MEDNER.DE: Medicinal Product Entity Recognition in German-Specific Contexts

Named entity recognition (NER) is an area of significant interest in the field of medicine, particularly for identifying specific terms such as medicinal products, medical events and diagnoses from doctors’ letters. However, training NER models in languages other than English presents challenges, including limited corpora and annotated training data, and existing models for extracting medicinal products often perform poorly when applied to real-life pharmacovigilance-related use cases. This study investigates the effectiveness of combining multiple techniques to achieve better performance compared to baseline models for medicinal product entity recognition in German-specific contexts. We propose a strategy that utilizes automatically annotated data that was generated using a rule-based approach, leveraging transfer learning along with further pre-training on language models (LMs) resulting in a set of models we named MedNER.de. Our models are designed for German-specific contexts and show a 3.3% improvement in F1 score compared to the best baseline models, GERNERMED++ and GPTNERMED. The best-performing MedNER.de model is publicly available at: https://huggingface.co/pei-germany/MEDNER-de-fp-gbert