Cross- & multi-lingual medication detection: a transformer-based analysis

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Medical Informatics

Background:

• Extracting medication mentions from unstructured medical texts is difficult without language-specific annotated data.
• Multilingual machine learning models offer a solution for cross-lingual knowledge transfer.
• Leveraging existing English resources is key when target languages lack training data.

Purpose of the Study:

• To investigate the effectiveness of a multilingual transformer model for drug name extraction.
• To evaluate multilingual and cross-lingual performance in German, English, French, and Spanish.
• To provide empirical evidence on the benefits of cross-lingual knowledge transfer in medical text analysis.

Main Methods:

• Fine-tuning a multilingual transformer model for named entity recognition (NER).
• Applying the model in both multilingual and cross-lingual settings.
• Evaluating performance on published datasets across four European languages.
• Conducting a qualitative error analysis to identify sources of prediction errors.

Main Results:

• The multilingual transformer model achieved competitive performance in drug name extraction across all tested languages.
• Cross-lingual transfer of medical knowledge proved effective for German, English, French, and Spanish.
• Annotation inconsistencies and vague entity labels were identified as sources of errors.

Conclusions:

• Multilingual transformer models are a viable approach for cross-lingual drug name extraction in the medical domain.
• Effective transfer of medical NLP capabilities is possible between European languages.
• Improving annotation quality and guidelines is crucial for enhancing model accuracy.