Forget More to Learn More: Domain-Specific Feature Unlearning for Semi-supervised and Unsupervised Domain Adaptation

Basak, Hritam; Yin, Zhaozheng

doi:10.1007/978-3-031-72920-1_8

We can easily evaluate similarities between concepts within semantic domains, e.g., doctor and nurse, or violin and piano. Here, we show that people are also able to evaluate similarities across domains, e.g., aligning doctors with pianos and nurses with violins. We argue that understanding how people do this is important for understanding conceptual organization and the ubiquity of metaphorical language. We asked people to answer questions of the form "If a nurse were an animal, they would be a(n)…" (Experiment 1 and 2), and asked them to explain the basis for their response (Experiment 1). People converged to a surprising degree (e.g., 20% answered "cat"). In Experiment 3, we presented people with cross-domain mappings of the form "If a nurse were an animal, they would be a cat” and asked them to indicate how good each mapping was. The results showed that the targets people chose and their goodness ratings of a given response were predicted by similarity along abstract semantic dimensions such as valence, speed, and genderedness. Reliance on such dimensions was also the most common explanation for their responses. Altogether, we show that people can evaluate similarity between very different domains in predictable ways, suggesting that either seemingly concrete concepts are represented along relatively abstract dimensions (e.g., weak-strong) or that they can be readily projected onto these dimensions.

More Like this