Online conversations: A study of their toxicity

Abstract

Social media platforms are essential spaces for modern human communication. There is a dire need to make these spaces most welcoming and engaging to their participants. A potential threat to this need is the propagation of toxic content in online spaces. Hence, it becomes crucial for social media platforms to detect early signs of a toxic conversation. In this work, we tackle the problem of toxicity prediction by proposing a definition for conversational structures. This definition empowers us to provide a new framework for toxicity prediction. Thus, we examine more than 1.18 million X (made by 4.4 million users), formerly known as Twitter, threads to provide a few key insights about the current state of online conversations. Our results indicated that most of the X threads do not exhibit a conversational structure. Also, our newly defined structures are distributed differently than previously thought of online conversations. Additionally, our definitions give a meaningful sign for models to start predicting the future toxicity of online conversations. We also showcase that message-passing graph neural networks outperform state-of-the-art gradient- boosting trees for toxicity prediction. Most importantly, we find that once we observe the first two terminating conversational structures, we can predict the future toxicity of online threads with ≈88 % accuracy. We hope our findings will help social media platforms better curate content in their spaces and promote more conversations in online spaces.