Detecting and Categorizing Islamophobic Hate Speech on Spanish Twitter/X

Abstract

Islamophobic hate speech on social media includes overt toxicity and contextual frames that associate Muslims with invasion, criminality, cultural threat, or racialized exclusion. This study examines 21,326 Spanish-language tweets about Islam and Muslims posted in Spain between 2010 and 2022. We combine human annotation, exploratory discursive coding, Perspective API scores, and a multilingual BERT classifier. Human coders classified 28.2% of tweets as Islamophobic hate speech. Explicit hostility was the most frequent pattern, followed by invasion, cultural threat, criminality, race/ethnicity, and socioeconomic burden. Although hate-labeled tweets received significantly higher toxicity-related scores than non-hate tweets, many remained below a 0.70 high-probability threshold. The BERT classifier achieved 0.89 accuracy, 0.97 precision, 0.85 recall, and 0.91 F1 for the hate class. The findings show that generic toxicity measures capture overt aggression more effectively than contextual hostility, supporting target-specific, context-sensitive detection.