Figure 5. Early responses reflect parallel activation of all context models, later responses selectively reflect activity in the sentence-constrained model.

(A) Current magnitude of temporal response functions (TRFs) to phoneme surprisal for each level of context (mean and within-subject standard error [Loftus and Masson, 1994]; y-axis scale identical in all panels of the figure). To allow fair comparison, all TRFs shown are from the same symmetric region of interest, including all current dipoles for which at least one of the three context models significantly improved the response predictions. Bars indicate time windows corresponding to source localizations shown in panel G. (B) When plotted separately for each hemisphere, relative lateralization of the TRFs is consistent with the lateralization of predictive power (Figure 4). (C, D) TRFs to lexical cohort entropy are dominated by the sentence context model. (E, F) TRFs to phoneme entropy are similar between context models, consistent with parallel use of different contexts in predictive models for upcoming speech. (G) All context models engage the superior temporal gyrus at early responses, midlatency responses incorporating the sentence context also engage more ventral temporal areas. Anatomical plots reflect total current magnitude associated with different levels of context representing early (−50 to 150 ms), midlatency (150–350 ms), and late (350–550 ms) responses. The color scale is adjusted for different predictors to avoid images dominated by the spatial dispersion characteristic of magnetoencephalography source estimates.