Abstract: Knowledge graphs suffer from incompleteness, largely determining the application and development of downstream tasks. Therefore, it is necessary to improve the knowledge graphs to supplement the missed values, i.e., knowledge graph completion. Many existing knowledge graph completion models reshape entity and relation embeddings to capture the local interaction. However, such a method corrupts the original triplet structure and can only utilize the local interaction while ignoring the impact of global interaction between entities and relations. Therefore, a knowledge graph completion method called AGILI, which aggregates global interaction and local interaction, is proposed. This method introduces a self-attention mechanism to evaluate the information correlation degree between head entities and relations, and then generates embedding representations that incorporate global interaction information. Then, it employs a convolutional neural network to capture local interaction information from the embedding representations. In addition, a learnable interaction aggregator based on relationship weights is designed. When the two types of interactions are fused, their importance can be adaptively adjusted according to the relationship category, which improves the expressive ability of the method on multi-relational knowledge graphs. Experimental verification was carried out on the publicly available FB15K-237, WN18RR, and Kinship datasets by link prediction task. Compared with the latest Convolutional Neural Network-based model ConvD, the Hits@1 and Hits@3 metrics of the proposed method increase by 6.9% and 5.3% respectively on the FB15K-237 dataset. Experimental results demonstrate the superiority of the proposed method.

Key words: knowledge graph, knowledge graph completion, link prediction, self-attention mechanism, convolutional neural network

摘要： 知识图谱的不完整性严重影响了下游任务的应用与发展，因此，有必要对其进行改进以补充缺失值，即知识图谱补全。现有的知识图谱补全模型大多重组实体关系嵌入表示以捕获局部交互。但这种方法破坏了三元组的原有结构，只能利用单一的局部交互而忽略了实体关系间全局交互的影响。为此，提出一种聚合全局交互与局部交互的知识图谱补全方法AGILI。该方法首先引入自注意力机制获取头实体和关系间的信息关联程度，生成融入全局交互信息的嵌入表示，再采用卷积神经网络从新嵌入表示中提取局部交互信息，除此之外设计基于关系权重的可学习交互聚合器，在将全局交互与局部交互进行特征融合时，可以根据关系类别自适应地调整两种交互的重要程度，提高方法在多关系知识图谱上的表达能力。在公开数据集FB15k-237、WN18RR和Kinship上通过链接预测任务进行实验验证，实验结果表明，与最新的基于卷积神经网络的模型ConvD相比，所提出的方法在FB15K-237数据集上Hits@1、Hits@3指标分别提高了6.9%、5.3%，证明了所提出方法的优越性。

关键词: 知识图谱, 知识图谱补全, 链接预测, 自注意力机制, 卷积神经网络