대규모 언어 모델 기반 정보 이론적 복잡도 지표와 언어 처리의 신경 상관성

Abstract

This study explores the relationship between information-theoretic metrics from large language models and neural activity during natural sentence reading. Analyzing EEG data, we found that surprisal was associated with reduced lower-beta power during initial processing, reflecting updates to an existing predictive model. Furthermore, entropy correlated with increased broadband neural power, primarily over left-hemisphere regions. In contrast, entropy reduction was associated with increased high-beta and gamma power, linked to information integration. These findings demonstrate that different information-theoretic metrics map onto distinct neural signatures of predictive processing and cognitive load. While the results provide strong evidence for these links, the fixation-locked analysis method suggests a need for future research to capture the continuous, dynamic time-course of meaning integration.