Back

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

Rouhizadeh, H.; Yazdani, A.; Zhang, B.; Teodoro, D.

2025-02-27 health informatics
10.1101/2025.02.27.25323007 medRxiv
Show abstract

Over the past few years, discriminative and generative large language models (LLMs) have emerged as the predominant approaches in natural language processing. However, despite significant advancements, there remains a gap in comparing the performance of discriminative and generative LLMs in cross-lingual biomedical concept normalization. In this paper, we perform a comparative study across several LLMs on the challenging task of cross-lingual biomedical concept normalization via dense retrieval. We utilize the XL-BEL dataset covering 10 languages to evaluate the models capacity to generalize across various linguistic contexts without further adaptation. The experimental findings demonstrate that e5, a discriminative model, exhibited superior performance, whereas BioMistral emerged as the top-performing generative LLM. The code for reproducing the experiments is available at: https://github.com/hrouhizadeh/zsh_cl_bcn.

Matching journals

The top 3 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.