RAG를 활용한 실시간 개인방송 참여자 반응 요약 생성 방법에 관한 연구

With the rapid proliferation of live-streaming platforms, users face increasing difficulty in comprehensively viewing and understanding all braodcast content. While previous research has primarily focused on generating video highlights, the chat has been largely overlooked in summarization efforts. this study proposes a novel method for generating summaries of live-streaming broadcasts based on audience feedback, explicitly addressing the unique characteristics of real-time personal broadcasting environments. The proposed approach leverages Retrieval-Augmented Generation (RAG) technology to overcome critical limitations of traditional Large Language Model (LLM). The methodology involves three key stages: data extraction using chunking of chat data, document retrieval with the multitask embedding model and vector database, and summary generation through the LLM model. An experimental evaluation using BLEU and BERT Score metrics demonstrates the superior performance of the RAG-based approach compared to LLM-only methods. The proposed system achieves higher semantic similarity scores while reducing token usage and decreasing generation time. These improvements in both accuracy and efficiency make the approach particularly suitable for real-time broadcasting environments where contextual understanding and rapid response generation are essential for enhancing user engagement and content accessibility.