With the rapid development of information technology, student generated content gradually presents multi-modal characteristics, including text, images, and other forms. These data contain rich emotional information, which is of great significance for improving teaching quality and achieving personalized education. Traditional emotion analysis methods are limited by a single mode, and it is difficult to fully capture the emotional diversity. Therefore, this article proposes a new dynamic fusion model that combines knowledge graph and multi-modal sentiment analysis, which achieves dual accurate representation of learning content and learner state, aiming to more accurately perceive learners' emotional cognitive behavior, such as learning interest, motivation, and focus. In experimental verification, the model effectively improved the accuracy and comprehensiveness of emotional cognitive analysis, providing a more personalized and accurate learning resource recommendation and feedback mechanism for online learning platforms.