Copilot and Gemini’s Translation of English and Arabic "Half" -Based Expressions

Abstract

This study compares the translation accuracy of two AI models: Microsoft Copilot (MC) and Google Gemini (GG), in translating English and Arabic half‑expressions, the direction that is easier, types of translation errors they commit and the translation strategies they utilize. A sample of 100 Arabic and 100 English half‑expressions was randomly selected from Al‑Maany Dictionaries and Google searches, representing both general‑language and specialized‑domains. Results showed that in the Arabic half‑expressions set, MC and GG produced 80% correct equivalents, consisting of 61% identical equivalents in form and meaning and 19% correct equivalents with variant wording. Additionally, GG supplied 12% extra correct equivalents not produced by MC, raising its total accuracy to 92%, compared to 80% for MC. In the English half‑expressions set, MC and GG produced 56% correct equivalents, including 21% identical equivalents in form and meaning and 35% correct equivalents with variant wording. GG provided 24% additional correct equivalents, while MC supplied 9% additional correct equivalents. Overall, GG achieved 80% accuracy, compared to 65% for MC. Error analysis revealed consistent MC and GG failures in culture‑specific expressions, social and colloquial Arabic expressions, commercial‑domain terminology (furniture, fabrics, vehicles), expressions with polysemous words, word‑order errors, and cases where GG produced mixed correct and incorrect equivalents. The most common translation strategies utilized by MC and GG included literal translation, explanation (specifying the domain, context in which the expression is used, and a definition), giving multiple equivalents, and English equivalents containing semi, hemi, mid, middle, and demi. Transliteration was the least frequent. Arabic‑to‑English translation was easier than English‑to‑Arabic as reflected by a higher percentage of correct equivalents and differences in lexical density and synonymy. Compared with the author’s prior studies, half‑expressions were significantly easier for AI to translate than zero‑expressions, Abu/Umm‑brand names, animal and plant names, and Gaza–Israel war terms, and Arabic grammatical terms used metaphorically. Explanations are provided for the highly identical equivalents given by MC and GG, why Arabic-English translation was easier than English-Arabi translation, and why GG outperformed MC. Recommendations for AI developers and translation students are given as well.