Figure 4. All context models significantly contribute to predictions of brain responses.

(A) Each context model significantly improves predictions of held-out magnetoencephalography (MEG) data in both hemispheres (t_max ≥ 6.16, p ≤ 0.005). Black bars below anatomical plots indicate a significant difference between hemispheres. The white outline indicates a region of interest (ROI) used for measures shown in (B), (C), and (E). Brain regions excluded from analysis are darkened (occipital lobe and insula). (B) Surprisal and entropy have similar predictive power in each context model. Each dot represents the difference in predictive power between the full and a reduced model for one subject, averaged in the ROI. Cohort- and phoneme entropy are combined here because the predictors are highly correlated and hence share a large portion of their explanatory power. Corresponding statistics and effect size are given in Table 1. A single left-handed participant is highlighted throughout with an unfilled circle. LH: left hemisphere; RH: right hemisphere. (C) Even when tested individually, excluding variability that is shared between the two, cohort- and phoneme entropy at each level significantly improve predictions. A significant effect of sentence-constrained phoneme entropy is evidence for cross-hierarchy integration, as it suggests that sentence-level information is used to predict upcoming phonemes. (D) Predictive power of the acoustic feature representations. (E) The lateralization index ${\displaystyle (LI=R/(L+R))}$ indicates that the sublexical context model is more right-lateralized than the sentence context model. Left: LI = 0; right: LI = 1. Significance levels: *p ≤ 0.05; **p ≤ 0.01; ***p ≤ 0.001.