About · Research · Publications · CV · News · Contact

Doctoral Research

My doctoral research investigates how variation in clinical language affects the development and evaluation of NLP systems, with a focus on German clinical texts.

Research Questions

• Clinical language variation: How do linguistic, terminological, and stylistic features differ across clinical specialties?
• Synthetic text evaluation: How closely do synthetic clinical texts resemble human-produced texts, and which evaluation methods capture meaningful differences?
• Clinical NLP methods: How does this variation affect named entity recognition, entity linking, and relation extraction, particularly for German texts and SNOMED CT annotation?

Supervision

I'm supervised by Prof. Martin Boeker and mentored by Dr. Diego Frassinelli. Luise Modersohn and Dr. Jacqueline Lammert advise me on my project.

Selected Research Outputs

GeMTeX: Legal Compliance and Semantic Enrichment

Problem: German clinical NLP requires shareable resources based on real clinical documents.

Contribution: Our LREC 2026 paper describes the GeMTeX corpus, its approach to European regulatory requirements, and semantic annotation using SNOMED CT.

My role: I lead the semantic annotation work package, develop annotation guidelines, coordinate consortium annotation activities, and supervise the local TUM annotation team.

Resources: Corpus overview · Paper · Poster · Public annotated example · Annotation agreement code

GerMedIQ: Simulated and Synthetic Medical Interviews

Problem: Limited German medical interview resources make it difficult to evaluate synthetic interview responses.

Contribution: GerMedIQ provides human-produced simulated responses to standardized medical history questions and responses generated by 18 LLMs.

Main findings: Structural and semantic evaluations revealed differences between human-produced and synthetic responses. In the study's acceptability evaluation, human raters preferred Mistral (124B) responses to the human-produced simulated responses. These results concern the models, questions, and evaluation setup used in the study.

My role: First author of the ACL 2025 Student Research Workshop paper.

Resources: Study overview · Paper · Poster · Data and code · Dataset