This study explores the dynamic integration and optimization path of teaching resources in the construction of smart housing with data mining technology as the core. By constructing a multi-level semantic model (LDA and semi-supervised SSGLDA), a BERT-TextCNN knowledge point linking model and an intelligent search framework integrating Elastic Search, semantic annotation, knowledge network construction and personalized retrieval of teaching resources are realized. The empirical analysis shows that the word frequency statistics reveal the core features of teaching resources, with “Automation” topping the list with 7116 times, but “Smart Meter” (1696 times) and “Power Consumption” (838) about energy management is only 10% of the high-frequency vocabulary of smart housing, which highlights the shortcomings of the construction of technical resources about energy monitoring and control. Multi-modal resource fusion significantly improves the model performance, with the F1-Score reaching 0.7046 at K=30, which is 144%, 36%, and 54% higher than that of single video (0.2890), PPT (0.5165), and PDF resources (0.4582), respectively. Sentiment analysis tracking shows that the sentiment tendency of smart housing residents is strongly correlated with the teaching nodes, with the sentiment value of the “writing” module dropping to 0.15 during the midterm exam, and the sentiment value of the “listening test” during the final review stage rising to 0.49 against the trend, confirming the impact of resource suitability on the learning experience. This confirms the influence of resource appropriateness on the learning experience. In the knowledge point association model, the F1 value of the Enhanced-EL-FM model based on fuzzy matching of Levenshtein Distance reaches 85.76%, which is 3.72% higher than that of the perfect matching method, which verifies the optimization value of the algorithm’s fault-tolerance for practical application. Data mining technology can effectively drive the semantic association and dynamic optimization of teaching resources, and the multimodal fusion and fuzzy matching strategy can significantly improve the accuracy of resource integration.