The continuous expansion of web content requires robust methods for systematically categorizing and retrieving web-based multimedia information. While traditional classifiers predominantly focus on textual features, they often overlook the intricate relationships between text and visual content, leading to suboptimal classification performance. In this study, we introduce a multi-modal topic fusion framework to improve semantic consistency between textual and visual data representations. Additionally, we incorporate adaptive relevance weighting mechanisms, allowing for more precise web document classification in different multimedia environments. Our experimental results, conducted on an expanded dataset with complex real-world web content, show a further increase in classification robustness, particularly in dynamic and heterogeneous domains. By refining generative techniques for web crawling and information retrieval, our work contributes a substantial leap forward in multimedia classification efficiency.