Proceedings of the Annual Meeting of the Cognitive Science Society

Generalisation of a linguistic rule can be facilitated by certain distributional characteristics. Previous work has shown that a rule is better generalised if it applies to items that (i) follow a skewed frequency distribution, or (ii) follow a uniform frequency distribution over many distinct item types. These two observations cannot be unified under explanations of rule generalisation that are based on entropy of the frequency distributions (since skewed distributions have low entropy, while a greater type count increases the entropy), nor explanations that focus on one highly-frequent type providing a basis for analogical extension (since all types in uniform distributions are equally frequent). Using an artificial language learning experiment and an agent-based model, we show that participants' generalisation behaviour is best matched by a model encoding preferential generalisation of rules containing long-tailed distributions—that is, containing a greater number of low-frequency types.