🌍 Multilingual Summarizer (Bengali + English)

A multilingual text summarizer built with Hugging Face Transformers that supports both Bengali and English.
This project demonstrates state-of-the-art abstractive summarization using mT5-small, optimized with ONNX runtime and 8-bit quantization for faster inference.

---

✨ Model Highlights

• ✅ Supports two languages: Bengali + English
• ✅ Transformer-based summarization using MT5ForConditionalGeneration
• ✅ Evaluated with ROUGE metrics (best ROUGE-L = 0.0654)
• ✅ Optimized with ONNX runtime + 8-bit quantization
• ✅ Post-processing filters for cleaner summaries
• ✅ Inference Speedup – Reduced runtime from 71s → 21s

---

📊 Datasets

• Bengali: Prothom Alo News (Kaggle)
• English: CNN/DailyMail (Kaggle)

Data columns:

• description → Input text
• summary → Target summary

---

⚙️ Model & Training

• Tokenizer: MT5Tokenizer
• Base Model: MT5ForConditionalGeneration (google/mt5-small)
• Training Args:
  • learning_rate = 2e-5
  • num_train_epochs = 6
  • per_device_train_batch_size = 8
  • evaluation_strategy = "epoch"
• Best Validation Loss: 0.8312 (after 6 epochs)

---

📈 Results

Metric	Value
Validation Loss	0.8312
ROUGE-L	0.0654

---

⚡ Optimization

• ✅ ONNX Runtime – reduced from 71.4s → 58.5s
• ✅ 8-bit Quantization – reduced further to 21.4s

---

🛠️ Post-Processing

• filter_short_sentences() → Removes sentences with < 3 words
• Beam search decoding parameters:

  max_length = 500
  num_beams = 4
  early_stopping = True

---

📂 Repository Structure

QuickSum/
├── training/
│   └── training.ipynb           # Model training experiments, Evaluation & ROUGE scoring
│   
├── optimization/
│   ├── onnx_optimize.ipynb         # ONNX conversion + quantization
│   └── postprocess.ipynb           # Post-processing filters
│
├── static/
│   ├── css/                   # cascaded style file
│   └── js/                    # javascript file
│
├── templates/
│   └── index.html             # front end skeleton
│
├── .gitignore
├── LICENSE
├── README.md
└── requirements.txt

Setup Instructions

1. Install Dependencies: Make sure you have Python installed on your system. Install the required Python libraries by running the following command:

pip install -r requirements.txt

2. Run the FastAPI Server: To start the server, use the following command:

uvicorn server:app --reload

Once the server is running, you can access the API at:

• http://127.0.0.1:8000/ (Main endpoint)

📸 Demo

Summarizer API	Bengali Text	English Text

🧑‍💻 Tech Stack

• Python
• Hugging Face Transformers
• PyTorch
• ONNX Runtime
• FastAPI

---

🎯 Key Takeaways

This project highlights:

• Using Transformer-based architectures
• Handling multilingual NLP tasks
• Model training + evaluation + optimization
• Applying quantization for production-ready inference

---

📬 Contact

If you'd like to discuss NLP, transformers, or optimization techniques, feel free to connect! 🚀