What Makes Large Language Models Interpret Context-Free Cross-Lingual Novel Idioms?

Idioms are particularly pervasive figurative expressions with the property of being non-compositional. This means that the meaning of an idiom cannot be deduced from the combination of its individual lexical entities. In most cases, the meaning of an idiom is understood through conversational context and learned through repeated exposure, with connections between literal and figurative senses often made post-hoc. Our work argues that interpretation does not arise solely from context or familiarity, but that an idiom's literal sense lends to its understanding through two key dimensions: decomposability (the syntactic mapping of the idiom's constituents to those of its figurative meaning) and transparency (the metaphorical motivation linking the literal and figurative senses of the idiom). To control for the effect of familiarity, we constructed a dataset of 80 idioms literally translated from Russian, Mandarin Chinese, Hindi, and Spanish (20 from each language). Through free-form and multiple-choice response experiments with 100 monolingual English speakers and 11 state-of-the-art large language models - which have demonstrated strong performance on reasoning tasks but have not been thoroughly tested with linguistic analysis tasks - we examined how decomposability and transparency influence idiom interpretation. Additionally, we developed annotation frameworks to systematically analyze these semantic properties, while also investigating the role of conceptual metaphors - underlying mappings between concrete source domains and abstract target domains - in facilitating interpretation. Results demonstrated that both decomposability and transparency have a significant effect on idiom interpretation. Decomposable idioms were 15.5% more likely to be correctly interpreted by humans and 20.3% more likely by LLMs compared to non-decomposable ones. For transparent idioms, the odds of correct interpretation were approximately 5-6 times higher than for opaque ones, with both culturally and universally transparent idioms showing similar advantages. Idioms with higher levels of transparency were also more frequently associated with near-universal conceptual metaphors and rarely contained specific cultural elements or hyper-specific conceptual mappings. The source language of idioms had minimal effect on interpretation accuracy, suggesting that semantic analyzability plays a more crucial role than language-specific features in determining an idiom's interpretability. From a computational perspective, larger LLMs consistently outperformed both smaller models and human participants, suggesting that increased model size may lead to improved linguistic reasoning abilities or latent embedded cross-lingual knowledge, rather than mere memorization of fixed idiom phrases and their use. In conclusion, our findings have implications for better understanding figurative language processing across cognitive, educational, and technical systems, particularly across language learning, building more detailed models of reasoning, and creating artificial systems that can interface conversationally with people across cultures and languages.