Visual Similarity in Adult Word Extension and Artificial Word Learning

Visual Similarity in Adult Word Extension and Artificial Word Learning. By Patricio M. Vera, from Neurocreaciones. Published on the tenth of November.

When two objects share a name, how much of that is explained by how alike they look, and what do the teaching examples add? The answer offered here comes in two parts. Visual similarity accounts for part of how generalization is organized, in adults and in an artificial learner alike. The composition of the examples contributes something else, the part that the visual distance used here does not capture.

The first line of evidence is a reanalysis of trial-level data from a large adult word-extension study by Lewis and Frank. For each test image, the distance to the examples shown was computed in the space of a frozen self-supervised visual encoder, DINOv2. Distance was associated with whether people extended the word: log odds of minus three point two three per standard deviation, across five hundred and twenty people. An exploratory post hoc analysis suggested that this association persisted within the taught domain, after controlling for taxonomic level and teaching condition.

Example composition mattered as well, and by a wide margin. After three examples, adults extended a word to one point one percent of superordinate items when the examples were subordinate. When the examples were superordinate, that rose to seventy-seven point five percent.

The second line uses an artificial learner built on the same encoder. Words joining two categories were learned less accurately the farther apart those categories were. Naming accuracy was eighty-four point seven percent for near pairs, sixty-five point four percent for mid pairs, and fifty-three percent for far pairs. The encoder's distances also correlated with human similarity and confusability measures from three datasets, with rho between zero point five zero one and zero point seven nine five.

Several limits are stated, and they shape how far the claim reaches. That representational alignment with human judgments did not translate into conclusive evidence that human-derived distances predict the learner's difficulty. A second encoder, trained on one child's experience, reproduced the ordering across near, mid and far tiers, but not the effect at the level of individual pairs. The within-domain result in the adult data came from an exploratory post hoc analysis. And the secondary human data were analyzed under an internal plan fixed before the confirmatory analyses, which was not externally preregistered.

Picture the two forces side by side. Distance in a visual encoder's space predicts, to a degree, where a word gets extended and how hard a word is to learn. But the jump from one point one to seventy-seven point five percent came from changing which examples were shown, not from how similar the test items looked. Similarity sets part of the map, and the examples draw much of the boundary.