Would Your Tweet Invoke Hate on the Fly? Forecasting Hate Intensity of Reply Threads on Twitter

Curbing hate speech is undoubtedly a major challenge for online<br/>microblogging platforms like Twitter. While there have been studies<br/>around hate speech detection, it is not clear how hate speech finds<br/>its way into an online discussion. It is important for a content<br/>moderator to not only identify which tweet is hateful, but also<br/>to predict which tweet will be responsible for accumulating hate<br/>speech. This would help in prioritizing tweets that need constant<br/>monitoring. Our analysis reveals that for hate speech to manifest<br/>in an ongoing discussion, the source tweet may not necessarily be<br/>hateful; rather, there are plenty of such non-hateful tweets which<br/>gradually invoke hateful replies, resulting in the entire reply threads<br/>becoming provocative.<br/>In this paper,we define a novel problem – given a source tweet and<br/>a few of its initial replies, the task is to forecast the hate intensity of<br/>upcoming replies. To this end, we curate a novel dataset constituting<br/>∼ 4.5𝑘 contemporary tweets and their entire reply threads. Our preliminary<br/>analysis confirms that the evolution patterns along time<br/>of hate intensity among reply threads have highly diverse patterns,<br/>and there is no significant correlation between the hate intensity of<br/>the source tweets and that of their reply threads. We employ seven<br/>state-of-the-art dynamic models (either statistical signal processing<br/>or deep learning based) and show that they fail badly to forecast the<br/>hate intensity. We then propose DESSERT, a novel deep state-space<br/>model that leverages the function approximation capability of deep<br/>neural networks with the capacity to quantify the uncertainty of<br/>statistical signal processing models. Exhaustive experiments and<br/>ablation study show that DESSERT outperforms all the baselines<br/>substantially. Further, its deployment in an advanced AI platform<br/>designed to monitor real-world problematic hateful content has improved<br/>the aggregated insights extracted for countering the spread<br/>of online harms.