Institute of Formal and Applied Linguistics

The combination of massively multilingual LLMs and various types of multilingual linguistic resources presents a massive opportunity for computational typology as well as for interpretability. In this talk, I will present three ongoing projects from my group in this direction. First, I will show how typological theory can be used to investigate structural overgeneralisation to English in multilingual LLMs. Next, I will present ongoing work on the MultiBLiMP benchmark: using decision trees on UD to model word order and agreement in a fine-grained and data-driven way, and creating new sets of minimal pairs. Finally, I will preview an exciting new project on using LLMs to infer parse trees from glossed data.

*** The talk will be delivered in person (MFF UK, Malostranské nám. 25, 4th floor, room S1) and will be streamed via Zoom. For details how to join the Zoom meeting, please write to sevcikova et ufal.mff.cuni.cz ***