Apppublic

anshul32/legal-summarizer

sourceHugging Facemitupdated 4mo agoView on Hugging Face

0likes

App README

⚖️ Legal Contract Summarizer

Hybrid Computer Vision + NLP Legal Contract Analysis System

Fine-tuned T5-small on LEDGAR · ROUGE-1: 0.93 · EasyOCR · OpenCV

🚀 Features

Feature	Description
🤖 Structured AI Summary	6-section legal report: Overview, Financial, Clauses, Risks, AI narrative, Assessment
📑 Agreement Classification	Auto-detects NDA, Employment, Service, Lease, and 8+ agreement types
🎯 Risk Dashboard	Rule-based 0–100 score with categorized severity flags
📋 18+ Clause Types	Termination, NDA, IP, Arbitration, Indemnity, Force Majeure, and more
🔬 CV Pipeline	8-panel visualization — Units 1, 2, 4, 5, 6 (perspective + bilateral filter)
👁️ OCR + Heatmap	EasyOCR with colour-coded confidence bounding boxes
✍️ Signature Detection	Contour solidity analysis
🔵 Stamp Detection	Hough circles + HSV color ink detection
📐 Skew Correction	Hough line angle estimation + affine warp
⬇️ JSON Export	Full structured export of analysis results

🧠 NLP Pipeline

Contract Text / OCR Output
    ↓
Text Cleaning + Normalization
    ↓
Chunked Summarization (T5-small, max 400 words/chunk)
    ↓
Regex Field Extraction (parties, dates, payment, governing law...)
    ↓
Clause Detection (18 clause types via regex patterns)
    ↓
Risk Scoring (missing clauses + red flag language)
    ↓
HTML Report + JSON Export

🔬 Computer Vision Pipeline

Step	OpenCV Operation	Syllabus Unit
RGB → Grayscale	`cvtColor(COLOR_RGB2GRAY)`	Unit 5
Perspective Correction	Contour quad + `warpPerspective`	Unit 2
Skew Correction	Hough lines + `warpAffine`	Unit 2
Histogram Equalization	`equalizeHist()`	Unit 1
Noise Removal	`bilateralFilter` + `GaussianBlur`	Unit 1
Binarization	`threshold(THRESH_OTSU)`	Unit 6
Edge Detection	`Canny(50, 150)`	Unit 6
Contour Segmentation	`findContours()`	Unit 6
Harris Corners	`cornerHarris()`	Unit 4
HOG Features	`HOGDescriptor`	Unit 4
Signature Detection	Contour solidity heuristics	Unit 6
Stamp Detection	Hough circles + HSV masking	Unit 5, 6
OCR	EasyOCR + bbox confidence	Multimodal

📥 Input Formats

—Tab 1: Paste raw contract text directly
—Tab 2: Upload .txt or digital .pdf
—Tab 3: Upload scanned contract image (.jpg, .png) — runs full CV → OCR → NLP

🏗️ Architecture

                    ┌─────────────────────────────────┐
                    │   Gradio Web Interface (3 tabs)  │
                    └───────────┬─────────────────────┘
                                │
              ┌─────────────────┼─────────────────┐
              ▼                 ▼                  ▼
       [Text Input]      [File Upload]     [Image Upload]
              │                 │                  │
              │                 │           CV Pipeline
              │                 │          (8 steps, OpenCV)
              │                 │                  │
              └─────────────────┴──────────────────┘
                                │
                         EasyOCR (if image)
                                │
                         NLP Pipeline
                         (T5 + Regex)
                                │
                    ┌───────────┴───────────┐
                    ▼                       ▼
              HTML Report              JSON Export

🛠️ Tech Stack

—Model: anshul32/t5-legal-summarizer (T5-small fine-tuned on LEDGAR)
—OCR: EasyOCR
—CV: OpenCV (opencv-python-headless)
—UI: Gradio 4+
—PDF: PyMuPDF (fitz)
—Visualization: Matplotlib

⚠️ Disclaimer

This tool is for educational and informational purposes only. It is not a substitute for professional legal advice. Always consult a qualified attorney for legal matters.

👤 Author

Anshul Kumar · HuggingFace · Kaggle

Computer Vision + NLP Academic Project