Legal Document Summarization Flowchart

Extractive Models Execution

Four extractive models run in parallel:

TextRank

Graph-based ranking algorithm for text

TF-IDF

Term frequency-inverse document frequency

TF-IDF with BERT

Enhanced with contextual embeddings

Luhn's Algorithm

Frequency-based sentence scoring

Abstractive Models Execution

Six models run in total (2 standalone + 4 hybrid):

Standalone Fine-tuned LLMs

Legal LED - BillSum

Fine-tuned for US bills and legal documents

Legal LED - ILC

Fine-tuned for Indian Legal Cases

Hybrid Models

TF-IDF + Legal LED - BillSum

Extractive → Abstractive pipeline

TF-IDF with BERT + Legal LED - ILC

Enhanced extractive → Abstractive

TF-IDF (fine-tuned) + Legal LED - BillSum

Optimized extractive → Abstractive

Luhn's Algorithm (fine-tuned) + Legal LED - ILC

Frequency-based → Abstractive

Hybrid Model Process

Generate condensed version using extractive logic
Feed extractive output to fine-tuned LLM
LLM produces polished abstractive summary

Summary Output

All model outputs displayed for comparison:

Gradio Web Interface

Clean, scrollable display format
Summaries clearly labeled by model
Side-by-side comparison enabled
Download options for each summary

All processing uses GPU acceleration (T4 on Kaggle)