BioHiCL: Hierarchical Multi-Label Contrastive Learning for Biomedical Retrieval with MeSH Labels

A team of researchers has introduced BioHiCL (Biomedical Retrieval with Hierarchical Multi-Label Contrastive Learning), a novel AI framework designed to significantly improve biomedical information retrieval. Developed by Mengfei Lan, Lecheng Zheng, and Halil Kilicoglu, the system addresses a key limitation in existing biomedical generative retrievers: their reliance on coarse binary relevance signals, which often fail to capture nuanced semantic overlap between texts. BioHiCL instead leverages the structured, hierarchical annotations provided by MeSH (Medical Subject Headings) to train more semantically aware models through multi-label contrastive learning.

The researchers released two model variants: BioHiCL-Base with 0.1 billion parameters and BioHiCL-Large with 0.3 billion parameters. These compact yet powerful models demonstrate promising performance across several critical tasks, including biomedical document retrieval, sentence similarity assessment, and question answering. A key advantage is their computational efficiency, making them suitable for practical deployment where resource constraints are a concern. The paper detailing this work has been accepted for presentation at the ACL 2026 Main Conference, highlighting its significance in the natural language processing and information retrieval communities.

By moving beyond simple binary relevance, BioHiCL's approach allows AI systems to understand the complex, multi-faceted relationships inherent in biomedical literature. This hierarchical understanding is crucial for applications like literature review, clinical decision support, and drug discovery, where precision and context are paramount. The use of MeSH labels provides a rich, standardized source of supervision that guides the model to learn more accurate representations of biomedical concepts and their interconnections.