The proliferation of user-generated content on film review platforms has heightened reliance on user comments for movie selection. However, the misuse of comment sections to spread illegal, misleading, or offensive content poses serious moderation challenges. Traditional detection methods, primarily based on textual and statistical features, often fail to address evasion tactics such as homophonic substitutions and metaphorical expressions. To overcome these limitations, we propose KSDF, a Knowledge-based Sensitivity Detection Framework that integrates BERT with a Knowledge Graph (KG). KSDF employs a Graph Convolutional Network (GCN) to extract deep knowledge features from a sensitive word KG and utilizes BERT to capture semantic representations of review texts. A semantic association mechanism aligns knowledge-based and text-based features, constructing a multi-dimensional joint discrimination model. Additionally, we introduce an entity embedding enhancement strategy, incorporating named entities from review texts as auxiliary features to refine detection performance. Extensive experiments on real-world datasets demonstrate that KSDF achieves superior sensitivity detection accuracy, outperforming state-of-the-art baselines.