Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Decoding anticipated semantic and visual word features in Spanish-English heritage speakers
Poster Session D, Saturday, October 26, 10:30 am - 12:00 pm, Great Hall 3 and 4
Timothy Trammel1, Elaina Jahanfard1, Matthew J. Traxler1, Tamara Y. Swaab1; 1University of California, Davis
Predictive coding models (Rao & Ballard, 1999) have recently gained popularity as potential architectures for the role of prediction during language comprehension (Kuperberg & Jaeger, 2016). These models suggest that predictions are continuously generated from higher cortical levels about upcoming input from lower-levels of processing. As bottom-up input is encountered by each level of processing, prediction error is computed by comparing the input with the top-down prediction. To date, most research on prediction during language processing has focused on native speakers of one language. However, as more than 50% of the world speaks at least two languages (Grosjean, 2019) there is a need to better understand how bilinguals predict within both their native language and their second language (L2). Currently, little is known about which linguistic features are anticipated when bilinguals process L2 and whether this is done in a top-down fashion as hypothesized by predictive coding. The present study aimed to examine prediction and anticipation of information during visual word recognition in L2 of Spanish-English bilinguals. We predicted that bilinguals would be able to select the context appropriate language and tested the two crucial assumptions of the predictive coding theory 1) bilinguals anticipate semantic features (concreteness) before sub-lexical features (word length) during visual word processing in a priming, and 2) failed prediction would be evident from increased prediction error. During a visual word priming task, Spanish-English bilingual participants’ (n=29) electroencephalogram (EEG) was recorded while they read a prime and target words in English. For every prime-target trial, they were instructed to try to predict the upcoming target word based on the meaning of the prime word before target onset. They were asked to indicate via button press to respond with yes if the visual target word was identical to the word they had predicted, and no when it was not. Two-thirds of the word pairs were related, and those remaining were unrelated. The combination of prediction accuracy and relatedness of the target stimuli resulted in three conditions: predicted – participants predicted the actual related target word, unpredicted – participants did not predict the actual related target word, and unrelated – participants could not predict the target word because it was unrelated. We analyzed the entire prime-locked trial epoch (3600ms; with target onset at 2000ms) using support vector machine (SVM) EEG decoding – to classify either concreteness or word length of the target word – and mass-univariate ERP analysis. Statistical testing for both analyses were performed using cluster-based permutation testing. The SVM EEG decoding results showed that target word features – both concreteness and word length – were reliably decodable at greater than chance-level (50%) prior to target word onset when the target word was predicted. The mass-univariate analyses showed reduced N400 to predicted target words relative to unpredicted but related or unrelated target words, suggesting an increase in prediction error during unsuccessful predictions. These results suggest that bilinguals predict in L2 in a manner consistent with predictive coding accounts of language processing.
Topic Areas: Multilingualism, Methods