Idiomatic Expressions in English-to-Persian Translation: Human vs. AI Performance

Abstract

This study evaluates and compares the performance of two AI-powered large language models (LLMs), namely ChatGPT and DeepSeek, in translating English idiomatic expressions into Persian. Drawing on a typology of idiom translation strategies and employing a mixed-methods approach, this research study integrates both quantitative analysis of translation accuracy and qualitative examination of strategy use by both AI- and human-translators. A total number of 75 idioms, selected from five thematic categories, Age, Beauty, Family, Food, and Clothes, were translated by the two AI models and benchmarked against expert-validated human translations. The idioms were sourced from a frequency-sorted thematic index and evaluated for acceptability by five Iranian experts in English language with final accuracy checks by a Persian literature specialist. The research findings indicate a clear advantage for ChatGPT over DeepSeek in both accuracy and strategic alignment with human translators. Quantitatively, ChatGPT has produced a significantly higher rate of acceptable translations, while DeepSeek has generated over twice as many non-acceptable and often irrelevant translations. Qualitatively, both models have employed various translation strategies, yet ChatGPT has shown more alignment with human translation norms and superior capacity in handling the complexities of idiomatic language than DeepSeek.