EmoAgent: A Dual-Agent Framework for Short Text Emotion Classification based on Retrieval-Augmented Generation and Chain-of-Thought Reasoning

Abstract

Emotion classification in short texts is still an important but very difficult task in emotion calculation because of the lack of contextual clues and implied semantic expressions. In order to solve the problem of large-scale language model (LLM) in a domain-specific zero-shot scene, we propose a new dual-agent framework named EmoAgent, which includes a search enhanced generation (RAG) agent and a question query (QA) agent. Specifically, as a knowledge extractor, RAG Agent looks for similar examples in a structured database, while QA Agent uses the Chain-of-Thought (CoT) hint to infer from the discovered context step by step. We created a balanced emotional data set to test the framework using DeepSeek and Doubao models. Many experimental results show that EmoAgent improves the classification accuracy from 37.6% to 80.7%, which proves that it is better to combine sparse search method with deep reasoning mechanism.