Exploring Zero-Shot Cross-Lingual Biomedical Concept Normalization via Large Language Models.

Area of Science:

• Natural Language Processing
• Biomedical Informatics
• Machine Learning

Background:

• Large Language Models (LLMs) are key in NLP, with discriminative and generative approaches dominating.
• A gap exists in comparing these LLM types for cross-lingual biomedical concept normalization.

Purpose of the Study:

• To conduct a comparative analysis of discriminative and generative LLMs for cross-lingual biomedical concept normalization.
• To evaluate LLM generalization across 10 languages using the XL-BEL dataset without adaptation.

Main Methods:

• Utilized dense retrieval for cross-lingual biomedical concept normalization.
• Evaluated multiple LLMs on the XL-BEL dataset, covering 10 languages.
• Assessed model performance across diverse linguistic contexts.

Main Results:

• The discriminative model e5 demonstrated superior performance in cross-lingual biomedical concept normalization.
• BioMistral achieved the highest performance among the generative LLMs.
• Performance was evaluated across 10 languages to assess generalization capabilities.

Conclusions:

• Discriminative LLMs, specifically e5, show strong potential for cross-lingual biomedical concept normalization.
• Generative LLMs like BioMistral also show promise, though e5 currently leads.
• Further research can build on these findings for improved cross-lingual biomedical NLP tools.