projects

2022–2024 — COLEX: Automatic Analysis and Summarization of German Legal Correspondence

Legal professionals routinely navigate dense webs of interconnected documents: statutes reference other statutes, court rulings cite precedent, and a single case file can span hundreds of pages. COLEX tackled this by building computational tools for German legal text that respect this structure rather than treating documents as flat text.

The project developed a graph-based retrieval augmentation system that leverages citation relationships in German legal documents to provide richer context for downstream analysis. Alongside this, we created a multi-perspective summarization corpus for court rulings that distinguishes between the perspectives of the court (tenor), the plaintiff, and the defendant — reflecting how legal professionals actually read these texts. Both components were integrated into an end-to-end prototype with a web frontend, built in cooperation with DATEV and funded by the BMBF via Software Campus.

2023–2025 — Misinformation Analysis for EMF and 5G Communication

Public discourse around electromagnetic fields and 5G has been persistently shaped by misinformation, but the structure and dynamics of these narratives across platforms were poorly understood. This project, funded by the German Federal Office for Radiation Protection (BfS), set out to map that landscape computationally.

We collected and analyzed approximately 400,000 social media posts across multiple platforms, characterizing how misinformation narratives form, spread, and evolve around EMF topics. The resulting analysis fed into an interdisciplinary effort combining communication science and risk perception research, ultimately producing a recommendation action plan for evidence-based public communication by the BfS. The work also contributed to a forthcoming book chapter in the Research Handbook on Public Communication (Edward Elgar).

2021–2023 — HUILE: Human-in-the-Loop Techniques for Enterprise Use Cases

Enterprise NLP systems need to be reliable enough for business-critical deployment, but the domains they operate in shift continuously making static models insufficient. HUILE explored how human feedback can be systematically integrated into NLP pipelines to keep these systems aligned with evolving requirements, investigating techniques for human-model collaboration and feedback-driven learning. The project was an industry collaboration with Siemens AG.