Legal Multi-Document Summarizer

A. Dataset — URL / Source

1. Multi-LexSum Dataset (Primary)

URL: https://multilexsum.github.io/

Multi-LexSum is a dataset of 9,280 legal case summaries designed to support research on summarization of civil rights litigation. It provides summaries at multiple levels of granularity:

• One-sentence "extreme" summaries
• Paragraph-length summaries
• Long multi-paragraph summaries (500+ words)

Each case includes multiple source documents: complaints, court opinions, orders, settlement agreements, etc. Summaries are expert-authored following detailed guidelines, making this a reliable benchmark for long document summarization, multi-document summarization, and controllable summarization.

2. Supreme Court of India Dataset (Secondary / In Progress)

URL: https://www.sci.gov.in/case-status-case-no/

To validate the model on Indian legal documents, documents are collected from the Supreme Court of India website. Each case consists of multiple documents. Gold summaries for these cases are being produced in collaboration with law domain professionals.

B. Software and Hardware Requirements

Hardware Requirements

Training Phase (resource-intensive):

Component	Specification
CPU	Intel Core i7 (8th Gen or higher) or equivalent AMD
GPU	NVIDIA Tesla T4 or equivalent (minimum 15 GB VRAM)
RAM	12 GB minimum
Storage	100 GB disk space

The GPU is crucial for acceleration of graph construction, LLM pre-training, and summarization. RAM is required for handling large in-memory graph matrices and document embeddings.

Execution / Inference Phase: A modern CPU is sufficient for generating summaries after training.

Software Requirements

• Python: 3.9 or higher
• OS: Windows / macOS / Linux

Category	Library / Framework
NLP Framework	PyTorch 2.1.2, TensorFlow
NLP Libraries	Hugging Face Transformers 4.40.0
NLP Libraries	sentence-transformers 2.7.0
NLP Libraries	spaCy 3.x, NLTK
Graph Processing	NetworkX, torch-geometric
Data Handling	NumPy 1.26.4, Pandas
PDF Processing	PyMuPDF (fitz)
Web Framework	FastAPI, Uvicorn, Streamlit
Utilities	tqdm, scikit-learn, dateparser, evaluate, rouge-score, bert-score, python-multipart, requests

All specific library versions are pinned in requirements.txt to ensure reproducibility. Do not upgrade versions without testing compatibility.

---

C. Instructions to Execute the Source Code

Prerequisites

• Python 3.9+ installed on your system
• The source code folder should contain:

web-app/
├── app.py
├── backend.py
├── legal_summarizer.py
└── requirements.txt

Step-by-Step Execution

Step 1 — Navigate to the web-app directory

cd path/to/web-app

Step 2 — Create a Python virtual environment

python -m venv venv

Step 3 — Activate the virtual environment

On Windows:

venv\Scripts\activate

On macOS / Linux:

source venv/bin/activate

You should see (venv) appear in your terminal prompt.

Step 4 — Install all dependencies

pip install -r requirements.txt

This may take several minutes as it downloads PyTorch, Transformers, and other large libraries.

Step 5 — Download the spaCy English language model

python -m spacy download en_core_web_sm

Step 6 — Start the backend server

python3 backend.py

Wait until you see the following message before proceeding:

INFO: Uvicorn running on http://127.0.0.1:8000

The first run will download AI model weights (BART, SentenceTransformer). This may take a few minutes. Keep this terminal open.

Step 7 — Start the frontend (in a NEW terminal window)

Open a second terminal, navigate to the web-app folder, activate the venv (Steps 1 & 3), then run:

streamlit run app.py

The app will open in your browser at http://localhost:8501

Both terminals must remain open while using the application.

---

How to Use the Application

1. Upload one or more legal PDF documents using the sidebar
2. Select the desired summary length: short or long
3. Click Generate Summary and wait for the AI pipeline to complete
4. The generated summary will appear in the main panel

---

Troubleshooting

Error	Fix
ModuleNotFoundError: No module named 'spacy'	pip install spacy
Can't find model 'en_core_web_sm'	python -m spacy download en_core_web_sm
Form data requires python-multipart	pip install python-multipart
Failed to connect to backend in Streamlit UI	Ensure backend.py is running and shows the Uvicorn message before using the frontend
Bus error / crash on macOS during classification	Set the environment variables in the macOS section above and ensure torch==2.1.2 is installed