"A Virus Has No Religion": Analyzing Islamophobia on Twitter During the COVID-19 Outbreak

Mohit Chandra, Manvith Reddy, Shradha Sehgal, Saurabh Gupta, Arun Balaji Buduru, Ponnurangam Kumaraguru

TL;DR

This paper introduces "CoronaBias," a large-scale dataset of over 410,000 tweets analyzing Islamophobia on Twitter during the COVID-19 pandemic. It reveals a significant correlation between real-world events and online anti-Muslim rhetoric, identifying prevalent hateful topics and characterizing users who spread such content.

What Problem Does This Paper Address?

The COVID-19 pandemic, while a global health crisis, also unfortunately fueled a surge in xenophobia and discrimination. While research has explored Sinophobia (anti-Chinese sentiment) during this period, Islamophobia received less attention. This paper addresses this gap by investigating the rise of anti-Muslim sentiment on Twitter, particularly in the Indian subcontinent, during the pandemic. The authors highlight how events like the Tablighi Jamaat gatherings were exploited to spread misinformation and fuel hate speech, leading to real-world consequences. They aim to provide a quantitative understanding of this phenomenon to inform interventions.

Methods

The researchers employed a multi-faceted approach:

  • Dataset Creation: They curated the "CoronaBias" dataset by filtering tweets related to COVID-19 and Islam using a comprehensive list of keywords.
  • Temporal Analysis: Using the PELT algorithm, they identified significant change points in tweet frequency and sentiment, correlating these with real-world events.
  • Semantic Similarity: Word2Vec (CBOW) models were trained monthly to track how the semantic context around "Muslim" and "Virus" evolved, indicating a shift towards associating Muslims with the virus.
  • Topic Modeling: Non-Negative Matrix Factorization (NMF) was used for both macro-level (overall prevalent topics) and micro-level (topics around specific events) analysis to identify themes in hateful and non-hateful tweets.
  • User Analysis: Word clouds from user bios and network graph analysis were used to characterize users based on their posting behavior (percentage of hateful tweets) and identify ideological leanings.
  • External URL Analysis: The toxicity of external content linked in tweets was analyzed using the Perspective API to understand the nature of information being shared.

Datasets

  • CoronaBias Dataset: A newly introduced dataset comprising 410,990 tweets from 244,229 unique users, collected from February 1, 2020, to May 31, 2020. This dataset is publicly available.
  • External Content Datasets: Data was also collected from YouTube (764 videos), OpIndia (231 articles), and BBC (82 articles) for external URL analysis.

Key Insights and Ideas

  • Offline Events Drive Online Hate: The study strongly demonstrates that real-world events, particularly those involving the Tablighi Jamaat, directly triggered spikes in anti-Muslim rhetoric on Twitter.
  • Semantic Shift: The meaning of "Muslim" on Twitter gradually became associated with "Virus" and "COVID-19" as the pandemic progressed, indicating a deliberate conflation.
  • Patriotism as a Shield for Hate: User bio analysis revealed that terms related to nationalism and Hinduism were often used by those posting hateful content, suggesting religion was weaponized as a symbol of patriotism to label Muslims as "anti-national."
  • Echo Chambers of Hate: Network analysis showed that users posting a high percentage of hateful tweets formed tightly clustered communities, suggesting the formation of echo chambers.
  • Toxic External Sources: A significant portion of tweets linked to external news sources, with OpIndia being a prominent referrer of Islamophobic content, highlighting the role of biased media in amplifying hate.

Key Results

  • The CoronaBias dataset contains 410,990 tweets, with a significant number of hateful tweets identified.
  • Temporal analysis revealed that events like the Tablighi Jamaat congregation announcement on March 31st/April 1st, 2020, led to a global maxima in hateful tweets, even surpassing non-hateful tweets for a period.
  • Semantic similarity experiments showed a clear increase in the association between "Muslim" and "Virus" from February to May 2020.
  • Topic modeling indicated that hateful topics related to the Tablighi Jamaat and anti-Abrahamic sentiments were more prevalent than non-hateful topics.
  • User analysis showed a progression from general terms to nationalism and Islamophobic rhetoric in bios as users were categorized into higher "hateful tweet" percentages.
  • External URL analysis found that approximately 86% of tweets linking to OpIndia fell into higher toxicity score bins, compared to BBC where most links had low toxicity.

Limitations and Future Work

  • Manual Annotation Limitations: While semi-automatic annotation was used, the sheer volume of data means a portion was classified by models, potentially introducing some error.
  • Focus on Indian Subcontinent: The study primarily focuses on the Indian subcontinent, and findings might not generalize globally without further research.
  • Future Work: The authors suggest exploring mitigation strategies for Islamophobia and expanding the analysis to other platforms and regions. They also propose further investigation into the role of bots and the impact of misinformation on real-world hate crimes.

Summary updated

BibTeX
@inproceedings{chandra2021virus,
  author = {Mohit Chandra and Manvith Reddy and Shradha Sehgal and Saurabh Gupta and Arun Balaji Buduru and Ponnurangam Kumaraguru},
  title = {{"A Virus Has No Religion": Analyzing Islamophobia on Twitter During the COVID-19 Outbreak}},
  booktitle = {HT},
  year = {2021},
  doi = {10.1145/3465336.3475111},
}

Loading related content...

Figure