Trecca, F. 1 , Callaway, F. . 2 , McCauley, S. M. 2 & Christiansen, M. H. 1, 2
1 Center for Child Language, Department of Language and Communication, University of Southern Denmark
2 Department of Psychology, Cornell University
Statistical learning mechanisms in language acquisition are generally considered to be robust and reliable. However, the cues on which these operate are highly probabilistic and language-specific, rather than universal. Recent studies of Danish (Bleses et al.2011) hint at a possible influence of phonology on statistical learning. To explore this possibility, we first tested 112 adult native speakers of Danish (N=56) and US English (N=56) on an artificial learning task based on Saffran et al.(1996). Subjects were randomly assigned to one of two language conditions: a contoid Language and a vocoid language, where the C in each CV-syllable was respectively a contoid (plosive) or a vocoid (semivowel). Danish subjects scored significantly higher than American subjects in both conditions, and showed faster reaction times to words in the vocoid condition. Next, we simulated our experiment by training 112 simple recurrent networks (with similar architecture to Christiansen et al. 1998) on either Danish or English child-directed speech corpora. The networks were then exposed to and tested on the same familiarization and test stimuli as our human subjects. The networks? performance reflected that of our participants. Our results indicate that both native- and target-language phonology influence statistical learning in artificial speech segmentation.