Learning word meaning with little means: An investigation into the inferential capacity of paradigmatic information

AbstractTo what extent can the similarity structure of categories be inferred based on paradigmatic vs syntagmatic information? We explore this question in two studies that aim to capture paradigmatic information directly: first by having participants generate near-neighbors to exemplars from 15 basic categories, and second by having them partially rank the most similar exemplars. After constructing neighborhood graphs of the items in each category, we derived a local measure (based on direct neighbors) and a global measure (including indirect paths as well) of paradigmatic information. Both measures predict independently-obtained human pairwise similarities for each category, but incorporating indirect information substantially improves this prediction. In a third study, we contrast these measures with syntagmatic information obtained from a vast semantic network derived from 3 million judgments. The paradigmatic graphs are better predictors of similarity despite only encoding a fraction of these data. Broad implications for word learning and meaning are discussed.

Return to previous page