Multilingual Computational Models Capture a Shared Meaning Component in Brain Responses across 21 Languages.

Area of Science:

• Neuroscience
• Computational Linguistics
• Cognitive Science

Background:

• Understanding how the brain processes multiple languages is a key question in neuroscience.
• Multilingual neural network models provide a framework for investigating cross-lingual linguistic representations.
• Previous research has explored neural processing of language, but cross-lingual similarities require further investigation.

Purpose of the Study:

• To evaluate the similarity of linguistic representations across speakers of 21 languages.
• To test if multilingual models can predict brain activity in the language network.
• To determine if neural encoding models can be transferred zero-shot across languages.

Main Methods:

• Combined existing and newly collected functional magnetic resonance imaging (fMRI) data from 21 languages across 4 language families.
• Utilized multilingual neural network models to generate linguistic representations.
• Developed and tested encoding models to predict brain activity from model representations.

Main Results:

• Multilingual model representations reliably predicted brain activity within individual languages.
• Encoding models demonstrated successful zero-shot transfer across languages, predicting brain responses in unseen languages.
• A shared cross-lingual component in neural processing was identified.

Conclusions:

• The human brain utilizes shared neural representations for processing multiple languages.
• These shared representations are linked to a common underlying meaning space.
• Findings support the existence of universal linguistic processing mechanisms in the brain.