Utilizing a Cultural Translation Evaluation Framework for Semantic Drift in AI-Based Multilingual Translation
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This study investigates how three state-of-the-art AI translation models—GPT-4, DeepL, and NLLB-200—handle semantic drift, cultural fidelity, and affective fidelity in the translation of 100 culturally embedded Korean-language sentences. Semantic drift, defined as the diachronic or contextual evolution of word meanings, serves as a critical lens through which the preservation of cultural nuance in translation is assessed. Experimental findings reveal that GPT-4 demonstrates the highest level of performance in both contextual interpretation and emotional expressiveness, effectively preserving the cultural logic of the source text. DeepL excels in grammatical precision but falls short in cultural interpretation. NLLB-200, while offering broad multilingual coverage, exhibits limited cultural sensitivity and fails to adequately capture symbolic meaning. Based on these results, the study proposes a complementary qualitative evaluation framework that addresses the limitations of conventional quantitative metrics such as BLEU. It further outlines a future direction for AI translation systems that prioritizes not only linguistic accuracy but also cultural receptivity and ethical interpretability in multilingual,
Abstract
This study investigates how three state-of-the-art AI translation models—GPT-4, DeepL, and NLLB-200—handle semantic drift, cultural fidelity, and affective fidelity in the translation of 100 culturally embedded Korean-language sentences. Semantic drift, defined as the diachronic or contextual evolution of word meanings, serves as a critical lens through which the preservation of cultural nuance in translation is assessed. Experimental findings reveal that GPT-4 demonstrates the highest level of performance in both contextual interpretation and emotional expressiveness, effectively preserving the cultural logic of the source text. DeepL excels in grammatical precision but falls short in cultural interpretation. NLLB-200, while offering broad multilingual coverage, exhibits limited cultural sensitivity and fails to adequately capture symbolic meaning. Based on these results, the study proposes a complementary qualitative evaluation framework that addresses the limitations of conventional quantitative metrics such as BLEU. It further outlines a future direction for AI translation systems that prioritizes not only linguistic accuracy but also cultural receptivity and ethical interpretability in multilingual,