Generalistic or Specific Embeddings, Which is Better? An Empirical Study on Search for Clinical Coding in Non-English Languages

In three linesComparative study of generic vs domain-specific embeddings for multilingual clinical search (ICD-10-CM). A bi-encoder fine-tuned on Gemini-generated synthetic data (6 languages) outperforms BioBERT-ST: R@5=0.822 vs 0.790, with major gains in Portuguese (+0.115). Open recipe for LLM-based medical retrievers.

Summary generated by Claude — human-verified