This study proposes a novel approach for analyzing learning behaviors in Chinese language education by integrating generative attention mechanisms and generative state transition equations. This method dynamically adjusts attention weights and models real-time changes in students’ emotional and behavioral states, addressing key limitations of existing approaches. A central innovation is the introduction of a generative loss function, which jointly optimizes sentiment prediction and behavior analysis, enhancing the adaptability of the model to diverse learning scenarios. This study is based on empirical experiments involving student behavior tracking, sentiment analysis, and personalized learning path modeling. Experimental results demonstrate this method’s effectiveness, achieving an accuracy of 90.6%, recall of 88.4%, precision of 89.3%, and F1-score of 88.8% in behavioral prediction tasks. Furthermore, this approach attains a learning satisfaction score of 89.2 with a 94.3% positive feedback rate, significantly outperforming benchmark models such as BERT, GPT-3, and T5. These findings validate the practical applicability and robustness of the proposed method, offering a structured framework for personalized teaching optimization and dynamic behavior modeling in Chinese language education.