This project is a Natural Language Processing (NLP) pipeline designed to classify legal case reports into their respective areas of law (e.g., Civil Procedure, Criminal Law, Company Law, etc.).

It supports inference via a FastAPI endpoint where users can submit the full body of a legal judgment (full_report) and receive the predicted legal category.

• Clean and preprocess dataset
• Build a pipeline that maps legal case reports to legal categories.
• Compared traditional ML methods (TF-IDF + Logistic Regression) vs transformer-based deep learning models (DeBERTa).
• Serve the final transformer model as an API.

• full_report: The body of the legal judgment.
• introduction: Used to extract the area of law (labels).
• Other fields: case_title, suitno, facts, issues, decision.

Labels include:

• Civil Procedure
• Criminal Law and Procedure
• Enforcement of Fundamental Rights
• Company Law
• Election Petition, etc.

• Text Vectorization: TF-IDF (TfidfVectorizer)
• Model: Logistic Regression
• Label Extraction: Regex pattern matching from introduction
• Evaluation:
  • Accuracy: ~35%
  • F1 Score: Low for imbalanced classes
• Limitation: Unable to understand contextual meaning, struggled with nuanced legal phrasing.

• Model Used: distilbert-base-uncased fine-tuned on full_report
• Tokenizer: Hugging Face tokenizer
• Label Encoder: LabelEncoder() for mapping categories
• Training: Done on CPU (adapted for low-resource environment)
• Evaluation:
  • Higher F1 and accuracy
  • Better understanding of complex language
  • Handled rare labels more gracefully

LawPavilion/
├── app/
│   ├── api.py             # FastAPI routes and logic
│   └── model_utilis.py    # Load model/tokenizer/label_encoder
├── models/
│   ├── config.json, tokenizer, model weights, saved_model.pkl
├── main.py                # Entrypoint for FastAPI app
├── requirements.txt
└── Legal_Classifier.ipynb # Jupyter notebook with traditional + BERT training

git clone https://github.com/shashacode/NLP_Code_Challenge.git
cd LawPavillion

pip install -r requirements.txt

uvicorn main:app --reload

Visit: http://127.0.0.1:8000/docs to access the Swagger UI.

Endpoint: POST /predict

Request Body:

{
  "full_report": "The appellant was tried for armed robbery and sentenced to life imprisonment. The appeal concerns improper identification and denial of fair hearing."
}

Response:

{
  "predicted_area_of_law": "Criminal Law and Procedure"
}

This solution was developed as part of the LawPavilion Legal NLP Challenge, aimed at advancing AI solutions in the legal domain in Nigeria.