Smart brands, trustworthy people, and evil words: Sub-lexical distributional patterns capture human semantic...

We show that given names, pseudowords, and made-up company names elicit robust associations across four semantic dimensions (femininity, evilness, smartness, and trustworthiness) in a best-worst scaling survey administered to Dutch native speakers. Associations were strongest for semantic dimensions associated with conventional morphological markers (gender, evilness) but did not differ substantially across item types. Critically, items varied in the degree to which they belong in native speakers' linguistic experience: names are experienced, pseudowords and made-up company names are not, but the former were designed to be indistinguishable from Dutch words and thus obey orthotactic constraints, whereas the latter need not. We also used a sub-lexical distributional semantic model to extract an unsupervised metric quantifying the degree of association between a string and a semantic dimension, which we then correlated with human ratings. When reliable sub-lexical patterns were available, participants could exploit them in rating strings (robust correlations for pseudowords on all four target dimensions), but relied on other sources of information both when experience with the full form was available and the semantic dimension was not associated with conventional morphological markers (given names and smartness/trustworthiness), or when strings diverged from typical sub-lexical patterns (made-up company names). These results further characterize the relation between word form and lexical meaning and the mechanisms that support semantic associations for known and novel strings.