LegalSumAI: Robust and Representative Legal Summarization with Large Language Models

Purvi Sehgal, Sahithi Ankireddy, Saumya Chauhan

Keywords: Large Language Models

TLDR: LegalSumAI introduces a two-stage Large Language Model (LLM) framework that generates accurate, interpretable legal case summaries

Abstract: Legal disputes are complex, high-stakes challenges, yet 56% of US households encounter legal issues annually. The jar- gon, length, and complexity of legal documents often make them inaccessible to the general public. This work introduces LegalSumAI, a novel framework that generates robust legal summaries from case documents using Large Language Mod- els (LLMs). We investigate whether LLMs can produce ac- curate, comprehensible summaries of legal cases and opin- ions without hallucinations. Our two-step pipeline first gener- ates structured CSV fact sheets, capturing case details via the IRAC framework (Issue, Rule, Application, Conclusion) and metadata on the involved parties. These fact sheets are then converted into natural language summaries prompted with Chain of Density (CoD) techniques. We leverage the Multi- LexSum dataset, which provides expert-authored summaries at three different granularity levels (tiny, short, long) to evalu- ate the generated summaries using ROUGE and BERTScore metrics. Results show high semantic accuracy, with aver- age BERTScore improvements of 157% over Multi-LexSum baselines, demonstrating that our structured reasoning and CoD prompting mitigate LLM hallucinations and improve le- gal summarization results. This research demonstrates the po- tential for interpretable LLM pipelines to democratize access to legal knowledge.

Submission Number: 60