Speaker
Jonas Mayer Martins
(University of Göttingen (CIDAS))
Description
Children acquire language not just by listening, but through interacting with others in their social context. Inspired by this active learning, we ask: can language models become better storytellers if they learn not only from next-word prediction, but also from high-level, cognitively-inspired feedback? We train a student model to generate stories, with a teacher model rating each attempt based on grammar, narrative coherence, and creativity. By varying the proportion of self-supervised text versus teacher feedback in training, we assess the impact of this social interactive learning on formal and functional language competence. We thus aim to connect principles of human language acquisition with computational modeling.
Author
Jonas Mayer Martins
(University of Göttingen (CIDAS))
Co-authors
Ali Hamza Bashir
(University of Göttingen)
Muhammad Rehan Khalid
(University of Göttingen)
Lisa Beinborn
(University of Göttingen (CIDAS))
