NewsSelect

🚀 NewsSelect — AI-Powered News Summarization App

📌 Demo & Code

An end-to-end, AI-powered web application for real-time, abstractive news summarization using a Seq2Seq LSTM model with Attention — integrated with Django backend, web scraping for live news, and a responsive frontend UI.

📌 1️⃣ Problem Statement

In the digital age, readers face information overload from countless news sources. Manually scanning lengthy articles is inefficient and impractical.

✅ Solution: Build an intelligent system that can automatically fetch live news articles, summarize them using an AI model, and display them to users through a clean web interface.

📌 2️⃣ Technology Stack & Rationale

Technology	Purpose	Why?
Python 3.x	Core language for ML and backend development	High ecosystem support for ML, data, and web frameworks.
TensorFlow 2.x / Keras	Deep learning framework for model implementation	Native support for Seq2Seq + Attention, easy TPU integration.
Pandas / NumPy	Data manipulation and numerical computation	Efficient, reliable, and industry standard.
Matplotlib	Visualizing training metrics	Quick and effective for loss/accuracy plots.
Contractions	Text normalization library	Handles text cleaning and standardization.
Django	Backend web framework	Secure, scalable, Python-native framework with ORM.
BeautifulSoup / Requests	Web scraping of live news articles	Light, flexible scraping for dynamic news updates.
HTML/CSS/JS (Bootstrap)	Frontend for UI rendering	Rapidly build responsive, clean user interfaces.
TPU Runtime (Colab/Kaggle)	Accelerated training of deep learning models	Dramatically reduces training time for large NLP models.

📌 3️⃣ High-Level Design (HLD)

+-----------------+         +----------------+        +-------------------+
|    Preprocess    |  --->   |    Encoder      |  --->  |   Attention Layer  |
|   Clean + Token  |         |   (Bi-LSTM)     |        | (Context Vector)   |
+-----------------+         +----------------+        +-------------------+
                                                            |
                                                            V
                                                   +----------------+
                                                   |    Decoder      |
                                                   |   (LSTM + FC)   |
                                                   +----------------+
                                                            |
                                                            V
                                                   +----------------+
                                                   |    Summary       |
                                                   +----------------+
                                                            |
                                                            V
                                                   +----------------+
                                                   |  Django Backend  |
                                                   | (Model Serving & |
                                                   |   Web Scraping)  |
                                                   +----------------+
                                                            |
                                                            V
                                                   +----------------+
                                                   |   Frontend UI    |
                                                   | (Responsive App) |
                                                   +----------------+

Workflow:

Load and preprocess dataset — 42,000 articles from Kaggle (news_summary.csv)
Clean, normalize and tokenize text (handled contractions, removed punctuations).
Train a Seq2Seq Encoder-Decoder model:
- Encoder: Bi-directional LSTM
- Attention Layer for context alignment
- Decoder: LSTM + Dense for sequence prediction
Achieved:
- Final Training Accuracy: 89.62%
- Final Validation Accuracy: 74.08%
- Final Training Loss: 0.7421
- Final Validation Loss: 2.0726
- AUC Score: 0.79
- F1 Score: 0.73
Model trained on Kaggle TPUs in 180 minutes over 100 epochs.
Deployed via Django backend providing summarization API.
Integrated web scraping module fetching latest news articles.
Served through a Bootstrap-based responsive frontend UI.

📌 4️⃣ Features

✅ Automatic real-time summarization of live news.
✅ Deep learning-based abstractive summarization.
✅ Custom preprocessing: lowercasing, contractions fixing, cleaning.
✅ Django-based REST APIs for ML model access.
✅ Live web scraping for dynamic news input.
✅ Responsive, clean UI to view articles and summaries.
✅ TPU-accelerated model training.
✅ Final Training Accuracy: 89.62%, Validation Accuracy: 74.08%

📌 5️⃣ Results & Metrics

Final Training Accuracy: 89.62%
Final Validation Accuracy: 74.08%
Final Training Loss: 0.7421
Final Validation Loss: 2.0726
AUC Score: 0.79
F1 Score: 0.73
Trained in 180 minutes on Kaggle TPU over 100 epochs.
Visualized through:
- 📈 Accuracy vs Epoch
- 📉 Loss vs Epoch
Real-time summaries accessible via Django APIs and frontend interface.

📌 6️⃣ Local Setup

Prerequisites:

Python 3.8+
Django
TensorFlow 2.x
BeautifulSoup4
Requests

Steps:

Clone repository
Install dependencies via pip install -r requirements.txt
Run Django backend: python manage.py runserver
Launch frontend via index.html
Access real-time summarizer and live news updates

📌 7️⃣ Project Structure

├── major-project-model-final.ipynb    # ML model development & training
├── backend/                           # Django backend serving model & scraping
├── frontend/                          # Web frontend for users
└── README.md                          # Project documentation

📌 8️⃣ Future Improvements

📝 Support multilingual summarization.
📊 Integrate cloud-based model deployment (AWS/GCP).
📲 Build mobile-friendly PWA interface.
📈 Use real-time analytics on summary usage trends.
🗞️ Extend to additional content categories (sports, finance, health).