Analyzing Islamophobic Discourse Using Semi-Coded Terms and LLMs

Raza Ul Mustafa, Roi Dupart, Gabrielle Smith, Noman Ashraf, Nathalie Japkowicz · 2025

Read Analyzing Islamophobic Discourse Using Semi-Coded Terms and LLMs by Raza Ul Mustafa, Roi Dupart, Gabrielle Smith, Noman Ashraf, Nathalie Japkowicz (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-18273.txt.

Introduction

Raza Ul Mustafa and co-authors used language models to decode semi-coded Islamophobic speech online in a 2025 paper. Extremist platforms circulate slurs respelled to look harmless out of context. Human moderators and automated filters both miss them.

The team gathered posts from platforms such as 4chan, Gab, and Telegram that use these terms. They first tested whether large language models understand the coded words, and found that the models do grasp them. A toxicity service also scored Islamophobic posts higher than other hate categories. Topic modeling then mapped the recurring themes of this discourse. The results argue models can help moderation, though detection still needs work.

Catalog record

CREATORRaza Ul Mustafa, Roi Dupart, Gabrielle Smith, Noman Ashraf, Nathalie Japkowicz
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-18273

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Analyzing Islamophobic Discourse Using Semi-Coded Terms and LLMs Raza Ul Mustafa, Roi Dupart, Gabrielle Smith, Noman Ashraf, Nathalie Japkowicz Abstract Islamophobia 1 1 1 Warning: This paper contains examples of racist and Islamophobic/Antisemitic statements that may be disturbing to the reader. started evolving into a global phenomenon by attracting followers across the globe, particularly in Western societies. Thus, understanding Islamophobia’s global spread and online dissemination is crucial. This paper performs a large-scale analysis of specialized, semi-coded Islamophobic terms such as (muzrat, pislam, mudslime, mohammedan, muzzies) floated on extremist social platforms, i.e., 4Chan, Gab, Telegram, etc. First, we use large language models (LLMs) to show their ability to understand these terms. Second, using Google Perspective API, we also find that Islamophobic text is more toxic compared to other kinds of hate speech. Finally, we use BERT topic modeling approach to extract different topics and Islamophobic discourse on these social platforms. Our findings indicate that LLMs understand these Out-Of-Vocabulary (OOV) slurs; however, measures are still required to control such discourse. Our topic modeling also indicates that Islamophobic text is found across various political, conspiratorial, and far-right movements and is particularly directed against Muslim immigrants. Taken altogether, we performed the first study on Islamophobic semi-coded terms and shed a global light on Islamophobia. I Introduction There has been a significant increase in hate speech on social media platforms, e.g., X, 4Chan, Gab, etc. In the past, offensive language was used on these platforms to incite Islamophobia [ 1 ] ; however, the use of coded, semi-coded and out-of-vocabulary (OOV) language is currently experiencing a significant surge. Coded language is ambiguous, presenting itself as neutral while embedding hateful meanings. To better explain coded and OOV words, the following post was posted on the Bitchute social media platform on August 10, 2024. • “muzrats are gonna be getting it any day now. pack your bags abdul” – Bitchute (2024-08-10) In the above post, it is clear that an Islamophobic connotation is implied. Muzz is a British marriage and dating app for Muslims. However, (muzrats) are two words, muz – muzz and rat are various medium-sized, long-tailed rodents. Similarly, abdul is also used in Islamophobic connotation, a neutral but hateful slur against Muslims. To add more context to neutral text floating on social media platforms, consider this example posted on 4Chan, dated November 5, 2024, on the US election day, 2024. • “Your idiotism won us elections, tank you Abdul . Now get lost.” In the above text, the Islamophobic connotation is quite clear; however, the words tank and thank both have different meanings. A tank is an armored fighting vehicle intended as a primary offensive weapon in front-line ground combat, whereas thank is an expression of gratitude. Moreover, by adding context to the message, abdul is an Islamic name; however, it is a neutral word used in both ways. Similarly, on extremist platforms, we find other terms like mohammedan , pislam , shitskin and Muzzies . These term, while not offensive in its literal meaning (except Shitskin ) is used derogatorily to imply something negative about the Muslim identity, differentiating and alienating Muslims in a subtle but damaging way. Thus, awareness of these terminologies, similar to abdul and OOV words such as pislam, muzrats, mohammedan are necessary for accurate hate speech detection against Muslims. The recent advancements in Natural Language Processing (NLP), including Large Language Models (LLMs), (e.g., GPT, LLama, etc.), provide a contextual understanding of the text [ 2 ] . We also argue that LLMs like GPT-4 and recent versions understand these OOV terminologies (slur) used against Muslims. Therefore, we recommend that social platforms integrate these OOV and coded terms and related expressions into their content moderation strategies to create a more inclusive and safer space for all users. We believe that whenever these slurs are used against Muslims, their purpose is not to discuss a specific type of trope but to blame Muslims for various issues worldwide. Finally, we argue that the toxicity of words expressed against Muslims is more severe than hate speech against other religions. Therefore, in this research, we try to verify the following hypotheses: • H1: We hypothesize that advanced LLMs understand the OOV words used against Muslims. Moreover, in recent years, the most floated trope on extremist social media platforms is Muslim immigration to white countries . • H2: Hate speech against Muslims is more violent with toxic content, such as the use of abusive and racial remarks. • H3: The social users who share Islamophobic content discuss various topics and tropes, and the content changes over time and is linked to global tensions. To verify our hypothesis H1 , we use GPT-4 to classify OOV and neutral text. We selected a random sample from our labeled dataset, which is benign without toxic language but contains these terms. For hypothesis H2 , we compare the toxicity of Antisemitism with Islamophobia. We select Antisemitism because it targets Jewish people based on their religious, ethnic, or cultural identity, in this respect, is similar to Islamophobia. Finally, for hypothesis H3 , we used BERT topic modeling, a state-of-the-art transformer architecture with contextual understanding to extract topics. The choice of BERT is due to its ability to understand the contextual representation of text rather than giving a fixed meaning to words. The rest of the paper is organized in the following way: In section II , we provide background and state-of-the-art work, followed by dataset collection for our study in section III . Next, Section IV discusses the results of all the hypotheses. We conclude our paper in Section V and discuss future work directions. II Background and State-of-the-art Hate speech often employs slurs and stereotypes to demean race, religion, or cultural practices [ 3 ] . Attacks on race, for instance, may generalize all Muslims as belonging to specific racial groups, such as Arabs, South Asians, or Africans, using derogatory terms rooted in colonial or racial hierarchies. Similarly, religious hate speech shifts focus to Islam as a belief system, often mocking its practices or associating it with violence or extremism. Likewise, anti-Muslim rhetoric has recently surged, especially directed at Muslim immigrants in Western nations, where racial tensions are already high. This rhetoric often includes xenophobic language, dehumanizing Muslims as creatures or accusing them of stealing a country. Another recurring theme in Islamophobia is to target the Prophet Muhammad, which frequently involves defamatory claims designed to discredit him as a religious figure. Common tropes include references to his multiple wives, participation in wars, or accusations about the ages of his wives. For instance, accusing Prophet of being a pedophile or rapist attempting to delegitimize him through inflammatory language. A phrase like piss be upon him mocks the Islamic phrase “peace be upon him” , aiming to insult the Prophet and, by extension, the faith he represents. Finally, there are stereotypes about Muslim women, particularly surrounding modesty practices, polygamy, and cultural traditions. These religious themes often contain OOV and neutral terms such as muzrat, pislam, mudslime, mohammedan and muzzies, to discuss various political tensions and target Muslims. These terms are introduced in a way to convey a disrespectful and toxic message to Muslims. We show an example of each OOV term in

Read the complete text