Characterizing directional dynamics of semantic prediction based on inter-regional temporal generalization.

Abstract

The event-related potential/field component N400(m) is a widely accepted neural index for semantic prediction. Top-down input from inferior frontal areas to perceptual brain regions is hypothesized to play a key role in generating the N400, but testing this has been challenging due to limitations of causal connectivity estimation. We here provide new evidence for a predictive model of speech comprehension in which IFG activity feeds back to shape subsequent activity in STG/MTG. Magnetoencephalography (MEG) data was obtained from 21 participants (10 men, 11 women) during a classic N400 paradigm where the semantic predictability of a fixed target noun was manipulated in simple German sentences through the preceding verb. To estimate causality, we implemented a novel approach, based on machine learning and temporal generalization, to test the effect of inferior frontal gyrus (IFG) on temporal regions. A support vector machine (SVM) classifier was trained on IFG activity to classify less predicted (LP) and highly predicted (HP) nouns and tested on superior/middle temporal gyri (STG/MTG) activity, time-point by time-point. The reverse procedure was then performed to establish spatiotemporal evidence for or against causality. Significant decoding results were found in our bottom-up model, which were trained at hierarchically lower level areas (STG/MTG) and tested at the hierarchically higher IFG areas. Most interestingly, decoding accuracy also significantly exceeded chance level when the classifier was trained on IFG activity and tested on successive activity in STG/MTG. Our findings indicate dynamic top-down and bottom-up flow of information between IFG and temporal areas when generating semantic predictions.Significance Statement Semantic prediction helps anticipate the meaning of upcoming speech based on contextual information. How frontal and temporal cortices interact to enable this crucial function has remained elusive. We used novel data-driven MEG analyses to infer information flow from lower to higher areas (bottom-up) and vice versa (top-down) during semantic prediction. Using "earlier" MEG signals in one area to decode the "later" in another, we found that inferior frontal cortices feed predictions back to temporal cortices, to help decipher bottom-up signals going to the opposite direction. Our results provide experimental evidence on how top-down and bottom-up influences interact during language processing.