Scholarlink

<p>This work explores provision-level retrieval and neural reranking for UK primary and secondary legislation. We introduce <a href="http://doi.org/10.15128/r14x51hj064"><span>UK-StatuteCorpus</span></a>, a corpus of recent UK Acts and statutory instruments from <a href="https://www.legislation.gov.uk/"><i>legislation.gov.uk</i></a>, together with a 100-query evaluation set of practitioner-style questions whose graded relevance judgements distinguish legally operative, supporting and contextual provisions. Using BM25 and an MPNet-based dense retriever to build candidate sets, we evaluate ten neural rerankers, including transformer cross-encoders, a late-interaction reranker, an LLM-based listwise reranker and proprietary APIs. Across both sparse and dense pools, neural reranking consistently improves normalized Discounted Cumulative Gain (nDCG) and Mean Reciprocal Rank (MRR) over first-stage retrieval. We further distil a proprietary Voyage reranker into a ModernBERT-based cross-encoder, Distilled-Voyage-ModernBERT, which approaches the teacher’s effectiveness and outperforms other open rerankers on our benchmark. Results are based on 100 expert-validated queries, each linked to three graded provisions from a single UK instrument, so they characterise single-instrument, provision-level retrieval over recent UK legislation.</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

Neural reranking for UK statutory retrieval: Provision-level evaluation and an open distilled model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月5日

查看全文
  • Amal Saad Alshehri,
  • Can Eken,
  • Nelly Bencomo,
  • Amir Atapour-Abarghouei

摘要

This work explores provision-level retrieval and neural reranking for UK primary and secondary legislation. We introduce UK-StatuteCorpus, a corpus of recent UK Acts and statutory instruments from legislation.gov.uk, together with a 100-query evaluation set of practitioner-style questions whose graded relevance judgements distinguish legally operative, supporting and contextual provisions. Using BM25 and an MPNet-based dense retriever to build candidate sets, we evaluate ten neural rerankers, including transformer cross-encoders, a late-interaction reranker, an LLM-based listwise reranker and proprietary APIs. Across both sparse and dense pools, neural reranking consistently improves normalized Discounted Cumulative Gain (nDCG) and Mean Reciprocal Rank (MRR) over first-stage retrieval. We further distil a proprietary Voyage reranker into a ModernBERT-based cross-encoder, Distilled-Voyage-ModernBERT, which approaches the teacher’s effectiveness and outperforms other open rerankers on our benchmark. Results are based on 100 expert-validated queries, each linked to three graded provisions from a single UK instrument, so they characterise single-instrument, provision-level retrieval over recent UK legislation.