Prior research suggests that preschoolers can generalize object properties based on category information conveyed by semantically-similar labels. However, previous research did not control for co-occurrence probability of labels in natural speech. The current studies re-assessed children's generalization with semantically-similar labels. Experiment 1 indicated that adults made category-based inferences regardless of co-occurrence probability; however, 4-year-olds generalized with semantically-similar labels that co-occurred in child-directed speech (e.g., "bunny-rabbit") but not with non-co-occurring labels (e.g., "crocodile-alligator"). Experiment 2 indicated that generalization with semantically-similar labels increased gradually between 4- and 6-years of age. These results are discussed in relation to theories of early learning. (Contains 3 tables and 5 figures.)