Introduction
The digital age has seen a significant rise in online hate speech, yet computational efforts to detect and mitigate this phenomenon are often unevenly distributed across different religious and ethnic groups. While tools for identifying Islamophobia and Sinophobia have seen substantial advancement, Hinduphobia—the irrational fear, hatred, or prejudice against Hindus—remains a critically under-explored domain in natural language processing (NLP).
During the COVID-19 pandemic, this gap became particularly evident as social tensions escalated. Discriminatory discourse often targeted religious communities, blaming them for the spread of the virus. General-purpose hate speech models frequently fail to identify these nuances because Hinduphobic rhetoric often relies on specific cultural tropes, such as mocking traditional practices or using derogatory references that require deep contextual understanding. To address this, researchers developed HP-BERT, a specialized framework designed for the longitudinal analysis of Hinduphobia.
Methodology and Data Pipeline
The core of the HP-BERT framework is a sophisticated multi-stage fine-tuning and data curation process. Recognizing that "hate" is not a monolith, the researchers employed a Human-in-the-Loop (HITL) strategy to ensure high-quality training data.
The pipeline involved several critical stages:
- Data Curation: A hybrid annotation strategy where 2,000 tweets were manually labeled to fine-tune GPT-3.5 Turbo. This model then generated "silver-standard" labels for 15,000 additional tweets, which were subsequently verified by human experts to produce a "gold standard" dataset of 8,000 tweets.
- Model Architecture: Built on the BERT (Bidirectional Encoder Representations from Transformers) base, HP-BERT was first fine-tuned for binary classification (Hinduphobic vs. Non-Hinduphobic).
- Sentiment Integration: The model underwent a second stage of fine-tuning using the SenWave dataset. This allowed the model to perform multi-label sentiment analysis, detecting specific tones such as "annoyed," "joking," or "pessimistic."
- Longitudinal Deployment: The final model was used to analyze a massive corpus of 27.4 million tweets from six countries (India, Brazil, Australia, Indonesia, Japan, and the UK) to track sentiment evolution from April 2020 to January 2022.
Datasets
The study's contributions include the release of the Hinduphobic COVID-19 X Dataset, consisting of 8,000 annotated and manually verified tweets (4,600 Hinduphobic, 3,400 neutral/positive). This serves as a benchmark for future research in the field. Additionally, the researchers utilized the SenWave dataset for sentiment tuning and a global corpus of 27.4 million tweets to provide a macro-level view of online discourse during the pandemic.
Key Insights: The Context Gap and Pandemic Amplification
The research revealed a significant "Context Gap" in existing NLP tools. General hate speech models like HateBERT achieved only 67.21% accuracy on the Hinduphobia dataset, whereas HP-BERT reached 94.72%. This disparity underscores the necessity of training models on culturally specific tropes.
Furthermore, the longitudinal analysis identified a moderate correlation ($r = 0.312$ to $0.428$) between surges in COVID-19 cases and the volume of Hinduphobic content. Public health crises appear to act as catalysts for religious scapegoating. Notably, Hinduphobic tweets were characterized by a "stark absence of empathy," frequently targeting religious gatherings like the Kumbh Mela by re-packaging colonial-era caricatures for the digital age.
Results and Performance
HP-BERT's performance was robust, outperforming baseline Transformer models by margins ranging from 27.51% to 36.38%. Even when those baselines were fine-tuned on the same specialized data, HP-BERT maintained a performance lead of approximately 5.7%.
Geographically, India accounted for the highest volume of detected Hinduphobic content, followed by the UK and Australia. Statistical analysis (One-way ANOVA) confirmed that the "during-COVID" period saw significantly higher volumes of hate speech compared to pre- or post-peak periods ($p < 0.001$).
Limitations and Future Directions
While HP-BERT represents a significant step forward, the authors note several limitations. The study is confined to the X platform, which may not reflect the broader population's views. Additionally, the model still faces challenges with "Hinglish" (a mix of Hindi and English) and complex linguistic features like deep irony or sarcasm. Future work aims to expand the framework to platforms like Reddit and Facebook and incorporate multi-modal analysis to detect Hinduphobic memes and imagery.
Summary updated