Earticle

Generative AI has revolutionized machine translation (MT) by enabling customizable outputs through prompt engineering. Unlike traditional neural machine translation (NMT) systems such as Google Translate and DeepL, which rely on predefined translation patterns, models like GPT-4 allow users to refine translation quality through different prompting strategies. This study examines the impact of Zero-shot, Few-shot, Style-specific, and Step-by-step prompting on translation quality and evaluates the effectiveness of Auto Prompt Optimization (APO) in further enhancing performance. Using GPT-4, translations were conducted for Korean-to-English, Japanese, Chinese (Ko→En, Ja, Zh) language pairs. The study assessed BLEU, TER, and COMET scores to measure accuracy, fluency, and semantic consistency. Results show that Few-shot and Step-by-step prompting significantly improve translation quality, while APO further enhances BLEU scores and reduces TER values across all prompting types. Findings also indicate that prompt effectiveness varies by language pair, emphasizing the need for language-specific prompt strategies. This study provides empirical evidence of how prompt engineering influences AI translation quality and offers strategic guidelines for optimizing translation performance.