Translation of English and Arabic “Sleep” Terms and Formulaic Expressions by Artificial Intelligence: A Comparison of Copilot and DeepSeek

Abstract

This study examined how Microsoft Copilot (MC) and DeepSeek (DS) translate sleep terms vs formulaic expressions from English to Arabic and Arabic to English, compared MC and DS’s translations, identified the percentage of accurate equivalents, types of errors made, the translation strategies used, and AI translation error causes. Analysis of 130 English sleep terms, 91 English sleep formulaic expressions and 110 Arabic terms and formulaic expressions revealed that MC and DS rendered 91% correct equivalents to English sleep idioms, 79% and 71.5% correct equivalents to English formulaic expressions respectively as She wept herself to sleep بكت حتى نامت, and 48% and 49% of the Arabic terms and formulaic expressions respectively as نوم قلق restless sleep. The most common translation strategy was literal, word for word translation as غرق في النّوم He drowned in sleep, instead of he is fast asleep); نمت كالقتيلة I slept like the dead instead of slept like a log. DS gave an explanation or annotation following the equivalent in 14% and MC in 3% of the Arabic items as inطار النَّومُ من عينيه Sleep flew from his eyes (couldn't sleep). The tendency to Arabic sleep items literally based on surface meaning because is due to an AI default literal translation strategy. AI systems tend to flatten nuance. Higher percentages of correct equivalents to English sleep items were given in this study than medical terms, zero-expressions, expressions of impossibility, Gaza-Israel War terminology, Arabic grammatical terms used metaphorically, Abu & Umm medical folk terms, Abu brand names and Abu & Umm metonymic animal and plant folk names. English-Arabic translation was easier than Arabic-English translation due to training bias, as most AI models are trained on English-dominant corpora, with English-Arabic translation receiving more attention and refinement than Arabic-English translation. Additionally, Arabic-English translation is less represented in the AI corpora, especially for domain-specific or idiomatic content, leading to lower performance. The study gives recommendations for enhancing Arabic idiom recognition and disambiguation by AI models, adding Arabic dictionaries, such as Almaany Dictionaries, to the AI corpora and for translators to make the best use of AI in translation.