한국어 무주어 구문의 영어 번역 양상: 인간번역, 구글번역, 챗GPT 간의 차이를 중심으로

• Published in: 언어학 Vol. 32; no. 3; pp. 1 - 22
• Main Authors: 임진, Yim Jin
• Format: Journal Article
• Language: Korean
• Published: 대한언어학회 30.09.2024
• Subjects: Korean-English translation; large language model; machine translation; neural machine translation; zero subject; 기계번역; 대규모언어모델; 무주어; 신경망기계번역; 언어학; 한영번역
• ISSN: 1225-7141; 2671-6283
• DOI: 10.24303/lakdoi.2024.32.3.1

This paper investigates the translation strategies employed in human and machine translations of Korean zero-subject sentences into English. The author translated 343 zero-subject segments from management forewords in business reports using Google Translate (NMT) and GPT-3.5 (LLM) and compared the results with quality human translation, seeking to investigate the patterns of three corpora’s translation strategies―subject restoration or structural modification. It was found that all three corpora-human translation (HT), NMT, and LLM translation-the dropped subject was most commonly replaced by personal pronouns rather than other nouns. Two statistically significant differences emerged among the corpora. First, HT exhibited a higher frequency of proper or general noun subjects, likely reflecting translators' efforts to avoid repetitive use of the first-person plural pronoun "we" in adjacent sentences. In contrast, NMT and LLM translations frequently adopted "we," leveraging it as a safe choice to enhance reader engagement in this genre. Second, NMT showed an overuse of short passive constructions without an agent, a choice underrepresented in LLM translations. While short passives can be effective when the subject is omitted in the source text, they may weaken the connection between action and agent, thereby altering the original discourse effect. This study contributes to the MT literature by expanding the scope to include genre-specific features, LLM translation tendencies, and particular translation challenges.