Proceedings of the Annual Meeting of the Cognitive Science Society

A well-known hypothesis for language learning is that learners are predisposed to learn patterns attested in human language, while eschewing alternative hypotheses that are linguistically unattested. Despite the popularity of this hypothesis, supporting experimental evidence has remained scarce. We aim to fill this gap by measuring the effectiveness with which adult learners can infer a linguistically attested pattern (a hierarchically nested constituency structure) in comparison to an unattested one (a constituency structure based on non-consecutive constituents) within a controlled artificial language learning task. We also investigate whether any observed learning biases extend to a non-linguistic, general puzzle-solving task. We find that our study participants are better at learning linguistically attested structures over unattested ones across the two types of learning domains. More broadly, our study serves as a methodological proof-of-concept extensible to testing the strength of other kinds of learning biases in linguistic and non-linguistic learning tasks.