AugAbEx: Bridging Abstractive and Extractive Legal Summarization

2026-07-23 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, extended

Summary

AugAbEx introduces a novel, transparent pipeline designed to bridge the gap between abstractive and extractive legal summarization by generating high-quality extractive gold standard summaries from existing abstractive versions. Addressing the cognitive burden and potential inaccuracies of abstractive methods in legal contexts, this approach circumvents the high cost of manual annotation for extractive summaries. The pipeline transforms original abstractive gold (OAG) summaries into transformed extractive gold (TEG) summaries across seven diverse legal datasets from India, the U.S., U.K., and Australia. It employs a two-stage process: selecting top-k candidate sentences using ROUGE metrics, then applying Maximum Marginal Relevance (MMR) with λ=0.5 to ensure relevance and diversity. Extensive evaluation confirms TEG summaries maintain strong structural, lexical, and semantic fidelity to OAGs, preserving expert opinions and outperforming LSA-generated extractive summaries in legal information richness. The augmented datasets will be publicly released.

Key takeaway

For NLP Engineers and AI Scientists developing legal summarization systems, this research offers a validated method to generate high-quality extractive gold standard summaries from existing abstractive datasets. You should consider leveraging the AugAbEx pipeline to augment your own datasets, reducing costly manual annotation. Prioritize algorithms that explicitly preserve legal entities and provisions, as these are critical for practitioner utility and summary fidelity, as demonstrated by the evaluation framework.

Key insights

A pipeline converts abstractive legal summaries into expert-aligned extractive gold standards, overcoming manual annotation costs.

Principles

• Extractive summaries are preferred for legal precision.
• Preserving expert-identified saliency is key for extractive quality.
• Legal entities are critical for practitioner utility.

Method

The pipeline selects top-k ROUGE-similar sentences from judgments, then applies MMR with λ=0.5 to balance relevance and diversity for extractive summary creation.

In practice

• Employ ROUGE metrics for candidate sentence selection.
• Apply MMR (λ=0.5) to balance summary relevance and diversity.
• Assess legal summaries across domain, semantic, lexical, and structural attributes.

Topics

• Case Summarization
• Extractive Summarization
• Legal NLP
• Data Augmentation
• LegalBert
• Maximum Marginal Relevance

Code references

• Law-AI/summarization
• Legal-NLP-EkStep/legal_NER
• ra-MANUJ-an/CivilSum

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• UlyssesLegalNER-Br: from Legislative to Legal, a comprehensive corpus of Brazilian legal documents for Named Entity Recognition — UlyssesLegalNER-Br offers a comprehensive, publicly available corpus for Brazilian legal Named Entity Recognition.
• Abstractiveness Metrics for Evaluating Text Summarization: A Refined Formulation with Empirical Validation — New metrics (RA, SA, AR) quantify summary abstractiveness, distinguishing extractive from abstractive models and flagging potential hallucination.
• I Built an AI Pipeline for Kindle Highlights — Automate book summarization from Kindle highlights using Python and a local LLM for efficient knowledge retention.
• From Articles to Premises: Building PrimeFacts, an Extraction Methodology and Resource for Fact-Checking Evidence — Decontextualizing hyperlink-anchored evidence from fact-checks significantly improves automated retrieval and verification performance.
• Discovery of Legal Patterns in Civil Petitions via LLM-Based Fact Extraction and Density Clustering — LLM-based fact extraction significantly improves legal document clustering by reducing noise and enhancing semantic coherence.
• Extracting Problem and Method Sentence from Scientific Papers: A Context-enhanced Transformer Using Formulaic Expression Desensitization — Formulaic expression desensitization and context-enhanced transformers improve problem/method sentence extraction from scientific papers, outperforming baselines.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.