The paper introduces the RoleCraft framework, which focuses on advancing personalized role-playing experiences with large language models (LLMs). The key aspects of the framework are:
Emotion-Driven Character Profiling: The framework utilizes a detailed emotion classification strategy to annotate dialogues with emotional labels, enabling LLMs to create character profiles that accurately reflect diverse emotional states and personality traits.
Contextual Q&A Generation: The system employs LLMs to generate contextually coherent questions and answers, ensuring that the dialogues are consistent with the established character profiles and the ongoing scenario.
Hybrid Instruction-Based GLM Refinement: The framework combines general instructions with character-specific Q&A pairs to train the LLM, striking a balance between flexibility and character integrity in dialogue generation.
The paper presents a novel dataset, RoleInstruct, which features a diverse set of non-celebrity characters with detailed emotional annotations. Experiments demonstrate that the RoleCraft-GLM model outperforms mainstream models like GPT-4 in various role-playing evaluation metrics, including dialogue authenticity, emotional accuracy, and contextual relevance.
Ke Bahasa Lain
dari konten sumber
arxiv.org
Wawasan Utama Disaring Dari
by Meiling Tao,... pada arxiv.org 04-05-2024
https://arxiv.org/pdf/2401.09432.pdfPertanyaan yang Lebih Dalam