DA-BioNER is a context-preserving data augmentation framework for biomedical named entity recognition that couples few-shot learning with distant supervision: multiple base NER models trained on only a handful of examples produce coarse annotations, which are then refined by a knowledge-guided large language model within the original sentences -- boosting entity diversity and label quality without collecting new manual annotations.
Named entity recognition is a foundational step for structured knowledge extraction in biomedicine, feeding downstream tasks such as relation extraction, literature mining, and knowledge base construction. Yet its effectiveness in emerging or specialized domains is repeatedly bottlenecked by the same problem: high-quality, domain-specific annotated corpora are expensive and slow to build, requiring expert curators to label entities such as diseases, chemicals, and genes.
To sidestep costly manual annotation, prior work has turned to data augmentation and distant supervision. However, these strategies come with well-known trade-offs. Generation-based augmentation can introduce limited entity diversity or hallucinate implausible text, distant supervision from external resources tends to inject noisy labels, and sentence-rewriting approaches often disrupt the syntactic and semantic integrity of the original context -- all of which limit generalization precisely in the low-resource settings where they are needed most.
Key Challenge: Biomedical NER models must be trained under severe annotation scarcity, but the usual remedies each degrade the training signal in a different way -- too little entity diversity, too much label noise, or broken context. DA-BioNER is designed to enrich the training data and preserve contextual integrity at the same time, so that few-shot models can generalize instead of overfitting to a tiny seed set.
DA-BioNER treats data augmentation not as a text-generation problem but as an annotation-refinement problem over existing, unlabeled biomedical sentences. It combines few-shot learning to bootstrap initial annotators, distant supervision to expand coverage, and a carefully constrained LLM refinement stage to clean up the result.
Manual annotation is the single most expensive ingredient in building biomedical NER systems, and it becomes a hard blocker whenever a new entity type, disease area, or emerging domain appears faster than curators can label it. By recycling abundant unlabeled text into high-quality supervision, DA-BioNER reduces the reliance on hand-labeled corpora and makes it feasible to stand up competent NER models from only a few examples.
Just as importantly, the framework demonstrates a disciplined way to use large language models in a scientific pipeline: as a knowledge-guided refiner operating within real sentences, rather than a free-running generator. This design keeps the augmented data faithful to the source text and mitigates hallucination -- a practical template for trustworthy LLM-assisted data augmentation in high-stakes domains like biomedicine.