Research

The aim is that measurements produced by algorithms meet the standards of validity and reliability expected of established instruments, so that organisations can base decisions on them.

Employee strain, customer trust, risk perception, the strength of evidence behind a causal claim or the state of a crisis are rarely measured directly. They are inferred from text, often by large language models (LLMs), and the validity of such inferences is seldom established. My research develops data science methods that turn these inferences into measurements: grounded in theory, validated, explicit about uncertainty, and suitable for downstream analysis and decision-making.

Methodologically, I combine measurement and validity theory with representation learning and explicit knowledge representations, and I evaluate the resulting methods in business, health care and public administration. The agenda has three connected lines.

1 · LLM-based measurement of latent constructs

Psychometrics provides a mature theory of latent constructs, but its instruments depend on costly surveys. With Kai R. Larsen (University of Colorado Boulder) and colleagues, I co-developed ALIGNS, a fine-tuned language model that represents survey indicators semantically. Latent structure is recovered from the embeddings through principal component analysis with oblique rotation. ALIGNS reconstructs nomological networks across more than 550,000 indicators from psychology, medicine and the social sciences and makes construct, content and nomological validity assessable without new data collection (prototype: nomologicalnetwork.org).

As part of ALIGNS, a PROMIS-fine-tuned model decomposed 17,180 behavioural medicine indicators and validated NIH PROMIS scales (Advances in Patient-Reported Outcomes, 2025). Related work includes the ITEM Ontology on the semantic anatomy of indicators (Information Systems Research, 2026) and a successful NIH proposal with Columbia University and CU Boulder on AI-assisted ontology development. Results of this line are incorporated in PsyProxy.

Next questions

• Reliability of LLMs as measurement instruments: decomposing score variance into facets such as model, prompt and sampling with generalisability theory, and calibrating LLM ratings against established scales.
• Hierarchical construct structure: inferring construct hierarchies with formal concept analysis and distributional-inclusion methods.
• Valid downstream inference: keeping regressions and causal analyses valid when LLM-derived scores replace survey measures, by combining measurement-error models with approaches such as prediction-powered inference.

2 · Knowledge and causal-claim extraction

My group developed the Information Systems Ontology, introduced at DESRIST 2022 with 2,725 entities and 383,101 synonyms on 11 hierarchy levels, and an ontology-grounded pipeline that annotates research articles with theories, methods, topics and technologies (ICIS 2024, Best Design Science Paper, second runner-up). In a 2024 comparison with ChatGPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro, the pipeline extracted on average 3.1 times as many domain terms at an error rate of 0.2%. Two prototypes put this into practice: the semantic search engine SemanticIS, which indexes more than 57,000 IS articles and outperformed an established metasearch engine in re-identifying the studies of a published literature review (International Journal of Information Management, 2025), and the Knowledge Synthesis Dashboard, which helps identify research gaps and trends.

Fine-tuned language models help keep the ontology up to date, identifying synonyms of existing entities with 98% accuracy (HICSS 2025). This builds on earlier work that extracts causal claims from text (HICSS 2018, 2019) and from graphical research models (Image2Theory, WI 2023), and on Theory-Data Maps that link causal theories to empirical evidence (HICSS 2015, 2016).

Next step: combining LLMs with ontological constraints so that each extracted element carries its provenance and a calibrated confidence. This enables computational evidence synthesis, such as living evidence maps and semi-automated meta-analysis, with outputs that can be audited.

3 · Validity and uncertainty of data-science claims

With Kai R. Larsen and an international team, I co-developed Validity in Design Science (MIS Quarterly, 2025). Its conference precursor received the Herbert A. Simon Award at DESRIST 2020. The framework specifies which validity types support which claims about artefacts, including machine-learning systems. Fruitfulness in Design Science (DESRIST 2026) adds a complementary criterion: contributions are evaluated not only by how well they solve a given problem, but by what they make possible for future design. Complementary work models uncertainty explicitly: through fuzzy-logic methods for integrating inconsistent estimates, including the solution of fuzzy equation systems, and through a patented Bayesian-network mechanism for risk-adaptive authentication developed with Bundesdruckerei.

Next steps

• Uncertainty-aware explanations that distinguish data, model and inference uncertainty and communicate them so that users can calibrate their trust in model outputs.
• Uncertainty propagation through interdependent systems. In KASKADE, LLMs translate the emergency plans of hospitals, utilities and authorities into composable decision objects that carry meaning and uncertainty. AgentSpace Berlin applies the same principles to controllable agentic AI.

Research programme

I am building a research group on Computational Measurement and Valid Inference. It investigates three questions: how language models can serve as reliable measurement instruments, how explicit knowledge representations can improve inference, and how validity and uncertainty can be built into AI-supported decisions. These questions connect to machine learning, social media analytics, operations research, business process management, empirical IS research and digital government, and beyond Information Systems to psychology and to medicine through patient-reported outcomes.

• Publications: leading IS journals (MIS Quarterly, Information Systems Research), broadening into computational linguistics (ACL, EMNLP) and quantitative methodology (e.g. Psychological Methods, Behavior Research Methods).
• Open science: models, benchmarks and code are released openly wherever data protection permits, in line with the transparency standards I helped implement as Transparency Editor of MIS Quarterly.

Third-party funding

Secured or managed (2021–2026)

• NIH project Stress Ontologies and Novel AI-Assisted Tools for Ontology Development, co-author of the successful proposal, with Columbia University, the Science of Behavior Change Resource and Coordinating Center and the University of Colorado Boulder, 2023–2024 (approx. USD 850,000).
• Internal research funding for AI-based automated literature review methods, 2021–2023 (research associate position for three years).

Submitted or under review

• AgentSpace Berlin: Innovation Space for Sovereign, Secure and Controllable Agentic AI (BMFTR). Lead applicant, with 11 partners from industry and the public sector, including n8n, Cavorit, Technologiestiftung Berlin, Aiio and the Berlin Senate Chancellery (approx. EUR 2 million requested).
• KASKADE: KRITIS Decision Intelligence Platform for Critical Infrastructures (BMDS, Digital Tech-to-Product). Project lead; project sketch submitted in October 2026, planned 2027–2028, pilot partner Helios Schwerin.

In preparation

• DFG: The Perceptual Atlas: An AI-Powered Platform for Integrating and Querying Human Perceptions.
• NSF with the University of Colorado Boulder: Automated Extraction and Integration of Survey Items Using Large Language Models.

Earlier funding (selection): BMWi project Ko-RFID with DaimlerChrysler, HU Berlin, IBM, SAP and TU Berlin (2006–2008, approx. EUR 5 million in total); ISETI with Thales, TNO and others (EFRO, 2009); EU project PARMENIDES; cloud and GPU grants from Amazon, Microsoft and IFAF.

Methods and technical expertise

• Statistics and psychometrics: latent-variable models, factor and principal component analysis, measurement validation, Bayesian networks, fuzzy systems, causal modelling and causal machine learning.
• Machine learning and NLP: supervised and unsupervised learning (classification, regression, clustering, outlier detection) with scikit-learn, deep learning with PyTorch, fine-tuning of large language models, embeddings and transformers, ontologies and knowledge graphs, graph and vector databases.
• Data engineering: data modelling, SQL, data integration, data warehousing, big-data and streaming architectures, MLOps and process mining.
• Programming: Python, SQL, JavaScript and Java, including full-stack web development.