News Summarization AI

Overview

A production-grade news summarization system that leverages state-of-the-art transformer architectures to generate concise, accurate summaries of Indian news articles. The system combines PRIMERA (Pyramid-based Masked Sentence Pre-training for Multi-document Summarization) and LED (Longformer Encoder-Decoder) models to handle long-form news content effectively.

The project demonstrates end-to-end deep learning pipeline design — from data preprocessing and model fine-tuning to containerized deployment with a RESTful API interface.

Tech Stack

Architecture

The system follows a modular architecture with clear separation between data processing, model inference, and API serving layers.

Model Architecture

The project evaluates two long-document transformer architectures optimized for summarization tasks:

• PRIMERA — Built on the Longformer architecture with pyramid-based pre-training, enabling efficient processing of documents up to 4,096 tokens. Uses global attention on sentence-level tokens for multi-document awareness.
• LED (Longformer Encoder-Decoder) — Extends the Longformer's local+global attention mechanism to an encoder-decoder framework. Handles long input sequences with linear complexity through windowed local attention combined with task-motivated global attention.

Evaluation Results

The models were evaluated using industry-standard summarization metrics on Indian news datasets:

• ROUGE-1: 71.43% — Measures unigram overlap between generated and reference summaries. This score indicates strong lexical alignment with human-written summaries.
• BERTScore: 0.93 — Uses contextual BERT embeddings to measure semantic similarity. A score of 0.93 demonstrates that generated summaries capture the meaning of the source text with high fidelity.
• ROUGE-2 & ROUGE-L — Additional bigram and longest common subsequence metrics were used to validate consistency across different evaluation dimensions.

Deployment

The system is fully containerized using Docker for consistent deployment across environments:

• FastAPI — Async REST API with automatic OpenAPI documentation. Endpoints accept raw news text and return structured JSON summaries.
• Docker — Multi-stage Dockerfile with optimized image size. Model weights are loaded at container startup for fast inference.
• Batch Processing — Supports both single-article and batch summarization through the API interface.

# Run with Docker
docker build -t news-summarizer .
docker run -p 8000:8000 news-summarizer

# API Usage
curl -X POST http://localhost:8000/summarize \
  -H "Content-Type: application/json" \
  -d '{"text": "Your news article text here..."}'

Key Features

• Handles long-form news articles (up to 4,096 tokens) without truncation
• Dual-model evaluation pipeline for comparing PRIMERA vs LED performance
• Comprehensive evaluation using ROUGE-1/2/L and BERTScore metrics
• Production-ready Docker deployment with FastAPI REST endpoints
• Focused on Indian English news for domain-specific summarization quality