Cross-Cultural Idiom Understanding in Large Language Models

Tue, 13 Oct 2026 00:00:00 +0000

Context

An idiom is an expression whose meaning does not follow from its words, such as “break a leg”. Large language models (LLMs) are neural models trained on large text collections, and they often read idioms literally.

Idioms also vary across cultures. Some idioms are shared across languages, with the same image and meaning. Others keep the meaning but change the image: the Italian “in bocca al lupo” and the English “break a leg” both wish good luck. Others belong to one culture only. Existing datasets, including SemEval shared tasks, already pair idioms with their meanings in many languages.

Objective

This proposal tests whether LLMs grasp shared, image-different, and culture-specific idioms, without new annotation. The student builds the three groups from existing datasets and compares models across languages.

The three groups of idioms this proposal compares, on examples chosen for illustration rather than taken from a dataset. In each group, an Italian idiom and its English counterpart point to one meaning, and the culture-specific idiom has no English counterpart.

Linguistics | Language Technologies Lab

Cross-Cultural Idiom Understanding in Large Language Models

Context

Objective