German Medical NER with BERT and LLMs: The Impact of Training Data Size

Area of Science:

• Computational linguistics
• Medical informatics

Background:

• Medical Named Entity Recognition (NER) is challenging due to complex terminology.
• Low-resource settings lack sufficient annotated data for training NER models.
• Evaluating advanced language models in specialized domains is crucial.

Purpose of the Study:

• To assess the performance of a large language model (LLaMA3.1) and a BERT-based model for German medical NER.
• To investigate the impact of varying training data sizes on model performance.
• To simulate and evaluate performance under lower-resource conditions.

Main Methods:

• Utilized two German annotated corpora for Named Entity Recognition tasks.
• Trained and evaluated LLaMA3.1 and a BERT-based model.
• Systematically varied the amount of training data to simulate low-resource scenarios.

Main Results:

• Both models demonstrated comparable performance on the German medical NER datasets.
• LLaMA3.1 exhibited a slight performance advantage, particularly when trained with smaller datasets.
• The findings indicate the effectiveness of these models even with limited annotated medical text.

Conclusions:

• LLaMA3.1 and BERT-based models are viable for German medical NER.
• LLaMA3.1 shows promise for low-resource medical NER tasks due to its efficiency with less data.
• Further research can explore model optimization for specialized, data-scarce domains.