Conversational content is organized across multiple timescales in the brain

Area of Science:

• Neuroscience
• Linguistics
• Cognitive Science

Background:

• Conversation enables complex thought and emotion exchange.
• Meaning construction integrates speech across timescales (words, sentences, discourse).
• Neural underpinnings of interactive sense-making in conversation are not well understood.

Purpose of the Study:

• To investigate the neural mechanisms of interactive sense-making during spontaneous conversations.
• To model neural representations of conversational content at varying timescales.
• To understand how shared and distinct neural processes contribute to language production and comprehension.

Main Methods:

• Functional magnetic resonance imaging (fMRI) was used to record brain activity during hours of spontaneous conversations.
• Contextual embeddings from a large language model (GPT) were employed to model neural representations of conversational content.
• Neural representations were analyzed across multiple timescales (words, sentences, discourse).

Main Results:

• Linguistic representations in the brain are both shared and distinct between speech production and comprehension.
• Shared representations, primarily in language-selective regions, were found at shorter timescales (words, sentences).
• Modality-specific representations showed timescale differences: shorter for production, longer for comprehension, indicating distinct integration mechanisms.

Conclusions:

• Conversational meaning arises from the interaction of shared linguistic codes and modality-specific temporal integration.
• Distinct neural mechanisms support context-dependent comprehension and adaptive speech production.
• Findings illuminate the neural basis of how humans process and generate language in real-time conversation.