Short video fake news often employ a strategy of stitching together cross-modal real footage and form a systematic false narrative by cognitively stitching credible local materials. Existing detection methods struggle to capture their complex semantic relationships, mainly due to the lack of a semantic relationship modeling approach with topic constraints, the oversight of differences between semantic spaces in different modalities, and the inability to effectively eliminate semantic noise unrelated to the topic. To address these challenges, we propose a novel disentangled framework for multimodal semantic relationships in short video fake news detection, which captures multiple semantic relationships focusing on the topic within a shared semantic space. Specifically, we first utilize contrastive learning to build a shared semantic space that mitigates representational discrepancies across modalities, then construct a topic-driven hierarchical graph network with topic modality as the core node. Subsequently, a multi-level influence attention mechanism is designed to simultaneously capture explicit and implicit semantic relationships. Ultimately, the final integration of multi-relation node features is employed to make the classification decision. Extensive experiments on two public benchmark datasets demonstrate the effectiveness of our method.