🏩 JudgerAI — Your Dream Legal AI Assistant

Table of Contents:

• 🚀 Overview
• 📺 Demo
• ⚙️ Installation
• 📊 Dataset
• 🧹 Project Modules
• 🤖 AI Models
• 🧪 Experiments & Evaluation
• 🏅 Training Approach
• 📊 Ensemble Voting
• 📚 Notebooks & Further Reading

🚀 Overview

JudgerAI is an innovative NLP application that predicts legal case outcomes with impressive accuracy by analyzing past cases, precedents, and case facts. It empowers legal professionals to:

• 📈 Increase prediction accuracy
• ⏱️ Save valuable time on case research
• 🧠 Make informed, data-driven decisions

---

📺 Demo

Watch JudgerAI in action:

JudgerAI.2.0.-.Project.Demo.-.Trim.mp4

---

⚙️ Installation

1️⃣ Clone the repository

git clone https://github.com/MohammedAly22/JudgerAI.git

2️⃣ Download GloVe embeddings (50‌-dim) from Kaggle and save as:

./GloVe/glove.6B.50d.txt

3️⃣ Download pre-trained models (heavy files) and place them in models/. Download Models from Here

4️⃣ Directory structure:

JudgerAI/
├── csvs/
├── dataset/
├── GloVe/
├── models/
├── src/
└── *.ipynb

5️⃣ Run the app:

streamlit run src/main.py

---

📊 Dataset

• Total cases: 3,464
• Key columns: ID, name, href, first/second_party, winning_party, winner_index (0/1), facts
• Input: facts → Output: winner_index

Here is the dataset summary:

column	datatype	description
ID	int64	Defines the case ID
name	string	Defines the case name
href	string	Defines the case hyper-reference
first_party	string	Defines the name of the first party (petitioner) of a case
second_party	string	Defines the name of the second party (respondent) of a case
winning_party	string	Defines the winning party name of a case
winner_index	int64	Defines the winning index of a case, 0 => the first party wins, 1 => the second party wins
facts	string	Contains the case facts that are needed to determine who is the winner of a specific case

---

🧹 Project Modules

Modular structure for maintainability and clarity:

Module	Location	Responsibilities
Preprocessing	src/preprocessing.py	Tokenization, balancing, anonymization, vectorization
Plotting	src/plotting.py	Visualizing performance, confusion matrices, ROC-AUC, heatmaps
Utils	src/utils.py	Training helpers, k-fold CV, accuracy/loss summary builders
Streamlit App	src/main.py	Frontend UI for demo and deployment
Deployment Utils	src/deployment_utils.py	Model loader, sample picker, vectorizer generator, highlights words

---

🤖 AI Models

JudgerAI incorporates 7 different models:

• Doc2Vec – Documents as dense vectors
• 1D-CNN – Convolutional features over text
• TF-IDF + TextVectorization – Weighted bag-of-words embedding
• GloVe – Global co-occurrence embeddings
• FastText – Subword-enhanced embeddings
• LSTM – Memory‌-capable sequences
• BERT – Contextual pre-trained transformer

A mix of traditional and modern architectures to maximize coverage.

---

🧪 Experiments & Evaluation

Three core preprocessing decisions were evaluated:

1. Preprocessing steps – stopword removal, stemming, etc.
2. Data anonymization – replacing party names with _PARTY_
3. Label imbalance – strategies for balanced classes

This results in 2³ = 8 experiments, each run with 4-fold cross-validation, giving thorough analysis across 32 total runs per model.

Combination #	Preprocessing	Data Anonymization	Label Class Imbalance
1	No	No	No
2	No	No	Yes
3	No	Yes	No
4	No	Yes	Yes
5	Yes	No	No
6	Yes	No	Yes
7	Yes	Yes	No
8	Yes	Yes	Yes

---

🏅 Training Approach

• 80/20 train/test split
• 4-fold CV on training set
• Best combination selected per model based on accuracy

Workflow:

• Train each model × 8 preprocessing setups × 4 CV folds
• Evaluate and select best-performing model

---

📊 Ensemble Voting

Final predictions are generated through an ensemble voting method across all tuned models to ensure robustness.

---

📚 Notebooks & Further Reading

In-depth exploration available in the following notebooks:

• BERT_experiments.ipynb
• cnn_experiments.ipynb
• doc2vec_experiments.ipynb
• FastText_experiments.ipynb
• glove_experiments.ipynb
• LSTM_experiments.ipynb
• tf_idf_experiments.ipynb
• voting_experiments.ipynb

---

🤝 Contributing & License

Contributions welcome! Feel free to contribute to JudgerAI.

---

🙏 Thank You!

Thank you for exploring JudgerAI—ushering in a smarter future for legal decision-making with AI-powered precision.

---