HILGEN introduces a novel approach that leverages both structured medical knowledge and LLMs to generate synthetic training data for biomedical named entity recognition tasks.

• Utilizes UMLS hierarchical structure to expand training with related medical concepts
• Employs GPT-3.5 to generate contextually-rich examples for rare medical entities
• Demonstrates significant performance improvements on multiple biomedical datasets
• Addresses the critical challenge of data sparsity in specialized medical domains

This research offers a practical solution for healthcare AI systems that need to accurately identify medical entities in clinical text, potentially improving clinical decision support, research, and medical information extraction.

HILGEN: Hierarchically-Informed Data Generation for Biomedical NER Using Knowledgebases and Large Language Models