Close encounters of the word kind: Attested distributional information boosts statistical learning

Statistical learning, the ability to extract regularities from input (e.g., in language), is likely supported by learners' prior expectations about how component units co-occur. In this paper, we investigated how adults' prior experience with sublexical regularities in their native language influences performance on an empirical language learning task. Forty German-speaking adults completed a speech repetition task, in which they repeated eight-syllable sequences from two experimental languages: one containing bisyllabic words comprised of frequently occurring German syllable transitions (naturalistic words), and the other containing words made from unattested syllable transitions (non-naturalistic words). Participants demonstrated learning from both naturalistic and non-naturalistic stimuli. However, learning was superior for the naturalistic sequences, indicating that participants used their existing distributional knowledge of German to extract naturalistic words faster and more accurately than non-naturalistic words. This supports theories of statistical learning as a form of chunking, whereby frequently co-occurring units become entrenched in the long-term memory