Shared neural geometries for bilingual semantic representations

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Linguistics

Background:

• The human brain's capacity for multilingualism is a complex cognitive function.
• Understanding the neural basis of how bilingual individuals process and represent meaning across languages is crucial.

Purpose of the Study:

• To investigate the neural mechanisms underlying semantic representation in bilingual individuals.
• To identify how the brain distinguishes or integrates meaning across different languages.

Main Methods:

• Examined hippocampal neuron activity in bilingual participants during passive listening, speaking, and conversation in English and Spanish.
• Utilized correlational analysis to identify neural responses to equivalent words across languages.
• Employed representational similarity analysis to compare semantic geometry between languages.

Main Results:

• Identified putative "translation neurons" with correlated responses to equivalent words (e.g., "tierra" and "earth").
• Observed language-specific semantic tuning in neurons, indicating distinct neural implementations for each language.
• Found preserved distances between neural responses for equivalent words, forming a shared semantic geometry.
• Demonstrated that this shared geometry is implemented by the same neurons but with distinct readout axes, potentially preventing cross-language interference.

Conclusions:

• The hippocampus encodes a language-independent internal model for meaning.
• Shared semantic geometry with distinct readout axes may be a general principle for multilingual representation, observed in both the brain and multilingual language models (e.g., mBERT).
• Distinct readout mechanisms help maintain language separation while preserving a unified semantic space.