Short videos are becoming increasingly popular around the world, serving as an essential aspect of entertainment, daily life, and social interaction. But this also leads to the spread of negative information. Existing works focus on the image-text field, lacking a comprehensive dataset for short videos and the utilization of multimodal information in videos. In order to address this issue, we establish a Chinese short video dataset named NCSV, which contains videos with negative content such as discrimination, hatred, and malicious satire. The dataset covers multimodal information about these videos, including video frames, audio, text, and user replies. Then, we propose a multimodal framework that combines features of video content and social context to detect whether a video contains negative information. Extensive experimental results show that this new research direction is feasible and our method shows promising performance. The dataset and corresponding code will be released in https://github.com/1292863248/NCSV .