hessian.AI Junior Research Group Leader Philipps University Marburg & Interim Professor University of Mannheim

University of Marburg

“Factual Coherence in Language Models”

When the Queen of England is not England’s Queen: Despite well-known challenges, such as hallucinations, today’s language models encode a rich set of factual knowledge. While most research is focused on how to quantify and extract this knowledge, the talk addresses a complimentary aspect, namely whether knowledge is represented coherently in language models. A fact is coherently represented, if both directions of the relation yield the same fact, regardless of its correctness. For example, “Queen Elizabeth is England’s monarch” and “England’s monarch is Queen Elizabeth” is coherent (though not correct). Spoiler: language models lack factual coherency.


Jörg Schlötterer leads a hessian.AI junior research group on perspectivism in natural language processing at the Philipps-University Marburg and is an interim professor for data science in the economic and social sciences at the University of Mannheim. Before he was a junior research group leader at the institute for artificial intelligence in medicine at the University of Duisburg-Essen. His research is about understanding, extracting and adapting the knowledge and reasoning represented in machine learning models.