Presentation

Alice Hodapp¹, Alma Lindborg¹, Milena Rabovsky¹; ¹University of Potsdam

Introduction. In the typical N400 paradigms an attenuation of the ERP is recorded when stimuli are congruous with their preceding context e.g., sentences or words. There is an active debate concerning the functional basis of N400 amplitudes. One proposal is that N400 amplitudes might reflect unpredicted semantic information and update of the internal predictive model, which is continuously adjusted based on statistical regularities in the environment (e.g., Bornkessel-Schlesewsky & Schlesewsky, 2019, Front Psychol; Kuperberg, 2016, LCN; Rabovsky et al. 2018, NHB; Rabovsky & McRae, 2014, Cognition). Here, we test this idea using a Bayesian sequential learner model on EEG data obtained during a semantic oddball task with stimuli from different semantic categories and an additional manipulation of transition probabilities between categories. Methods. Specifically, participants will be exposed to a semantic oddball task while their EEG is being recorded. In the task a sequence of nouns from the same semantic category (land formations, clothing, vegetables, tools, or musical instruments) will be presented to the participants one-by-one on a screen. The sequence is followed immediately by a series of nouns from a different category with the categories repeating to create a continuous sequence. We expect a significantly weaker N400 ERP amplitudes for the last word in a stimulus sequence (commonly called standard) than for the first word of the new category (deviant). Crucially, we additionally manipulate the predictability of sematic information via the transition probabilities between categories: a category is followed by one other category with an 85% probability and each of the other categories with a 5% probability. Various statistical learning paradigms have shown that transition probabilities can be (implicitly) learned and influence behavior as well as EEG signals (e.g., in the auditory domain: Koelsch et al, 2016, Sci Rep). Our experimental design allows us to investigate which statistics can be learned in a semantic oddball task and whether this learning modulates N400 amplitudes. To this end, different Bayesian sequential learner models will be compared: (1) The main Bayesian model of interest will infer the transition probabilities of the categories combined with a finite memory. It will be compared to a Bayesian model that (2) learns the (local) occurrence frequency of categories and (3) a null model that captures the category switches only. The change in a model’s probability distribution represents its semantic surprise (i.e., the Bayesian surprise; Itti & Baldi, 2009, Vision Res) which can be fitted to the single-trial ERP response for model evaluation. We will also investigate the time windows and electrodes in which the model’s surprise readouts best predict ERP activity to compare the hypothesized influence of transition probabilities on the N400 to possible other modulation of later (anterior or posterior) positive ERP components. Discussion. A correspondence of changes in the N400 to Bayesian learning of transition probabilities would not only further support notions that this component indexes surprise from a probabilistic internal semantic model but would also explicitly demonstrate that this probabilistic information can be (implicitly) learned over time from the statistical regularities of the environment.

Topic Areas: Computational Approaches, Meaning: Lexical Semantics