Reinforcement Learning-Based Generative Text Regulation for Guiding Youth Values

With the widespread application of generative artificial intelligence, text generation models have demonstrated significant potential in intelligent dialogue systems and social media recommendation. However, existing models may introduce biases or extremist expressions, making it challenging to ensure that generated content aligns with mainstream values, particularly in guiding youth values. To address this issue, this paper proposes a reinforcement learning from human feedback (RLHF)-based text regulation method, which leverages policy optimization and contextual knowledge embedding to enhance value alignment while maintaining linguistic naturalness and diversity. The experiments utilize multiple real-world datasets, such as Reddit public comments and news comment datasets, and are validated in intelligent dialogue systems and social media recommendation systems. The results show that, compared to traditional fine-tuning methods and rule-based text filtering, the RLHF approach significantly improves value alignment, text fluency, and user acceptance. Specifically, value alignment increased by $15 \%$, text diversity improved by $10 \%$, extremist expressions were reduced by $20 \%$, and user acceptance reached $82 \%$, demonstrating the effectiveness of the proposed method. This study provides a novel technical approach for generative text regulation, advancing the application of reinforcement learning in the alignment of human values.