# Legal Case Document Summarization: Extractive and Abstractive Methods and their Evaluation

**Authors:** Abhay Shukla, Paheli Bhattacharya, Soham Poddar, Rajdeep Mukherjee, Kripabandhu Ghosh, Pawan Goyal, Saptarshi Ghosh  
**arXiv:** [2210.07544](https://arxiv.org/abs/2210.07544)  
**Date:** 2022-10-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2210.07544v1/

## Key points

- Legal NLP faces challenges in summarizing legal case judgement documents
- Different types of summarization models, such as extractive and abstractive, need analysis in the legal field
- Transformer-based abstractive summarization models have limitations with lengthy legal documents
- Various extractive and abstractive summarization methods were tested across three legal datasets
- General domain-agnostic methods often outperformed domain-specific approaches in legal document summarization tasks

## Summary

, , , , 

In the field of Legal NLP, summarizing legal case judgement documents poses a significant challenge. There is a lack of analysis on how different types of summarization models, such as extractive and abstractive, perform when applied to legal case documents. This issue is particularly crucial due to the limitations of recent transformer-based abstractive summarization models in handling the lengthy nature of legal documents. Additionally, there is a need to determine the most effective way to evaluate legal case document summarization systems. To address these challenges, extensive experiments were conducted using various extractive and abstractive summarization methods, both supervised and unsupervised, across three legal summarization datasets. The study included evaluation by law practitioners and yielded valuable insights into legal summarization practices and long document summarization techniques in general. The research explored a range of summarization methods, including unsupervised extractive approaches like LexRank, DSDR, and PacSum, supervised extractive methods such as SummaRunner and BERT-SUMM, as well as supervised abstractive models like BART and Longformer. Surprisingly, general domain-agnostic methods often outperformed domain-specific approaches in legal document summarization tasks. Furthermore, the study highlighted the benefits of domain-specific training and fine-tuning using pre-trained models like Legal-Pegasus for improved performance. Various strategies for generating legal data for training supervised models were compared to enhance model effectiveness. One key challenge addressed was how to handle long legal documents with existing abstractive summarizers that have limited input capacity. Three approaches were tested: utilizing long document summarizers like Longformer designed for lengthy texts; employing short document summarizers like BART along with chunking techniques; and combining extractive and abstractive methods for efficient summary generation. The chunking-based approach showed promising results for legal documents with fine-tuning proving beneficial. Additionally, the evaluation methodology emphasized not only assessing full-document summaries but also evaluating how well summaries represented different logical segments within a legal case document (e.g., Facts, Final Judgment). Document-wide automatic evaluations alongside segment-wise assessments were conducted alongside evaluations by law practitioners to ensure comprehensive analysis of summary quality. Overall, this comprehensive study sheds light on effective strategies for legal document summarization while providing valuable insights applicable to long document summarization tasks in diverse domains.