Scholarlink

<p>Inductive Knowledge Graph Reasoning (IKGR) aims to complete missing relations for emerging entities in knowledge graphs (KGs), with the core challenge lying in constructing a semantically rich yet structurally compact reasoning context. Existing approaches primarily follow two paradigms, each with inherent limitations: path-based methods leveraging Pretrained Language Models (PLMs) can harness rich textual semantics but struggle to effectively model the structured constraints and complex topology of KGs; GNN-based methods operating on local subgraphs, while capable of capturing structural information, face a dilemma whereby the use of subgraphs leads to exponential growth in KG size and the introduction of irrelevant entities, whereas progressive sampling often suffers from sampling bias due to the lack of semantic guidance. To unify semantic understanding with structural reasoning, we propose PINS, a novel <b>P</b>LM-guided <b>I</b>ncremental <b>N</b>ode <b>S</b>ampling framework. PINS innovatively incorporates an Ontology Matching Module (OMM) into the progressive sampling process. This module leverages the deep semantic knowledge from PLMs to quantify the relevance between candidate nodes and the query, thereby achieving semantics-aware precise sampling. Extensive experiments on benchmark datasets demonstrate that PINS achieves competitive results against baseline methods. It further optimizes the model’s ability to reason about triple relations, significantly reduces subgraph extraction time and the number of irrelevant nodes sampled, and effectively balances efficiency and accuracy in IKGR scenarios.</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

PLM-guided incremental node sampling for inductive knowledge graph reasoning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月12日

查看全文
  • Ling Wang,
  • Yan Liu,
  • Jicang Lu,
  • Zhipeng Li

摘要

Inductive Knowledge Graph Reasoning (IKGR) aims to complete missing relations for emerging entities in knowledge graphs (KGs), with the core challenge lying in constructing a semantically rich yet structurally compact reasoning context. Existing approaches primarily follow two paradigms, each with inherent limitations: path-based methods leveraging Pretrained Language Models (PLMs) can harness rich textual semantics but struggle to effectively model the structured constraints and complex topology of KGs; GNN-based methods operating on local subgraphs, while capable of capturing structural information, face a dilemma whereby the use of subgraphs leads to exponential growth in KG size and the introduction of irrelevant entities, whereas progressive sampling often suffers from sampling bias due to the lack of semantic guidance. To unify semantic understanding with structural reasoning, we propose PINS, a novel PLM-guided Incremental Node Sampling framework. PINS innovatively incorporates an Ontology Matching Module (OMM) into the progressive sampling process. This module leverages the deep semantic knowledge from PLMs to quantify the relevance between candidate nodes and the query, thereby achieving semantics-aware precise sampling. Extensive experiments on benchmark datasets demonstrate that PINS achieves competitive results against baseline methods. It further optimizes the model’s ability to reason about triple relations, significantly reduces subgraph extraction time and the number of irrelevant nodes sampled, and effectively balances efficiency and accuracy in IKGR scenarios.