研究提出结合 Mistral-Instruct 与 TRF 的分析框架以量化社会语境对语音感知的神经影响

Social information plays an important role in everyday speech communication, yet existing neurophysiological approaches provide limited means of isolating its contribution during naturalistic speech perception. Here, we present an analytical framework that combines the large language model (LLM) Mistral-Instruct, fine-tuned on conversational language, and temporal response functions (TRFs) to investigate the influence of social context on word-level speech processing. The framework disentangles local context (LC) from speaker-specific social context (SC) when generating LLM-derived lexical expectation measures, enabling such predictions to be generated under either matching or mismatching speaker contexts for the same speech material. This provides a direct way of isolating the contribution of social information to lexical predictions. We first show that speaker-related social context carries information about speech register and conversational role beyond that available from previous local context alone by interrogating Mistral-Instruct. We then demonstrate that lexical expectations generated from socially congruent contexts produce small but reliable improvements in EEG prediction relative to expectations generated from incongruent social contexts. These effects were observed despite identical acoustic input and neural recordings, indicating that the influence of social context on lexical expectations explains dissociable variance in neural responses to speech. Together, these findings provide a proof of concept for using LLM-derived lexical expectations to isolate and measure the influence of social context on speech processing. More broadly, the proposed framework offers a flexible approach for generating and testing hypotheses about how different forms of contextual knowledge shape neural responses during naturalistic language comprehension.