Recently, in the field of natural language processing, there has been an increasing emphasis on enhancing the language expressiveness of generative pre-training models. To address this challenge, this paper proposes an approach that involves specified identities or roles and evaluation metrics. By introducing specified identities or roles, the model can adapting various communicative roles tailored to specific contexts and needs, thereby better adapting to different scenarios. In terms of model evaluation, we used ten sample models and provided each model with 3000 questions. Other models and humans rated the answers given by the model on a scale of 0 to 10. The average score was then obtained. This average score is then provided as feedback to the model, encouraging it to reflect and provide more accurate answers. Finally, the paper explores the potential application prospects of this approach in human-computer dialogues, personalized Q&A systems, and other domains, demonstrating its value in enhancing natural language processing technology.