Statistics All the Way Down: How is Statistical Learning Accomplished Using Varying Productions of Novel, Complex Sound Categories?

For statistical learning to aid in language learning, learners must resolve statistical information along multiple dimensions of the same linguistic signal. Given that infants show evidence of lexical knowledge while they are still learning how to categorize speech, infant learners are likely presented with at least two statistical learning problems simultaneously. In an effort to approximate this scenario, we presented adult participants with multiple exemplars of sounds from 4 experimenter-defined categories. These sounds were novel and thus, adult have not developed specialized processing for these sounds. Stimuli were presented in a regular, continuous stream containing statistical structure between sound-category types with variable exemplars (i.e. pairs of sound categories but with variable exemplars of each category presented instead of just one). Participants were tested for familiarity with high probability pairs. We found that participants can learn from statistical structure based on varying exemplars of novel sounds but they learn based on the perceptual grouping biases that they bring into the experiment and not based on the experimenter-defined categories (groupings they would have to form ad hoc in the experiment). We discuss these results in relation to language learning.