When Figures Speak with Irony: Investigating the Role of Rhetorical Figures in Irony Generation with LLMs

Irony poses a persistent challenge for computational models because it depends on context, implicit meaning, and pragmatic cues. This study investigates the ability of Large Language Models (LLMs) to generate ironic content by focusing on rhetorical figures—pragmatic devices that may shape and signal ironic intent. Using two datasets, TWITTIRÒ-UD and the Italian subset of MultiPICo, we fine-tune multilingual LLMs for rhetorical figure classification and evaluate their capacity to generate ironic Italian texts. Our work addresses two main questions: (1) how accurately LLMs can classify rhetorical figures in ironic Italian texts, and (2) whether such training supports the generation of irony that reflects human-like rhetorical usage. Human evaluation shows that LLMs achieve fair agreement with annotators in rhetorical figure classification, indicating a partial but promising alignment with human judgment. By leveraging rhetorical figures as a bridge between irony detection and generation, our results suggest that such training improves the stylistic control and interpretability of LLM-generated ironic language.