Legal NLP Guide

A skill for applying natural language processing techniques to legal texts. Covers legal document classification, named entity recognition for legal entities, contract clause extraction, case law similarity search, and court opinion summarization using modern NLP tools.

Legal Text Characteristics

Legal language presents unique NLP challenges:

• Long documents: Court opinions average 5,000-20,000 tokens; contracts can exceed 50,000
• Domain-specific vocabulary: Terms of art with precise legal meanings (e.g., "consideration", "estoppel")
• Complex syntax: Multi-clause sentences with nested qualifications and cross-references
• Citation networks: Dense cross-referencing between cases, statutes, and regulations
• Temporal reasoning: Effective dates, amendments, and retroactivity

Legal Text Classification

Document Type Classification

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Legal-BERT: domain-adapted BERT for legal text
model_name = "nlpaueb/legal-bert-base-uncased"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(
    model_name, num_labels=5
)

# Legal document categories
labels = ["contract", "court_opinion", "statute", "regulation", "brief"]

def classify_legal_document(text: str, max_length: int = 512) -> dict:
    """
    Classify a legal document into predefined categories.
    For long documents, use the first 512 tokens (typically the
    preamble/introduction which contains strong classification signals).
    """
    inputs = tokenizer(
        text, return_tensors="pt",
        max_length=max_length, truncation=True, padding=True
    )
    with torch.no_grad():
        logits = model(**inputs).logits
    probs = torch.softmax(logits, dim=-1).squeeze()
    predicted = labels[probs.argmax().item()]
    return {
        "predicted_class": predicted,
        "confidence": probs.max().item(),
        "all_scores": {l: p.item() for l, p in zip(labels, probs)},
    }