Scholarlink

<p>Large Language Models (LLMs) serve as the core engine of natural language processing capabilities within intelligent cloud services, but their constrained factual stores and ongoing hallucination issues impair effectiveness in intricate reasoning scenarios. Knowledge graphs (KGs), offering organized data, serve as a solid basis for reasoning, yet current approaches relying on KGs typically view them as fixed repositories, overlooking their relational architecture. This results in suboptimal use of the information, potentially introducing spurious knowledge, thereby compromising the accuracy of the reasoning. To address these drawbacks, we introduce PDR, namely a Prompt-driven KG-enhanced LLM Reasoning, which combines LLMs and KGs in a cohesive manner while refining prompt to boost reasoning reliability and interpretability. PDR consists of two phases: (1) Subgraph Retrieval, where a refined PageRank algorithm align queries with the KG structure and a document retrieval is integrated to extend graph boundaries, thereby generating subgraphs with maximal answer coverage and high relevance; and (2) Reasoning, where task-specific prompts embed subqueries as reasoning cues, directing the LLM to generate chain-of-thought and candidate KG paths, followed by a stepwise filtering process that evaluates semantic coherence and structural alignment with the KG to identify the most pertinent answer. Thorough experiments show that PDR surpasses state-of-the-art baselines on both simple and multi-hop reasoning tasks, yielding more accurate and interpretable results.</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

Advancing faithful KBQA services via prompt-driven knowledge graph-enhanced LLMs in cloud

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月25日

查看全文
  • Zishun Rui,
  • Shengjie Chen,
  • Shucun Fu,
  • Wenzheng Sun,
  • Shengjun Xue

摘要

Large Language Models (LLMs) serve as the core engine of natural language processing capabilities within intelligent cloud services, but their constrained factual stores and ongoing hallucination issues impair effectiveness in intricate reasoning scenarios. Knowledge graphs (KGs), offering organized data, serve as a solid basis for reasoning, yet current approaches relying on KGs typically view them as fixed repositories, overlooking their relational architecture. This results in suboptimal use of the information, potentially introducing spurious knowledge, thereby compromising the accuracy of the reasoning. To address these drawbacks, we introduce PDR, namely a Prompt-driven KG-enhanced LLM Reasoning, which combines LLMs and KGs in a cohesive manner while refining prompt to boost reasoning reliability and interpretability. PDR consists of two phases: (1) Subgraph Retrieval, where a refined PageRank algorithm align queries with the KG structure and a document retrieval is integrated to extend graph boundaries, thereby generating subgraphs with maximal answer coverage and high relevance; and (2) Reasoning, where task-specific prompts embed subqueries as reasoning cues, directing the LLM to generate chain-of-thought and candidate KG paths, followed by a stepwise filtering process that evaluates semantic coherence and structural alignment with the KG to identify the most pertinent answer. Thorough experiments show that PDR surpasses state-of-the-art baselines on both simple and multi-hop reasoning tasks, yielding more accurate and interpretable results.