Human Neural Dynamics of Elements in Natural Conversation – A Deep Learning Approach

Human conversations are one of the most common methods for communication and social interactions. Engaging in a conversation, our brains are able to flexibly transit between speaker-listener alternations, and actively process language comprehension or production planning with ease. Previous work showed a series of interconnected brain areas that support speech production planning, articulation, and speech comprehension, but little is known about the neural relations between these modalities. Here, we use intracranial depth electrode to record local field potentials (LFP) from participants actively participated in natural conversations. We applied pre-trained natural language processing (NLP) models capable of capturing the unique sequence of words and their structures within phrases and sentences to study the neural activities of elements in conversation. Our results showed a remarkable convergence between deep learning embeddings and brain activities in frontotemporal areas at a wide range of frequencies. The neural dynamics diverges between articulation planning and comprehension, suggesting that one is not simply reverse mirroring the other. Finally, we discovered that alpha-beta activities in frontotemporal areas increased at the speaker-listener transition, and these response electrodes overlapped by a large degree to electrodes that are correlated to NLP, suggesting that these neural activities are actively process language information during transitions. Together, our work revealed a remarkable orchestration of language processing across multiple areas and frequencies in our brain, serving as the building blocks underlying smooth conversation and human communication.