German Medical Named Entity Recognition Model and Data Set Creation Using Machine Translation and Word Alignment:...

Area of Science:

• Natural Language Processing (NLP)
• Medical Informatics
• Computational Linguistics

Background:

• Medical data analysis often relies on unstructured text, necessitating robust NLP tools.
• A scarcity of German medical training data limits the development of open-source named entity recognition (NER) models.
• Legal restrictions on proprietary data hinder the publication of medical datasets and models.

Purpose of the Study:

• To develop and publicly release a novel German medical NER model and a synthetic dataset.
• To demonstrate the feasibility of creating medical NLP resources using exclusively public data.
• To bypass legal restrictions by avoiding internal, proprietary datasets.

Main Methods:

• A synthetic German dataset was generated via translation and word alignment of an English dataset.
• A simple, low-computational-requirement network architecture was employed for the NER model.
• The model was trained and evaluated on the synthetic dataset, with performance assessed on an external dataset.

Main Results:

• The generated dataset comprised 8,599 sentences with 30,233 annotations across 7 NER types.
• The NER model achieved a class frequency-averaged F1 score of 0.82 on the test set.
• Artifacts in the synthetic data were identified, and model performance on external data was discussed in comparison to a baseline.

Conclusions:

• The study successfully demonstrated a method for creating a German medical NER dataset and model using only public data.
• The approach addresses the challenge of data scarcity and legal restrictions in medical NLP.
• Limitations were discussed, with proposed future work to mitigate remaining issues.