Proceedings of the Annual Meeting of the Cognitive Science Society

A question that has long interested cognitive scientists is how to best represent the different emotions we experience and attribute to others. For example, constructionist and appraisal theories propose that differences between emotions can be captured in part by their variation along a set of appraisal dimensions. More recently, researchers have used language models to capture the differences across different emotion terms. Both approaches allow us to represent emotions as occupying different locations in high-dimensional representational spaces. To ask how well these different approaches capture the similarity between emotion concepts, we collected pairwise similarity and appraisal feature ratings for 58 different emotion concepts and then employed representational similarity analysis to investigate the overlap between people’s pairwise similarity judgments and emotion similarity in a 14-dimensional appraisal space and three word embedding spaces from two word2vec models (300 dimensions) and the newer GPT-3 model (12288 dimensions). The results indicate that while there is a high correlation between appraisal feature-based similarity and pairwise similarity judgments, word embedding-based similarity exhibits lower correlations, though GPT-3 showed much better performance than the word2vec models. Finally, characterizing the errors made by word embedding models showed that they can be largely attributed to an over-reliance on the valence of emotion concepts.