人称表現の日韓･日中翻訳から見る機械翻訳と生成AIの課題: 誤訳様相を中心に

Abstract

This study investigates the challenges faced by machine translation systems and generative AI in translating Japanese address forms into Korean and Chinese, through comparison with human translation. When Korean is the target language, Papago, DeepL, and ChatGPT share common difficulties in handling interpersonal relationships and translating personal names. Papago, in particular, exhibits significant issues in aligning referents and selecting appropriate demonstratives between Japanese and Korean. When Chinese is the target language, Baidu, DeepL, and ChatGPT encounter similar problems with referent alignment and name translation, while DeepL frequently produces recognition errors when processing long sentences. To address these issues, this study proposes pre-processing and post-processing strategies for human translator intervention. For Korean translations, pre-processing includes tagging honorific systems and adding phonetic guides for names, while post-processing involves validating honorific appropriateness, checking name representation, correcting referent alignment, and reviewing address form translation. For Chinese translations, pre-processing includes semantic tagging of address forms, converting Japanese phonetic names into Chinese characters, establishing sinicization guidelines, and segmenting long sentences; post-processing focuses on checking naturalness of address forms, verifying name representation, and confirming translation accuracy. These strategies aim to enhance human–machine collaboration and improve translation quality in complex linguistic contexts.