Information extraction across different genres of medical text can yield valuable new insights, e.g., tracking the time it takes for new research results to be translated into clinical practice. Although similarities in medical terminology are shared across various types of documents, like scientific publications, clinical guidelines, and text from healthcare routine, there is a critical difference: the primary research literature is mostly published in English, whereas practitioners and patients communicate in many other languages. Therefore, medical language technology needs to accommodate a variety of languages to be useful for analyzing text documents across genre boundaries. Unfortunately, most languages other than English lack essential language resources for such cross-lingual investigations. As annotated text corpora and terminological resources have been predominantly curated in English, most software tools and machine learning models are optimized for English texts as well. Therefore, this thesis aims to address the lack of language resources in multilingual biomedical natural language processing through multiple contributions. First, I describe the curation and linguistic annotation of a novel language resource: GGPONC, a large, semantically annotated corpus of German oncology guidelines. Building upon gold-standard annotations in GGPONC, machine learning models for nested named entity recognition and resolution of elliptical compounds in German medical language are built and evaluated in detail. Second, I introduce xMEN, a modular software system for cross-lingual medical entity normalization. Since annotated training data for this task are often unavailable, xMEN combines unsupervised cross-lingual candidate generation and weakly supervised re-ranking, enabling high-performance entity normalization, even in low-resource scenarios. When training data is available, fully supervised re-rankers obtain new state-of-the-art performance across a wide range of multilingual benchmark datasets. Moreover, I present a novel approach for improving candidate generation performance for complex entity mentions through the integration of generative language models into a generate-and-rank entity normalization framework. Building upon these advances, I describe an application of cross-lingual text analysis for evidence-based medicine. Extracted metadata from multiple sources is integrated into a comprehensive database of medical evidence, which is used to estimate time lags in translation of new findings from clinical trials into clinical guidelines. To support this translation process, I show that the same database can be used to identify signal publications, i.e., clinical trials with the potential to trigger a guideline update. Overall, the thesis demonstrates the utility of clinical guidelines as a language resource, the value of semantically interoperable metadata across medical text genres in different languages, as well as the potentials for applying natural language processing to support evidence-based decision-making in healthcare.
Florian Borchert (Thu,) studied this question.