Mapping the neuronal building blocks of human language with language models

Key Points:

• The study involved eight adult epilepsy patients undergoing cortical surface grid recordings during language production tasks, with microelectrode arrays implanted in language-related brain regions to record single-neuron and local field potential (LFP) activity.
• Neuronal responses were analyzed for selectivity and modulation to various linguistic features derived from naturalistic, de novo sentence production, using advanced parsing techniques to label parts of speech, syntactic dependencies, and hierarchical structure.
• Decoding analyses demonstrated that neuronal populations could reliably predict linguistic categories, with rigorous controls confirming the generalizability and specificity of neural tuning to language features across different contexts and participants.
• Multiple language models, including syntactic, semantic, and contextual embeddings (notably Vicuna-7B), were used to predict neuronal firing patterns, revealing that contextual models best captured the temporal dynamics and complexity of neuronal language representations.
• Extensive statistical and control analyses ensured that neural selectivity was not confounded by low-level word properties, collinearity among features, or non-linguistic factors, supporting the conclusion that specific neuronal populations encode high-order linguistic information during natural speech production.