RoleCraft introduces a framework aimed at enhancing role-playing experiences in large language models by focusing on authentic and non-celebrity characters and emotional depth.
Abstract: The development of large language models(LLMs) has initiated a new chapter in complex tasks such as role-playing, enhancing user interaction experiences by enabling models to imitate various characters.However, LLMs are somewhat lacking in their ability to portray lesser-known characters, especially in aspects of dialogue delivery and scriptwriting skills. To this end, we aim to swiftly acquire essential language skills for character development, greatly enhancing role-playing comfort. In this work, we present RoleCraft, an innovative framework designed to enrich personalized role-playing experiences. Central to this framework is RoleInstruct, a distinctive dataset featuring emotional annotations, transitioning from traditional celebrity-focused roles to more authentic, daily non-celebrity roles,each accompanied by carefully crafted character descriptions. We combined RoleInstruct with open-source instructions from the general domain, employing a hybrid instruction tuning strategy to create RoleCraft-GLM. Experiments in role-playing demonstrate that our model excels in generating dialogue that accurately reflects character traits and emotions, outperforming most mainstream LLMs, including GPT-4. Our datasets are available at https://anonymous.4open.science/r/RoleCraft-GLM-30C8/.
Paper Type: Long
Research Area: Dialogue and Interactive Systems
Research Area Keywords: Large language models, role-playing, dialogue generation
Contribution Types: Model analysis & interpretability, Data analysis, Theory
Languages Studied: English, Chinese
Submission Number: 1276
Loading