Check out this new article by Rawat, Chaturvedi and Samant, which shifts the focus from what hate speech says to who is spreading it. The aim of the paper is clear: to identify whether users who propagate hate speech on social media have distinct behavioural, linguistic and interaction patterns compared to other users who do not spread hate.
The study builds a custom dataset of Twitter posts collected between January 2020 and September 2024 using public scraping tools. It starts with a large pool of tweets filtered through 1,186 hate-related keywords drawn from established lexicons. From this pool, only clearly hateful tweets are retained through a strict annotation process. Three trained annotators independently label each tweet, and only cases with agreement are included. The final English-language dataset includes 3,789 tweets, of which 917 are classified as hate speech. Users are then labelled as “hate propagators” if they posted at least one high-confidence hateful tweet. The analysis combines multiple dimensions: account metadata (e.g. account age, followers), activity patterns (posting frequency over time), interaction types (replies, quotes, retweets), engagement metrics (likes, views), and linguistic features (sentiment, topics, syntax).
The results show that hate propagators are not random users. They tend to have newer accounts, likely due to bans and re-creation cycles, and they are more active across the day. They engage more through replies and quote tweets, suggesting direct and often confrontational interactions, rather than passive sharing. Their posts receive fewer likes and retweets, but similar levels of replies, indicating that they trigger reactions rather than approval. Linguistically, their content is more negative and focused on identity-based topics such as race, religion and gender. Topic modelling also shows recurring clusters around highly polarising issues. Importantly, much of the hate is not explicit, but subtle or coded, which makes detection harder.
From a policy perspective, the study suggests that platforms should move beyond content-only moderation. Behavioural signals such as new accounts, high reply rates, and low media use can act as early warning indicators. This supports more targeted and proportional interventions, such as flagging high-risk accounts for review rather than blanket censorship.
For research, the paper opens a clear direction: combining behavioural, linguistic and network data can improve detection systems. It also highlights the need for longitudinal and cross-platform studies, given that hate dynamics are event-driven and context-dependent.