【作者】 吴强；

【导师】 何清；

【作者基本信息】 郑州大学 ， 计算机科学与技术， 2024， 硕士

【摘要】 知识图谱是一种结构化的信息表达方式,主要用于描绘各个实体之间的联系。它由节点(实体)和边(关系)组成,形成了一个图形结构。现有的知识图谱涵盖了大量知识,并且在问答、推荐系统等领域应用广泛,但这些知识图谱通常存在不完整的情况,即其中包含许多未知的实体和关系。链路预测任务的目标便是填补知识图谱中缺失的信息,从而使其更为完善。当前主流的知识图谱链路预测模型依赖于知识图谱嵌入技术。这种方法充分利用了知识图谱中的三元组结构信息,将实体与关系有效地嵌入到低维且密集的向量空间中。通过这种方式,我们能够得到实体和关系的向量表示,进而利用特定的得分函数来执行链路预测任务。尽管现有的链路预测模型已经取得了显著的成果,但它们仍然面临一些亟待解决的挑战。首要的挑战是如何进一步提升实体和关系的建模精确度,并充分利用三元组结构信息。其次,知识图谱中除了包含三元组结构信息外,还蕴藏着诸如文本描述、实体属性等多源信息,这些信息对于完善知识图谱具有不可忽视的作用。鉴于近年来图卷积神经网络在图结构数据建模上的卓越表现,本文决定采用图卷积神经网络对知识图谱进行建模。通过这种方法,我们能够更有效地捕获图结构信息,强化实体和关系之间的语义关联,从而优化知识图谱的链路预测效果。本文的主要研究内容聚焦在以下两个方面:(1)提出了一种基于注意力机制和关系信息的链路预测模型,旨在更精准地捕捉实体与关系之间的深层次信息。相较于传统的图卷积神经网络方法,我们的模型在整合邻居节点和关系信息时,创新性地引入了循环相关算子和自注意力向量。通过这两种机制,模型能够生成邻域信息的中间表示,并经过聚合处理得到实体的最终表示。最终,利用得分函数对三元组进行评估,实现知识图谱的链路预测。为了验证模型的性能,我们选取了 TransE、ConvE等传统方法作为基准,并在FB15K-237和WN18RR这两个公共数据库上进行了实验对比。实验结果表明,我们的新模型在链路预测任务上展现出了明显的优势,充分证明了其在实际应用中的有效性。(2)提出了基于实体描述信息和关系信息的链路预测模型。该方法针对实体语义信息缺失的问题,使用预训练语言模型编码实体描述信息,并与实体的初始嵌入进行融合,在聚合邻域信息阶段,本文将关系和尾实体进行有效的组合,使实体信息中包含关系的信息,从而丰富实体的嵌入表示,最后通过得分函数评估三元组,实现知识图谱的链路预测。本文在FB15K-237和WN18RR公开数据集上与TransE、ConvE等经典模型进行实验对比,实验结果表明所提出模型在各项指标上都有着优秀的表现,进而验证模型在链路预测任务中的有效性。更多还原

【Abstract】 Knowledge graph is a structured way of expressing information,mainly used to depict the connections between various entities.It is composed of nodes(entities)and edges(relations),forming a graphical structure.Existing knowledge graphs cover a vast amount of knowledge and are widely used in areas such as question answering and recommendation systems.However,these knowledge graphs often suffer from incompleteness,meaning that there are many unknown entities and relations within them.The objective of link prediction tasks is to fill in the missing information in knowledge graphs,thus making them more comprehensive.Currently,the mainstream knowledge graph link prediction models primarily rely on knowledge graph embedding.This approach leverages the triplet structural information in the knowledge graph,effectively embedding entities and relations into a low-dimensional and dense vector space.Through this method,we can obtain vector representations of entities and relations,and then use specific scoring functions to perform link prediction.Despite the significant progress made by existing link prediction models,they still face some urgent challenges.The primary challenge is how to further enhance the accuracy of entity and relation modeling while fully utilizing triplet structural information.Secondly,in addition to triplet structural information,knowledge graphs also contain multi-source information such as textual descriptions and entity attributes,which play an indispensable role in improving knowledge graphs.Given the excellent performance of graph convolutional neural network in modeling graph-structured data in recent years,this paper decides to adopt graph convolutional neural network for modeling knowledge graph.By doing so,we can more effectively capture graph structural information,strengthen the semantic associations between entities and relations,and thereby optimize the link prediction effectiveness of knowledge graphs.The main research contents of this paper focus on the following two aspects:(1)We propose a link prediction model based on attention mechanisms and relational information,aiming to capture deeper information between entities and relations more accurately.Compared to traditional graph convolutional neural network methods,our model innovatively introduces recurrent correlation operators and selfattention vectors when integrating neighbor nodes and relational information.Through these two mechanisms,the model can generate intermediate representations of neighborhood information and obtain the final representation of entities through aggregation processing.Finally,a scoring function is used to evaluate triplets and perform link prediction in knowledge graphs.To validate the performance of our model,we selected traditional methods such as TransE and ConvE as baselines and conducted experimental comparisons on the two public databases,FB15K-237 and WN18RR.The experimental results show that our new model exhibits significant advantages in link prediction tasks,fully demonstrating its effectiveness in practical applications.(2)We propose a link prediction model based on entity description information and relational information.This method addresses the issue of missing semantic information for entities by using pre-trained language models to encode entity description information and fusing it with the initial embeddings of entities.In the phase of aggregating neighborhood information,we effectively combine relations and tail entities to enrich the embedding representation of entities by incorporating relational information into the entity information.Finally,a scoring function is used to evaluate triplets and perform link prediction in knowledge graphs.We conducted experimental comparisons with classic models such as TransE and ConvE on the public datasets FB15K-237 and WN18RR.The experimental results show that the proposed model performs excellently on various metrics,further validating its effectiveness in link prediction tasks.更多还原