Verifying data about companies goes multi-modal

From banks to policymakers, many decisions rely on knowing what companies actually do. Are they manufacturers, service providers, or part of critical infrastructure? These classifications shape economic statistics, financial risk assessments, and even responses to climate change.
But in reality, this information is often incomplete, inconsistent or just outdated. As a result, classification still relies heavily on manual work and expert judgment–time-consuming, costly, and prone to error.

Connecting language, space, and real-world context

This is where the research of Prof. Ivan Habernal and his collaborators from TU-Darmstadt and the German Federal Bank comes in. Their paper MONETA: Multimodal Industry Classification through Geographic Information with Multi Agent Systems, co-authored by Arda Yüksel (TU-Darmstadt), Gabriel Thiem (Deutsche Bundesbank), Patrick Felka (Deutsche Bundesbank), Susanne Walter (Deutsche Bundesbank), and Gabriela Alves Werb (Deutsche Bundesbank and Frankfurt University of Applied Sciences), explores a new way of approaching this challenge.
Instead of relying on text alone, the team combines multiple sources of information: website content, knowledge bases such as Wikipedia or Wikidata, and even geographic signals like satellite imagery and OpenStreetMap data. The idea is simple but powerful: understanding what a company does requires more than words–it requires context.

To bring these different perspectives together, the researchers design a system of specialized AI agents. Each agent analyzes a specific type of data and extracts clues about a company’s activities. These clues are then combined into a final decision, making the process more transparent and robust.

Why this matters beyond research

Reliable industry classification is essential for understanding economic activity and making informed decisions at the heart of the German federal and EU financial sector.
By improving precise data verificaion, systems like MONETA could support:

• financial institutions in assessing risks and investments
• policymakers in monitoring economic developments

In contexts where data is sparse or ambiguous, combining multiple signals can make the difference between guesswork and informed analysis.

From Bochum to the global NLP stage

The paper has been accepted to the main conference of ACL 2026, the 64th Annual Meeting of the Association for Computational Linguistics, taking place July 2–7, 2026 in San Diego, USA. As the annual gathering of the international natural language processing community, ACL is one of the central venues where new ideas and advances in language technologies are discussed and shaped. Being part of the main conference marks the work as a contribution at the forefront of current research in the field.
For Ivan Habernal, Professor at Ruhr University Bochum, this work reflects the core mission of his research. At the Faculty of Computer Science, he leads the Trustworthy Human Language Technologies (TrustHLT) group and is a member of the Research Center Trustworthy Data Science and Security (RC Trust). His research focuses on making language technologies more reliable, privacy-aware, and applicable in sensitive domains such as law and public decision-making.
The MONETA project exemplifies this approach: combining technical innovation with a clear focus on real-world impact.