Comprehensive Study on German Language Models for Clinical and Biomedical Text Understanding

TL;DR

This study addresses domain adaptation for German biomedical NLP by comparing continuous pre-training on private German clinical data and translation-based Germanization of English medical text. It pre-trains multiple German biomedical/clinical language representations on 2.4B translated tokens and 3B clinical tokens, then evaluates on NER, multi-label classification, and extractive QA. The results show domain-specific pre-training typically yields gains over general-domain baselines, with translation-based approaches often matching or approaching the performance of data from private clinical sources, highlighting a privacy-friendly path for large-scale domain adaptation. The findings support a practical strategy for deploying German medical NLP systems, balancing data availability, privacy, and task-specific demands, and they provide publicly released translation-based resources to the research community.

Abstract

Recent advances in natural language processing (NLP) can be largely attributed to the advent of pre-trained language models such as BERT and RoBERTa. While these models demonstrate remarkable performance on general datasets, they can struggle in specialized domains such as medicine, where unique domain-specific terminologies, domain-specific abbreviations, and varying document structures are common. This paper explores strategies for adapting these models to domain-specific requirements, primarily through continuous pre-training on domain-specific data. We pre-trained several German medical language models on 2.4B tokens derived from translated public English medical data and 3B tokens of German clinical data. The resulting models were evaluated on various German downstream tasks, including named entity recognition (NER), multi-label classification, and extractive question answering. Our results suggest that models augmented by clinical and translation-based pre-training typically outperform general domain models in medical contexts. We conclude that continuous pre-training has demonstrated the ability to match or even exceed the performance of clinical models trained from scratch. Furthermore, pre-training on clinical data or leveraging translated texts have proven to be reliable methods for domain adaptation in medical NLP tasks.

Figures (1)

• Figure 1: Workflow for Continuous Pre-training of different Publicly Available General Domain Models with Private and Public Datasets: The private dataset comes from a major German hospital and yields 25M documents. The public dataset, sourced from English PubMed abstracts and MIMIC-III clinical notes, is preprocessed, tokenized into sentence segments, and translated into German via the Fairseq WMT'19 English to German model, resulting in approximately 45M documents. Both model setups are subject to model fine-tuning and evaluation across various biomedical and clinical German language downstream tasks.