Legal Text Classification Project

📌 Introduction

A hybrid legal text classification system combining:

• Rule-based keyword matching
• Unsupervised machine learning (TF-IDF + K-means)
• Dimensionality reduction (UMAP)

Motivation: As a pre-law student, I developed this to solve the challenge of efficiently finding relevant legal cases for research across different domains (family law, criminal law, etc.).

📂 Dataset

Source: Kaggle Legal Text Dataset
Contents: 24,985 Australian legal cases with:

• Case ID (unique identifier)
• Outcome (judicial decision)
• Title (case name)
• Text (full narrative)

🛠️ Methodology

1. Preprocessing Pipeline

# Sample preprocessing steps
text = text.lower()
text = re.sub(r'[^\w\s]', '', text)  # Remove punctuation
tokens = word_tokenize(text)          # Tokenization
lemmatized = [lemmatizer.lemmatize(t) for t in tokens]
clean_text = [t for t in lemmatized if t not in stopwords]

2. Hybrid Classification

legal_categories = {
    'family': ['custody', 'divorce', 'marriage', ...],
    'criminal': ['theft', 'murder', 'fraud', ...],
    # 8 other domains...
}

Unsupervised Learning:

• TF-IDF Vectorization (1-3 ngrams)
• UMAP Dimensionality Reduction (n_components=20)
• K--means Clustering (k=19 via silhouette score)

3. Performance Metrics

Stage	Silhouette Score	Calinski-Harabasz
Initial	0.036	159.75
After UMAP	0.36	7607.45

📊 Results

Cluster Analysis

Top Business Law Terms:

['case', 'court', 'agreement', 'party', 'contract', 'ltd']
['fca', 'decision', 'applicant', 'immigration', 'tribunal']

🔍 Key Findings

• Successfully categorized 60% of cases via initial keyword matching
• Cluster quality improved 10x after UMAP reduction
• Identified limitations in business/financial law separation

📝 Discussion

Advantages

✅ Efficient - Quick categorization of obvious cases
✅ Transparent - Clear reasoning for classifications
✅ Adaptable - Expandable keyword dictionary

Limitations

⚠️ Keyword dependence - May miss niche terminology
⚠️ Cluster overlap - Some domain boundaries unclear
⚠️ Resource constraints - Limited BERT implementation

🌟 Future Work

• Web interface for legal researchers
• Dynamic keyword expansion
• BERT fine-tuning with better hardware
• Hierarchical classification system

👩‍💻 Contributor

Yoshita Aligina
Rutgers University CS 2026

📚 References

1. Nghiem et al. (2022) - Transformer-based legal classification (https://aclanthology.org/2022.lrec-1.504.pdf)
2. UMAP Documentation
3. HuggingFace Transformers
4. https://www.geeksforgeeks.org/removing-stop-words-nltk-python/
5. https://stanfordnlp.github.io/CoreNLP/lemma.html