Semantic Control Regions Enhance Information Processing Through Active Predictions

Predictions facilitate the processing of future events and can be actively generated depending on the current goal and context. However, the neural process behind it remains only partially understood. Potential candidates for fulfilling this predictive function are semantic-control regions, known for their role in retrieving task-relevant semantic information and therefore also potentially suited for guiding the selection of predictable upcoming stimuli. Hence, in two fMRI experiments, we tested whether semantic-control regions are recruited for generating active predictions and which regions exhibit facilitated processing due to those predictions. In experiment 1 (N=20), participants read sentences describing two protagonists interacting with items from three semantic categories (place, object, animal). Sentences alternated following a regular pattern based on taxonomic relationships, allowing strategic anticipation of the next item in half trials, while the other half remained unpredictable. In experiment 2 (N=25), we introduced masked sentences where the item with which the protagonists interact was replaced by a series of “X”, allowing us to dissociate anticipatory from post-stimulus mechanisms. Additionally, since normal sentences are richer in semantic content than masked ones, the contrast between these two conditions allowed to localize areas within the semantic processing network. To identify areas generating active predictions, in experiment 1 we compared predictable and unpredictable sentences, finding a left-lateralized network partially overlapping with semantic- control regions (IFG, pMTG, IPS, dmPFC-SMA), which we termed Active Prediction Network (APN). To locate areas exhibiting facilitated processing of semantic content, we performed the opposite contrast, finding less neural activity for predictable sentences in default mode network areas, precuneus and left AG. Experiment 2 replicated and extended these results. The APN was more active for predictable masked sentences, confirming its anticipatory role. The semantic network localizer identified regions such as precuneus, left AG, vmPFC, left ATL and left hippocampus, all showing less neural activity for predictable sentences. In experiment 2, to understand how APN generates prediction, we performed a decoding analysis to test its sensitivity to semantic content (discriminability of item semantic categories). As expected, the semantic network was always sensitive to semantic content regardless of predictability, confirming its function in processing semantic information within this design. Instead, APN was only sensitive to the semantic content of predictable sentences. This suggests that generating predictions about a sentence belonging to a semantic category requires the APN to encode the semantic representation of the predicted stimulus. Collectively, using a novel paradigm we were able to clearly dissociate neural substrates generating active predictions, the APN, from their repercussions, namely more efficient processing of semantic content within semantic processing areas. These results shed new light on studies concerning predictions in general, ranging from automatic semantic priming to predictive coding.