BioELX: Cross-lingual Biomedical Entity Linking via Alias-based Retrieval and LLM Ranking

Researchers introduce BioELX, a two-stage cross-lingual biomedical entity linking system that maps medical mentions across languages to knowledge base identifiers without requiring task-specific training data. The framework combines multilingual alias-enriched retrieval with LLM-based ranking, achieving state-of-the-art results across five benchmarks with substantial improvements for low-resource languages.

BioELX addresses a critical bottleneck in biomedical NLP: the scarcity of expert-annotated training data for entity linking in non-English languages. Traditional cross-lingual biomedical systems rely heavily on English-centric knowledge bases and SapBERT retrievers trained predominantly on English aliases, creating a generalization gap that degrades performance for underrepresented languages. This research tackles the problem through two complementary innovations that eliminate dependency on costly supervised training.

The framework's first stage enriches the retriever by incorporating Wikidata-derived multilingual aliases into SapBERT training, directly addressing the language bias that hampers existing systems. The second stage leverages pre-trained large language models as context-aware rankers, enabling sophisticated disambiguation without task-specific annotation. This design choice reflects a broader industry trend toward zero-shot and few-shot learning paradigms that reduce human annotation overhead.

The experimental results demonstrate substantial practical impact, particularly for underserved languages—Thai shows a 30.8-point improvement in Recall@1, while Turkish and Korean each gain over 21 points. Results on established benchmarks like XL-BEL, EMEA, and Patent confirm the approach generalizes across different biomedical domains. For healthcare NLP applications serving global populations, improved cross-lingual entity linking directly enhances clinical decision support systems, drug discovery pipelines, and multilingual medical literature analysis.

The release of code and resources will accelerate adoption in biomedical AI development, particularly benefiting low-resource language communities historically underserved by NLP technology. Future work should focus on scaling these methods to emerging biomedical knowledge bases and evaluating performance on extremely low-resource language pairs.

• →BioELX eliminates need for task-specific annotated training data by combining multilingual aliases with LLM-based ranking
• →Average Recall@1 improvement of 19.2 on XL-BEL benchmark with exceptional gains for low-resource languages (21.6-30.8 points)
• →Framework achieves state-of-the-art performance across five different biomedical entity linking benchmarks
• →Wikidata-derived multilingual aliases directly address language bias endemic to English-centric biomedical knowledge bases
• →Open-source release enables rapid adoption in clinical NLP and biomedical research applications globally