Affective Content Analysis of Online Video Clips with Live Comments in Chinese

Recent years have witnessed an increasing interest in online video affective content analysis, since having a better understanding of the emotions of videos can facilitate many applications including video retrieval and classification. Research in video affective computing requires ground truth data for training and evaluation. The live commentary (also known as 'danmaku', 'barrage', or 'bullet comment') is quite popular in recent years, but few researchers have paid attention to the commentary information in the affective analysis of online videos. In this paper, we build a dataset of online video clips, namely DaLC (Dataset with Live Comments), for affective content analysis and related applications. In contrast to existing datasets with only video clips, DaLC consists of not only 204 good quality online video excerpts but also their live comments with a large content diversity. In this paper, we first introduce a multi-dimensional emotional descriptor, which can accommodate the presence of multiple and even ambivalent emotions in a video clip, and is used as the labeling model for the dataset. Moreover, we extract effective features for online video affective content analysis with live comments in Chinese. In addition, to highlight the importance of live comments for online video sentiment analysis, we further conduct several experiments for emotion prediction on the DaLC dataset using different methods. Experimental results show that live commentary features have significantly improved the effectiveness of online video affective analysis.