HiLegalSum: Hybrid Extractive Summarization of Long Legal Texts
Year: 2026
Type: conference-paper
Abstract: Legal documents are complicated by their nature, lengthy, and full of jargon that is specific to the field, thus making their manual review a time-consuming and error-prone process. Most of the traditional text summarization methods in Natural Language Processing (NLP) fail to consider such documents properly as they impose input length restrictions and have issues with retaining legally significant and contextually accurate clauses. This work, therefore, presents HiLegalSum, which is a hybrid legal-aware extractive summarization framework that leverages semantic graph centrality, positional weighting, Term Frequency - Inverse Document Frequency (TF-IDF) relevance, and domain-specific legal keyword analysis to generate not only fluent but also contextually appropriate summaries from the long legislative texts. For the purpose of evaluation, the model employs the BillSum dataset and uses Sentence Bidirectional Encoder Representations from Transformers (SBERT) (all-mpnet-base-v2) embeddings for capturing the semantics of the sentence. In comparison experiments with standard extractive baseline methods of the class such as Lead-3, Random, LexRank, and TextRank, HiLegalSum is always superior to these methods as it obtains 0.3876 Recall Oriented Understudy for Gisting Evaluation (ROUGE)-1, 0.1765 ROUGE-2, 0.2243 ROUGE-L, and 0.8329 BERTScore (F1) over$\mathbf{5 0 0}$test documents. The results demonstrated the benefit of blending semantic similarity, positional heuristics, TF-IDF weighting, and the legal keyword sensitivity to both raise the quality and the trustworthiness of the summarization of long legal documents. The proposed approach contributes to the development of more efficient, interpretable, and domain-adapted summarization systems that are targeted at the legal sector.
Institution: VIT-AP University
Language: English
Subfield: Law
Field: Social Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7167474367