Abstract/Summary

Financial sentiment analysis (FSA) has seen substantial advancements with the use of large language models (LLMs). Previous research highlighted the effectiveness of retrieval-augmented generation (RAG) and multiagent LLMs for FSA as these approaches alleviate the problems of hallucination, a lack of factual knowledge, and limited complex problem-solving capability. Despite this, the interplay and potential synergies between these two methods remain largely unexplored. This study presents a notable leap forward by introducing a retrieval-augmented multiagent system (RAMAS) to enhance LLM-based FSA performance. An RAMAS is specifically designed to deepen understanding of the critical factors that are inherent in FSA and mimic human-like consensus-making processes by adaptively learning from semantically similar few-shot samples and engaging in conversations among the generator, discriminator, and arbitrator agents. Our evaluation of RAMASs demonstrates improved accuracy and F1-score across multiple established FSA benchmark datasets.