Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on Twitter

Online hate speech represents an escalating challenge for social media platforms, posing direct risks to user safety due to its established connection to real-world hate crimes. Platform moderation currently relies heavily on manual human review, which is difficult to scale, introduces subjective individual bias, and struggles to balance content removal with free speech protections. The article addresses the urgent need for standardized definitions and automated methods to identify hate speech accurately, specifically examining whether incorporating extra-linguistic user information improves automated detection.

The article evaluated the effectiveness of character-level textual features alongside demographic and geographic data for detecting hate speech on Twitter. Specifically, it set out to determine whether identifying user demographics, such as gender and geographic location, improves identification performance over text-based models alone.

To conduct this evaluation, the researchers established an 11-point annotation guide grounded in critical race theory to standardize what constitutes hate speech. Using these criteria, they collected and manually labeled an unbalanced dataset of 16,914 English-language tweets gathered over two months, categorizing them into sexist, racist, or non-offensive content. They used logistic regression classifiers with ten-fold cross-validation to compare various feature combinations, assessing character sequences (character n-grams) against word-level tokens, text lengths, geographic location, and inferred user gender.

The investigation yielded several critical findings regarding feature effectiveness. First, models using character sequences significantly outperformed traditional word-based models, improving the F1 performance score from 64.58 to 73.89. Second, adding extra-linguistic features generally degraded classification performance: incorporating user location reduced the F1 score to 73.62, and incorporating text length reduced it to 73.66. Third, while adding inferred gender produced the highest overall F1 score of 73.93, this minor improvement was not statistically significant. Fourth, the demographic analysis revealed a strong gender imbalance among identified perpetrators, with men accounting for 50.24% of sexist tweets and 33.33% of racist tweets, while women accounted for under 2.3% in each category.

These findings suggest that automated moderation systems should prioritize sub-word character analysis over demographic profiling and metadata. Because abusive language frequently repurposes ordinary words or introduces spelling variations, character-level modeling provides greater accuracy and resilience against sparse text data. Furthermore, relying on user metadata introduces noise that can harm system performance while adding unnecessary computational complexity and data governance burdens.

For organizations developing automated content moderation pipelines, the immediate recommendation is to deploy character-based sequence modeling as the baseline text analysis tool rather than relying solely on static keyword blocklists or demographic metadata. Operational teams should also establish formal, structured annotation guidelines, such as the criteria introduced in the article, to reduce annotator subjectivity during human review. Before demographic signals can be reliably used in production systems, further work is required to develop more comprehensive user profile inference methods.

Decision-makers should view these results with caution regarding demographic factors due to major data coverage limitations. Gender could only be identified for 52.56% of users in the dataset, and location metadata was similarly scarce. Additionally, the racist subset within the dataset was concentrated among a very small pool of users, which may limit generalizability across wider social media environments.