maitry30/verilang-muril

🔍 VeriLang — Vernacular Misinformation Detector

AI-powered fake news detector for WhatsApp forwards in Hindi, Gujarati, Marathi and Telugu using Google's MuRIL model.

---

🎯 Problem Statement

WhatsApp misinformation in Indian regional languages is a serious problem affecting millions of people daily. Existing fake news detectors work only for English. VeriLang solves this by detecting misinformation in 4 major Indian languages — Hindi, Gujarati, Marathi and Telugu.

---

✨ What Makes This Unique

Feature	Other Projects	VeriLang
Languages	English only	4 Indian languages
Model	Generic BERT	MuRIL — Indian language specialist
Target	General news	WhatsApp forwards specifically
Explainability	None	SHAP word importance
Deployment	Jupyter only	Live web app

---

📊 Dataset

• Source: Extracted 4 different language dataset from Zenodo — Multilingual Fake News Detection (https://zenodo.org/records/11408513)
• Size: 49,426 articles
• Labels: Fake (0) / Real (1)
• Languages: Hindi, Gujarati, Marathi, Telugu

Language	Total	Fake	Real
Hindi	15,051	7,599	7,452
Gujarati	14,830	6,145	8,685
Telugu	11,424	4,795	6,629
Marathi	8,121	707	7,414

---

🔬 Methodology

Phase 1 — Data Cleaning

• Removed URLs, HTML tags, extra spaces
• Kept unicode scripts intact for Indian languages
• No duplicates, no missing values

Phase 2 — Baseline ML Models

• TF-IDF with character n-grams (2-4 grams)
• Character n-grams work better than word n-grams for Indian scripts
• Trained Logistic Regression, SVM, Naive Bayes

Phase 3 — MuRIL Fine-tuning

• Used google/muril-base-cased from HuggingFace
• MuRIL is trained on 17 Indian languages
• Fine-tuned for 3 epochs on Google Colab T4 GPU
• Applied class weights to fix Marathi imbalance

---

📈 Model Performance

Overall Accuracy

Model	Accuracy
Logistic Regression	99.63%
Naive Bayes	97.77%
SVM	99.84%
MuRIL (final)	99.92%

Per Language Accuracy

Language	SVM	MuRIL
Hindi	99.66%	99.80%
Gujarati	99.86%	99.93%
Marathi	99.94%	100.00%
Telugu	99.96%	100.00%

Overfitting Check

Model	Train	Test	Gap	Status
Logistic Reg	99.60%	99.63%	0.03%	No overfit
SVM	99.99%	99.84%	0.15%	No overfit
MuRIL	—	99.92%	—	CV std 0.05%

---

🛠 Tech Stack

Category	Tools
Deep Learning	PyTorch, HuggingFace Transformers
Model	google/muril-base-cased
ML Baseline	Scikit-learn, TF-IDF
Explainability	SHAP
Web App	Streamlit
Model Hosting	HuggingFace Hub
Training	Google Colab T4 GPU

---

💻 How to Run Locally

git clone https://github.com/Maitry09/verilang-misinformation-detector
cd verilang-misinformation-detector
pip install -r requirements.txt
streamlit run app.py

Add .streamlit/secrets.toml:

HF_TOKEN = "your_token_here"

---

⚖️ Ethical Considerations

• Tool is for awareness only — not a final judgment system
• Always verify from official sources like PIB, ANI, PTI
• Model may have bias toward certain writing styles
• Marathi fake detection improved with augmentation but still has fewer training samples

---

🚀 Future Work

• Add Bengali, Tamil, Kannada support
• Work on mixture of language
• Build WhatsApp bot integration using Twilio API
• Add image-based misinformation detection
• Real-time news verification via fact-check APIs
• Mobile app for direct WhatsApp forward checking

---

👤 Author

Maitry

• GitHub: @Maitry09
• Live App: [verilang.streamlit.app]

---

📄 License

MIT License

---

⭐ Star this repo if you found it useful!