R&R: A Role-playing Model Enhanced by Retrieving and Reflecting
Abstract: Role-playing is one of the essential capabilities of large language models (LLMs). However, existing role-playing models make it challenging to fully immerse oneself in a character. They do not understand the limitations of knowledge expected in their current role, nor do they possess the appropriate mindset, which makes it easily apparent that they are not truly fulfilling their role. To solve this, we propose R&R, a role-playing model enhanced by retrieving and reflecting. Before generating responses, our model first retrieves relevant role knowledge and similar dialogues based on the questions asked. Then, it uses reflections extracted from historical dialogues to understand the context. Finally, by establishing knowledge boundaries and inputs for these reflections, our model can produce replies that accurately represent the current role's perspective. To assess the effectiveness of our approach, we build a new dataset and compare our model with other models in "Values", "Personality", "Hallucination", "Stability" and "Mindset" five dimensions. The results demonstrate that the average performance of our model improves by 8\% over ChatacterLLM.
Paper Type: Long
Research Area: Dialogue and Interactive Systems
Research Area Keywords: spoken dialogue systems, retrieval, knowledge augmented
Contribution Types: Data resources, Data analysis, Theory
Languages Studied: Chinese, English
Submission Number: 481
Loading