-task learning framework enables the model to mine fine-grained signals from entity boundaries without requiring additional labeled data. RESULTS: Extensive evaluations were conducted on four benchmark biomedical datasets: BC2GM, BC5CDR-chem, BC5CDR-disease, and NCBI-disease. BOAT achieved F1 scores of 85.00%, 93.68%, 85.86%, and 89.62%, respectively, ranking first on BC5CDR-chem and NCBI-disease and second on BC2GM and BC5CDR-disease among the compared models. These results demonstrate the competitive performance of BOAT across different types of biomedical entities. CONCLUSIONS: The BOAT model effectively enhances entity boundary perception and language order awareness. By improving the precision of biomedical terminology extraction, this approach provides a robust methodological foundation for downstream tasks such as clinical decision support and medical knowledge graph construction. METHODS: We propose the Boundary and Order Aware Transformer (BOAT), a novel model specifically optimized for biomedical text. The methodology incorporates two key innovations: (1) integrating relative position encoding into the self-attention mechanism to better capture inherent language order, and (2) introducing an auxiliary task of Span Entity Classification (SEC) based on pointer annotation. This multi-task learning framework enables the model to mine fine-grained signals from entity boundaries without requiring additional labeled data. RESULTS: Extensive evaluations were conducted on four benchmark biomedical datasets: BC2GM, BC5CDR-chem, BC5CDR-disease, and NCBI-disease. BOAT demonstrated state-of-the-art performance, achieving F1-scores of 85.00%, 93.68%, 85.86%, and 89.62%, respectively. Notably, our model significantly outperformed established baselines in identifying complex, multi-word entities and disambiguating morphologically similar acronyms. CONCLUSIONS: The BOAT model effectively enhances entity boundary perception and language order awareness. By improving the precision of biomedical terminology extraction, this approach provides a robust methodological foundation for downstream tasks such as clinical decision support and medical knowledge graph construction.
Authors
- Chao Che (ORCID: https://orcid.org/0000-0003-2978-5430)
- Chengjie Zhou (ORCID: https://orcid.org/0000-0002-9275-6096)
- Xiaoming Ren (ORCID: https://orcid.org/0009-0005-6019-7700)
- Zhuoya Liu
Institutions
- Dalian University of Foreign Languages (CN)
- Dalian University (CN)
Publication Details
- Journal
- Methods of Information in Medicine
- Published
- 2026-11-16
- DOI
- https://doi.org/10.1055/a-2990-3399
- Primary Topic
- Biomedical Text Mining and Ontologies
- Type
- article
- Field-Weighted Citation Impact
- 0.00