SNS 데이터 마이닝 및 RAG를 활용한 AI 페르소나 기반 대화형 에이전트 설계

초록

This study aims to address the limitations of the standardized personas often found in conventional Large Language Model (LLM)-based conversational agents (hereafter “chatbots”) by presenting a chatbot that incorporates personas reflecting the linguistic characteristics, personality traits, and interests of real user groups. Using SNS data from teenagers, the research designed a Retrieval-Augmented Generation (RAG)–based chatbot that generates and applies such personas, and compared its responses to those of a general LLM-based chatbot when presented with the same questions. Conversations and posts collected from Instagram and X (Twitter) over a three-month period in the first half of 2025 were processed through Okt morphological analysis and keyword frequency analysis to extract interests and Big Five personality traits. Multiple persona profiles were generated based on these results. With the RAG structure, the SNS data were embedded using KR-SBERT and stored in a FAISS vector database to preserve linguistic styles, which were retrieved via similarity search to reflect the speaking style in the chatbot’s responses. GPT-4o was used for dialogue generation, and the outputs were compared to those of a GPT-4o chatbot without persona integration. The persona-integrated chatbot’s responses displayed teenage-specific language habits-such as the use of emojis, colloquial expressions, and empathetic remarks-enhancing familiarity and engagement. In contrast, the general GPT-4o chatbot tended to produce more formal, explanatory, and literary responses. The findings suggest that integrating user group–specific personas into chatbots can enhance familiarity, immersion, and trust, making them effective for counseling, education, and activities tailored to those groups or generations.

키워드