Proceedings of the Annual Meeting of the Cognitive Science Society

To what extent can the similarity structure of categories beinferred based on paradigmatic vs syntagmatic information?We explore this question in two studies that aim to captureparadigmatic information directly: first by having participantsgenerate near-neighbors to exemplars from 15 basic categories,and second by having them partially rank the most similar ex-emplars. After constructing neighborhood graphs of the itemsin each category, we derived a local measure (based on di-rect neighbors) and a global measure (including indirect pathsas well) of paradigmatic information. Both measures predictindependently-obtained human pairwise similarities for eachcategory, but incorporating indirect information substantiallyimproves this prediction. In a third study, we contrast thesemeasures with syntagmatic information obtained from a vast se-mantic network derived from 3 million judgments. The paradig-matic graphs are better predictors of similarity despite onlyencoding a fraction of these data. Broad implications for wordlearning and meaning are discussed.