DE-Lemma: A Maximum-Entropy Based Lemmatizer for German Medical Text

Area of Science:

• Computational linguistics
• Natural Language Processing (NLP)
• Biomedical Informatics

Background:

• Processing German text, especially in the biomedical domain, requires accurate lemmatization of inflected words.
• Medical language, with Latin and Greek roots, poses challenges for standard lemmatization and stemming techniques.
• Existing methods may yield inaccurate results for specialized German medical texts like discharge letters and electronic health records (EHR).

Purpose of the Study:

• To develop and evaluate a machine learning-based lemmatization approach for German biomedical text.
• To train Apache OpenNLP lemmatizer models using publicly available German treebanks.
• To assess the performance of these models on real-world medical data.

Main Methods:

• Utilized Apache OpenNLP to train lemmatizer models.
• Employed publicly available German treebanks for model training.
• Evaluated four "DE-Lemma" models against a dataset of biomedical nouns from discharge letters.

Main Results:

• The developed "DE-Lemma" models were tested on a sample of biomedical nouns.
• The most effective DE-Lemma model demonstrated a high accuracy of 88.0%.
• The best model achieved an F1 score of .936, indicating robust performance.

Conclusions:

• Machine learning offers a viable solution for German biomedical text lemmatization.
• The trained DE-Lemma models show significant promise for improving NLP tasks in the medical field.
• Accurate lemmatization is crucial for effective processing of electronic medical and health records (EMR, EHR).