One major aspect of successful language acquisition is the ability to organize words into form class categories and generalize from properties of experienced items to novel items. Furthermore, learners must often determine how to use a new word, when there is very sparse information regarding its acceptable contexts. In this work we employ an artificial language learning paradigm to explore how adult learners, under circumstances of varying distributional cues to category boundaries, apply their knowledge of category properties to a new word. We find that in cases of strong category cues and strong category learning, adults readily generalize all of the distributional properties of the learned category to a word that shares just one context with the other category members. However, as the distributional cues regarding the target category become sparser and contain more systematic gaps, learners show more conservatism in generalizing the allowable distributional properties to the novel word. Taken together, these results show striking flexibility in learners tendency to generalize, depending on the distributional properties of the input corpus, in a probabilistically rational way.