How do predictions in the brain incorporate the temporal unfolding of context in our natural environment? We here provide evidence for a neural coding scheme that sparsely updates contextual representations at the boundary of events. This yields a hierarchical, multilayered organization of predictive language comprehension. Training artificial neural networks to predict the next word in a story at five stacked time scales and then using model-based functional magnetic resonance imaging, we observe an event-based "surprisal hierarchy" evolving along a temporoparietal pathway. Along this hierarchy, surprisal at any given time scale gated bottom-up and top-down connectivity to neighboring time scales. In contrast, surprisal derived from continuously updated context influenced temporoparietal activity only at short time scales. Representing context in the form of increasingly coarse events constitutes a network architecture for making predictions that is both computationally efficient and contextually diverse.

When speech intelligibility is reduced, listeners exploit constraints posed by semantic context to facilitate comprehension. The left angular gyrus (AG) has been argued to drive this semantic predictability gain. Taking a network perspective, we ask how the connectivity within language-specific and domain-general networks flexibly adapts to the predictability and intelligibility of speech. During continuous functional magnetic resonance imaging (fMRI), participants repeated sentences, which varied in semantic predictability of the final word and in acoustic intelligibility. At the neural level, highly predictable sentences led to stronger activation of left-hemispheric semantic regions including subregions of the AG (PGa, PGp) and posterior middle temporal gyrus when speech became more intelligible. The behavioural predictability gain of single participants mapped onto the same regions but was complemented by increased activity in frontal and medial regions. Effective connectivity from PGa to PGp increased for more intelligible sentences. In contrast, inhibitory influence from pre-supplementary motor area to left insula was strongest when predictability and intelligibility of sentences were either lowest or highest. This interactive effect was negatively correlated with the behavioural predictability gain. Together, these results suggest that successful comprehension in noisy listening conditions relies on an interplay of semantic regions and concurrent inhibition of cognitive control regions when semantic cues are available.

46Sentence comprehension requires the rapid analysis of semantic and syntactic 47 information. These processes are supported by a left hemispheric dominant fronto-48 temporal network, including left posterior inferior frontal gyrus (pIFG) and posterior 49 superior temporal gyrus/sulcus (pSTG/STS). Previous electroencephalography 50 (EEG) studies have associated semantic expectancy within a sentence with a 51 modulation of the N400 and syntactic gender violations with increases in the LAN 52 and P600. Here, we combined focal perturbations of neural activity by means of short 53 bursts of transcranial magnetic stimulation (TMS) with simultaneous EEG recordings 54 to probe the functional relevance of pIFG and pSTG/STS for sentence 55 comprehension. We applied 10 Hz TMS bursts of three pulses at verb onset during 56 auditory presentation of short sentences. Verb-based semantic expectancy and 57 article-based syntactic gender requirement were manipulated for the sentence final 58 noun. We did not find any TMS effect at the noun. However, TMS had a short-lasting 59 impact at the mid-sentence verb that differed for the two stimulation sites. 60Specifically, TMS over pIFG elicited a frontal positivity in the first 200 ms post verb 61 onset whereas TMS over pSTG/STS was limited to a parietal negativity at 200-400 62 ms post verb onset. This indicates that during verb processing in sentential context, 63 frontal brain areas play an earlier role than temporal areas in predicting the upcoming 64 noun. The short-living perturbation effects at the mid-sentence verb suggest a high 65 degree of online compensation within the language system since the sentence final 66 noun processing was unaffected. 67 68 69 70

How can anticipatory neural processes structure the temporal unfolding of context in our natural environment? We here provide evidence for a neural coding scheme that sparsely updates contextual representations at the boundary of events and gives rise to a hierarchical, multi-layered organization of predictive language comprehension. Training artificial neural networks to predict the next word in a story at five stacked timescales and then using model-based functional MRI, we observe a sparse, event-based “surprisal hierarchy”.The hierarchy evolved along a temporo-parietal pathway, with model-based surprisal at longest timescales represented in inferior parietal regions. Along this hierarchy, surprisal at any given timescale gated bottom-up and top-down connectivity to neighbouring timescales. In contrast, surprisal derived from a continuously updated context influenced temporo-parietal activity only at short timescales. Representing context in the form of increasingly coarse events constitutes a network architecture for making predictions that is both computationally efficient and semantically rich.

No abstract