Visual similarity partly explains adult and model word extension
Original title: Visual Similarity in Adult Word Extension and Artificial Word Learning
DINOv2 distance tracked adult extension, and an artificial learner was less accurate for farther category pairs.
A reanalysis of Lewis and Frank's 2018 adult word-extension data found DINOv2 visual distance associated with whether adults extended a word. Example composition also mattered: after three subordinate examples, adults extended to 1.1% of superordinate items, versus 77.5% after superordinate examples. An artificial learner using the same encoder named near, mid, and far category pairs at 84.7%, 65.4%, and 53.0%.
Unconfirmed: Human-derived distances did not conclusively predict learner difficulty, and secondary analyses were not externally preregistered.
- Distance linked to extension at log odds -3.23 per standard deviation across 520 people.
- Encoder distances correlated with human similarity measures at rho 0.501 to 0.795.
- A child-experience encoder reproduced tier order but not the pair-level effect.
- Authors say example composition explains generalization that distance does not capture.
From the original article
1 min read at huggingface.co
When two objects share a name, how much does their visual similarity explain, and what do the teaching examples add? First, we reanalyzed trial-level data from a large adult word-extension study (Lewis & Frank, 2018), computing each test image's distance to the examples shown in the space of a frozen self-supervised visual encoder (DINOv2). Distance was associated with extension (log odds -3.
By Patricio M. Vera · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.