Title: IdiomX: A Multilingual Benchmark for Idiom Understanding, Retrieval, and Interpretation

Abstract

Idiomatic expressions continue to pose a significant obstacle for natural language processing, primarily due to their non-compositional nature, context sensitivity, and the difficulty of aligning meanings across different languages. Current idiom resources frequently suffer from limitations in scale, contextual variety, or multilingual scope, which hinders their effectiveness for contemporary language models. To address these gaps, we present IdiomX, a comprehensive multilingual benchmark designed for idiom understanding, retrieval, and interpretation. This dataset was developed using a reproducible, multi-stage pipeline that integrates lexical resource extraction, large-scale normalization, controlled enrichment via large language models, and structured validation.

The resulting dataset comprises more than 190,000 contextualized examples covering over 12,000 idioms. It features aligned semantic representations in English, Arabic, and French, alongside labels for both idiomatic and literal usage, as well as extensive linguistic metadata. Leveraging this resource, we establish a unified four-task benchmark that encompasses idiom detection, context-to-idiom retrieval, Arabic-to-English idiom retrieval, and idiom interpretation. This framework expands evaluation capabilities beyond figurative recognition to include semantic grounding and the retrieval of explainable meanings.

Our experiments reveal that contextual transformer models markedly enhance idiom detection performance. Furthermore, hybrid retrieval and reranking architectures substantially improve both monolingual and cross-lingual idiom retrieval. The results also indicate that idiom interpretation can be effectively treated as a semantic retrieval task, thereby introducing interpretability as a novel dimension for benchmarking. Ultimately, IdiomX offers a scalable framework for investigating idiomatic language as a continuum from detection to retrieval and semantic interpretation, with a modular design that can be extended to additional languages and figurative reasoning tasks.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC