MedCOD: Enhancing English-to-Spanish Medical Translation of Large Language Models Using Enriched Chain-of-Dictionary...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Accurate medical translation is crucial for global healthcare access.
• Large Language Models (LLMs) show promise but struggle with domain-specific medical terminology.
• Existing translation methods lack robust integration of structured medical knowledge.

Purpose of the Study:

• To introduce MedCOD (Medical Chain-of-Dictionary), a novel hybrid framework for improved English-to-Spanish medical translation.
• To enhance LLMs by integrating structured domain knowledge from UMLS and LLM-KB paradigms.
• To evaluate the effectiveness of MedCOD in improving translation quality across multiple open-source LLMs.

Main Methods:

• Constructed a parallel corpus of 2,999 English-Spanish MedlinePlus articles.
• Developed a 100-sentence test set with structured medical contexts.
• Employed structured prompting with multilingual variants, synonyms, and UMLS definitions.
• Utilized LoRA-based fine-tuning on four open-source LLMs (Phi-4, Qwen2.5-14B, Qwen2.5-7B, LLaMA-3.1-8B).

Main Results:

• MedCOD significantly improved translation quality across all evaluated LLMs.
• Phi-4 with MedCOD and fine-tuning achieved superior BLEU (44.23), chrF++ (28.91), and COMET (0.863) scores.
• Both MedCOD prompting and model adaptation independently enhanced performance, with combined use yielding maximal gains.
• Performance surpassed strong baseline models like GPT-4o and GPT-4o-mini.

Conclusions:

• Structured knowledge integration via MedCOD substantially enhances LLM performance for medical translation.
• The MedCOD framework offers a viable strategy for improving the accuracy and reliability of medical language translation.
• This approach holds significant potential for advancing AI applications in global health communication.