Video Review Analysis via Transformer-Based Sentiment Change Detection

Zilong Wu, Siyuan Huang, Rui Zhang, Lin Li

Abstract: With the popularization of video websites, huge amounts of reviews have been generated by the audience online. These reviews usually reflect people's sentiment changes which are valuable information for understanding the market and guiding social media operation. A number of studies analyzing sentiment changes have made great progress. However, most of them are not sensitive enough to detect sentiment changes thoroughly and subtly. To solve this problem, this paper proposes a transformer-based sentiment change detection model by using the Anomaly Clearing Cumulative SUM (ACCSUM) model to analyze review streams and distinguish sentiment change points therein. Our proposed model can detect sentiment change points precisely and thoroughly by excluding abnormal reviews from review streams. Experiment results show that our model can find sentiment change points with greater sentiment gap.

External IDs:dblp:conf/mipr/WuHZL20