Developing a discourse space for analysing online discourse

Understanding the dynamics of online discourse is crucial for dealing with disinformation, radicalisation and hate speech. However, there are few formal models of how commenters orient their messages to each other to create online discourse. We introduce the concept of a ‘discourse space’—a novel conceptual framework that serves as an abstract meta-representation of discourse. It provides an opportunity to quantify discourse and explore its dynamics by leveraging a range of possible discourse strategies, spanning four key aspects: cohesion, attitude, logic quality and coherence. With this view, discourse strategies emerge as generalised techniques for linguistically shaping thoughts based on the social context. To construct an empirical space from real data, 1,684 message pairs from 50 YouTube video comment sections were tagged for 25 discourse strategies. Using an advanced dimension-reduction method (t-distributed stochastic neighbour embedding, t-SNE), we demonstrate that a systematic discourse space can be constructed from the data. Specifically, the relations between individual social media messages can be positioned within the discourse space and that messages which attempt to derail the discourse occupy a specific part of this space. Furthermore, there are distinct patterns of discourse derailment within this discourse space that an automatic system could detect.