A statistical learning bias predicts second-language reading patterns

Onnis, L. 1 , Frank, S. 2 , Yun, H. 3 & Lou-Magnuson, M. 1

1 Nanyang Technological University
2 Radboud University
3 Konkuk University

A previous Artificial Grammar Learning study found a language-specific 'statistical learning (SL) bias': English-speaking adults segment an ambiguous speech string such that words' transitional probabilities are consistent with the head-initial structure of English, whereas Korean-speaking adults favour a different segmentation, consistent with the head-final structure of Korean.

In the current study, we investigate how this SL bias affects second-language processing. Using the above mentioned AGL task, we assessed the SL bias of 58 adult Koreans who were advanced speakers of English as a second language. The same participants then performed a self-paced reading task on a general sample of English sentences. Word-reading times were analysed by linear mixed-effects regression, including as predictors: base frequency and forward transitional probability of the word, participant's SL bias and L2 proficiency, as well as several covariates.

We found that individuals with a more 'English like' SL bias more efficiently incorporated statistical regularities during online reading, in that their word-reading times were significantly more sensitive to the words' forward transitional probability as opposed to base frequencies. These findings further support the view that statistical learning skills underlie not only language learning in childhood, but also second language processing in adults.