BAIL-PREDICTION SYSTEM

An AI-powered decision support system that predicts bail outcomes using Hindi legal documents from Indian district courts. Developed as part of the Software Development Project - II at SSN College of Engineering, this system combines advanced NLP techniques with judicial insight to automate and interpret bail decisions.

---

Overview

• Domain: LegalTech, NLP, AI for Social Good
• Language: Hindi (processed via HLDC – Hindi Legal Documents Corpus)
• Outcome: Binary classification – 1: Bail Granted, 0: Bail Denied
• Model: Fine-tuned IndicBERT (Transformers-based)
• Explainability: Confidence scoring and highlighting of influential text segments

---

Problem Statement

The Indian judiciary faces massive delays due to a backlog of cases, with a significant portion involving bail decisions in district courts—often documented in Hindi. This project builds an automated bail prediction system to:

• Process Hindi bail applications
• Predict bail outcomes
• Provide confidence-based decision support

---

System Capabilities

• Real-time processing of Hindi bail applications
• Binary bail outcome prediction (grant or deny)
• Confidence score with interpretable highlights
• User interface for legal practitioners

---

AI Solution Design

Model Pipeline

1. Text Preprocessing: Tokenization, normalization, segmentation
2. Feature Extraction: IndicBERT embeddings trained on Indian legal language
3. Binary Classification: Transformer-based classification layer
4. Post-Processing: Thresholding & interpretability (highlighted facts)

Model Metrics

Metric	Value
Accuracy	79.84%
F1 Score	0.798
Loss	0.433

---

Dataset

• Source: HLDC – Hindi Legal Documents Corpus
• Format: .json or .csv
• Fields: id, district, text (facts + arguments + opinion), label (0/1)

---

Tools & Technologies

Type	Tools Used
Language	Python 3.9
Libraries	Transformers, IndicNLP, Scikit-learn, TensorFlow, Pandas, NumPy
Model	IndicBERT (via HuggingFace Transformers)

---

Installation & Usage

Requirements

• Python 3.9+
• At least 8GB RAM (GPU recommended) [Only for Training the model]
• HLDC formatted data

⚙️ Setup

pip install -r requirements.txt

Running the System

python app.py

Output

• 0 → Bail Denied
• 1 → Bail Granted
• Confidence score with highlighted decision text

---

Note on Model Folder

The model/ folder is not included in this Git repository. It contains the fine-tuned model files required for inference. For educational and research purposes, it can be shared upon request. Please contact ashuwin2210335@ssn.edu.in.

---

Acknowledgements

Special thanks to Dr. S. Karthika (Client & Guide), and the faculty of the Department of IT, SSN College of Engineering.

References

• HLDC: Hindi Legal Documents Corpus – [IEEE NLPKE 2024]
• IL-TUR: Indian Legal Text Understanding and Reasoning – arXiv:2407.05399