Abstract

The traditional approach to summarizing dense, complex, and jargon-filled legal documents is done manually and is very time-consuming, labor-intensive, prone to human biases, and applicability-driven. This study proposes the use of the T5 (Text-to-Text Transfer Transformer) for automatic summarization of legal documents with 100% text extraction, semantic preservation, and high readability. The Indian Legal Database Corpus (ILDC) is put forward as the primary dataset, which has case judgments written with expert-written summaries that can serve as an ideal benchmark for training and evaluating legal document summarization models. The proposed system uses NLP and is fine-tuned on this dataset which allows it to produce brief, legally accurate, and understandable summaries. All processes include comprehensive preprocessing, smart document chunking, and employing optimized techniques to enhance summary generation which solves the issue of large legal text processing. To ensure the model’s robustness and generalizability, further validation is conducted on legal texts beyond ILDC. Testing the model on different legal datasets, such as international case law, Supreme Court cases, or regulatory documents from various jurisdictions, allows for evaluating its adaptability across different legal traditions. Additionally, cross-domain testing on legal contracts, corporate policies, and government regulations is performed to assess its performance beyond case law. Exploring multilingual legal documents is also considered for expanding the model’s applicability in diverse legal settings. The evaluation process entrusts different scoring systems, like ROUGE for textual overlap, BLEU for fluency and coherence, and semantic similarity score evaluated with Sentence- BERT to ensure meaning retention. A model of T5 fine-tuning on ILDC is professed as effective in this study to produce an efficient and better-summarizing system in the area of legal texts.