RoleRAG: Enhancing LLM Role-Playing via Graph Guided Retrieval

ACL ARR 2025 February Submission7946 Authors

Abstract: Large Language Models (LLMs) have been used to imitate various characters to make conversations more engaging and immersive. However, LLMs fail to accurately capture the extensive knowledge specific to a given character, often generating hallucinated content that is irrelevant or inconsistent with the character's known information. To overcome this, we propose RoleRAG, a retrieval-based approach that includes (1) a graph-based indexing module that extracts the target role's experiences and relationships from a vast knowledge corpus and (2) an adaptive retrieval module that efficiently retrieves relevant information from the indexing system to ensure responses are accurate and contextually appropriate. We conduct extensive experiments on role-playing benchmarks and demonstrate that RoleRAG's calibrated retrieval enables both general LLMs and role-specific LLMs to exhibit knowledge that is more aligned with the given character and reduce hallucinated responses. Our code is available at \url{https://github.com/AnonymousSub123/RoleRAG}.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: retrieval,knowledge augmented,spoken dialogue systems

Contribution Types: NLP engineering experiment

Languages Studied: English, Chinese

Submission Number: 7946