Brains and language models converge on a shared conceptual space across different languages

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Linguistics

Background:

• Human languages exhibit vast diversity in sounds, scripts, and syntax.
• Despite structural differences, languages convey comparable meanings.
• The neural basis for this shared conceptual meaning across languages remains largely unexplored.

Purpose of the Study:

• To investigate whether different languages converge on a shared neural substrate for conceptual meaning.
• To determine if artificial intelligence language models (LMs) can predict brain activity related to meaning across different languages.

Main Methods:

• Utilized naturalistic functional magnetic resonance imaging (fMRI) on native speakers of English, Chinese, and French.
• Trained language models (LMs) on distinct languages to analyze their representational spaces.
• Developed voxelwise encoding models to align LM embeddings with neural responses and tested cross-lingual generalization.

Main Results:

• Language models trained on different languages converged onto similar embedding spaces, particularly in middle layers.
• Encoding models trained on neural data from one language successfully predicted brain activity in listeners of other languages.
• Generalization of predictive models occurred across high-level language and default-mode brain regions.

Conclusions:

• Neural representations of meaning are shared across speakers of diverse languages.
• Language models trained on different languages capture this shared semantic space.
• Shared meaning emerges from human interaction and a common world, transcending linguistic diversity.