BeLink: Generative Re-Ranking for Biomedical Entity Linking
BeLink is a modular end-to-end system for Biomedical Entity Linking (BEL) that integrates a generative re-ranker. The system uses instruction-tuned open-source large language models at the re-ranking stage to improve efficiency and accuracy. A set-wise instruction-tuning formulation enables fast candidate selection. On multiple BEL benchmarks, BeLink achieves 3% to 24% improvement in linking accuracy while reducing inference time compared to state-of-the-art methods. The work addresses the computational inefficiency of LLM-based BEL in practical settings.
Key facts
- BeLink is a modular end-to-end system for Biomedical Entity Linking.
- It uses instruction-tuned open-source generative models for re-ranking.
- Set-wise instruction-tuning enables fast candidate selection.
- Achieves 3%-24% improvement in linking accuracy.
- Reduces inference time compared to state-of-the-art.
- Addresses computational inefficiency of LLM-based BEL.
- Designed for practical real-world BEL applications.
- Published on arXiv under Computer Science > Computation and Language.
Entities
Institutions
- arXiv