Video Content Classification Using Time-Sync Comments and Titles

The Time-Sync Comment (TSC) is a novel comment used by online video websites and has been gradually recognized by Internet users. Attracted by the active viewer-creator interaction, numerous content creators upload their videos to online TSC video websites. Facing the increasing number of videos, it is difficult to efficiently classify video in traditional manual review methods. Given the close comment-video correlation, achievements of natural language processing can be applied to the TSC video classification task, processing texts instead of images. In this paper, a new method of video classification based on TSCs and titles is proposed. It combines the BERT (Bidirectional Encoder Representation from Transformers) model with the machine learning classifier, and obtains classification results via analyzing TSCs and titles. In particular, this method can work with few TSCs and bypass the limit of the BERT for the input sequence length. The experimental results on the real-world dataset show that BERT-SVM can achieve a better performance than the baseline methods, and the maximum accuracy is up to 0.9396. This research can help online video websites to manage TSC videos more efficiently and intelligently, and provide novel ideas for researchers to study classical tasks.