• conventional metaphors
• distributional semantic
• language model
• metafore
• metafore convenzionali
• metafore creative
• metaphors
• modelli computazionali
• novel metaphors
• plausibilià semantica
• semantica distribuzionale

Metaphor is a widespread linguistic and cognitive phenomenon. Many studies were carried out to investigate how humans understand and produce metaphors. A key aspect of the phenomenon is the difference between frozen and creative metaphors. It was shown that humans interpretate in a different way conventional metaphors and novel metaphors and that humans are sensible to this difference. Our study confirms that result. Furthermore, metaphors, especially creative ones, are difficult to model computationally. Recent progress has been made in metaphor identification, also thanks to the contextualized embeddings from models like BERT. To test what BERT, RoBERTa and GPT2 know about metaphors, we challenge them with a new dataset of conventional and creative metaphors accompanied by various types of human judgments. We find that the models can "recognize" metaphors and shows interesting abilities like that of predicting creative metaphors. At the same time, we show that the models still struggle in "interpreting" metaphorical language, even if it outperforms traditional static vectors. Our findings confirm previous claims about the abilities and limitations of the models. Furthermore, they show that RoBERTa has better performance than BERT and GPT2 in the first experiment and that BERT large is almost good in "interpreting" metaphors in the upper-intermediate layers as suggested by the results of the second experiment. Finally, the models seem to show some similarities with humans, but still miss a significant part of human intuitions about the meaning of metaphors.