Scholarlink

<p>This paper introduces multi-granular corrective retrieval-augmented generation (MG-CRAG), a novel framework that enhances response quality in retrieval-based systems by processing text at multiple levels of granularity. Building on recent CRAG approaches that mitigate hallucinations in large language models through irrelevant content filtering, our method addresses the limitations of heuristic labeling via a weakly supervised, four-stage pipeline that combines manual annotation with autoencoder-guided pseudo-labeling. The framework employs a sequential passage-level retrieval evaluator and sentence-level retrieval evaluator, both based on efficient T5 architectures, to hierarchically refine documents. The short-answer datasets used in this study include ARC-Challenge, PubHealth, and PopQA. MG-CRAG achieves state-of-the-art performance on ARC-Challenge (68.85% accuracy) and PopQA (59.89% accuracy), while delivering equal results on the PubHealth dataset despite a lower web search rate. Key advantages include significantly reduced dependence on web search, minimal labeled data requirements, and customizable inference modes (strict/moderate/lenient) that optimize performance across different dataset characteristics. The framework also enables tunable trade-offs between accuracy and web search usage, demonstrating that multi-granular processing enhances focus on relevant content, substantially improving answer accuracy while maintaining computational efficiency.</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

MG-CRAG: fusion of multi-granular retrieval evaluators in corrective RAG with weakly supervised fine-tuning

  • 期刊论文

  • 发表日期：

    2026年5月5日

查看全文
  • Negin Masoumi,
  • Omid Davar,
  • Mahdi Eftekhari

摘要

This paper introduces multi-granular corrective retrieval-augmented generation (MG-CRAG), a novel framework that enhances response quality in retrieval-based systems by processing text at multiple levels of granularity. Building on recent CRAG approaches that mitigate hallucinations in large language models through irrelevant content filtering, our method addresses the limitations of heuristic labeling via a weakly supervised, four-stage pipeline that combines manual annotation with autoencoder-guided pseudo-labeling. The framework employs a sequential passage-level retrieval evaluator and sentence-level retrieval evaluator, both based on efficient T5 architectures, to hierarchically refine documents. The short-answer datasets used in this study include ARC-Challenge, PubHealth, and PopQA. MG-CRAG achieves state-of-the-art performance on ARC-Challenge (68.85% accuracy) and PopQA (59.89% accuracy), while delivering equal results on the PubHealth dataset despite a lower web search rate. Key advantages include significantly reduced dependence on web search, minimal labeled data requirements, and customizable inference modes (strict/moderate/lenient) that optimize performance across different dataset characteristics. The framework also enables tunable trade-offs between accuracy and web search usage, demonstrating that multi-granular processing enhances focus on relevant content, substantially improving answer accuracy while maintaining computational efficiency.