Research on Five-Category Classification of Comment Emotions Based on the Multi-agent RAG Framework

Abstract

Because text comment sentiment analysis plays an essential role in applications such as online public opinion monitoring, user experience analysis, and intelligent customer service, it is rightly considered a core task of natural language processing with both great practical value and considerable research significance. However, it is also well known that current mainstream single-text sentiment classification models have inherent limitations.Because existing approaches suffer from superficial semantic understanding, insufficient logical reasoning capabilities, vulnerability to model hallucinations, and low precision in fine-grained sentiment classification, they are not suitable for complex comment sentiment recognition tasks. Hence, this paper naturally and elegantly proposes a dual-agent RAG framework combined with Coherence Theory (CoT), where a Q-Agent is charged with senthe paper presents a framework combining sentiment analysis and reasoning guidance with an RAG-Agent for knowledge retrieval and information enhancement, thus naturally and elegantly achieving precise five-category sentiment classification. More importantly, comparative experiments on a balanced sentiment dataset show that the proposed framework attains optimal improvements on all evaluation metrics relative to baseline models, and therefore achieves substantially better sentiment classification performance. As such, the work provides new technical insights for fine-grained text sentiment classification and clear practical guidance for agent technology applications in natural language understanding.