Scholarlink

<p>Document retrieval aims to identify the target documents with the highest semantic relevance within a corpus. However, capturing hierarchical relations between documents is difficult, leading to semantic misalignment. While graph-based methods are effective, traditional graph structures struggle to model hierarchical relations. This paper proposes a Hierarchical Graph Semantic Alignment (HGSA), which improves retrieval accuracy by aligning semantic structures. Firstly, named entity recognition and relation extraction are used to construct document graphs. Next, all the document graphs are pruned to reduce redundancy, followed by the aggregation of the hierarchical relation features between the document graph pairs. Finally, the restructured document graph is utilized for document retrieval calculation. HGSA leverages structural and hierarchical information, mitigating conceptual deviations and ensuring consistent semantic frameworks. Experiment results show that HGSA achieves an average Recall@1 of 25.38% across all datasets, improving the accuracy of identifying relevant documents across diverse domains. From a computational perspective, the combination of transformer encoding and graph neural propagation imposes substantial computational demands at scale. The quadratic attention complexity and iterative message passing require GPU-accelerated high-performance computing for efficient large-scale deployment, aligning HGSA with supercomputing-oriented semantic processing tasks.</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

Enhancing document retrieval using semantic alignment with hierarchical graph matching

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Jingxuan Liu,
  • Yihan Huang,
  • Jialuoyi Tan,
  • Zhen Hua

摘要

Document retrieval aims to identify the target documents with the highest semantic relevance within a corpus. However, capturing hierarchical relations between documents is difficult, leading to semantic misalignment. While graph-based methods are effective, traditional graph structures struggle to model hierarchical relations. This paper proposes a Hierarchical Graph Semantic Alignment (HGSA), which improves retrieval accuracy by aligning semantic structures. Firstly, named entity recognition and relation extraction are used to construct document graphs. Next, all the document graphs are pruned to reduce redundancy, followed by the aggregation of the hierarchical relation features between the document graph pairs. Finally, the restructured document graph is utilized for document retrieval calculation. HGSA leverages structural and hierarchical information, mitigating conceptual deviations and ensuring consistent semantic frameworks. Experiment results show that HGSA achieves an average Recall@1 of 25.38% across all datasets, improving the accuracy of identifying relevant documents across diverse domains. From a computational perspective, the combination of transformer encoding and graph neural propagation imposes substantial computational demands at scale. The quadratic attention complexity and iterative message passing require GPU-accelerated high-performance computing for efficient large-scale deployment, aligning HGSA with supercomputing-oriented semantic processing tasks.