Summarizing legal case judgments presents a significant challenge within the realm of Legal Natural Language Processing (NLP). There is a notable gap in understanding the effectiveness of different summarization models, such as extractive and abstractive techniques, particularly in the context of legal documents. With approximately 40 million pending cases in the Indian judicial system, this study tackles the arduous task of manually summarizing legal texts. It introduces both supervised and unsupervised models for extractive and abstractive summarization, demonstrating their efficacy through evaluations based on ROUGE metrics and BERT scores. Models such as BART, T5, PEGASUS, Legal-PEGASUS, and Legal-BERT are employed for abstractive summarization, while TextRank, LexRank, LSA, Summarizer BERT, and KL-Summ are utilized for extractive summarization. Additionally, Longformer and Bert-Legal Pegasus are considered for summarization tasks. The study leverages hybrid abstractive-extractive techniques to generate summaries. This is the accepted manuscript version accepted for publication in AIP Conference Proceedings on December 16, 2025. Final publisher version pending publication.
Lyngdoh et al. (Tue,) studied this question.