Scholarlink

<p>Reasoning to answer logical queries over a knowledge graph has been paid attention to research recently. Previous studies have only used either distance measurement or machine learning approaches processed logical operators such as conjunction, disjunction, and negation but considering the entity description information of knowledge graph yet. This study proposes a model, named QUERY2TREE, to combine a graph representative learning model and large language model to augment semantic for embedded entities of the knowledge graph such that these entities can handle logical operations more accurately during the reasoning process. Specifically, the QUERY2TREE used the graph neural network (GNN) to embed entities and Gemma model, a family of Google Gemini model, to embed the entity's descriptive information into low-dimensional space. Then, these embedded entities were indexed by K-D tree according to each relation over the knowledge graph. Next, we used nearest neighbor search algorithm of K-D tree to build logical operations such as projection, intersection, union, and negation. Finally, we applied these created logical operations to infer the answer to complex questions. We experimented with our model over benchmark knowledge graphs such as FB15k, FB15k-237, and NELL995. As a result, the QUERY2TREE model improved the accuracy better than the baseline models QUERY2BOX, ConE, CQD-CO, MLP, SILR, GNN-QE, CKGR, LACK, and LACT by up to 69.6% of MRR and by up to 69.4% of hits@3.</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

QUERY2TREE: a reasoning model for answering logical queries based on knowledge graph embedding and large language model

  • 期刊论文

  • 发表日期：

    2025年3月14日

查看全文
  • Truong H. V. Phan,
  • Phuc Do

摘要

Reasoning to answer logical queries over a knowledge graph has been paid attention to research recently. Previous studies have only used either distance measurement or machine learning approaches processed logical operators such as conjunction, disjunction, and negation but considering the entity description information of knowledge graph yet. This study proposes a model, named QUERY2TREE, to combine a graph representative learning model and large language model to augment semantic for embedded entities of the knowledge graph such that these entities can handle logical operations more accurately during the reasoning process. Specifically, the QUERY2TREE used the graph neural network (GNN) to embed entities and Gemma model, a family of Google Gemini model, to embed the entity's descriptive information into low-dimensional space. Then, these embedded entities were indexed by K-D tree according to each relation over the knowledge graph. Next, we used nearest neighbor search algorithm of K-D tree to build logical operations such as projection, intersection, union, and negation. Finally, we applied these created logical operations to infer the answer to complex questions. We experimented with our model over benchmark knowledge graphs such as FB15k, FB15k-237, and NELL995. As a result, the QUERY2TREE model improved the accuracy better than the baseline models QUERY2BOX, ConE, CQD-CO, MLP, SILR, GNN-QE, CKGR, LACK, and LACT by up to 69.6% of MRR and by up to 69.4% of hits@3.