Child's play: A computational method for generating child-specific valence norms

Research on affective processing depends on the availability of word norms. Norms for adults are widely available and can be computed for whole-language vocabularies by mapping semantic similarities from a semantic vector space to a list of label words. While such methods are abundantly studied for adults, comparable methodologies for children are few and far between. In this article, we explored whether computational methods can be used to estimate word norms for children using valence as an example. We systematically investigated the impact of two main methodological choices relevant for the computation: the source of the semantic space used to measure semantic distance and the selection of the words in the label list. In Study 1, we correlated the computational estimates from all combinations with valence ratings from adults and children. Our results indicate that norms derived from adult vector space models and label lists outperform other combinations. Norms computed based on these parameters correlate with children's ratings at $r=.71$ and with adults' ratings at $r=.74$ . In Study 2, we used estimates and human ratings to predict valence effects in a lexical decision task. Again, we found that norms derived from vector space models and label lists for adults outperformed other combinations and approximated the functional form of children's and adults' valence effects best. We discuss the practical implications of these findings. Additionally, a new set of valence ratings from children (mean age = 12.5 years) for 535 German words is made available.