인간·AI 번역주체별 문화 중개 차이와 인간중심 포스트에디팅 : 한영 문화 텍스트 번역 오류 및 중개 양상 비교분석 및 포스트에디팅 지침 개발

다국어 초록 (Multilingual Abstract)

Amid the rapid development of machine translation and artificial intelligence (AI), translation—once regarded as an exclusively human activity—is increasingly transforming into a form of human–machine collaboration. With the advent of neural machine translation (NMT) and generative AI translation, the quality of machine translation output has improved markedly. At the same time, pessimistic views have emerged suggesting that human translators may retreat from direct “production” to the auxiliary task of “editing” machine-generated output, or even be replaced altogether.
This study argues that, in the context of these technological advances, the key to preserving human translators’ expertise and agency lies in culture—the most fundamentally human domain. Accordingly, it seeks to identify differences in translation patterns observed in the Korean–English translation and post-editing of culturally specific texts. More specifically, the study examines differences in error types and cultural mediation patterns and strategies across human and machine agents, with the aim of developing practical post-editing guidelines and educational curricula for cultural texts.
The study is structured around three analytical axes: (1) the density of cultural references, (2) the distinction between language for special purposes (LSP) and language for general purposes (LGP), and (3) translating agents. Among these, cultural reference density and the LSP/LGP distinction relate to the characteristics of source texts (ST), while translating agents determine how target texts (TT) are produced.
On this basis, the study proposes “cultural texts” as an analytical category, contrasted with the “technical texts” that have traditionally dominated machine translation and post-editing research. Here, a cultural text is defined as a “text whose interpretation and translation depend on culture-specific norms, shared knowledge, and value systems of a linguistic-cultural community,” such that meaning cannot be fully recovered from linguistic form alone and requires culturally informed mediation. The notion includes both LGP texts, where culture emerges through everyday practice, pragmatics, and lived experience, and LSP texts, where culture is embedded in domain-specific knowledge systems and culturally situated modes of knowledge transmission. A cultural text in this sense is defined not by the density of cultural references but by the need for cultural mediation: Even when explicit cultural references appear minimal, a text qualifies as cultural if it requires active interpretation and mediation of culture-specific assumptions, values, or knowledge.
The scope of the study is limited to the Korean–English translation direction. It adopts the MQM-Core model—a streamlined version of the DQF-MQM Framework that has become a de facto standard in translation quality assessment—while introducing additional categories required for the analysis of cultural texts. Based on this customized framework, the study focuses on two analytical targets: “errors,” defined as clear defects such as mistranslation and omission, and “mediation,” defined as non-error shifts that bridge cultural gaps through expressions differing from the source text.
The research questions of this study are organized into five broad thematic areas. The first set of questions examines “errors in the Korean–English translation of cultural texts,” asking: (1) Do errors occur in the Korean–English translation of cultural texts? (2) If so, what types of errors occur? (3) What factors give rise to these errors? (4) Are there differences across translating agents? (5) If differences exist, how do they manifest?
A second line of inquiry turns to “cultural mediation in the Korean–English translation of cultural texts.” Here, the study addresses the following questions: (1) Does cultural mediation occur in the Korean–English translation of cultural texts? (2) If so, what types of cultural mediation occur? (3) What factors give rise to cultural mediation? (4) Are there differences across translating agents? (5) If differences exist, how do they manifest?
The third area of investigation concerns “post-editing patterns and strategies adopted by humans and machines.” In this context, the study explores two questions: (1) Are there differences between human and machine post-editing patterns applied to Korean–English machine translation output of cultural texts? (2) If differences exist, can these patterns be systematically typologized in strategic terms?
Attention then shifts to the “design of post-editing guidelines for Korean–English cultural texts.” The guiding question in this area is: How should post-editing guidelines for Korean–English machine translation of cultural texts be structured?
Finally, the pedagogical dimension of the study focuses on “undergraduate translation and interpreting (T&I) training.” From this perspective, the study asks: (1) What implications do differences in error types and cultural mediation across translating agents have for T&I training? (2) How can these implications be applied to future translation and post-editing training? (3) How should a Korean–English cultural postediting course for undergraduate T&I training be designed?
To address these research questions, the study adopts a mixed-method research design consisting of comparative text analysis, post-editing experiments, and participant interviews. Phase 1 involves a review of previous studies and professional guidelines to establish the error and mediation classification framework. Phase 2 consists of qualitative and quantitative analyses of the Korean–English translations of five cultural texts produced by different agents (one human translation, two NMT outputs, and two generative AI outputs), followed by coding, typological classification, and statistical analysis. Phase 3 comprises post-editing experiments and interviews examining error correction and mediation patterns and strategies. In Phase 4, Korean–English cultural post-editing guidelines and a course design are presented as the final outputs of the study.
The findings from comparative text analysis, post-editing experiments, and interviews suggested clear differences in error occurrence, cultural mediation, and post-editing strategies across human translators, NMT systems, and generative AI in the Korean–English translation of cultural texts. Errors were observed in all ST and TT types, with those related to accuracy occurring most frequently. Typical cases included mistranslation of terms and proper names, distortions of logical or temporal relations, number and reference errors, and hallucinations in machine-generated output. Statistical analysis hinted that cultural density was a significant driver of error occurrence, particularly in accuracy-related categories, while the LSP/LGP distinction showed limited explanatory power overall. Instead, error patterns were more strongly associated with cultural specificity, translation direction, and translating agent. NMT produced the highest number of errors, generative AI showed noticeable improvement but still underperformed human translation, and human translators consistently demonstrated the lowest error rates.
Cultural mediation was also observed across all texts and translating agents, again most frequently within accuracy-related categories. Texts with higher cultural density prompted substantially more mediation, especially in LSP texts containing culturally specific knowledge unfamiliar to target readers. Translating agent differences were pronounced: human translations exhibited overwhelmingly more cultural mediation, generative AI occupied an intermediate position, and NMT showed minimal mediation. These findings indicate that cultural understanding and contextual inference remain core human strengths, while generative AI differs qualitatively from conventional NMT in its ability to approximate mediation, albeit without genuine cultural understanding.
Post-editing experiments further revealed distinct strategic orientations. Human participants prioritized accuracy and cultural appropriateness under time constraints, actively engaging in transliteration, explanation, and selective expansion, but occasionally missed machine-specific errors or necessary mediation due to limited postediting experience and research time. Machine post-editors—generative AI in this case—demonstrated strong performance in error detection and contextual research and showed unexpectedly active intervention, yet also produced culturally inappropriate substitutions occasionally. These patterns suggest a complementary collaboration model in which machines support error detection and information retrieval, while humans retain responsibility for culturally informed judgment and final validation.
Based on these findings, the study developed Korean–English cultural postediting guidelines for culturally specific texts, drawing on ISO 18587:2017 for process and quality requirements and using MQM-Core categories to structure error prioritization and agent-specific post-editing strategies. Finally, the study proposes an undergraduatelevel Korean–English cultural post-editing course design, arguing that translation training should explicitly address differences across translating agents, integrate error typologies and prioritization strategies, and systematically train cultural mediation skills to prepare students for effective human–machine collaboration.

목차 (Table of Contents)

• 제1장 서론 1
• 1.1 연구의 배경 및 필요성 1
• 1.2 연구 목적 및 연구 질문 5
• 1.2.1 연구의 목적과 주요 변수 5
• 1.2.2 연구 질문 8
• 1.3 연구 방법 9
• 1.4 논문의 구성 10
• 제2장 이론적 배경 12
• 2.1 번역의 주체 12
• 2.1.1 인간번역과 번역학 14
• 2.1.2 기계번역 17
• 2.1.3 인간과 기계의 언어와 번역 경향성 21
• 2.2 문화적 텍스트 23
• 2.2.1 문화와 문화소 23
• 2.2.2 일반언어와 전문언어 25
• 2.2.3 문화적 텍스트의 정의 27
• 2.3 번역오류와 문화중개 29
• 2.3.1 번역오류 29
• 2.3.2 문화중개 34
• 2.4 포스트에디팅 42
• 2.4.1 포스트에디팅 지침 43
• 2.4.2 포스트에디팅 교육 48
• 제3장 연구 방법 56
• 3.1 연구 설계 56
• 3.2 데이터 및 참여자 구성 57
• 3.2.1 텍스트 비교분석 57
• 3.2.2 포스트에디팅 실험 및 인터뷰. 68
• 3.3 연구 절차 72
• 3.3.1 텍스트 비교분석 72
• 3.3.2 포스트에디팅 실험 및 인터뷰. 81
• 3.3.3 포스트에디팅 지침과 강의 설계 제안 86
• 3.4 소결 87
• 4장 분석 결과 89
• 4.1 텍스트 비교분석 89
• 4.1.1 텍스트의 구성과 코드 적용 내역 89
• 4.1.2 분류체계의 구성과 정의 93
• 4.1.3 오류 분포 101
• 4.1.4 중개 분포 112
• 4.1.5 주요 테마별 오류·중개 분포 120
• 4.1.6 텍스트 비교분석 소결 137
• 4.2 포스트에디팅 실험 139
• 4.2.1 오류 분포 139
• 4.2.2 중개 분포 143
• 4.2.3 주요 테마 146
• 4.3 인터뷰 151
• 4.3.1 인간 참여자 인터뷰 152
• 4.3.2 기계 참여자 인터뷰 157
• 4.3.3 포스트에디팅 실험 및 인터뷰 소결 161
• 4.4 전체 소결 163
• 4.4.1 텍스트 비교분석 163
• 4.4.2 포스트에디팅 실험 및 인터뷰 164
• 제5장 논의 165
• 5.1 문화적 텍스트의 한영번역과 오류 165
• 5.1.1 문화소 165
• 5.1.2 LSP/LGP 구분 171
• 5.1.3 번역 주체 176
• 5.2 문화적 텍스트의 한영번역과 문화중개 181
• 5.2.1 문화소 182
• 5.2.2 LSP/LGP 구분 188
• 5.2.3 번역 주체 193
• 5.3 인간과 기계의 포스트에디팅 양상 및 전략 197
• 5.3.1 번역오류에 대한 대응 양상 및 전략 197
• 5.3.2 문화소의 처리 양상 및 전략 201
• 5.3.3 인간과 기계의 포스트에디팅 역량 비교 205
• 5.4 포스트에디팅 실무 및 교육을 위한 시사점 209
• 5.4.1 포스트에디팅 역량 모델과 교육 209
• 5.4.2 번역오류 분류체계와 포스트에디팅 지침의 유용성 210
• 5.4.3 포스트에디팅을 통한 인간과 기계의 협업 211
• 제6장 한영 문화 포스트에디팅 지침과 강의 설계 212
• 6.1 한영 문화 포스트에디팅 지침 212
• 6.1.1 지침의 목적과 차별점 212
• 6.1.2 지침의 구성과 내용 214
• 6.2 한영 문화 포스트에디팅 지침을 활용한 교육과정 설계 예시 214
• 6.2.1 학부 수준의 포스트에디팅 교육이 필요한 이유 214
• 6.2.2 한영 문화 포스트에디팅 수업의 설계 224
• 제7장 결론 229
• 7.1 연구 결과 종합. 229
• 7.1.1 문화적 텍스트의 한영번역과 오류 229
• 7.1.2 문화적 텍스트의 한영번역과 문화중개 230
• 7.1.3 인간과 기계의 포스트에디팅 양상 및 전략 231
• 7.1.4 문화적 텍스트의 한영 포스트에디팅 지침 설계 233
• 7.1.5 학부 통번역 교육을 위한 한영 문화 포스트에디팅 강의 설계 234
• 7.2 AI 번역 시대, 인간 주체의 미래 236
• 7.3 연구의 의의와 한계 237
• 참고문헌 240
• ABSTRACT 265