【作者】 王楠；

【导师】 潘鹏；

【作者基本信息】 华中科技大学 ， 计算机技术（专业学位）， 2024， 硕士

【摘要】 近年来,归纳知识图谱补全由于更符合现实场景受到广泛研究,其内容根据不可见实体与原始知识图谱有无链接,可以分为半归纳和全归纳两种任务。现有方法主要目标是在不重新训练模型的情况下,获取不可见实体或子图的准确表示,目前仍存在提取邻域或子图信息量不足、关系信息建模不充分以及可解释性弱的问题。针对半归纳补全仅提取一阶邻域忽略多跳邻域信息、仅聚合局部邻域信息忽略关系信息的问题,提出一种基于邻域采样和关系语义理解的半归纳知识图谱补全模型。首先设计基于锚节点、邻居节点和相邻关系的二阶邻域采样聚合方法,得到实体的丰富局部表示。然后通过基于超图神经网络的超关系特征建模方法,构建实体的关系语义增强全局表示,并与局部表示融合。最后设计多尺度卷积网络增加实体融合表示与查询关系表示的特征交互,通过基于关系先验分布的条件生成对抗网络,约束查询关系下实体表示与对应关系先验相匹配。针对全归纳补全子图信息量不足且存在噪声、未充分捕捉结构信息建模关系以及可解释性弱的问题,提出一种基于局部感知和关系语义增强的全归纳知识图谱补全模型。首先设计基于局部感知聚类的方法去除噪声信息,获取查询关系相关子图,并通过上述半归纳邻域采样方法收集实体邻域信息初始化子图各实体特征。然后设计关系增强的子图表示,构建以子图中关系路径为基础的关系上下文图,通过交互图注意力网络分别对子图和关系上下文图进行结构编码,以充分捕捉局部结构信息得到关系路径表示,将其与子图表示组合;通过基于局部-全局互信息的对比学习方法,以捕捉全局图结构信息更新关系路径表示,进而优化子图表示。此外,设计对关系路径表示进一步量化分析的方法,为补全结果提供多方面的解释依据。在3个公开数据集的10个版本上进行了大量对比实验,半归纳模型在Hits@10指标上最高提升了3.7%,全归纳模型在Hits@10指标上最高提升了3.1%。此外,消融实验、参数敏感性实验和可解释性分析从多角度验证了各部分工作的有效性。更多还原

【Abstract】 Inductive knowledge graph completion has gained widespread research attention in recent years due to its relevance to real-world scenarios.This area is divided into semiinductive and fully inductive tasks,based on whether there are connections between the unseen entities and the original knowledge graph.The primary goal of existing methods is to obtain accurate representations of unseen entities or subgraphs without retraining the model.However,there are ongoing challenges such as insufficient extraction of neighborhood or subgraph information,inadequate modeling of relational information,and poor interpretability.To address the issue of semi-inductive completion,which only extracts first-order neighborhood information and overlooks multi-hop neighborhood details,and aggregates only local neighborhood information while ignoring relational data,a Semi-Inductive Completion Model based on neighborhood sampling and relational semantic understand ing has been proposed.It starts by designing a second-order neighborhood sampling and aggregation method based on anchor nodes,neighbor nodes,and adjacent relationships to achieve rich local representations of entities.Then,a super-relational feature modeling method based on hypergraph neural networks is introduced to construct global representations of entities based on relational semantics,which are then fused with local representations.Finally,a feature interaction between the fused entity representation and the query relation representation is performed using a multi-scale convolutional network,and a conditional generative adversarial network based on the prior distribution of relations is designed to match the entity representation under the query relation with the corresponding relational prior.For fully-inductive completion,which suffers from insufficient and noisy subgraph information,a Fully-Inductive Completion Model based on local perception and enhanced relational semantics has been proposed.It starts by designing a noise-reducing local perception clustering method to obtain subgraphs related to the query relation,and uses the aforementioned semi-inductive neighborhood sampling method to initialize features of entities in the subgraph.Then,an relation-enhanced subgraph representation is proposed,constructing a relational context graph based on the relationship paths within the subgraph and employing an interactive graph attention network to structurally encode the subgraph and relational context graph.This captures local structural information to derive a representation of the relational paths,which is then combined with the subgraph representation.A contrastive learning method based on local-global mutual information is designed to capture global graph structure information,update the relational path representation,and further optimize the subgraph representation.Additionally,quantitative analysis of the relational path representations provides a multi-faceted basis for interpreting the completion results.A large number of comparative experiments were conducted on 10 versions of 3public datasets.The semi-inductive model achieved a maximum improvement of 3.7% on the Hits@10 metric,and the fully inductive model achieved a maximum improvement of 3.1% on the Hits@10 metric.Additionally,ablation experiments,parameter sensitivity experiments,and interpretability analysis were conducted from multiple perspectives to validate the effectiveness of each part of the work.更多还原