PO-RAG: Memory-Enhanced and Self-Optimizing System for Opinion Mining in Public Opinion

With the widespread use of social media and online news platforms, the spread of public opinion events has accelerated. It is increasingly important to accurately uncover multiple opinions within these events. Existing retrieval-enhanced generation (RAG) methods can handle long-text viewpoint extraction, but the current RAG system often generates redundant content and struggles to extract deeper insights. When processing complex events with multiple conversations and interactive responses, it tends to lose contextual consistency and has difficulty tracking information accurately, resulting in unreliable outputs. In this regard, this paper proposes a public opinion system (PO-RAG), which aims to effectively refine multi-dimensional public opinion views. Building on vanillaRAG, a progressive, confidence-compensated generation method is proposed to optimize the generated content, and a memory-enhanced module and a multi-hop query strategy are introduced to improve opinion mining in complex scenarios. While extracting multi-dimensional public opinion views, we also refine specific content and identify the perspectives of key figures. We conducted experiments across eight evaluation dimensions using the LangSmith tool, and the results show that PO-RAG outperforms vanilla RAG in handling the depth and breadth of complex event analysis. This study not only offers new insights and applications for the field of public opinion but also contributes to promoting public safety and a healthy public opinion environment.