This finding validates our approach of extracting formal measures from corpus-based language models. Moreover, the relation between linguistic accuracy and amount of explained variance makes it very unlikely that the effect on the N400 is in fact due to a confounding variable rather than to surprisal per se. This is because such a confound would need to explain not only the effect of surprisal but also the effect of linguistic accuracy. The relation between N400 and word surprisal is further confirmed by the results of a recent fMRI study in which participants listened to spoken narratives (Willems, Frank, Nijhof, Hagoort, & Van den Bosch, 2014). Words with

higher surprisal resulted in increased RG7422 datasheet activation of the left temporal lobe, an area that has repeatedly been identified SB431542 price as an important source for the N400 (Service et al., 2007, Simos et al., 1997 and Van Petten and Luka, 2006). N400 effects are usually investigated on content words only; Dambacher et al. (2006), too, excluded function words in their study of the relation between cloze probability and the N400. However, several studies have found that less predictable function words also result in increased N400 size ( DeLong et al., 2005, Martin et al., 2013 and Wicha et al., 2003). Separate analyses on content and function words revealed that, in our data, the effect is mostly (if not exclusively) present on content words.

One reason why we failed to find a reliable N400 effect on function words might simply be that natural language (as captured in our sample of sentences) does not display much variance in function-word surprisal. The question remains why word surprisal would be predictive of N400

size. Two functional interpretations of the N400 that have been proposed are that Adenosine triphosphate it reflects semantic integration (e.g., Hagoort et al., 2009 and Kuperberg, 2007) or the retrieval of lexical information from memory (e.g., Brouwer et al., 2012 and Kutas and Federmeier, 2000), with increased integration or retrieval difficulty resulting in a larger N400. We do not propose a third account but take the effect of surprisal to be subsumed by the memory-retrieval account: More predictable words can be pre-activated, thereby making it easier to retrieve their lexical information. In contrast, it is less clear why a more surprising word would be harder to semantically integrate into its sentence context, in particular when surprisal is estimated by language models that are only minimally (if at all) sensitive to semantics, as was the case here. The word probabilities estimated by our models arise from statistical word-order patterns, which depend much more on syntactic than on semantic factors. Gouvea, Phillips, Kazanina, and Poeppel (2010) argue that surprisal and entropy reduction, being ‘one dimensional measures of syntactic processing cost’ (p.