Scholarlink

<p>This paper examines the integration of large language models (LLMs) with knowledge graphs (KGs) through a systematic four-layer framework that includes data, model, technology, and application dimensions. We analyze the capabilities and limitations of LLMs in natural language processing along with the strengths and challenges of KGs in knowledge representation. We address fundamental weaknesses in each approach and identify complementary integration methods. Our analysis reveals that LLMs excel at contextual understanding and generation but struggle with factual consistency and reasoning transparency. In contrast, KGs provide structured and verifiable knowledge but lack adaptability to unstructured inputs. We review integration strategies, including knowledge injection techniques, retrieval-augmented generation, and neuro-symbolic approaches. The combined methods demonstrate significant performance improvements. Through the case study of the GLM architecture, we demonstrate how the integration of KGs and LLMs improves accuracy, interpretability, and factual grounding in specialized domains and also shows substantial performance improvements in knowledge-intensive tasks (15–20% on MedQA and 14–17% on the medical MMLU benchmarks). The resulting hybrid systems offer concrete advantages in critical applications requiring precision and adaptability, including healthcare diagnostics, financial compliance, and educational technology. Lightweight knowledge representation, adaptive update mechanisms, and unified cross-modal frameworks are promising research directions to advance KG–LLM integration.</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

A review on synergizing knowledge graphs and large language models

  • 期刊论文

  • 发表日期：

    2025年6月6日

查看全文
  • Zhenyao Yang,
  • Sha Yuan,
  • Zhou Shao,
  • Wenfa Li,
  • Runzhou Liu

摘要

This paper examines the integration of large language models (LLMs) with knowledge graphs (KGs) through a systematic four-layer framework that includes data, model, technology, and application dimensions. We analyze the capabilities and limitations of LLMs in natural language processing along with the strengths and challenges of KGs in knowledge representation. We address fundamental weaknesses in each approach and identify complementary integration methods. Our analysis reveals that LLMs excel at contextual understanding and generation but struggle with factual consistency and reasoning transparency. In contrast, KGs provide structured and verifiable knowledge but lack adaptability to unstructured inputs. We review integration strategies, including knowledge injection techniques, retrieval-augmented generation, and neuro-symbolic approaches. The combined methods demonstrate significant performance improvements. Through the case study of the GLM architecture, we demonstrate how the integration of KGs and LLMs improves accuracy, interpretability, and factual grounding in specialized domains and also shows substantial performance improvements in knowledge-intensive tasks (15–20% on MedQA and 14–17% on the medical MMLU benchmarks). The resulting hybrid systems offer concrete advantages in critical applications requiring precision and adaptability, including healthcare diagnostics, financial compliance, and educational technology. Lightweight knowledge representation, adaptive update mechanisms, and unified cross-modal frameworks are promising research directions to advance KG–LLM integration.