📄 Django-Based Bangla Document Classification Using Deep Learning

Enhancing Bangla Document Classification Using a Hybrid Ensemble of Bangla-BERT and Bi-LSTM Models
(IDAA 2025 – Published Research)

This repository contains a production-ready Django web application for Bangla document classification using a hybrid ensemble of Bangla-BERT and Bi-LSTM models.
The system is designed for research, academic demonstration, and real-world deployment of Bangla NLP models.

---

🚀 How to Run the Project

Option 1: Docker (Recommended for Deployment) 🐳

This project is production-ready and fully containerized with Docker.

Build the Docker Image

docker build -t bangla-classifier .

Run Locally

docker run -p 8000:8000 bangla-classifier

Open in Browser

http://localhost:8000

Deploy to Cloud (DigitalOcean, AWS, etc.)

# Tag for Docker Hub
docker tag bangla-classifier your-username/bangla-classifier
docker push your-username/bangla-classifier

# Or for DigitalOcean Container Registry
docker tag bangla-classifier registry.digitalocean.com/your-registry/bangla-classifier
docker push registry.digitalocean.com/your-registry/bangla-classifier

---

Option 2: Local Development (Without Docker)

1️⃣ Clone the Repository

git clone https://github.com/Mahedi9/Django-Bangla-Document-Classification-Using-Deep-learning.git
cd Django-Bangla-Document-Classification-Using-Deep-learning

2️⃣ Install Dependencies

pip install -r requirements.txt

3️⃣ Run Database Migrations

python manage.py migrate

4️⃣ Start the Django Server

python manage.py runserver

5️⃣ Open in Browser

http://127.0.0.1:8000/

---

🧠 Project Overview

The application classifies Bangla news articles into eight categories using a hybrid ensemble learning strategy:

• Bi-LSTM for sequential text representation
• Bangla-BERT for contextual semantic understanding
• weighted voting ensemble with confidence-based decision logic

The system produces:

• Final prediction with confidence
• Model-wise predictions
• Probability distribution visualizations
• Ambiguity warnings when predictions are uncertain

---

✨ Core Features

• Hybrid ensemble of Bi-LSTM + Bangla-BERT
• Confidence threshold–based ambiguity detection
• Model-wise prediction cards with probabilities
• Probability bar charts for each model
• Class-wise probability table
• Minimum input validation (≥ 20 Bangla words)
• Clean, responsive UI using Bootstrap
• Modular Django ML architecture

---

🧠 Supported News Categories

• Economy
• Education
• Entertainment
• International
• National
• Politics
• Science_Technology
• Sports

---

🧩 System Architecture

bangla_classifier_django/
├── bangla_classifier_django/
│ ├── settings.py
│ ├── urls.py
│ ├── asgi.py
│ └── wsgi.py
│
├── classifier/
│ ├── ml/
│ │ ├── loader.py (Model loading)
│ │ ├── preprocess.py (Text cleaning & stopword removal)
│ │ └── predict.py (Inference & ensemble logic)
│ │
│ ├── templates/classifier/
│ │ └── index.html (UI template)
│ │
│ ├── views.py (Controller logic)
│ ├── urls.py
│ └── apps.py
│
├── models/
│ ├── bilstm_model.keras
│ ├── bilstm_tokenizer.pkl
│ ├── bilstm_label_encoder.pkl
│ └── bangla_bert_model/
│
├── static/
├── Dockerfile
├── manage.py
├── requirements.txt
├── .gitignore
└── .gitattributes

---

🔍 Prediction Workflow

1. User submits a Bangla news article (minimum 20 words)
2. Text is preprocessed (cleaning, stemming, stopword removal)
3. Bi-LSTM and Bangla-BERT generate class probabilities
4. Probabilities are combined using soft voting
5. Confidence threshold determines final decision
6. Results are visualized and displayed

---

📊 Experimental Results (From Published Paper)

• Dataset: Potrika Bangla News Dataset
• Total Articles: 329,110
• Number of Classes: 8
• Hybrid Ensemble Accuracy: 97.16%

---

📄 Research Publication

Title
Enhancing Bangla Document Classification Using a Hybrid Ensemble of Bangla-BERT and Bi-LSTM Models

Conference
International Conference on Intelligent Data Analysis and Applications (IDAA 2025)

Venue
Daffodil International University, Dhaka, Bangladesh

---

👤 Author

Mahedi Hasan Emon
Researcher | Bangla NLP | Deep Learning
Published Author – IDAA 2025

---

🙏 Acknowledgements

• Potrika Bangla News Dataset
• Hugging Face Transformers
• TensorFlow
• PyTorch
• bnltk
• Daffodil International University

---

📌 License

This project is released for research and educational use.
For commercial usage, please contact the author.

---