Natural language processing models reveal neural dynamics of human conversation

Area of Science:

• Neuroscience
• Computational Linguistics
• Artificial Intelligence

Background:

• Understanding the neural basis of human language, particularly speech production and comprehension during natural conversation, remains a significant challenge.
• Existing research often focuses on isolated language tasks, limiting insights into real-time conversational dynamics.

Purpose of the Study:

• To identify neural signals reflecting speech production, comprehension, and their transitions during natural human conversation.
• To investigate the role of deep learning models in decoding neural activity related to language.

Main Methods:

• Utilized intracranial neuronal recordings from individuals engaged in natural conversation.
• Employed pretrained deep learning natural language processing (NLP) models to analyze neural data.
• Correlated neural activity patterns with specific words and sentence structures.

Main Results:

• Linguistic information encoding was broadly distributed across frontotemporal areas and multiple frequency bands.
• Neural activity was specific to conveyed words and sentences, influenced by context and word order.
• Partial overlap was observed in neural patterns during language production and comprehension, with distinct neural changes during listener-speaker transitions.

Conclusions:

• Neural activity supporting language production and comprehension exhibits a dynamic organization during conversation.
• Deep learning models are effective tools for uncovering neural mechanisms of human language.
• Findings advance our understanding of the neural underpinnings of real-time conversational language processing.