This paper studies the classification and generation of irony using large-scale language models. We note that existing models struggle with the subtle nature of irony, and propose the Sarc7 benchmark, which classifies seven types of irony (self-deprecating, gloomy, neutral, polite, offended, furious, and manic) based on the MUStARD dataset. We evaluate classification performance using zero-shot, few-shot, Chain of Thought (CoT), and a novel emotion-based prompting technique. We then identify key elements of irony (incongruity, shock value, and context dependence) and propose an emotion-based generation method. Experimental results show that the Gemini 2.5 model using emotion-based prompting outperforms other configurations, achieving an F1 score of 0.3664. Furthermore, human raters prefer emotion-based prompting over zero-shot prompting (a 38.46% increase in successful generation rate).
