Irony Detection, Reasoning, and Understanding in Zero-Shot Learning
Peiling Yi, Yuhan Xia, Yunfei Long · IEEE Transactions on Artificial Intelligence · 2025
Irony is a powerful figurative language (FL) on social media that can potentially mislead various NLP tasks, such as recommendation systems, misinformation checks, and sentiment analysis. Understanding the implicit meaning of this kind of subtle language is an essential step to mitigate the negative impact of irony in NLP tasks. However, existing efforts are limited to domain-specific datasets and struggle to generalize across diverse real-world scenarios. Moreover, reasoning for model decisions that accurately capture semantic and affective meaning remains underexplored. To address these limitations, this paper proposes a conceptual framework called IDADP, which leverages Large language models(LLMs)’ in-context learning capabilities to detect irony and generate human-like explanations across diverse datasets and platforms without prior training on ironic samples. Extensive experiments on six widely used irony detection datasets, utilising two large language models (GPT and Gemini), demonstrate that IDADP consistently outperforms six competitive zero-shot baselines and approaches the performance of three fine-tuned supervised learning baselines. Additionally, we examine GPT’s ability to understand the true intent behind ironic text within the IDADP framework, highlighting its strong potential to recognize and interpret statements where the intended meaning differs from or contrasts with the literal meaning. Furthermore, we conduct qualitative analyses to identify remaining challenges. This work, in turn, opens an avenue for transparent decision-making in irony detection.