Entity-level sentiment prediction in Danmaku video interaction

Sentiment analysis in Danmaku video interaction aims at measuring public mood in respect of the video, which is helpful for the potential applications in behavioral science. Once these sentiments are discovered, this feedback can help video creators improve the video quality and greatly enhance online users' watching experience. Predicting these entity-level sentiments is challenging because there is no publicly available dataset about entity-level sentiment analysis of Danmaku-enabled video comments. Furthermore, the targeted entity with skewed unbalance distribution in real-world scenarios, making the task more challenging, especially when the target entity only has positive (negative) emotional comments. In this case, applying previous aspect-level sentiment analysis models directly will introduce entity bias. In this paper, we propose a large-scale Chinese video comments dataset containing time-sync Danmaku comments and traditional video comments, targeting multiple entities and sentiments associated with each entity from popular video websites. We also propose a framework of entity-level sentiment analysis with two de-biasing models: hard-masking de-bias model and soft-masking de-bias model. This framework is defined by parallel neural networks to learn the representation of comments sentences. Based on the representations, our model learns a masking strategy for entity words to avoid overfitting and mitigate the bias. Our experiments on Danmaku-enabled video datasets show that the soft-masking model significantly outperforms comparable baselines, with a relative F1-score improvement of 9.33% compared to AEN-BERT and a relative F1-score improvement of 45.77% compared to Td-LSTM. Furthermore, experiments on different distribution bias of entity demonstrate that our proposed model can achieve competitive performances. The findings of this research have implications for measuring public sentiment for entities mentioned in a specific video domain. It can also be used as a benchmark dataset for aspect entity sentiment detection methods.