Research in theoretical linguistics has shown that human languages require abstract and highly detailed grammaticalrepresentations. However, we understand surprisingly little about the mechanisms through which these representations areacquired. What kinds of statistical relationships would learners need to compute to construct representations like those positedby linguistic theory? We created miniature languages containing patterns found in natural languages and also patterns notfound in natural languages. We showed that complex word-order contingencies are acquired only when they correlate withmorphological patterns like those in natural languages. We then asked how learning changes when the statistical evidence forthese patterns is manipulated. These experiments illuminate the nature of learners’ computations and the units over which theyare performed.

We report a series of artificial language learning experiments designed to test child and adult learners abilities to acquirethree types of phonological variation. Previous work on experimental morphology (Hudson Kam & Newport 2005, 2009;Schuler, Yang & Newport, 2016) has found that young children turn inconsistent input into an invariant rule, while adultsreproduce and match variation in their input. Here we investigate whether phonological variation of three different types(deterministic conditioning, unconditioned variation, and probabilistic variation) exhibits a similar age pattern. We finda clear effect of age in grammatically-conditioned variability, with the youngest children showing a strong tendency toregularize to the stem form, adults probability-matching, and intermediate-aged children learning correct conditioning butnot matching the input probabilities. These results suggest, in accord with previous findings on morphology, that variationis not readily learned by young children and may instead be acquired as a separate process.

In natural languages, closed-class items predict open-classitems but not the other way around. For example, in English, ifthere is a determiner there will be a noun, but nouns can occurwith or without determiners. Here, we asked whether languagelearners’ computations are also asymmetrical. In threeexperiments we exposed adults to a miniature language withthe one-way dependency “if X then Y”: if X was present, Ywas also present, but X could occur without Y. We createddifferent versions of the language in order to ask whetherlearning depended on which of these categories was an open orclosed class. In one condition, X was a closed class and Y wasan open class; in a contrasting condition, X was an open classand Y was a closed class. Learning was significantly betterwith closed-class X, even though learners’ exposure wasotherwise identical. Additional experiments demonstrated thatthe perceptual distinctiveness of closed-class items driveslearners to analyze them differently; and, crucially, that theprimary determinant of learning is the mathematicalrelationship between closed- and open-class items and not theirlinear order. These results suggest that learners privilegecomputations in which closed-class items are predictive of,rather than predicted by, open-class items. We suggest that thedistributional asymmetries of closed-class items in naturallanguages may arise in part from this learning bias.