Isbilen, E. 1 , Frost, R. 2 , Monaghan, P. 3 & Christiansen, M. 1
1 Cornell University
2 Max Planck Institute for Psycholinguistics, Nijmegen
3 University of Amsterdam
A large body of work examines learners' ability to chunk sequential information for items that directly follow one another in speech. However, comparatively little research tests whether chunking ability similarly accounts for the statistical learning of nonadjacent dependencies (relationships between particular elements that occur together reliably, but are separated by intervening material). In two experiments, we trained adult participants on an artificial language containing words comprising nonadjacent dependencies, and tested the statistical learning of words and structure using a chunk-based memory recall task (the statistically-induced chunking recall task; SICR). In Experiment 1, participants recalled significantly more syllables, trigram words, and nonadjacent dependencies from test strings that conformed to the statistics of the artificial language compared to random controls, suggesting chunking of nonadjacent structure. In Experiment 2, we tested whether these results were due to the learning of positional information, rather than nonadjacent dependencies per se. Here, participants' reproductions of strings containing trained dependencies were compared to their reproductions of strings containing ungrammatical nonadjacent combinations (phantom words). Participants recalled significantly more trained dependencies than phantom words, but only on generalization trials, where bigram information was absent. This suggests that participants do chunk specific nonadjacent pairings rather than only acquiring positional information.