[Paper Review] The Virality of Hate Speech on Social Media

Abdurahman Maarouf, Nicolas Pröllochs+1 more|arXiv (Cornell University)|Oct 25, 2022

Hate Speech and Cyberbullying Detection4 citations

TL;DR

This study investigates the virality of hate speech on X (formerly Twitter) using a dataset of 25,219 retweet cascades, applying generalized linear models to identify author and content-level factors driving the spread of hateful vs. normal content. It finds that hate speech from verified users is 3.5 times more likely to go viral in terms of cascade size, 3.2 times longer in lifetime, and 1.2 times more structurally viral, highlighting the outsized role of verified status in amplifying hate speech.

ABSTRACT

Online hate speech is responsible for violent attacks such as, e.g., the Pittsburgh synagogue shooting in 2018, thereby posing a significant threat to vulnerable groups and society in general. However, little is known about what makes hate speech on social media go viral. In this paper, we collect N = 25,219 cascades with 65,946 retweets from X (formerly known as Twitter) and classify them as hateful vs. normal. Using a generalized linear regression, we then estimate differences in the spread of hateful vs. normal content based on author and content variables. We thereby identify important determinants that explain differences in the spreading of hateful vs. normal content. For example, hateful content authored by verified users is disproportionally more likely to go viral than hateful content from non-verified ones: hateful content from a verified user (as opposed to normal content) has a 3.5 times larger cascade size, a 3.2 times longer cascade lifetime, and a 1.2 times larger structural virality. Altogether, we offer novel insights into the virality of hate speech on social media.

Motivation & Objective

• To understand the determinants behind the virality of hate speech on social media, particularly on X (Twitter).
• To identify author and content characteristics that differentiate the spread of hateful content from normal content.
• To address the gap in understanding virality at the content level, as prior research focused only on user-level dynamics.
• To inform policy and detection strategies by identifying structural features that signal high-infectiousness hate speech cascades.

Proposed method

• Collected 25,219 retweet cascades from X (Twitter), including root tweets and all retweets, with human-annotated labels for hate speech.
• Extracted author-level features: number of followers, followees, verified status, tweet volume, and account age.
• Extracted content-level features: presence of media, hashtags, mentions, and tweet length.
• Defined three outcome variables: cascade size (total retweets), cascade lifetime (duration of retweets), and structural virality (measure of bursty spreading).
• Applied generalized linear models (GLMs) to estimate the impact of author and content features on the three spread metrics.
• Used statistical modeling to isolate the effect of verified status and other features on the relative spread of hateful vs. normal content.

Experimental results

Research questions

• RQ1 What are the key determinants that explain differences in the spreading dynamics of hateful versus normal content on X?
• RQ2 How does verified status influence the virality of hate speech compared to normal content?
• RQ3 How do content-level features such as mentions, hashtags, and media presence affect the spread of hate speech?
• RQ4 To what extent do structural properties of retweet cascades (e.g., size, lifetime, structural virality) differ between hateful and normal content?

Key findings

• Hateful content from verified users has a 3.5 times larger cascade size than normal content from verified users.
• Hateful content from verified users has a 3.2 times longer cascade lifetime than normal content from verified users.
• Hateful content from verified users exhibits 1.2 times higher structural virality than normal content from verified users.
• The presence of mentions and hashtags is associated with reduced spread of hate speech, suggesting targeted hate may be less viral than generalized hate.
• Hate speech is more contagious than normal content, with significantly larger cascade sizes, longer lifetimes, and higher structural virality.
• The absence of hashtags and mentions in viral hate speech challenges traditional NLP-based detection, suggesting cascade-level features may be more effective for identification.