생성형 AI의 근대한어 번역 성능 분석 — ≪老乞大≫와 ≪朴通事≫를 대상으로 = Analysis of Generative AI Performance in Translating Early Modern Chinese:Focusing on Nogeoldae(老乞大) and Baktongsa(朴通事)

다국어 초록 (Multilingual Abstract)

This study evaluated the performance of three generative AI models—OpenAI’s ChatGPT-5, Google’s Gemini 2.5, and China’s DeepSeek-V3—in translating sentences from Early Modern Chinese(EMC) conversation textbooks (Nogeoldae and Baktongsa). Evaluation focused on speaker identification and chunking, grammar/vocabulary explanation, idiom/quotation interpretation, and cultural background understanding. Quantified out of 40 points, DeepSeek-V3 scored highest (31), followed by Gemini 2.5 (28), and ChatGPT-5 (24.5). DeepSeek excelled in chunking and translation, possibly due to better access to rich Chinese EMC data. Gemini 2.5 showed strength in grammar and logical cultural inference, whereas ChatGPT-5 exhibited the most hallucinations. The findings confirm AI’s benefit in reducing translation time and effort. However, the study identifies critical limitations: potential errors from limited EMC training data and the risk of subtle, hard-to-find errors caused by hallucination. Utilizing multiple AIs for cross-validation, while acknowledging these limitations, is recommended for high-quality EMC research. This is the first study to review generative AI EMC translation across these diverse items.