Decoding the time course of predictive feature activation during speech comprehension from EEG

Poster Session B - Sunday, April 14, 2024, 8:00 – 10:00 am EDT, Sheraton Hall ABC

Timothy Trammel¹ (tgtrammel@ucdavis.edu), Matthew J. Traxler¹, Tamara Y. Swaab¹; ¹University of California, Davis

There is a large body of research providing evidence that prediction facilitates language comprehension. However, less is known about which features are retrieved during predictive processing of imminent words in speech. Predictive coding theory suggests that top-down predictions are made continuously and are compared to bottom-up sensory input to generate prediction error. Thus, features of predicted upcoming words in constraining sentence contexts should be present prior to encountering these words and, as the process is top-down, higher-level features should be predicted prior to lower-level features. This assumption is examined in the present EEG study. Thus far univariate measures of EEG have not shown evidence of a temporal hierarchy in the retrieval of conceptual, lexical and length features prior to the onset of predictable words. Machine learning classification allows us to decode the content of potentially pre-activated features in EEG and to decode the time course by which these features are retrieved. The present study uses a support vector machine (SVM) learning algorithm to decode these features from EEG data. EEG will be recorded when participants listen to sentences in which the predictability of critical nouns is manipulated. EEG will be decoded for animacy, phonological neighborhood and length in a 2000ms epoch, starting 1000ms prior to critical noun onset. If predicted features are pre-activated in a top-down fashion, then animacy features should be reliably decodable before critical nouns are heard, and they should be decodable before lexical and length features. This pattern of results would support predictive coding theory.

Topic Area: LANGUAGE: Semantic