Presentation
Search Abstracts | Symposia | Slide Sessions | Poster Sessions | Lightning Talks
Investigating the extent of error-based adaptation in language using a repetition paradigm: the effect of sentence context and prediction violation
There is a Poster PDF for this presentation, but you must be a current member or registered to attend SNL 2023 to view it. Please go to your Account Home page to register.
Poster A17 in Poster Session A, Tuesday, October 24, 10:15 am - 12:00 pm CEST, Espace Vieux-Port
Alice Hodapp1, Milena Rabovsky1; 1University of Potsdam
Introduction. Predictive processing theories of language suggest that predictions are continuously updated (Bornkessel-Schlesewsky & Schlesewsky, 2019, Front Psychol; Fitz & Chang, 2019, Cogn. Psychol.; Kuperberg, 2016, LCN; Rabovsky et al. 2018, NHB). The N400 ERP component, which is sensitive to unexpected input (Kutas & Hillyard, 1980, Science), allows an investigation of these theories. Previous research has demonstrated that N400 amplitudes are reduced when a previously unexpected word is repeated (Rommers & Federmeier, 2018, Cortex), presumably reflecting an implicit memory benefit and adaptation to the unpredicted word (Hodapp & Rabovsky, 2021, Eur. J. Neurosci.). However, most current language models assume that prediction errors are processed in context (e.g., Rabovsky et al., 2018, NHB), which implies an update not only of the repeated word but also of the predictive properties of the corresponding sentence contexts. Here, we experimentally investigate this assumption. Additionally, the violation of strong predictions (i.e., unexpected words in high versus low constraint contexts) could influence possible sentence-level adaptation effects and influence predictions when the sentence is encountered again. While prediction violation does not influence N400 repetition effects to the unexpected words themselves (Lai et al., 2021, Brain. Res.), the change in predictions made from sentence contexts after encountering unexpected information could depend on how certain the language comprehender was about this prediction before the error was encountered. Methods. To investigate these issues, this pre-registered EEG study employed a sentence repetition paradigm that manipulated the repetition of context as well as the violation of predictions. The sentences presented to the participants (n=42) ended with unexpected words. 20 sentences later, this unexpected critical word was repeated either in the same sentence or in a new sentence (context manipulation). Of the repeated sentences, half were highly constraining, and the other half weakly constraining (prediction violation manipulation). We employed Bayesian linear-mixed effect models to be able to investigate possible null effects of the manipulation. Results. Compared to words that were repeated in new contexts, words repeated in the original sentence contexts demonstrated a stronger repetition effect (β = 0.65 [0.06, 1.24], BF10 = 3.16). This indicates that the prediction error at the critical word was processed in a sentence context, updating predictive (sentence) cues, and allowing for more precise context-based predictions later in the experiment. Bayesian methods showed evidence against an effect of violated predictions (i.e., high versus low constraint) on downstream predictions made from the same context (β = -0.26 [-0.85, 0.33], BF10=0.46). This suggests that the predictive properties of the sentence are updated in proportion to the amount of unpredicted information rather than the certainty of the prediction. Discussion. Together, the results support accounts of the N400 as neural signature of a semantic prediction error that drives adaptation and extend the experimental evidence beyond word-level effects to sentence-level mechanisms. Crucially, this error-driven one-shot adaptation on a sentence level did not depend on the certainty of the predictions, supporting previous findings that prediction violations do not seem to be critical to adaptation in the language system.
Topic Areas: Meaning: Lexical Semantics,