N400 amplitudes reflect change in a probabilistic representation of meaning, even if induced by meaningless articles: A neural network model

Rabovsky , M. . 1 & McClelland, J. L. 2

1 Department of Psychology, Freie Universitaet Berlin, Germany
2 Department of Psychology, Stanford University, CA, USA

The N400 ERP component has aroused much interest because it is thought to provide an online measure of meaning processing. We present a computationally explicit account of this process, simulating N400 amplitudes as the stimulus-induced change in an implicit and probabilistic representation of meaning (corresponding to a kind of implicit prediction error). Using this approach, we accounted for sixteen empirical N400 effects. Here, we focus on a new simulation of a finding, which recently triggered an active debate on probabilistic prediction in language comprehension, due to a partially failed replication attempt (Nieuwland et al., 2017; but see Yan et al., 2017). Specifically, DeLong et al. (2005) exploited the fact that English indefinite articles are adjusted to fit the words they are preceding such that "an" is used prior to vowels while "a" is used prior to consonants. In sentences such as "The day was breezy so the boy went outside to fly", N400s were smaller to "a" (compatible with the high cloze continuation "kite") compared to "an" (to be followed by a lower cloze continuation, "airplane"). Because a and an do not differ in meaning, this N400 effect was taken to indicate probabilistic pre-activation of word forms. Assuming a deterministic relation between the articles and nouns, our model captures the original results. The simulation suggests that even though the statistics of natural language may not always support the effects (because of unreliable relationships between articles and nouns, e.g. "an old kite"), the effects should be observable in principle when the articles constitute reliable cues. Furthermore, our simulation demonstrates that even if the effects are observed, they do not necessarily indicate pre-activation at the level of word forms, but instead may reflect a change in a probabilistic representation of meaning, which is cued by the encountered word forms.