:Legal Document Summarization and Classification Using Transformer Models

Abstract: Legal documents are often verbose and complex, posing challenges for legal practitioners, clients, and automated systems. This paper investigates the application of transformer-based models—particularly BERT, LegalBERT, and Longformer—for summarization and classification of legal documents. Using legal corpora from Indian Supreme Court judgments and U.S. case law, we evaluate models on metrics including ROUGE, F1-score, accuracy, and interpretability with SHAP and LIME. Regression and predictive analysis are used to correlate input document length, term density, and performance metrics. The results indicate that domain-specific pretraining (e.g., LegalBERT) significantly boosts performance, with improvements of up to 14% in classification accuracy and 18% in ROUGE scores for summarization. Keywords: Legal Document Processing, Transformers, LegalBERT, BERT, Longformer, Summarization, Classification, SHAP, LIME