Scholarlink

<p>This legal document summarization model aims to produce concise and accurate summaries of long, complex legal documents. Traditional methods for summarization face several issues, such as producing repetitive sentences, failing to cover all necessary information, creating incoherent summaries, and missing critical details. To address these issues, this study proposes SENDE, a fast extractive summarization model for legal documents. SENDE adopts a fully parallelizable, GPU-accelerated convolutional architecture that satisfies high-performance computing constraints. During reconstruction, the model performs sentence-level denoising, sharpening its grasp of semantic nuance and yielding more accurate legal summaries. Its Oracle Sentence Extraction algorithm further accelerates processing by selecting candidate sentences in parallel and returning only their indices, curbing memory overhead while supplying high-quality pseudo-labels that boost training on unlabeled corpora. Experimental results on the BillSum dataset show that SENDE outperforms existing models, including BB25HLS-FoSum and MatchSUM, achieving higher ROUGE scores and providing quick, accurate summarization of legal documents. Specifically, in the BillSum test sets, our model shows an improvement of 12.04 for baseline models and 8.45 for state-of-the-art models in the ROUGE metric score while delivering a 25.57% speedup.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SENDE: extractive summarization of legal documents by sentence noising-reconstruction and dilated-gated convolutional networks

  • 期刊论文

  • 发表日期：

    2026年2月2日

查看全文
  • Tiejun Xi,
  • Rui Huang,
  • Zongtao Duan,
  • Junzhe Zhang

摘要

This legal document summarization model aims to produce concise and accurate summaries of long, complex legal documents. Traditional methods for summarization face several issues, such as producing repetitive sentences, failing to cover all necessary information, creating incoherent summaries, and missing critical details. To address these issues, this study proposes SENDE, a fast extractive summarization model for legal documents. SENDE adopts a fully parallelizable, GPU-accelerated convolutional architecture that satisfies high-performance computing constraints. During reconstruction, the model performs sentence-level denoising, sharpening its grasp of semantic nuance and yielding more accurate legal summaries. Its Oracle Sentence Extraction algorithm further accelerates processing by selecting candidate sentences in parallel and returning only their indices, curbing memory overhead while supplying high-quality pseudo-labels that boost training on unlabeled corpora. Experimental results on the BillSum dataset show that SENDE outperforms existing models, including BB25HLS-FoSum and MatchSUM, achieving higher ROUGE scores and providing quick, accurate summarization of legal documents. Specifically, in the BillSum test sets, our model shows an improvement of 12.04 for baseline models and 8.45 for state-of-the-art models in the ROUGE metric score while delivering a 25.57% speedup.