Language Models and the Magic of Metaphor: A Comparative Evaluation with Human Judgments

This study evaluates whether Italian-trained Large Language Models (LLMs) can interpret metaphors by comparing their performance to both human judgments and human-produced interpretations. Using three datasets containing metaphors, human interpretations, and implausible alternatives, we assess model performance via log-likelihood scores. Results show that LLMs partially replicate human understanding and are influenced by expression conventionality and linguistic context.