Thinking Before Speaking: A Role-playing Model with Mindset

ACL ARR 2024 December Submission920 Authors

Abstract: Role-playing involves making Large Language Models (LLMs) generate responses in the manner of a specific character. This task is relatively easy for LLMs, as they excel at simulating human behaviors. Existing works mainly focus on ensuring the consistency of a character's personality, information, and tone by fine-tuning the models or using specialized prompts. However, these models often lack the ability to fully embody the mindset of the character, making it difficult for them to generate responses that align with the character's way of thinking. This limitation leads to a poor user experience. To solve this problem, we propose a Thinking Before Speaking (TBS) model in this paper, which can mimic the character's logical reasoning process and generate reflections before answering a question. We enhance the training data for each set of dialogues by incorporating logical reasoning based on the character's profile and the contextual content of the conversations. This approach enables the model to learn and replicate the character's mindset. Additionally, we include a small number of questions beyond the character's knowledge scope to train the model on how to appropriately decline to answer. To verify the effectiveness of our model, we prepare new evaluation datasets and metrics. Experimental results show that the TBS model achieves best role-playing performance in terms of tone, information, and mindset.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: spoken dialogue systems, embodied agents

Contribution Types: Data resources, Data analysis, Theory

Languages Studied: English, Chinese

Submission Number: 920