Large Language Models and Linguistic Insights
Anette Frank (Heidelberg)

The linguistic abilities of LLMs are outstanding and continuously improving, so there is reason to ask whether the way in which they learn language is similar to how humans do; whether the generalizations they make match those of humans – and whether these generalizations align with the insights of theoretical, empirical and computational linguistics, for that matter. Due to the intransparent neural representations of LLMs, this is a difficult endeavor. All the more, such a research program could bring invaluable insights not only for computational linguistics and AI researchers, but also for theoretical and empirical linguistics. I will reflect on such research lines, take a brief look at recent work and insights that have been gained in this direction, and will extend these perspectives to neuroscientific research, which has a long tradition studying the ‘language network’ in the human brain. I will argue that integrating insights from all these areas could greatly impact future work in linguistics – be it theoretical, empirical, computational or neuroscientific. While most current research in this direction is devoted to syntax, I will showcase aspects of this emerging paradigm in our own work: studying the generalization abilities of LLMs in language-based reasoning tasks; investigating neural representations of Vision and Language Models, and how to integrate representations when combining language with Knowledge Graphs.