【作者】 袁琦；

【导师】 苏厚胜；

【作者基本信息】 华中科技大学 ， 控制科学与工程， 2023， 硕士

【摘要】 知识图谱以结构化的形式储存知识,被广泛应用于医疗、安全和电商等领域,如临床辅助决策、犯罪团伙跟踪与预警和商品智能化推荐等。然而在实际中,直接应用知识存在计算效率低下的问题,因此知识表示的研究具有重大意义。知识图谱嵌入通过将实体和关系映射到低维向量空间来实现知识的高效表示,具有储存效率高和计算复杂度低的优点。基于卷积神经网络的模型具有参数效率高和表现力强的特点,在预测任务上取得了优异的表现。基于图神经网络的模型能够挖掘实体的邻居语义,有效建模图结构数据。但以上两类模型在建模三元组结构和邻居语义上依然存在交互能力受限和实体表示单一的问题。基于上述分析,本文以三元组为主要研究对象,挖掘实体和关系的交互特征,并结合邻居丰富实体表示,对知识图谱嵌入技术展开研究,提升知识表示质量。主要研究内容如下:(1)针对卷积神经网络在不同空间位置上特征交互能力受限的问题,提出基于内卷积网络的知识图谱嵌入模型。内卷积根据不同空间位置上的输入信息自适应地生成核参数,有效地增强实体和关系的特征交互作用。在提取特征前,引入通道补充和边界补充,使内卷积交互特征更为丰富。在FB15k-237和YAGO3-10数据集上的链接预测实验结果表明,模型性能与基于卷积神经网络的模型相比有显著提升。(2)针对图神经网络在聚合邻居的过程中实体表示单一的问题,提出融合关键邻居边的动态知识图谱嵌入模型。模型通过Transformer同时提取预测序列和关键邻居边的信息,利用关键邻居边为预测序列中的实体补充特征,增强实体在不同预测序列中的表示。同时提出一种关键邻居边挑选策略,用余弦相似度方法挑选邻居关系,然后用Page Rank算法挑选邻居实体,实现为预测序列中的实体挑选与当前场景中最相关的邻居边。在FB15k-237和WN18RR数据集上的实验结果表明,模型在链接预测任务上的表现优于大多数基于图神经网络的模型。更多还原

【Abstract】 Knowledge graph stores knowledge in a structured form and is widely used in medical,security,and e-commerce fields,such as clinical assisted decision making,criminal gang tracking and early warning,and commodity intelligent recommendation.However in practice,directly applying knowledge has the problem of low computational efficiency,so the research of knowledge representation is of great significance.Knowledge graph embedding enables efficient representation of knowledge by mapping entities and relations into a low-dimensional vector space,with the advantages of efficient storage and low computational complexity.The model based on convolutional neural network has the characteristics of high parameter efficiency and strong performance,and has achieved excellent performance in the prediction task.The model based on graph neural network can mine the neighborhood semantics of entities and effectively model graph-structured data.However,the above two types of models still have the problems of limited interaction ability and single entity representation in modeling triple structure and neighborhood semantics.Based on the above analysis,this thesis takes triples as the main research object,mines the interaction characteristics of entities and relations,combines with neighbors to enrich entity representation,carries out research on knowledge graph embedding technology,and improves the quality of knowledge representation.The main research elements are as follows:Aiming at the problem that convolutional neural network has limited feature interaction ability at different spatial locations,a knowledge graph embedding model based on involutional network is proposed.The involution adaptively generates kernel parameters according to the input information at different spatial locations,which effectively enhances the feature interaction between entities and relations.Before feature extraction,channel supplement and boundary supplement are introduced to enrich the interaction features of involution.The experimental results of link prediction on FB15k-237 and YAGO3-10 datasets show that the model performance is significantly improved compared with the model based on convolutional neural network.Aiming at the problem of single entity representation in the process of neighbor aggregation in graph neural network,a dynamic knowledge graph embedding model fused with key neighbor edges is proposed.The model extracts the information of the prediction sequence and the key neighbor edges simultaneously through Transformer,and uses the key neighbor edges to supplement the features of entities in the prediction sequence to enhance the representation of entities in different prediction sequences.A key neighbor edges selection strategy is proposed,using the cosine similarity method to select neighbor relations and the Page Rank algorithm to select neighbor entities,which can select the most relevant neighbor edges for the entities in the prediction sequence with the current scene.Experimental results on the FB15k-237 and WN18 RR datasets show that the model outperforms most graph neural network-based models on the link prediction task.更多还原