Mechanisms of prediction updating - preprint

Abstract

One of the most compelling pieces of evidence for the presence of predictions in language comprehension comes from event-related potential (ERP) studies which show that encountering an adjective whose gender marking is inconsistent with that of a highly expectable noun leads to an effect at the adjective. Until now the mechanism underlying this effect has been unknown. The present study tests a novel hypothesis whereby the effect at adjectives reflects prediction updating, which in turn impacts the N400 generated by the upcoming noun. In two experiments, native speakers of Polish read short stories that made several nouns likely at the position of a direct object in the story’s final sentence. The noun was preceded by a gender-marked adjective which was shown in either of the two grammatical forms: a) informative about the following noun (adjectival gender marking constrained which nouns can follow), or b) less informative (the grammatical marking did not disambiguate the gender of the following noun). Importantly, the same stories and critical words were used in both conditions, which was possible thanks to exploiting syncretisms in Polish adjectival inflection. The results provide support for one or more mechanisms of prediction updating at the adjective that impact the amplitude of the N400 elicited by the following noun. Together, these results show that when only morphosyntactic cues are available, they are instantly used to update predictions about upcoming words.

Definitely saw it coming? An ERP study on the role of article gender and definiteness in predictive processing

“…If pre-nominal ERP effects reflect the detection of a mismatch between predicted and expected gender, and gender-mismatch is detected more easily or faster for 'de' than 'het', then 'de' should have elicited a larger effect than 'het'. However, this is not what we found, and we do not believe that N400 amplitude reflects a prediction mismatch detection process (see also Kutas & Federmeier, 2011;Rabovsky et al 2018;Szewczyk & Wodniecka, 2018;Van Berkum, 2009). Alternatively, unexpected 'het' may elicit a larger gender effect because 'het' is a lower-frequent word taking the place of the more frequent article 'de'.…”

Bilingual children do not transfer stress patterns: Evidence from suprasegmental and segmental analysis of L1 and L2 speech of Polish–English child bilinguals

“…Two non-word repetition tasks were used, one in Polish (Szewczyk & Wodniecka, 2012) and one in English (Szewczyk, Wodniecka, Otwinowska, & Chiat, 2012). The creation of the task was initiated as a part of the COST Action ISO804, which aimed at designing tools for cross-linguistic comparisons between bilingual children’s languages (Armon-Lotem et al, 2015).…”

A Framework for Decoding Event-Related Potentials from Text

“…Traditionally, ERP studies of language processing use coarse-grained predictors like cloze rates, which often lack the precision to differentiate different neural computational models (for discus-sion, see Yan et al, 2017;Rabovsky et al, 2018). To overcome such limitations, a main line of attack has been to extract measures from probabilistic language models and evaluate them against ERP amplitudes (Frank et al, 2015;Brouwer et al, 2017;Rabovsky et al, 2018;Delaney-Busch et al, 2019;Fitz and Chang, 2018;Szewczyk and Wodniecka, 2018;Biemann et al, 2015).…”