Earticle

This study examines how the rapid diffusion of AI interpreting will affect interpreting research and training, and proposes future directions. On the research side, there is an urgent need to build an evidence base through systematic collection of real-world AI interpreting outputs, robust quality-assessment frameworks, user experience analyses, and large-scale market surveys. In anticipation of human-in-the-loop service models, new roles must be defined for interpreters such as that of a multilingual communication manager. In interpreter training, program goals should be recalibrated around two pillars: (i) cultivating professional knowledge on AI interpreting technology, and (ii) deepening core human interpreting competences. Flexible academic governance and diversified faculty profiles, supported by collaboration with industry and tech-sector partners, will be essential. Overall, AI is poised to reconfigure and expand the interpreting service landscape, with evidence-based research and curricular redesign enabling a broader, evolving role for human interpreters.

This study addresses the evaluation of simultaneous speech translation (SimulST), a technology whose rapid development and growing commercialization have substantially expanded its reach well beyond that of human interpreting. On the basis of characterizing SimulST as a distinctive language processing technology that combines elements of both interpreting and translation, existing evaluation approaches for SimulST are critically reviewed and directions for assessing the quality of SimulST services are discussed in view of their growing visibility among general users. The study argues for moving beyond system-level performance metrics and linguistic quality assessment toward a more comprehensive evaluation framework that incorporates user experience in real-world communicative contexts. To this end, the study proposes a comprehensive evaluation framework for SimulST services grounded in the concept of contextualized robustness, which encompasses three steps: use case identification, authentic input data composition, and robustness testing of the output. The framework calls for the participation of multiple stakeholders — including clients, technical experts, interpreting professionals, and potential end users— at varying levels across these steps. Additionally, the study suggests that the concept of non-inferiority, adapted from medical research methodology, may provide a useful guiding principle for implementing the proposed evaluation framework.

This study explores the development, thematic patterns, and research gaps in AI-related interpreting studies in Korea. Drawing on content analysis, it examines 27 journal articles published since 2022, when AI-related discussions began to emerge visibly in Korean interpreting studies. The articles were classified into six thematic categories according to their primary research focus. The findings show a sharp increase in publications in 2025, suggesting that AI has moved from a peripheral technological issue to a significant research agenda in the field. The largest share of studies focused on AI-based interpreter education and instructional design, followed by research on interpreters’ professionalism, roles, and ethics in the AI era. Overall, Korean scholarship has tended to treat AI less as a direct substitute for human interpreters than as a tool for supporting interpreter education and interpreting performance. This tendency indicates that Korean interpreting studies have engaged with AI primarily through pedagogical adaptation, performance support, and professional redefinition, rather than through direct testing of replacement potential. The study concludes by identifying current limitations and proposing directions for future research, including empirical validation of educational applications, more refined quality comparisons, and closer examination of service acceptance and institutional implementation.

This study aims to design an AI-integrated model for Vietnamese-to-Korean interpreting classes. It also examines learners’ experiences of using AI and their perceptions of the roles of AI and the instructor in interpreting classes. The model was implemented for nine weeks with fourth-year Korean language majors at a university in Hanoi, Vietnam. After the course, 152 students participated in a survey, and 40 students took part in in-depth interviews. The class model consisted of three stages: preparation, performance, and self-evaluation. AI was mainly used to support the search for background knowledge, the construction of situational scripts, expression checking, and post-class self-evaluation. In contrast, the instructor played a central role in providing immediate feedback and guiding actual interpreting activities in class. The findings show that AI can support learners’ preparation and self-evaluation processes. However, it does not replace the instructor’s professional guidance. The study also highlights the need to develop learners’ critical AI literacy. Such literacy enables learners to verify, evaluate, and restructure AI-generated output into their own interpreting utterances.

Nghiên cứu này nhằm thiết kế mô hình lớp học phiên dịch Việt – Hàn ứng dụng trí tuệ nhân tạo (AI) cho sinh viên ngành ngôn ngữ Hàn Quốc tại một trường đại học ở Việt Nam, đồng thời phân tích trải nghiệm sử dụng AI và nhận thức của người học về vai trò của AI và giảng viên trong lớp học phiên dịch. Mô hình được triển khai trong 9 tuần với sinh viên năm thứ tư ngành ngôn ngữ Hàn Quốc tại một trường đại học ở Hà Nội. Sau khóa học, 152 sinh viên tham gia khảo sát và 40 sinh viên được phỏng vấn sâu. Mô hình lớp học được thiết kế theo ba giai đoạn: chuẩn bị, thực hiện và tự đánh giá, trong đó AI chủ yếu hỗ trợ tìm kiếm tri thức nền, xây dựng kịch bản tình huống, kiểm tra diễn đạt và tự đánh giá sau giờ học, giảng viên giữ vai trò trọng tâm trong phản hồi tức thời và điều chỉnh hoạt động phiên dịch thực tế. Kết quả nghiên cứu cho thấy AI có thể được sử dụng như một công cụ hỗ trợ quá trình chuẩn bị và tự đánh giá của người học, nhưng không thay thế vai trò hướng dẫn chuyên môn của giảng viên. Nghiên cứu cũng chỉ ra sự cần thiết của việc giáo dục năng lực sử dụng AI có phê phán, giúp người học kiểm chứng, đánh giá và tái cấu trúc kết quả do AI tạo ra thành phát ngôn phiên dịch của chính mình.

In undergraduate interpreter training, students are more likely to encounter bilateral interpreting between two working languages than conference interpreting, either during their studies or future careers. However, dialogue interpreting classes often present challenges due to students' limited linguistic competence and lack of subject knowledge. To address these difficulties, this study explores how students use generative AI to create dialogue scripts for interpreting practice, focusing on prompt literacy, AI-generated text features, and learners' perceptions. Students worked in small groups to generate weekly dialogue scripts on assigned topics using ChatGPT. These scripts were then enacted in class as role-plays, with each student taking a participant's role. Spontaneous interpreting was performed by other students selected by the instructor. Each group submitted full logs of their prompts and AI-generated outputs weekly and completed a survey at the end of the semester. The findings indicate that students' prompting strategies generally aligned with expert techniques and improved over time. The generated scripts proved suitable as source texts for interpreting exercises, requiring only minor revisions. Furthermore, the dialogues displayed a strong sense of realismand authenticity, enhancing learner immersion and engagement. Overall, students responded positively to the use of ChatGPT in script preparation. This study highlights the pedagogical potential of generative AI in undergraduate interpreter education.

This study systematically analyzes Korean scholarship on generative AI in translation and interpreting, examining 77 research articles published between 2023 and 2025 in three major Korean journals. Drawing on Christensen et al.‘s (2017) tripartite classification, the study develops an extended coding framework to map methodological trends and identify research gaps in the field. The findings reveal sharp increase in AI-related publications, rising from 14 articles in 2023 to 37 in 2025. Translation and interpreting education (37.7%) and quality evaluation (27.3%) dominate the research agenda, followed by theoretical and critical discussions (13.0%). Empirical studies constitute the majority (77.9%), with textual analysis and survey methods being the most prevalent. Korean-to-English language pair remains the most studied (70.4% of language-specific studies), while explicit theoretical grounding appears only in 27.2% of the articles. These patterns suggest the Korean translation and interpreting research in the generative AI era is largely practice-driven and pedagogy-centered in the era of AI. The study identifies key gaps — including the underrepresentation of prompt engineering, process-oriented and multilingual research — and calls for a more theoretically grounded and methodologically diversified research agenda.

In the era of artificial intelligence (AI), translation and interpreting (T&I) education is undergoing unprecedented change. This paper explores the transformation pathways of China’s T&I education in the age of AI and the challenges it faces. It first reviews the developmental trajectory of T&I education in China and, through an analysis of recent scholarship, reveals the limitations inherent in traditional pedagogical models. The paper then examines in detail the impact of AI technologies on T&I education, as well as the ethical and academic risks that may emerge in the process. Building on this foundation, it proposes concrete strategies for transformation along five dimensions: reorienting educational philosophy, reshaping instructional models, reconstructing the curriculum system, innovating teaching methods, and improving assessment mechanisms. Finally, the paper argues that these measures can substantially enhance the quality and effectiveness of T&I education and help cultivate a larger cohort of high-caliber translators who meet the demands of the new era. Future research should continue to focus on reforms and innovations in T&I education to better align with the developmental requirements of the Age of AI.

AI is increasingly utilized at conferences to mediate communications between different languages. Although the quality of interpreting using AI applications has undergone remarkable improvement, this study questions whether AI can substitute human interpreters - especially considering that comprehension of Source Language message require interpreters to draw from non-linguistic context such as contextual meaning and context of situation, not to mention from a pool of accumulated knowledge base. This study analyzes the Target Language renditions of human interpreters and AI in order to analyze the differences in contextual information in respective interpretation. Korean into English interpretation of political speeches by AI and human interpreters are analyzed. The result shows that AI can carry out conversion of voice-recognized speech in SL to TL in real time, but fails to recognize full scope of meaning in the message and contains higher number of ambiguities and meaning errors that are not found in human interpreting.

정보보안에서 AI 기술은 알려지지 않은 악성코드를 탐지하기 위해 사용한다. AI 기술은 높은 정확도를 보장하지만, 오탐을 필연적으로 수반하므로 AI가 예측한 결과를 해석하기 위해 XAI 도입을 고려하고 있다. 그러나, XAI는 단순한 해석결과만 제공할 뿐 그 해석을 평가하거나 검증하는 XAI 평가 연구는 부족하다. XAI 평가는 어떤 기술이 더 정확한지 안전성 확보를 위해 필수적이다. 본 논문에서는 악성코드 분야에서 AI 예측에 크게 기여한 feature로 AI 결과를 해석하고, 이러한 AI 결과해석에 대한 평가방안을 제시한다. 약 94%의 정확도를 보이는 tree 기반의 AI 모델에 두 가지 XAI 기술을 사용하여 결과해석을 진행하고, 기술 정확도 및 희소성을 분석하여 AI 결과해석을 평가한다. 실험 결과 AI 결과해석이 적절하게 산출되었음을 확인하였다. 향후, XAI 평가로 인해 XAI 도입 및 활용은 점차 증가하고, AI 신뢰성 및 투명성이 크게 향상될 것으로 예상한다.

In information security, AI technology is used to detect unknown malware. Although AI technology guarantees high accuracy, it inevitably entails false positives, so we are considering introducing XAI to interpret the results predicted by AI. However, XAI evaluation studies that evaluate or verify the interpretation only provide simple interpretation results are lacking. XAI evaluation is essential to ensure safety which technique is more accurate. In this paper, we interpret AI results as features that have significantly contributed to AI prediction in the field of malware, and present an evaluation method for the interpretation of AI results. Interpretation of results is performed using two XAI techniques on a tree-based AI model with an accuracy of about 94%, and interpretation of AI results is evaluated by analyzing descriptive accuracy and sparsity. As a result of the experiment, it was confirmed that the AI result interpretation was properly calculated. In the future, it is expected that the adoption and utilization of XAI will gradually increase due to XAI evaluation, and the reliability and transparency of AI will be greatly improved.

본 연구는 2026년 1월 시행 예정인 ｢인공지능 발전 및 신뢰기반 조성에 관한 기본법｣(이하 “AI 기본법”)이 저널리즘 영역에서 기존의 자율규제 단계를 넘어 언론의 AI 활용에 대해 법적 책임을 부과하는 새로운 규범 체계로 기능할 수 있는지를 검토하는 것을 목적으로 했다. 이를 위해 본 연구는 AI 기본법에 의한 규율의 핵심 판단 기준을 ‘위험성’에 두고, 뉴스의 생애주기(취재–작성–편집–유통) 전 단계에서 활용되는 주요 AI 기술들을 선별하여 그 작동 방식과 구조적 위험을 기술적으로 분석하였다. 나아가 이러한 기술별 위험 구조 및 위험도 평가를 바탕으로, AI 기본법상의 투명성 확보 의무와 고영향 인공지능 관리 규정이 저널리즘 현장에 어느 범위까지 적용될 수 있는지를 검토하였다. AI 기본법상의 투명성 확보 의무와 고영향 인공지능 관리 규정이 저널리즘 현장에 적용될 수 있는 범위를 검토한 결과, 일반적인 뉴스 제작･유통 과정에서 AI를 활용하는 언론사와 포털은 생성형 AI를 직접 개발･제공하는 예외적 경우를 제외하고는 인공지능개발사업자 또는 인공지능이용사업자로서의 법적 책임을 직접 부담하지 않는 것으로 판단했다. 또한 가이드라인(안) 체계상 저널리즘 영역은 1단계 법정 고영향 영역에 포함되지 않아, 시행령(안) 제24조에 따른 2단계 위험성 판단 역시 형식적으로는 요구되지 않는 구조임을 확인했다. 다만 저널리즘 영역의 AI 기술이 현행 고영향 인공지능 관리규정의 법적 개념에 포섭되지 않는다고 하여, 해당 기술들이 위험하지 않거나 기본권에 미치는 영향이 경미하다는 결론으로 귀결되는 것은 아니다. 실제로 본 연구는 조문 해석만으로도 AI 기본법이 저널리즘 영역을 원칙적으로 직접 규율하지 않는다는 결론에 이를 수 있었음에도, 뉴스 생애주기 각 단계에서 실제로 활용되고 있는 인공지능 기술들의 작동 메커니즘과 이에 기반한 위험도 평가를 수행했다. 이는 법적으로 규율되지 않는 AI 기술의 위험이 저널리즘 영역에 존재하는지, 그리고 그러한 위험이 국민의 기본권과 어떠한 방식으로 연결되는지를 검토하기 위함이었다. 고영향 인공지능 판단 가이드라인(안)이 제시한 2단계 위험 판단 요소(기능 중요도, 시스템 신뢰성, 오작동 시 잠재 위험성, 데이터 정확성･처리 능력, 자율성･의사결정 정도)를 분석적 준거로 차용하여 기술별 위험도를 평가한 결과 LLM 기반 기사 자동 생성, AI 영상 생성, 개인화 추천 알고리즘, 댓글 탐지･차단 기술은 법적으로는 고영향 인공지능에 해당하지 않더라도, 실질적으로는 ‘상’ 수준의 고위험군에 속하는 것으로 나타났다. 결론적으로, 내년 시행될 AI 기본법은 극히 예외적인 경우를 제외하고 저널리즘 영역을 원칙적으로 직접 규율하지 않는 법 체계임을 확인했다. 이는 저널리즘 영역에서 AI 위험 규율은 법률에 의한 직접 통제보다는, AI 기본법상 인공지능 윤리원칙(제27~28조) 규정 및 저널리즘 생태계 내에서의 자율규제 체계 내에서 형성돼야 한다는 것을 의미한다. 이러한 법 적용 결과는 표현의 자유와 언론의 자유에 대한 최소 규제 원칙을 구현한다는 점에서는 정당성을 가지는 반면, 공론장 구조와 정보 가시성에 중대한 영향을 미치는 고위험 기술들이 법적 규율의 사각지대에 놓이는 한계를 동시에 내포한다. 향후 규율의 핵심 과제는 고영향 인공지능의 법적 적용 범위를 무리하게 확장하는 데 있지...

This study examines whether the Basic Law on AI Development and the Establishment of Trust (the “AI Basic Law”), scheduled to take effect in January 2026, can function as a regulatory framework that imposes legal responsibility on the use of artificial intelligence (AI) in journalism beyond existing self-regulatory approaches. Taking risk as the core criterion for legal intervention, the study analyzes the operational mechanisms and structural risks of major AI technologies deployed across the news life-cycle, including newsgathering, production, editing, and distribution, and assesses the applicability of the AI Basic Law’s transparency obligations and high-impact AI provisions to journalistic practice. The analysis shows that, in ordinary news production and distribution processes, media organizations and portal platforms generally do not bear direct legal responsibility as AI developers or operators, except in exceptional cases where they directly develop and provide generative AI systems. Moreover, under the current draft guideline framework, journalism is excluded from the first-stage statutory high-impact AI domains, such that the second-stage risk assessment under Article 24 of the Enforcement Decree is not formally required for most journalistic AI technologies. Nevertheless, the study demonstrates that the absence of formal legal classification as high-impact AI does not imply the absence of significant risk. Applying the second-stage risk assessment criteria proposed in the draft guidelines?functional criticality, system reliability, potential harm from malfunction, data accuracy and processing capacity, and the degree of autonomy and decision-making?the study finds that LLM-based automated news writing, AI-generated video production, personalized news recommendation algorithms, and comment detection and filtering systems constitute a high-risk category in practice. In conclusion, the AI Basic Law, as currently structured, does not directly regulate the journalism sector except in highly limited circumstances. While this reflects the principle of minimal regulation in light of freedom of expression and press freedom, it simultaneously reveals a regulatory blind spot in which high-risk technologies that shape the public sphere and information visibility remain outside direct legal oversight. Accordingly, this study argues that the key regulatory task lies not in expanding the formal scope of high-impact AI, but in developing a collaborative, ecosystem-level governance framework. In particular, it suggests that the journalism sector should voluntarily adopt differentiated accountability and internal governance mechanisms for high-risk technologies comparable to those required for high-impact AI. This study aims to provide an analytical foundation for designing such a responsibility-centered and self-regulatory system.

In this study, a survey was conducted of users of interpreting services to find out which mode of interpreting they preferred the most. Forty Korean employees of a foreign financial institution in Korea, who work with interpreters on a regular basis, participated in this study. Participants were instructed to answer questions on two different settings of interpreting: conference and in-house interpreting. In the conference interpreting setting, participants were asked to choose between human interpreting, AI interpreting and subtitles after watching an English speech with the three modes of interpreting into Korean. They liked subtitles the most, followed by human interpreting, and AI was the least preferred. Accuracy of interpreting was found not to be a decisive factor in user preferences because AI was the least preferred despite being the most accurate. That said, accuracy still influenced their choice as participants preferred human interpreters more when they were more accurate. The reasons participants chose human interpreters over AI included natural tone and comfortable voice. In the in-house interpreting setting where the choice was between human and AI interpreting, human interpreters were favored over AI for their ability to engage in open discussions, their knowledge of the internal matters of the company, etc. These results show that user preferences are influenced by a variety of non-linguistic factors while users certainly value accuracy of interpreting. This suggests that it is the very human qualities that give human interpreters a competitive advantage over AI, and that there is still significant demand for these uniquely human roles.

This study examines how artificial intelligence (AI) is reshaping translation and interpreting and clarifies the continuing value of human language professionals. It surveys the enabling technologies behind AI translation and interpretation: neural machine translation (NMT) grounded in Transformer architectures, automatic speech recognition (ASR), and text-to-speech synthesis (TTS). Using a risk-based lens, the paper distinguishes language-mediation tasks that can be reliably automated (e.g., routine and low-stakes communication) from contexts where human expertise remains essential. It argues that in high-stakes settings, including legal proceedings, healthcare, diplomacy, and national security, human translators and interpreters are indispensable for accountable decision-making, real-time ethical judgment, contextual reasoning, and culturally informed communication.

이 논문은 방법론적 접근으로 AI 기술을 활용한 중국통번역 연구의 현황을 분석하였다. 최근 생성형 AI 기술은 자연어처리(NLP)와 멀티모달AI의 발전을 주도하며 다양한 산업과 응용 분야에서 활용되고 있다. ChatGPT는 인간의 지적 노동력을 보조하는 수준을 넘어, 소설이나 시를 쓰고, 신문 기사를 작성하는 등 인간 고유의 영역까지 넘보고 있으며, 통번역 분야도 예외는 아니다. 중국에서는 ChatGPT, 文心一言, 讯飞星火 등을 비롯한 생성형AI는 잇따라 번역 실무에 적용되면서 생성형AI 모델에 관련된 연구가 활발히 진행되고 있으며, 특히 2025년 1월 20일에는 DeepSeek-R1을 발표하여 중국 국내 뿐만 아니라 국제사회에서도 큰 주목을 받게 되었다. 이 논문에서는 중국에서 진행 중인 생성형 AI 모델 기반 연구의 동향을 분석하기 위해 사회학 방법론의 접근을 통해 이론적⋅실증적⋅사례 연구 차원에서 AI 기술을 활용한 중국어 통번역 연구가 어떤 주제를 다루고 어떤 결론을 도출해 왔는지 살펴보았다. 연구에 의해 전반적으로 초거대 언어 모델은 기술적으로 처리 속도가 빠르고 효율성이 높으며, 과학기술 번역에서 높은 정확도를 보이는 장점이 있으나 여전히 문학 번역 적용에는 한계가 뚜렷하며, 다른 분야에서도 인간 번역자의 개입이 필요하고, 즉 궁극적인 번역의 질은 프롬프트 입력의 우수성에 달려 있다는 결론을 얻었다.

This paper adopts a methodological approach to analyze the current landscape of AI-assisted Chinese interpreting and translation studies. Recent advances in generative AI technology?driven by developments in natural language processing (NLP) and multimodal AI?have permeated various industries and application domains. ChatGPT, for instance, has transcended its role as a mere aid to human intellectual labor, venturing into traditionally human-exclusive domains such as novel and poetry writing, news article composition, and notably, the field of translation and interpreting. In China, generative AI models?including ChatGPT, Wenxin Yiyan, and Xunfei Xinghuo?have been progressively integrated into translation practice, spurring active research on related AI models. Of particular note is the release of DeepSeek-R1 on January 20, 2025, which has garnered significant attention both domestically and internationally. To examine the trends in generative AI model-based research conducted in China, this study employs a sociological methodological lens to investigate the theoretical, empirical, and case-study dimensions of AI technology applications in Chinese interpreting and translation studies. The analysis focuses on the research themes explored and conclusions drawn in existing scholarship. The findings indicate that while large language models generally offer advantages such as rapid processing speed, high efficiency, and strong accuracy in scientific and technical translation, their application in literary translation remains markedly limited. Furthermore, human translator intervention remains necessary across various domains, underscoring that the ultimate quality of translation heavily depends on the excellence of prompt input.

본 연구는 생성형 AI 확산 맥락에서 한국어 통번역 교육의 적용 가능성과 한계를 실천적 교수설계 관점에서 검토하였다. AI 도입 이전의 한계인 교육 자원 접근의 제약, 교수 방법의 단일화, 디지털 리터러시 교육의 부족을 진단하였고, 이를 완화하기 위해 사례로 ChatGPT-4o를 채택하여 BROKE 기반 프롬프트 설계에 따라 설계–실행–평가 전 주기의 프롬프트와 산출물을 제시하였다. 설계 단계에서는 교수 방안·수업 활동·자료 확보를 구체화하였고, 실행 단계에서는 교사 설명 지원·상호작용 학습 촉진·실제 상황 시뮬레이션 활용을 구체화하였다. 평가 단계에서는 과제 평가 지원·학습자 자기평가 지원·교수효과의 체계적 축적을 구체화하였다. 또한 기술 의존, 환각·편향, 데이터 프라이버시와 저작권 보호 등 쟁점을 식별하고 대응 방향을 제안하였다. 결론적으로, 본 연구는 생성형 AI 기반 한국어 통번역 교육의 효율성 제고와 질적 고도화에 실천적 시사점을 제공한다.

From a practice-oriented instructional design perspective, this study examined the applicability and limits of integrating generative artificial intelligence (AI) into Korean translation and interpreting (T&I) education. We diagnosed three pre-AI constraints: limited access to instructional resources, homogenized pedagogy, and insufficient digital-literacy instruction. To address them, we adopted ChatGPT-4o as a case model and, using a Background?Role?Objectives?Key Result?Evolve (BROKE) prompt framework, produced prompts and deliverables for the full design?implementation?assessment cycle. The design phase specified instructional strategies, learning activities, and resource curation; the implementation phase supported teacher explanations, interactive learning, and authentic scenario simulations; and the assessment phase covered assignment evaluation, learner self-assessment, and systematic accumulation of instructional effectiveness. We also identified key issues?technological dependence, hallucination and bias, and data privacy and copyright?and proposed responses. The study offers practical implications for improving the efficiency and advancing the quality of AI-enabled Korean T&I education.

Traditional typologies of interpreting have been constructed around the human interpreter as the central agent, offering a theoretical framework suited to the pre-AI era of interpreting practice. However, the rapid advancement of AI and speech translation technologies is introducing new modalities and agents into the interpreting ecosystem, necessitating a fundamental reassessment of existing typologies. This study explores how AI-based interpreting services can be situated within or beyond the traditional interpreting landscape and proposes a new typology for the age of AI. By reviewing literature on interpreting typologies and analyzing emerging use cases of AI-based and automated interpreting services, the study argues that the interpreting market is undergoing multidimensional diversification. In response, it proposes a five-part typology of interpreting, mapped along a spectrum of human?technology integration. Unlike previous typologies, this new framework is defined by the degree and nature of human?machine involvement. The study also highlights challenges inherent in technology-integrated interpreting and calls for evidence-based research to address them.

This study explores the analysis of unethical text data in Korean using generative AI, particularly ChatGPT. It emphasizes the importance of constructing and understanding Korean unethical text data for AI data ethics. The research investigates the categorization of unethical text using Korean news and hate comments, employing methods like frequency analysis and topic modeling for objectivity. It also assesses ChatGPT’s ability to handle Korean unethical text, highlighting difficulties with contextspecific unethicality and new or unique Korean terms. The findings indicate the necessity for context-rich unethical text data to improve AI’s ethical responses, especially in subtle but malicious scenarios. Overall, the study underscores the significance of recognizing and addressing unethicality in the Korean language. This understanding is crucial for enhancing AI tools like ChatGPT and developing AI systems for detecting unethical behavior and verifying ethics, particularly in the Korean linguistic context.

본 연구는 4년제 대학생의 중도탈락을 예측하기 위해 딥러닝 기반의 심층신경망 모델을 구축하고, 설명가능한 인공지능 기법인 SHAP을 활용하여 주요 예측요인을 분석하였다. 또한,중도탈락이라는 극단적 불균형 데이터 문제를 해결하기 위해 비용 민감 학습과 적응형 합성샘플링을 적용하였으며, 더 나아가, Optuna를 활용한 초매개변수 최적화를 수행하면서 불균형 문제를 고려한 최적화 목표를 F2 score로 설정하여 모델의 예측력을 극대화하였다. 분석자료는 한국고용패널Ⅱ 3-4차 자료와 대학알리미 공시자료이다. 분석 결과, 장학금, 부모·본인 소득, 건강 상태, 대학소재지, 교수지도, 교우관계 등 다양한요인이 중도탈락 가능성을 예측하는 데 중요한 요인으로 나타났다. 이 중 장학금이 중도탈락을 예측하는 기여도가 가장 컸고, 기존 연구와 달리 GPA가 다른 요인들과 영향을 주고받으며 중도탈락을 예측하는 방향성이 달라지는 패턴이 관찰되었다. 또한 교수지도와 교우관계는상호 보완적 관계 속에서 학업 지속성을 결정하는 핵심 요인으로 작용하였다. 한편, 대학 규모가 작을수록 재학생 충원율이 증가해도 중도탈락 가능성이 완전히 상쇄되지 않는 것으로나타나, 학령인구 감소에 따른 대학구조조정에서 학생들의 안정적인 학업 환경을 제공할 세심한 장치들을 적용할 필요성을 보여주었다. 본 연구는 AI 기반의 중도탈락 예측 모델을 통해 기존 연구의 한계를 보완하고, 중도탈락가능성을 예측하는 다양한 요인들의 복합적 상호작용을 분석하여, 보다 실질적인 대학 정책및 학생 지원 전략 수립에 기여할 수 있는 근거를 제시하였다는 데 의의가 있다.

This study aims to predict 4 year university student dropout by developing adeep neural network model and utilizing explainable artificial intelligencetechniques, specifically SHAP, to analyze key predictive factors. To address theextreme imbalance in dropout data, cost-sensitive learning and adaptive syntheticsampling were applied. Furthermore, hyperparameter optimization wasconducted using Optuna, with the F2 score set as the optimization objective toenhance predictive performance while accounting for the data imbalance. Datafrom KEEPⅡ (waves 3-4) and the University Disclosure System were used. The analysis identified scholarship amounts, parental income, health status, facultyguidance, and peer relationships as significant predictors of student dropout. Among these, scholarship amounts had the highest predictive contribution. Notably,unlike previous studies, GPA exhibited varying directional relationships withdropout depending on interactions with other factors. Additionally, faculty guidanceand peer relationships played a complementary role in supporting academicpersistence. Moreover, smaller universities showed a persistent risk of dropout evenwhen enrollment rates increased, highlighting the need for structural adjustments inresponse to the decline in the college-age population. This study contributes to the literature by integrating an AI-based approach topredict student dropout, addressing the limitations of conventional linearanalyses, and providing empirical insights into the complex interplay of multipledropout-related factors. The findings offer valuable implications for universitypolicies and student support strategies aimed at improving academic retention.

학령인구의 감소로 대학의 교육수요자 대상이 변화하는 상황에 직면하였다. 그에 따른 방책으로 유학생 유치가 증대되었고, 수업 환경도 그에 맞추어 다변화되고 있다. 한국어 의사소통 수준이 높지 않은 유학생으로만 이루어진 수업 현장에서는 실시간 STT번역기의 활용이 불가피하게 되었다. 본고는 실제 수업 현장에서 사용된 STT번역기를 분석하여 번역 양상과 오류를 통해 음성인식 번역기가 가진 현재의 한계점을 살펴보았다. 그 결과, STT번역기는 어휘, 어음, 어법의 범주에서 각각 오류가 나타났다. 그중 어휘 범주에서는 맥락 부적합어, 동음이의어, 다의어, 전문용어, 관용어, 고유명사, 번역 불필요, 영문·숫자 혼용, 외래어 혼용, 줄임말의 오류 유형이 있었다. 어음 범주에서는 자음·모음 인식, 외래어 발음, 발화 중 휴지(休止)로 인한 오류가 나타났다. 이러한 오류는 텍스트 기계 번역에서는 발생하지 않지만, 음성인식 번역은 실시간으로 생성되는 구어 발화이기 때문에 어음 범주의 문제가 돌출된다. 어법 범주의 오류는 어휘나 어음 범주의 오류에 비해서는 많지 않았으나, 그 가운데 가장 두드러진 문제는 한국어에서 생략된 주어에 해당하는 대명사를 적절하게 선별하지 못하는 문제와 접속사, 어순에서 오류가 나타났다. 이러한 모든 범주의 오류를 한 마디로 표현 한다면 “맥락”에 대한 이해 부족이다. 텍스트에서의 맥락은 전체 글에서 제공하는 모든 정보로, 상대적으로 단순할 수 있지만, 실시간 구어 발화에서는 발화의 내용뿐만 아니라, 현장 상황 역시도 하나의 맥락 정보이다. 사람은 시각과 청각 등을 통해 음성, 표정, 제스처, 뉘앙스 파악 등 복합적인 상황 판단이 가능하지만, 그에 비해 기계는 입력된 정보를 확률과 연산에 의한 일차원적인 판단과 분석만 가능하여 이러한 다양한 맥락 처리에 한계가 있다. 따라서 향후 음성인식 분야는 계속적인 보완이 필요하다. 본 연구에서 언급한 언어학적 오류 현상이 교수자에게는 효율적인 활용을 위한 지침을 제공하고, 또한 음성인식 번역기 분야의 개선과 발전에도 일부 밑거름이 될 수 있기를 기대한다.

Amid the decline in the school-age population, universities are confronting a shift in the composition of their educational clientele. In response, efforts to recruit international students have intensified, accompanied by the diversification of instructional environments. In classroom settings composed primarily of international students with limited Korean communicative competence, the use of real-time speech recognition?based speech-to-text (STT) translation systems has become virtually indispensable. This study analyzes such systems as employed in actual classroom contexts, examining translation patterns and errors in order to identify their current limitations. The findings reveal that real-time STT translation systems exhibit errors across lexical, phonological, and grammatical categories. Within the lexical category, error types include contextually inappropriate word choices, homonyms, polysemy, technical terminology, idiomatic expressions, proper nouns, unnecessary translations, mixed usage of English and numerals, loanword interference, and abbreviations. In the phonological category, errors arise from the recognition of consonants and vowels, the pronunciation of loanwords, and pauses occurring during speech production. While such errors are typically absent in text-based machine translation, they emerge in speech recognition translation due to the real-time processing of spontaneous spoken language. Grammatical errors occur less frequently than lexical or phonological ones; however, prominent issues include the failure to appropriately resolve omitted subjects in Korean?particularly in selecting corresponding pronouns?as well as errors involving conjunctions and word order. These diverse error types can be attributed, in essence, to a lack of contextual understanding. In written texts, context is relatively constrained, consisting of the information provided within the discourse. In contrast, real-time spoken interaction involves not only linguistic content but also situational context. Humans can integrate multimodal cues?such as auditory signals, facial expressions, gestures, and nuanced intonation?to interpret meaning, whereas machines rely solely on probabilistic computation and the analysis of input data, limiting their capacity to process such complex contextual information. Therefore, continued advancements in speech recognition technology remain necessary. It is anticipated that the linguistic error phenomena identified in this study will provide practical guidelines for instructors seeking to utilize such systems effectively, and further contribute to the ongoing improvement and development of speech recognition translation technologies.