Understanding of, and Explanations with, Large Language Models

01/06/2024 to 31/05/2028

Large language models (LLMs), such as chatGPT, are powerful, forward-looking innovations. Their development is proceeding at breakneck speed, also driven by economic interests, so that a nuanced, scientific analysis of the effects and consequences of these technologies falls by the wayside. Modern systems are so complex that internal mechanisms and the behavior they generate are only insufficiently comprehensible. This is where science is called upon to play its part in determining the potential risks and opportunities of these technologies for society and to share this knowledge in a generally understandable way. This requires interdisciplinary expertise. In this Momentum project, the existing interdisciplinary profile of the professorship (linguistics, cognitive science) is to be complemented with technical and practical expertise in the field of machine learning.
Structurally, this project aims to combine insights from the humanities with research from cognitive science and artificial intelligence. In teaching, new, interdisciplinary, digitally supported course formats are to be developed to enable participants to acquire in-depth knowledge of modern language technologies and to form nuanced opinions on this basis. This content will also be made accessible to the wider public. Two new lines of research are to be pursued. The first ("ling2ML") explores the mechanisms and behavior of LLMs from a linguistic perspective,
for example with the help of psycholinguistic experiments. The second ("ML2ling") uses the possibilities
possibilities of modern LLMs to make information that is otherwise difficult to formalize (e.g. world knowledge, intuitive judgements) usable for transparent, explanatory models of human language use.