Hateful online content is a growing concern of Government, academia and commercial platforms. It inflicts harm on targeted users and their communities, creates and exacerbates divisions across society, and toxifies public discourse. Developing accurate tools to automatically detect online hate is a crucial step in monitoring, analysing and tackling it. This, in turn, requires deep understanding of what constitutes hate and why certain posts are considered hateful. Many theories in linguistics and social science suggest that meaning is constructed not only through what is said but also who speaks, who listens, the specific setting and the broader social context – features which are largely ignored in computational approaches to studying online hate. These insights are also supported by a growing body of empirical evidence showing how perceptions of online hate varies substantially between different people.

At present, there is a lack of evidence about what drives perceptions of hateful content. We address this gap through using social psychology experiments to quantify the effect of context on whether content is considered hateful. We examine three aspects of context:

1. Who speaks – analysing speakers’ social identity, past behaviour and social authority;
2. Who listens – analysing audiences’ social identity, past online experiences and personality type;
3. Which group is targeted by the hateful content.

In this talk, I provide a work-in-progress summary of what we have achieved so far and our future plans. All feedback is welcomed.