중-한 기계번역 자막의 품질 향상 연구 : FAR 모델을 활용한 드라마 자막을 중심으로

다국어 초록 (Multilingual Abstract)

This study aims to analyze the quality of machine-translated subtitles and to propose post-editing guidelines for improving automatically generated subtitles. The research focuses on the Korean machine-translated subtitles of the Chinese drama <!-- Not Allowed Tag Filtered --><Go Ahead> provided on YouTube, which were created not by human translators but through the platform’s neural machine translation system. A total of 462 Chinese subtitles and their corresponding Korean machine-translated subtitles were collected and compiled as the dataset.
Based on Pedersen’s FAR model, the subtitles were evaluated across three dimensions, which are equivalence, acceptability, and readability. To ensure the reliability of the evaluation, two evaluators independently conducted assessments according to the same criteria, and any discrepancies were resolved through discussion before determining the final error categories and penalty scores.
The results show that stylistic errors and semantic errors accounted for the highest proportion among all error types. Stylistic errors frequently involved inappropriate honorific usage, unsuitable terms of address, and excessive use of written or formal expressions in contexts requiring colloquial forms. Semantic errors commonly included missing contextual logic, distorted pragmatic meaning, or incomplete representation of speaker intention. In contrast, formal errors such as grammatical errors, punctuation errors, and segmentation or spotting issues appeared less frequently, indicating that contemporary neural machine translation systems have achieved a certain level of linguistic accuracy. Nevertheless, significant limitations remain in capturing contextual nuances and reproducing natural spoken language features.
Drawing on these findings, the study proposes post-editing guidelines across five dimensions: semantic accuracy, stylistic appropriateness, lexical and idiomatic usage, subtitle formatting, and syntactic or grammatical correctness. These guidelines are expected to enhance the quality of automatically generated subtitles and to support the development of hybrid human–machine translation workflows.
By examining subtitles produced in real viewing environments, this study offers practical insights into the strengths and limitations of machine-translated subtitles. The results may serve as foundational data for future research on improving Chinese–Korean machine subtitle translation and for advancing post-editing practices in audiovisual translation.

목차 (Table of Contents)

• 1. 서론 1
• 1.1. 연구 목적 및 필요성 1
• 1.2. 선행 연구 3
• 1.3. 연구 대상 및 방법 7
• 2. 이론적 배경 11
• 2.1. 기계번역의 발전과 자동 자막 11
• 2.2. FAR 모델의 개념 15
• 2.3. 기계번역 시대의 포스트에디팅 21
• 3. 기계번역 자막의 품질평가 설계 및 결과 28
• 3.1. FAR 모델 기반 자막 품질평가 기준 설계 28
• 3.2. 평가 항목 정의 및 설명 33
• 3.3. 평가 절차 40
• 3.4. 평가 결과의 정량적 분석 42
• 4. 기계번역 자막 오류의 정성적 분석 48
• 4.1. 등가성 중심의 번역 양상 48
• 4.1.1. 의미적 오류 48
• 4.1.2. 문체적 오류 60
• 4.2. 수용성 중심의 번역 양상 65
• 4.2.1. 문법적 오류 65
• 4.2.2. 용어 규범 오류 67
• 4.2.3. 관용적 오류 70
• 4.3. 가독성 중심의 번역 양상 72
• 4.3.1. 분할과 스파팅 오류 72
• 4.3.2. 구두점 사용 오류 74
• 5. 기계번역 자막의 품질 향상 방안 77
• 5.1. 기계번역 결과에 대한 종합적 분석 77
• 5.2. 포스트에디팅 가이드라인 제시 81
• 6. 결론 96
• 참고문헌 100
• 부록 106
• ABSTRACT 110