A shared model-based linguistic space for transmitting our thoughts from brain to brain in natural conversations

Area of Science:

• Neuroscience
• Computational Linguistics
• Cognitive Science

Background:

• Effective communication relies on shared understanding of word meaning.
• Tracking neural processes underlying natural conversation is challenging.

Purpose of the Study:

• To model the neural dynamics of linguistic information exchange during spontaneous conversation.
• To investigate if large language models (LLMs) can capture the shared meaning space in human communication.

Main Methods:

• Recorded electrocorticography (ECoG) brain activity from epilepsy patients during natural conversations.
• Developed a model-based coupling framework aligning speaker and listener brain activity to LLM embeddings.
• Tracked word-by-word linguistic information exchange between brains.

Main Results:

• Context-sensitive LLM embeddings successfully aligned neural activity between conversing individuals.
• Linguistic content emerged in the speaker's brain before articulation and re-emerged in the listener's brain after articulation.
• LLM contextual embeddings provided superior word-by-word neural alignment compared to syntactic and articulatory models.

Conclusions:

• LLM contextual embeddings offer a numerical model of the shared, context-rich meaning space used in human communication.
• This framework enables tracking of neural information flow during natural dialogue.
• Findings advance our understanding of the neural basis of semantic processing and inter-brain communication.