<p>Language comprehension involves the prediction of upcoming linguistic units across multiple timescales. However, how this prediction process is hierarchically implemented in the human brain remains unclear. Combining natural language processing (NLP) and functional magnetic resonance imaging (fMRI) in a narrative comprehension task, we first applied the group-based general linear model (gGLM) to identify the neural underpinnings associated with anticipating upcoming words and sentences. Our results revealed a cortical hierarchy supporting linguistic prediction, extending from the superior temporal cortices to the default mode network (DMN). Next, we investigated how the word and sentence levels interact by testing two rival hypotheses: the continuous updating hypothesis posits that higher-level regions are updated continuously as inputs unfold over time, while the sparse updating hypothesis states that higher-level regions are updated only at the boundaries of their preferred timescales. Using computational modeling and autocorrelation analysis, we found that the sparse model outperformed the continuous model, with updating occurred at the sentence boundaries. Together, our results extend evidence for linguistic prediction to longer timescales and elucidate the neurocomputational mechanisms of hierarchical information updating in the human brain.</p>