Predictive Neural Signals during Natural Mandarin Speech Comprehension

Language comprehension requires transforming continuous speech into hierarchical linguistic representations, from phonemes to words to sentences. Prediction has been proposed as a key mechanism supporting this process across languages. Here, we extend this framework to Mandarin Chinese, a tonal language with a distinct phonological structure. Using magnetoencephalography (MEG), we recorded neural responses from native speakers listening to an audiobook, and applied linear regression modeling to examine how linguistic features modulate brain activity. We found that listeners segmented continuous speech into multiple linguistic units, and that surprisal modulated neural responses at several representational levels, though not at the finest phonemic level. Tonal information was not independently predicted, but played a critical role when integrated with its segmental component. Finally, entropy reduction—an information-theoretic measure that quantifies how much a word reduces uncertainty about future words in a sentence—elicited a later and temporally distinct neural response from surprisal, indicating independent contributions of these two information-theoretic measures. Together, our findings suggest that predictive mechanisms in language comprehension are universal in their computational principles, but implemented in ways shaped by language-specific structural features, and that local surprisal resolution precedes the global updating of sentence interpretation.