From Sentiment to Valence in Metaphor: a Comparison of BERT-based Sentiment and Prompted Large Language Models

Abstract

Although the affective dimension is a key aspect of metaphor, computational studies of figurative language have largely overlooked psycholinguistic variables such as valence. This study investigates whether computational models can reliably estimate the affective aspects of Italian and German metaphors and whether metaphor valence is compositionally derived. Outputs of BERT-based sentiment analysis and a valence-prompted LLM were compared with human ratings. Results show that the former exhibit limited alignment with human judgments, whereas higher agreement is achieved when the explicit concept of valence is prompted in a LLM. Both humans and models rely on the combined valence of the individual lemmas, suggesting a compositional contribution to metaphor valence.

Resources

Details