A Review of Automated Content Moderation on Live-Streaming and Short-Video Platforms: From Offline Batch Processing to Real-Time Streaming Detection

Abstract

The rapid growth of live-streaming and short-video platforms has led to the large-scale aggregation of user-generated content under conditions of high concurrency, strong interaction, and short dissemination chains, thereby placing content governance under multiple constraints involving detection accuracy, processing latency, and computational cost. This article systematically reviews the evolution of automated content moderation technologies from offline batch processing to real-time streaming detection from three perspectives: processing sequence, modality coverage, and model paradigm. The review shows that early offline moderation approaches based on keywords, rule repositories, and traditional machine learning offered advantages in deployment cost and workflow standardization, but exhibited clear limitations in handling adversarial textual variants, complex semantic understanding, and cross-modal violation recognition. Deep learning methods promoted the transition from rule matching to representation learning and significantly improved unimodal recognition performance for text, images, and speech. Multimodal fusion and vision-language large models further enhanced the modeling of implicit semantics, scene relationships, and policy boundaries, but also introduced greater inference latency, higher resource consumption, and limited interpretability. Based on existing studies and platform practice, a relatively robust engineering path is a cascaded moderation architecture composed of lightweight online filtering, multimodal re-ranking, and human review as a safeguard. Future research should continue to advance lightweight inference for streaming scenarios, cross-domain generalization, explainable decision-making, and the coordination between technical systems and governance mechanisms.