Multimodal sensitive label identification method and system for short video content security

Abstract

The application provides a kind of multi-modal sensitive label identification method and system for short video content security, anchors short video sensitive representation subject and disassembles to form multidimensional fine-grained representation unit, excavates the content association logic chain between multidimensional fine-grained representation unit, obtains relevant set;Perform cross-dimension feature interlacing and link fusion processing to construct sensitive association feature network;Sensitive association feature network is matched with preset sensitive label item set to generate initial anchoring association relationship;Then, the initial anchoring set is executed feature network reverse tracing and label anchoring counteracting calibration processing, and the anchoring association deviation is corrected;Based on the calibrated label feature node anchoring set, extract sensitive label and execute label interrelation aggregation and hierarchical division processing, output the sensitive label set corresponding to short video.The application comprehensively utilizes short video multi-modal information, and improves the accuracy of sensitive label identification.

Images