TL;DR
This paper introduces HP-BERT, a novel framework and model for detecting and analyzing Hinduphobic sentiment on social media, particularly during the COVID-19 pandemic. By curating a new dataset and fine-tuning a BERT-based model, the authors demonstrate a significant correlation between pandemic stress and increased Hinduphobic discourse, offering crucial insights into online religious discrimination.
What Problem Does This Paper Address?
The COVID-19 pandemic, while a global health crisis, also exacerbated existing social divisions and fueled discriminatory sentiments against various religious groups. While research has addressed anti-Islamic and anti-Chinese sentiments, the escalation of Hinduphobia on social media during this period remained critically underexplored. Existing computational frameworks for hate speech detection often struggle with culturally specific nuances and code-mixed languages (like Hinglish), creating a gap in our ability to monitor and understand this form of online discrimination. This paper aims to fill that gap by developing a robust computational framework to systematically analyze Hinduphobic content on social media platforms.
Methods
The core of this research lies in the development of the Hinduphobic BERT (HP-BERT) model. This framework employs a multi-stage fine-tuning approach:
- Hinduphobic Content Detection: A pre-trained BERT model is first fine-tuned on a custom-built "Hinduphobic COVID-19 X (Twitter) Dataset." This stage focuses on accurately classifying tweets as either Hinduphobic or non-Hinduphobic/neutral.
- Sentiment Analysis: The HP-BERT model is then further fine-tuned using the "SenWave dataset." This step equips the model with multi-label sentiment analysis capabilities, allowing it to not only detect Hinduphobic content but also to categorize the sentiment (e.g., optimistic, annoyed, denial) associated with it.
The model architecture leverages a Transformer-based encoder to process tokenized tweets, including Hinglish, and a fully connected layer for classification. The framework also incorporates a custom weight strategy for calculating polarity scores, aiming for more nuanced sentiment representation than standard methods.
Datasets
- Hinduphobic COVID-19 X (Twitter) Dataset: A newly curated dataset of 8,000 manually annotated tweets collected from November 2019 to December 2022. This dataset is crucial for training and evaluating the HP-BERT model's Hinduphobia detection capabilities.
- SenWave Dataset: A large-scale dataset containing over 104 million tweets in six languages related to COVID-19, used for fine-tuning HP-BERT's sentiment analysis capabilities.
- Global COVID-19 X (Twitter) Dataset: A comprehensive collection of 27.4 million tweets from six countries (Australia, Brazil, India, Indonesia, Japan, and the UK) spanning March 2020 to February 2022. This dataset is used for longitudinal analysis and examining global sentiment trends.
Key Insights and Ideas
- Pandemic Stress Fuels Discrimination: The study provides quantitative evidence linking increases in COVID-19 cases to spikes in Hinduphobic content, suggesting that pandemic-related stress can amplify existing social divisions and lead to scapegoating.
- Culturally-Specific Models Matter: General-purpose hate speech models (like HateBERT) struggle with the nuanced and culturally specific nature of Hinduphobia, highlighting the need for domain-specific fine-tuning.
- Hinglish is a Challenge: The prevalence of code-mixed languages like Hinglish in social media discourse presents a significant challenge for NLP models, necessitating robust handling of linguistic variations.
- Stereotypes Persist and Evolve: The analysis reveals how colonial-era stereotypes about Hindus (e.g., association with cow urine, "superspreader" narratives) are amplified and adapted in the context of the pandemic.
- India as a Focal Point: While Hinduphobic discourse is global, India emerges as a primary hub for such content, likely due to its large Hindu population and the intense socio-political dynamics surrounding the pandemic.
Key Results
- HP-BERT Performance: The HP-BERT model achieved 94.72% accuracy in detecting Hinduphobic tweets, significantly outperforming five state-of-the-art transformer models by 27.51% to 36.38% over baseline applications.
- Correlation with COVID-19 Cases: Statistical analysis revealed moderate positive correlations (r = 0.312-0.428) between COVID-19 case increases and the volume of Hinduphobic content across six countries.
- Longitudinal Analysis: The study analyzed 27.4 million tweets, uncovering previously undocumented geographical and temporal patterns of anti-Hindu sentiment during the pandemic.
- Dominant Sentiments: "Annoyed" and "official report" were the most prevalent sentiments in Hinduphobic tweets, often driven by narratives critical of traditional practices or perceived bias.
Limitations and Future Work
- Dataset Bias: The reliance on Twitter data means the findings might not fully capture the diversity of public opinion, as Twitter's user base has specific demographic skews. The 280-character limit also restricts the depth of expressed sentiments.
- Nuance in Sentiment: The HP-BERT model, while advanced, still faces challenges in accurately detecting sarcasm, irony, and negation, which are crucial for nuanced sentiment analysis.
- Multilingual Challenges: While Hinglish is addressed, the model's effectiveness across a wider range of languages and dialects requires further exploration.
- Future Directions: The authors suggest expanding data sources to include platforms like Facebook, Instagram, and Reddit. Future work could also explore modular neural architectures, irony detection models, and integrating cultural, religious, historical, political, and economic factors for a more comprehensive understanding of Hinduphobic sentiments.
Summary updated