We outline and test a Bayesian model of the effects of evidencesampling on property induction. Our model assumes thatpeople are sensitive to the effects of different sampling framesapplied to sampled evidence. Two studies tested the model bycomparing property generalization following exposure tosamples selected because they belong to the same taxonomiccategory or because they share a salient property. Both studiesfound that category-based sampling led to broadergeneralization than property-based sampling. In line withmodel predictions, these differences were attenuated when amixture of positive and negative evidence was presented(Experiment 1) and when category-property relations wereprobabilistic rather than deterministic (Experiment 2).

Semantic structure in the mental lexicon is often assumed to follow a taxonomic structure grouping similar items. This study uses a network clustering analysis of a massive word association dataset that does not primarily focus on concrete noun categories, but includes the majority of the words used in daily life. At this scale, we found widespread overlap between thematically organized clusters, arguing against a discrete categoric view of the lexicon. An empirical analysis focusing on taxonomic categories confirmed the widespread thematic structure even for concrete noun categories in the animal domain. Overall, this suggests that applying network clustering to word association data provides valuable insight into how large-scale semantic information is represented. This analysis leads to a different, more thematic topology than the one inferred from idealized small-scale approaches that sample only specific parts of the lexicon