Using bots or machine learning approaches to identify and possibly deter online harassment shows promise, and would greatly reduce the burden on human moderators (both lay users and moderators employed by social-media platforms) to manually search for and/or respond to reports of harassment. However, no existing tools incorporate the affordances of predictive models to actively respond to harassment efforts. 

The work proposed here will use an existing model to identify conversations where we could de-escalate the situation through active participation (e.g., bots). The model forecasts both the presence and intensity of hostility in Instagram comment threads. The project will focus on two strategies for de-escalation: (1) automatically learning de-escalation behaviors and (2) applying linguistic strategies influenced by social theories of behavior change. In our preliminary work on predicting critical points in conversations, we observed that the language of the conversation was highly beneficial for predicting a critical transition to hostile behavior.