N400 amplitudes reflect change in a probabilistic representation of meaning: Evidence from a connectionist model

The N400 component of the event-related brain potential is widely used in research on language and semantic memory, but the cognitive functions underlying N400 amplitudes are still unclear and actively debated. Recent simulations with a neural network model of word meaning suggest that N400 amplitudes might reflect implicit semantic prediction error. Here, we extend these simulations to sentence comprehension, using a neural network model of sentence processing to simulate a number of N400 effects obtained in empirical research. In the model, sequentially incoming words update a representation capturing probabilities of elements of sentence meaning, not only reflecting the constituents presented so far, but also the model’s best guess at all features of the sentence meaning based on the statistical regularities in the model’s environment internalized in its connection weights. Simulating influences of semantic congruity, cloze probability, a word’s position in the sentence, reversal anomalies, semantic and associative priming, categorically related incongruities, lexical frequency, repetition, and interactions between repetition and semantic congruity, we found that the update of the predictive representation of sentence meaning consistently patterned with N400 amplitudes. These results are in line with the idea that N400 amplitudes reflect semantic surprise, defined as the change in the probability distribution over semantic features in an integrated representation of meaning occasioned by the arrival of each successive constituent of a sentence.