Low-Quality DanMu Detection via Eye-Tracking Patterns

With the development of online video platforms, a comment visualization system that inserts dynamic and contextualized comments on a video has become popular in Japan and China, known as DanMu, which provides a feeling of "virtual liveness". However, at the same time, it also brings some bad influences such as goal impediment and information overload, distraction problems, impolite and irrelevant comments. To solve this problem, there are several studies utilizing textual content for low-quality DanMu detection. However, they leave out the visual context and do not consider users' watching behavior. To this end, in this paper, we propose an end-to-end multimodal classification framework for low-quality DanMu detection. Specifically, we first design a lab-based user study to investigate users' watching patterns. Based on the discovered fixation patterns, we propose a new fusion method to fuse them with textual context. Moreover, visual content is also considered with a further fusion mechanism. Our model outperforms other baselines in almost all classification metrics in the real-world dataset.