Decoding Islamophobic Discourse: Using LLMs to Identify Tropes and Semi-Coded Hate Speech

Technical Summary: Decoding Islamophobic Discourse Using LLMs

Problem Statement
The paper addresses the escalating prevalence of Islamophobic hate speech on extremist social media platforms (e.g., 4Chan, Gab, Telegram). A primary challenge identified is the shift from overt hate speech to "semi-coded" and Out-Of-Vocabulary (OOV) terminology. These terms (e.g., muzrat, pislam, mudslime, mohammedan) are linguistically manipulated to appear lexically neutral or ambiguous outside specific contexts, thereby evading detection by both human moderators and traditional automated systems. The authors argue that while these terms may seem harmless in isolation, they carry specific derogatory connotations within extremist communities. Furthermore, the paper posits that current moderation strategies struggle to distinguish between benign usage and hate speech when context is subtle, and that the toxicity levels of Islamophobic discourse may differ significantly from other forms of hate speech, such as Antisemitism.

Methodology
The study employs a mixed-method approach combining large-scale data collection, manual annotation, and Large Language Model (LLM) evaluation.

1. Data Collection: The authors utilized the software Pyrra to scrape posts from various alt-social media platforms between 2024 and 2025, resulting in a dataset of approximately 63,628 instances. The data was filtered using seed terms including known OOV slurs (muzrat, pislam, etc.).
2. Annotation and Tropes: A subset of the data was manually labeled by undergraduate researchers and experts in Islamic terminology. The labeling process categorized content into five specific Islamophobic tropes:
   • Women's Oppression
   • Attacks on Prophet Muhammad
   • Hate against Islam (as a belief system)
   • Anti-Immigrant rhetoric
   • Racial Bias
3. Hypothesis Testing:
   • H1 (LLM Understanding of OOV Terms): The authors used GPT-4 to classify posts containing semi-coded terms, assigning a negativity score (0–1) to determine if the model could identify the hate speech without explicit toxic keywords.
   • H2 (Toxicity Comparison): Using the Google Perspective API, the study compared the toxicity scores of Islamophobic posts against a baseline dataset of Antisemitic posts (350 samples each).
   • H3 (Trope Classification): The authors evaluated GPT-4's ability to classify the specific tropes present in the discourse, comparing the model's predictions against human-labeled ground truth using precision, recall, and F-scores.

Key Results

• LLM Comprehension of OOV Terms: GPT-4 demonstrated a strong ability to identify the hateful intent behind semi-coded terms. The mean toxicity score assigned by GPT-4 to Islamophobic slurs was 0.8767, significantly higher than a neutral baseline of 0.5 (t = 10.78, p < 0.05). However, the model's accuracy relied heavily on context; for instance, the term mohammedan was sometimes rated as neutral when lacking other aggressive cues, and the name Abdul was occasionally misidentified as a benign reference rather than a slur.
• Toxicity Levels: The study found that Islamophobic discourse exhibits significantly higher toxicity than Antisemitic discourse. The mean toxicity score for Islamophobic samples was 0.67, compared to 0.40 for Antisemitic samples (t = 17.01, p ≈ 0). The most toxic tropes were identified as those targeting Muslim immigrants and the Prophet Muhammad.
• Trope Classification Limitations: While GPT-4 performed well in identifying tropes related to the Prophet Muhammad (F-score: 0.88) and hate against the religion of Islam (F-score: 0.76), it struggled with tropes involving Racial Bias, Immigration, and Women's Oppression. The confusion matrix revealed that the model often failed to distinguish between these categories, likely due to the reliance on specific keyword associations (e.g., "pedophile" for the Prophet trope) rather than a deep understanding of the underlying ideological narratives.

Significance and Claims
The paper claims to be one of the first large-scale studies focusing specifically on semantically coded and OOV terms in the context of Islamophobia. Its primary contributions are:

1. Demonstrating LLM Capability: Showing that state-of-the-art LLMs (specifically GPT-4) possess the contextual understanding necessary to recognize semi-coded slurs that traditional keyword-based filters might miss.
2. Toxicity Disparity: Highlighting that Islamophobic content on extremist platforms is statistically more toxic and violent than Antisemitic content, suggesting a need for differentiated moderation strategies.
3. Identification of Gaps: The authors modestly conclude that while LLMs can identify the terms, they still exhibit limitations in distinguishing the specific ideological tropes driving the discourse, particularly when the hate is subtle or relies on complex cultural stereotypes.

The authors suggest that social media platforms should integrate these specific OOV and semi-coded terms into their moderation strategies. They also note that future work should involve training specialized models to detect these terminologies and expanding the analysis to other coded slurs (e.g., pajeet, chink) to foster safer online environments.