Proceedings of the Annual Meeting of the Cognitive Science Society

The N400 component of the event-related brain potential iswidely used in research on language and semantic memory,but the cognitive functions underlying N400 amplitudes arestill unclear and actively debated. Recent simulations with aneural network model of word meaning suggest that N400amplitudes might reflect implicit semantic prediction error.Here, we extend these simulations to sentencecomprehension, using a neural network model of sentenceprocessing to simulate a number of N400 effects obtained inempirical research. In the model, sequentially incoming wordsupdate a representation capturing probabilities of elements ofsentence meaning, not only reflecting the constituentspresented so far, but also the model’s best guess at all featuresof the sentence meaning based on the statistical regularities inthe model’s environment internalized in its connectionweights. Simulating influences of semantic congruity, clozeprobability, a word’s position in the sentence, reversalanomalies, semantic and associative priming, categoricallyrelated incongruities, lexical frequency, repetition, andinteractions between repetition and semantic congruity, wefound that the update of the predictive representation ofsentence meaning consistently patterned with N400amplitudes. These results are in line with the idea that N400amplitudes reflect semantic surprise, defined as the change inthe probability distribution over semantic features in anintegrated representation of meaning occasioned by the arrivalof each successive constituent of a sentence.