Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Multilingual Computational Models Reveal Shared Brain Responses to 21 Languages

Andrea Gregor de Varda¹, Saima Malik-Moraleda, Greta Tuckute¹, Evelina Fedorenko¹; ¹Massachusetts Institute of Technology

Presenter: Andrea Gregor de Varda

How does the human brain process the rich variety of languages? Multilingual neural network language models (MNNLMs) offer a promising avenue to answer this question by providing a theory-agnostic way of representing linguistic content across languages. We combined existing and newly collected fMRI data from speakers of 21 languages to test whether MNNLM-based encoding models can predict brain activity in the language network. Across 20 models and 8 architectures, encoding models successfully predicted responses in the various languages, replicating and extending previous findings. Critically, models trained on a subset of languages generalized zero-shot to held-out ones, even across language families. This cross-linguistic generalization points to a shared component in how the brain processes language, plausibly related to a shared meaning space.

Topic Area: Language & Communication

Extended Abstract: Full Text PDF