Offline Detection of Violations in Chinese E-Commerce Live Streaming Content

With the rapid growth of the Chinese e-commerce live streaming industry, regulatory oversight of its content has become increasingly critical. This paper addresses the task of offline detection of content violations in recorded Chinese e-commerce live streaming videos. We introduce a new task, CLiveSVD, which categorizes content into three classes: compliance, suspected violation, and serious violation. To support research in this area, we constructed a high-quality dataset in collaboration with market supervision experts, comprising a training set, validation set, and two test sets. Given the semantic similarity between suspected and serious violations, we propose a two-stage classification framework that improves upon direct multi-class classification. Additionally, to address the challenges of data imbalance and the high cost of manual annotation, we leverage large language models (LLMs) to generate synthetic violation examples, enhancing both the diversity and volume of the training data. Experimental results show that our two-stage approach achieves superior performance, and that LLM-generated data further boosts the robustness and effectiveness of the violation detection system in offline settings.