Adaptive semantic prediction in dynamic linguistic environments: Computational-Neural Insights from Hierarchical Bayesian Inference

Abstract

Real-world language processing unfolds in environments of volatility and uncertainty, requiring comprehenders to continuously adapt semantic predictions as environmental statistics change. However, the fundamental computational-neural mechanisms supporting such hierarchical adaptive prediction remain incompletely understood. Critically, our goal was to examine how comprehenders update their beliefs about high-level environmental reliability as evidence accumulates across trials, and how this update modulates low-level semantic prediction. Here, we introduce a paradigm that extends traditional filler-based designs and allows for the examination of trial-by-trial adaptation and corresponding neural responses. We further implemented a computational neural analytical framework based on the Hierarchical Gaussian Filter (HGF) to model the trial-by-trial hierarchical inference process. We showed that environmental statistical regularities systematically modulate multiple neural signatures, from the well-established N400 to later ERP components, such as late anterior positivity (LAP) and P600, as well as cortical oscillatory components in the theta and beta bands. Model-based analyses showed that distinct computational variables were reflected in dissociable neural signatures. At the ERP level, N400 activity was associated with both high-level belief states and precision-weighted prediction error (pwPE2), whereas LAP was more strongly associated with pwPE2. At the level of neural oscillations, theta-band activity was associated with high-level belief states, whereas beta-band activity selectively tracks prior belief2. Together, these findings indicate that distinct neural signatures are differentially associated with complementary components of hierarchical Bayesian inference, supporting the ongoing estimation of the environmental statistical structure. More broadly, our results provide model-based evidence that hierarchical predictive coding operates in advanced language comprehension under uncertainty.