KJCKorea
Journal Central

In this paper, we propose an AI-based service called Baleunmalssami aimed at fostering a positive language culture among adolescents and preventing cyberbullying. This service utilizes a classification model based on LLM and RAG to accurately detect offensive language and provide real-time suggestions for alternative expressions or emojis suited to the context, thus protecting the privacy of minors and reducing the use of profanity. The service consists of a keypad and an app. The keypad replaces offensive language with emojis that match the tone and provides a real-time risk rating for unethical expressions. The app uses accumulated text data analyzed by the LLM to automatically generate reports on language habits and cyberbullying for both students and parents. The KoSim-CSE-BERT-multitask model used in the keypad delivers fast and accurate results on-device without exposing data, while reports generated with LLM and RAG include only select portions of actual conversations to protect the privacy of minors. As a result, this service provides students with real-time language correction and self-awareness opportunities, while offering parents insightful information about their children’s language habits. Baleunmalssami will play an important role in fostering healthy language habits and contributing to a safer cyber environment.

본 연구는 공공행정 분야에서 생성형 AI 도입의 가능성과 한계를 고찰하고 이를 통해 행정 효율성 증진 및 공공 서비스의 질 향상을 모색하기 위해 혼합 연구 방법(mixed-method approach)을적용하여 PoC(Proof of Concept) 사례를 분석하였다. AI 모델의 성능과 실질적 유효성을 평가하기위해 정량적 지표인 F1 스코어와 정성적 방법인 사용자 만족도 조사를 함께 활용하였으며, 최신데이터 반영을 위해 RAG(Retrieval-Augmented Generation) 기법을 적용하였다. 연구 결과, 생성형AI는 법령 검색 및 자료 조사와 같은 반복적 업무에서 공무원의 효율성을 유의미하게 향상시키며사용자 만족도 조사에서도 비교적 높은 점수를 획득하여 업무 경감에 실질적으로 기여하는 것으로 나타났다. 또한 TOE(기술-조직-환경) 프레임워크를 적용하여 공공부문 AI 기술 도입 시 발생할수 있는 기술적, 조직적, 환경적 과제를 체계적으로 분석함으로써 AI 도입 과정에서의 변화 관리와 사회적 수용성 확보의 필요성을 강조하였다. 본 연구는 공공부문에서 생성형 AI의 도입 타당성을 제고하며 향후 디지털 전환 촉진을 위한 체계적 변화 관리 전략, 윤리적·법적 준수 강화, 데이터 수급 및 품질 관리의 중요성에 대한 정책적 시사점을 제공한다.

In this paper, we proposes a CODA service application equipped with a sign language translation function based on generative AI. The application is characterized by its sign language translation technology and a user interface (UI) that minimizes text. The sign language translation technology consists of the Image2Pose, Pose2Gloss, Gloss2Text, and Image Generation modules. The Image2Pose module analyzes input sign language videos and converts them into images. The Pose2Gloss module translates the meaning of Gloss based on continuous Pose key points. The Gloss2Text module converts Gloss into text, while the Image Generation module converts text into images. The application undergoes field testing with actual CODA users and is refined based on their feedback. Additionally, the need for additional sign language features is identified, and these are implemented using Contrastive Learning. The proposed application is expected to contribute not only to smooth communication within CODA families but also to the advancement of sign language translation technology.