Earticle

The emergence of generative AI, such as ChatGPT, developed based on large language models, has brought significant changes to chatbot technology. Rapid technological advancements have substantially improved the performance and quality of generative AI services and their outputs. As a result, the scope of generative AI utilization by companies is gradually expanding. Building on these technological advancements, this study presents prompt writing strategies for effectively utilizing ChatGPT to develop entrepreneurial business models and derive affordance design strategies. An experimental study compared and analyzed the impact of providing affordances and prompt writing strategies on satisfaction with ChatGPT utilization. The study divided the tasks for developing entrepreneurial business models into data exploration and idea generation. The findings of this study offer important implications for entrepreneurs developing business models by providing practical prompt writing guidelines tailored to specific task characteristics. Additionally, it suggests affordance design strategies for companies offering generative AI services to optimize user experience.

This study explores the design of a generative AI-based MTPE course model for Korean translation education, focusing on instructor perceptions. Pre- and post-surveys with open-ended questions were administered by three instructors teaching undergraduate Korean translation courses at a university in Egypt. The results showed that instructors perceived chain-of-thought prompting as potentially useful for analyzing AI-generated translation drafts, identifying errors, carrying out post-editing procedures, and supporting self-review activities. However, the applicability of chain-of-thought prompting varied depending on learner proficiency, translation direction, and task difficulty. Concerns were also raised regarding student dependence on AI, ethical issues, and the need to simplify prompts. Based on these findings, this study proposes an instructor-centered MTPE course model consisting of translation brief design, AI draft analysis, post-editing guidance, and MQM-based evaluation and feedback.

본 연구는 생성형 AI의 확산으로 변화하는 대학 한국어 번역교육 환경에서 교수자 인식을 바탕으로 생성형 AI 기반 MTPE 수업 모델의 설계 방향을 탐색하는 데 목적이 있다. 이를 위해 이집트 대학 학부 한국어 번역 수업을 담당하는 교수자 3인을 대상으로 사전･사후 설문과 서술형 응답을 분석하였다. 분석 결과, 교수자들은 CoT 프롬프트가 AI 번역 초안 분석, 오류 식별, 후편집 절차 구조화, 자기 점검 활동에 활용될 가능성이 있다고 인식하였다. 다만 그 적용 가능성은 학습자 수준, 번역 방향, 과업 난이도에 따라 차이를 보였으며, AI 의존과 윤리 문제, 프롬프트 단순화 요구도 함께 나타났다. 이에 본 연구는 번역 브리프 설계, AI 초안 분석과 후편집 지도, MQM 기반 평가･피드백으로 구성된 교수자 중심 MTPE 수업 모델을 제안하였다.

대규모 언어 모델(Large Language Model, LLM)의 활용이 확산되면서 개인정보 보호 문제가 중요한 기술적 도전 과제로 부상했다. 특히 한국어 데이터셋은 초성, 중성, 종성 결합 구조와 비정형적 표현으로 인해 개인정보 탐지와 마스킹이 더욱 어렵다. 본 연구는 한국어 데이터셋의 특성을 반영하여 KcBERT에 조건부 랜덤 필드(Conditional Random Field)를 결합한 모델과 Chain-of-Thought(CoT) 프롬프팅을 통합한 딥러닝 기반 탐지 모델을 설계하고, 여기에 정규 표현식 및 키워드 기반 탐지를 결합한 하이브리드 모델을 제안한다. 이를 통해 개인정보 탐지 정확도와 효율성을 동시에 개선했다. 하이브리드 모델은 평균 정확도 0.99, 매크로 평균 정확도 0.96, 재현율 0.91, F-1 점수 0.94를 기록하여 그 효과성을 증명하였다. 본 연구는 한국어 특화 개인정보 보호 기술의 새로운 표준을 제시하며, 글로벌 개인정보 보호 규제 준수의 기술적 토대를 마련할 것으로 기대된다.

The widespread adoption of Large Language Models (LLMs) has elevated privacy protection as a critical technical challenge. Particularly, Korean datasets pose additional difficulties for privacy detection and masking due to their combination of initial, medial, and final consonants, as well as unstructured expressions. This study a hybrid model that reflects the characteristics of Korean datasets by integrating a deep learning-based detection approach, combining KcBERT with Conditional Random Fields (CRF) and Chain-of-Thought (CoT) prompting. In addition, it incorporates regular expression-based and keyword- based detection to enhance accuracy and efficiency. The hybrid model demonstrated strong performance, achieving an average accuracy of 0.99, macro-average precision of 0.96, recall of 0.91, and an F1-score of 0.94, confirming its effectiveness. This research establishes a new benchmark for Korean-specific privacy protection technologies and provides a solid technical foundation for complying with global privacy regulations.

AI 기반 자산운용은 투자 판단의 정교화와 자동화를 동시에 실현할 수 있는 새로운 패러다임으로 주목받고 있다. 본 연구는 LLM(ChatGPT)을 활용해 수치 중심 자산운용 접근을 보완하고, 뉴스 기반 비정형 정보를 통합하는 자산배분 프레임워크를 제안한다. 2024년 상반기 수집한 약 9만 건의 경제 뉴스 헤드라인을 대상으로, ChatGPT의 CoT(Chain-of-Thought) 프롬프팅을 활용해 핵심 키워드를 추출하고, 섹터별 감성과 중요도를 평가한 뒤 ETF 투자 비중을 산출하였다. 이 과정을 통해 구성된 포트폴리오 중 4주 단위 리밸런싱 전략은 KOSPI 200 및 국내 AI 운용 ETF 대비 우수한 성과(수익률 40.96%, 샤프 비율 2.76, 최대 낙폭 -3.22%)를 기록하며, LLM 기반 텍스트 해석의 투자 적용 가능성을 실증적으로 보여주었다. 본 연구는 프롬프트 기반 구조화 추론을 AI 자산운용에 통합한 사례로, 향후에는 프롬프트 최적화와 언어모델 비교, RAG 기법 적용 등을 통해 분석 프레임워크를 더욱 정교하게 고도화하고자 한다.

AI-based asset management is gaining attention as a new paradigm that enables refined and automated investment decisions. This study presents a framework that integrates unstructured, news-based information into ETF portfolio allocation using a large language model (ChatGPT). Based on 90,000 economic news headlines from early 2024, key keywords were extracted and sector-level sentiment assessed through Chain-of-Thought (CoT) prompting, allowing the derivation of ETF weights. A four-week rebalancing strategy outperformed KOSPI 200 and domestic AI-driven ETFs (return: 40.96%, Sharpe ratio: 2.76, max drawdown: -3.22%), demonstrating the practical value of LLM-based text analysis. The study highlights how prompt-based structured reasoning can be applied to AI asset management, with future potential in prompt optimization, model comparison, and RAG integration.

This study examines how a large language model performs inferential reasoning in Korean reading comprehension, using inferential reading items from the Test of Proficiency in Korean (TOPIK). The analysis focuses on two item types, sequence arrangement and sentence insertion, and compares model performance across three experimental conditions: zero-shot prompting, chain-of-thought (CoT) prompting, and supervised fine-tuning. In addition to answer accuracy, the study analyzes CoT-generated outputs at the item level, focusing on reasoning?answer alignment, internal reasoning consistency, and the use of evidential cues in the passages. Within the limited test set, supervised fine-tuning showed the highest overall accuracy at 88.89%, followed by CoT prompting at 85.19% and zero-shot prompting at 77.78%. Performance patterns, however, differed by item type. CoT prompting showed the highest accuracy for sequence-arrangement items, whereas supervised fine-tuning showed comparatively stronger performance on sentence-insertion items. This difference suggests that item types within the same broad inferential category may place distinct processing demands on the model. The item-level analysis of CoT-generated explanations also revealed cases of reasoning?answer misalignment, including an instance in which a correct answer was accompanied by unstable or insufficiently aligned reasoning. These findings indicate that final-answer accuracy alone does not necessarily serve as a reliable indicator of reasoning stability. Although cues related to sentence function and structure were frequently invoked, their use did not always lead to correct answer selection or alignment between reasoning and final answers. Overall, the results highlight the need to evaluate LLM performance not only in terms of accuracy but also in terms of reasoning stability and textual evidence use. The study contributes to ongoing discussions about the critical and pedagogically informed use of LLM outputs in Korean reading instruction, learner feedback, teacher support, and TOPIK item review.