Visual Similarity in Adult Word Extension and Artificial Word Learning
When two objects share a name, how much does visual similarity explain, and what do the teaching examples add? A reanalysis of trial-level data from a large adult word-extension study (Lewis & Frank, 2018; 520 people) measures each test image's distance to the taught examples in the space of a frozen DINOv2 encoder: greater distance goes with less extension (log odds −3.23 per SD), and example composition matters as well. In an artificial learner built on the same encoder, words joining two categories are learned less accurately the farther apart the categories are (84.7%, 65.4%, 53.0% naming accuracy for near, mid, and far pairs). The encoder's distances track human similarity measures (ρ = .501–.795), yet that alignment does not yield conclusive evidence that human-derived distances predict learner difficulty.
