Studies in novel word learning show that comparison settings (i.e., several stimuli introduced simultaneously) favor taxonomically-based generalization. Most comparison studies have been done with forced-choice designs. Here, we investigated, in a free-choice comparison design the type of stimuli four and five-year-old children chose in a novel noun generalization task, either from the same basic level category, or a near superordinate category, or a distant superordinate category, but also perceptual, thematic, and unrelated lures. Same basic level category items were more chosen than other taxonomically related items. Perceptual lures and near superordinate items did not differ, suggesting that children did not arbitrate between perception and taxonomy. Results are discussed in terms of different theoretical perspectives on stimulus generalization, lexical constraints, stimulus comparison and finally Bayesian approaches. They suggest that children integrate the results of their comparison rather than sampling probabilistic regularities