Linguistic Rule Generalisation Creates the Same Distributional Structure That Feeds It

Part of language's great expressivity comes from its users creating new forms by applying familiar rules to novel items. But linguistic rules aren't all created equal-some are more readily generalisable than others. In this paper, we focus on how rule generalisation is affected by certain properties of frequency distributions. In an artificial language learning experiment that asks adult learners to generalise using one of two suffixes, we find that they probability-match their input but slightly prefer whichever suffix they encountered with more low-frequency stems. Then with an urn model of learning, we show that previous explanations of generalisation that focus only on a distribution's type count or its skew fail to capture participants' behaviour-only the low-frequency preference yields convergent results. We model learners' behaviour in terms of rational Bayesian inference about how likely a rule is to apply to more word types than somebody has already encountered. Overall, we suggest that linguistic rule generalisation is a self-sustaining process: by creating novel and therefore low-frequency items, rule generalisation produces the very same distributional structure that feeds it.