Abstract
Language is central to human life; however, how our brains derive meaning from language is still not well understood. A commonly studied electrophysiological measure of on-line meaning related processing is the N400 component, the computational basis of which is still actively debated. Here, we test one of the recently proposed, computationally explicit hypotheses on the N400 - namely, that it reflects surprise with respect to a probabilistic representation of the semantic features of the current stimulus in a given context. We devise a Bayesian sequential learner model to derive trial-by-trial semantic surprise in a semantic oddball like roving paradigm experiment, where single nouns from different semantic categories are presented in sequences. Using experimental data from 40 subjects, we show that model-derived semantic surprise significantly predicts the N400 amplitude, substantially outperforming a non-probabilistic baseline model. Investigating the temporal signature of the effect, we find that the effect of semantic surprise on the EEG is restricted to the time window of the N400. Moreover, comparing the topography of the semantic surprise effect to a conventional ERP analysis of predicted vs. unpredicted words, we find that the semantic surprise closely replicates the N400 topography. Our results make a strong case for the role of probabilistic semantic representations in eliciting the N400, and in language comprehension in general.