Proceedings of the Annual Meeting of the Cognitive Science Society

Various studies have recently shown that the long-held claim that the relation between the sound of a word and its meaning is arbitrary needs to be revisited. In two computational studies we investigated whether word valence can be derived from sound features in English, Dutch and German. In Study 1, we identified the extent to which individual phonological features explained valence scores per language separately. In Study 2, we aimed to determine the optimal combination of cues that can predict valence scores across the three languages using two statistical classifiers and four machine learning classifiers. Our results showed that frequency and word complexity were the most reliable shared cues to predict valence for all three languages, obtaining a correct valence classification of about 60%. This percentage could be enhanced for individual or pairs of languages using additional relevant cues. These findings demonstrated that the claim that the relation between the sound of a word and its meaning is arbitrary is too strong.