84χ2>3.84 (the critical value at the α=.05α=.05 level). Selleck Enzalutamide As displayed in Fig. 1, the exploratory analysis identified four potential effects: Word surprisal seems to predict the amplitude of N400 and, to a much lesser extent, LAN;

Word entropy reduction may explain EPNP and, to a much lesser extent, PNP. There are no potential effects of the PoS information measures (see the supplementary materials for all exploratory results). Of the four potential effects, only the N400 survives in the Confirmatory Data (see Fig. 2). All model types reach χ2>11χ2>11 for this component, which corresponds to p<.001p<.001. Hence, we have reliable evidence for an effect of word surprisal on the N400 but selleck chemicals not for any other relation between word (or PoS) information and any ERP component. Having established that a word surprisal effect occurs in both the Exploratory and Confirmatory Data sets, we now take the full set of data to investigate whether the effect can indeed be considered an N400. To this aim, Fig. 3 plots average ERP wave forms at each electrode, separately for words with low (bottom third) and high (top third)

word surprisal as estimated by the 4-gram model because this model showed the strongest overall effect on the N400 (see Fig. 4). The high-surprisal words result in a more negative deflection than the low-surprisal words, in particular within the 300–500 ms time window and at central sites, Baricitinib as is typical for the N400. Hence, word surprisal indeed affects N400 amplitude. The corresponding regression coefficient ranges from -0.17-0.17 (for the n  -gram model) to -0.22-0.22 (for RNN), which is to say that one standard deviation increase in surprisal corresponds to an average increase in N400 amplitude of between 0.17 and 0.22 μV. Because nearly all studies that find N400 effects are concerned with content words only, it is of interest to perform separate analyses

for content (i.e., open-class) and function (closed-class) words, constituting 53.2% and 46.8% of the data, respectively. A word’s class was determined from its PoS tag, where nouns, verbs (including modal verbs), adjectives, and adverbs were considered content words, and all others were function words. As can be seen in Fig. 4, there is no reliable N400 effect on function words. Nevertheless, the effect is generally weaker when only content words (as opposed to all words) are included. Most likely, this is because function words on average have lower surprisal and elicit a smaller N400 than content words. In other words, part of the effect over all words is due to the difference between content and function words. Table 2 shows results of pairwise comparisons between the best models of each type, that is, those whose word surprisal estimates fit the N400 amplitude best (for a fair comparison with the RNN and PSG models, n-gram models trained on the full BNC were not included).