Medical concept normalization in French using multilingual terminologies and contextual embeddings

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Concept normalization links medical terms to standardized terminologies like UMLS®.
• Traditional methods struggle with non-English languages due to resource limitations.
• Developing effective concept normalization for diverse languages is crucial for global medical informatics.

Purpose of the Study:

• To present a novel system for concept normalization in French medical texts.
• To leverage multilingual terminologies and embedding models for improved accuracy.
• To achieve effective concept normalization without direct translation or supervision.

Main Methods:

• The system treats concept normalization as a highly-multiclass classification problem.
• Contextualized embeddings encode medical terms, classified using cosine similarity and softmax.
• A two-step training process involves finetuning embeddings and further training with hard negative selection.

Main Results:

• The proposed approach achieves good results on French medical corpora, even with no labeled data.
• Performance surpasses existing supervised methods that utilize labeled data.
• Training with both French and English terms significantly boosts performance on French benchmarks.

Conclusions:

• The developed distantly supervised method is applicable across various medical domains and document types.
• This approach offers a simpler and more effective multilingual strategy for processing non-English medical texts.
• The findings pave the way for broader advancements in multilingual medical text analysis.