This study investigated the dynamic resolution of lexical–semantic ambiguity during sentence comprehension, focusing on how uncertainty evolves as contextual information accumulates. Using time-resolved eye-tracking and a novel entropy-based measure derived from group-level semantic choice distributions, we quantified semantic uncertainty at a fine-grained temporal resolution for ambiguous words. By parametrically manipulating the semantic bias strength of the sentence context, we examined how context guides disambiguation over time. The results showed that semantic uncertainty declined gradually over temporal segments and dropped sharply following the onset of ambiguous words, reflecting both incremental integration and syntactic anchoring. A stronger contextual bias led to faster reductions in uncertainty, with effects following a near-linear trend. These findings support dynamic semantic processing models that assume continuous, context-sensitive convergence toward intended meanings. In contrast, a pretrained Chinese BERT model (RoBERTa-wwm-ext) showed similar overall trends in uncertainty reduction but lacked sensitivity to contextual bias. This discrepancy suggests that, while language models can approximate human-level disambiguation broadly, they fail to capture fine-grained semantic modulation driven by context. These findings provide a novel empirical characterization of disambiguation dynamics and offer a new methodological approach to capturing real-time semantic uncertainty. The observed divergence between human and model performance may inform future improvements to language models and contributes to our understanding of possible architectural differences between human and artificial semantic systems.
Understanding how multiple linguistic components are represented within the human language network, and how these representations change across development, remains an important open question. This study examines whether the language network integrates linguistic components through regions dedicated to single components (single-component coding) or through regions representing combined components (combination-based coding), and how these representational patterns differ across development. We recruited 64 participants (34 children and 30 adults) to complete a hierarchical reading task involving pseudowords, words, sentences, and stories, enabling assessment of neural substrates for individual language components (semantics, syntax, situation model) and their combinations. Based on this task structure, we constructed seven cognitive models targeting individual components and their combinations. Model-based multivariate regression revealed overlapping engagement of left frontal and temporal regions across all models in both groups, indicating multifunctional regions supporting combination-based coding. Compared to children, adults showed more spatially focal model correspondence within core language regions. Moreover, clustering based on model-fitting performance identified three functional clusters with marked developmental differences, broadly corresponding to right frontotemporal regions (Cluster 1), left anterior frontotemporal regions (Cluster 2), and left posterior frontotemporal regions (Cluster 3). Further statistical analyses showed that adults exhibited stronger Combination Code representations in Cluster 2, whereas children relied more on Single Code representations in Cluster 3, suggesting a developmental shift toward integrated coding in anterior networks. Together, these findings reveal that the language network supports combination-based coding through functionally organized clusters, and that the representational roles of anterior and posterior regions undergo reorganization from childhood to adulthood.