Yapay Zeka ile Şiir Çevirisinde Erişim Destekli Üretim (RAG) Mimarisi Yönteminin Çeviri Kalitesine Etkisi Üzerine Bir Değerlendirilme

The deployment of large language models in translation has increased substantially in recent years, yet their performance on literary texts, particularly poetry translation, remains contested. This study compares a Retrieval-Augmented Generation (RAG) based translation system against vanilla GPT-4o when translating classical Turkish Sufi poetry. Using Talât Sait Halman’s translations of Yunus Emre as reference corpus, the evaluation employed BLEU, METEOR, and ROUGE metrics across a test set of twenty-one poems. Results demonstrate statistically significant improvements across all metrics: 408 percent improvement in BLEU scores, 93 percent in METEOR, and 90 percent in ROUGE. Qualitative analysis revealed that the RAG system achieved perfect or near-perfect scores when retrieving appropriate reference translations from the vector database, though both systems struggled with mystical content and figurative language. These findings suggest that context-aware retrieval mechanisms can substantially enhance literary translation quality and offer a valuable assistive tool that could be integrated into translators’ workflows.