Creative Destruction: Can Language Models Interpret Oxymorons?

Fan Xu, Ziyun Zhu, Xiaojun Wan

Abstract: Oxymoron is a figurative language which combines seemingly contradictory words in a short phrase. It is used to create an impression, enhance a concept or entertain the readers. In this work, we propose a novel task named oxymoron interpretation, which requires destructing the contradiction and creating new senses to make the phrase comprehensible and harmonious with the context. We construct a dataset, Oxymoron with Context-Based Interpretation(OCBI) by gathering data from the internet and manually annotating it. Furthermore, some prevalent models and methods, including definition generation, GPT-2, T5, ChatGPT, LLaMA and Alpaca, are tested to explicitly generate interpretation. According to automatic and manual evaluation results, models except ChatGPT still struggle to generate plain and context-dependent interpretation, which brings us some insights of the capability gap between models. Our work reveals that oxymoron interpretation is a challenging task requiring common sense and contextual association ability. It can be used to assess how well language models comprehend figurative language. The OCBI dataset is publicly available (https://github.com/pku0xff/oxymoron_interpretation).