Testing the Effectiveness of Prompt Engineering Technique in the Translation of CSIs using LLMs: Naguib Mahfouz’s Midaq Alley (1947) a Case in Point | ||
| TANWĪR: A Journal of Arts & Humanities | ||
| Volume 2, Issue 3, November 2025, Pages 44-72 PDF (1.01 M) | ||
| Document Type: Original Article | ||
| DOI: 10.21608/tanwir.2025.464396 | ||
| Author | ||
| Noura Wael Aly Kamel | ||
| English Department, Faculty of Languages and Translation, Ahram Canadian University | ||
| Abstract | ||
| Large Language Models (LLMs) are the new architects of linguistic intelligence. Throughout the past two years, they are rewriting the boundaries of specialized translation. Translation of literary texts from Arabic into English, especially those exhibiting many culture specific items (CSIs) that are deeply embedded in the context of the source language, presents a unique challenge for automated systems because of the historical references, religious allusions, and local idioms. Technological techniques are now presented trying to preserve the cultural nuances and literary style. In this regard, this paper investigates the effectiveness of five prompt engineering methods in guiding the LLMs to accurately translate the CSIs found in selected texts of culturally dense passages in Mahfouz’s Zuqāq al-Maddaq [Midaq Alley] (1947). In addition, the paper aims at identifying the main drawbacks found in the translated text along with determining the translation strategies used. Two LLMs are used; the UAE TII Falcon-H1-7B-Instruct and Aya-expanse-8B, Cohere’s recently released state-of-the-art LLM. After applying the five methods on both LLMs, translated texts are evaluated based on cultural adequacy, and semantic meaning using five automatic evaluation metrics to assess the extent to which they correspond with the human analysis. Moreover, an analysis based on Davies’ (2003) translation strategies of CSIs and Newmark’s (1988) taxonomy is conducted to describe how both models have performed in dealing with translating different types of CSIs. Then, MQM is used to categorize the error according to their error typology. This paves the way for more advancements in the field of translation technology to achieve a contextually faithful and culturally nuanced machine-generated translated text. | ||
| Keywords | ||
| Large Language Models; Prompt Engineering; MQM; culture-specific items; Falcon-H1-7B-Instruct; Aya-expanse-8B | ||
|
Statistics Article View: 108 PDF Download: 14 |
||
