Scholarlink

<p>The retrieval-augmented generation (RAG) reduces hallucinations in large language models (LLM) for domain-specific question answering (Q&amp;A), but current retrieval-first approaches introduce noise. Such noise often leads to irrelevant information inclusion, erroneous associations, or reduced domain-specific relevance in generated answers, thereby undermining the accuracy and reliability of responses. To address this issue, we propose a mutual information-based retrieval-augmented generation (MI-RAG) method for domain Q&amp;A, which adopts a generate-then-retrieve strategy. MI-RAG draws on three types of data sources: knowledge graphs (KG), large language models (LLM), and external web knowledge. In the retrieval phase, two components—LLM and graph retrieval—adopt mutual information (MI) as the evaluation metric. This metric helps identify and retrieve the subgraph from the KG that is most relevant to the input question. After answer generation, LLM is further used to update the KG. Evaluations on domain Q&amp;A tasks demonstrate that MI-RAG outperforms existing state-of-the-art models and other retrieval-augmented generation methods significantly.</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

Mutual information-based retrieval-augmented generation for domain question answering

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Haitao Lu,
  • Dayu Zheng,
  • Luo Yang

摘要

The retrieval-augmented generation (RAG) reduces hallucinations in large language models (LLM) for domain-specific question answering (Q&A), but current retrieval-first approaches introduce noise. Such noise often leads to irrelevant information inclusion, erroneous associations, or reduced domain-specific relevance in generated answers, thereby undermining the accuracy and reliability of responses. To address this issue, we propose a mutual information-based retrieval-augmented generation (MI-RAG) method for domain Q&A, which adopts a generate-then-retrieve strategy. MI-RAG draws on three types of data sources: knowledge graphs (KG), large language models (LLM), and external web knowledge. In the retrieval phase, two components—LLM and graph retrieval—adopt mutual information (MI) as the evaluation metric. This metric helps identify and retrieve the subgraph from the KG that is most relevant to the input question. After answer generation, LLM is further used to update the KG. Evaluations on domain Q&A tasks demonstrate that MI-RAG outperforms existing state-of-the-art models and other retrieval-augmented generation methods significantly.