Toxicity in Evolving Twitter Topics

Tracking the evolution of discussions on online social spaces is essential to assess populations’ main tendencies and concerns worldwide. This paper investigates the relationship between topic evolution and speech toxicity on Twitter. We construct a Dynamic Topic Evolution Model (DyTEM) based on a corpus of collected tweets. To build DyTEM, we leverage a combination of traditional static Topic Modelling approaches and sentence embeddings using sBERT, a state-of-the-art sentence transformer. The DyTEM is represented as a directed graph. Then, we propose a hashtag-based method to validate the consistency of the DyTEM and provide guidance for the hyperparameter selection. Our study identifies five evolutionary steps or Topic Transition Types: Topic Stagnation, Topic Merge, Topic Split, Topic Disappearance, and Topic Emergence. We utilize a speech toxicity classification model to analyze toxicity dynamics in topic evolution, comparing the Topic Transition Types in terms of their toxicity. Our results reveal a positive correlation between the popularity of a topic and its toxicity, with no statistically significant difference in the presence of inflammatory speech among the different transition types. These findings, along with the methods introduced in this paper, have broader implications for understanding and monitoring the impact of topic evolution on the online discourse, which can potentially inform interventions and policy-making in addressing toxic behavior in digital communities.