Abstract: Aiming at the problem that qualifiers in the hyper-relational knowledge graph will introduce irrelevant noise into the main triple, the Hyper-Relational knowledge graph completion method with fused Noise Filtering (HRNF) was proposed.First, the feature enhancement module was constructed in order to effectively enhance the hyper-relational facts. Convolutional neural network was utilized to extract the ordinary triple features, and the complex relational structure in the hyper-relational fact was captured by heterogeneous graph neural network.Subsequently, these two features were fused to enhance the information of the main triple in the hyper-relational fact by utilizing the stability and reliability of the ordinary triple to reduce the effect of noise introduced by qualifiers. Then, a relevance-aware module was constructed, to fuse feature representations more accurately. The graph attention network was utilized to update the enhanced feature representation by dynamically learning the weights among different nodes.Finally,the semantic enhancement module was constructed , to capture complex semantic information. The Transformer model was utilized to generate the final predicted sequence by capturing the dependency between any two elements in the sequence through the self-attention mechanism. In addition, extensive experiments were conducted on two commonly used datasets, WikiPeople and JF17K,to validate the effectiveness of HRNF. Experimental results show that, compared to the optimal results of the baseline method, HRNF improves MRR, Hit@1, and Hit@10 by 0.6, 1.1, and 1.8 percentage points, when predicting the main triple entities on the Wikipeople dataset. On the JF17K dataset, MRR, Hit@1, and Hit@10 are improved by 0.5, 0.5, and 2.9 percentage points. This significant enhancement proves that the HRNF method can effectively reduce the noise problem caused by qualifiers in dealing with the task of hyper-relational knowledge graph completion.

Key words: noise filtering, qualifier, hyper-relationship fact, hyper-relational knowledge graph completion, ordinary triple

摘要： 针对超关系知识图谱中限定符会为主三元组引入无关的噪声问题，提出融合噪声过滤的超关系知识图谱补全方法（HRNF）。首先，为了有效增强超关系事实，构建特征增强模块。利用卷积神经网络提取普通三元组特征，并通过异构图神经网络捕获超关系事实中的复杂关系结构。随后，将这两种特征进行融合，利用普通三元组的稳定性与可靠性来增强超关系事实中主三元组的信息，减少限定符引入的噪声影响。其次，为了更准确地融合特征表示，构建相关性感知模块。利用图注意力网络，通过动态学习不同节点间的权重，更新增强后的特征表示。最后，为了捕获复杂的语义信息，构建语义增强模块。利用Transformer模型，通过自注意力机制捕获序列中任意两个元素之间的依赖关系，从而生成最终的预测序列。此外，为了验证HRNF的有效性，在两个常用的数据集WikiPeople和JF17K上进行了广泛的实验。实验结果表明，在Wikipeople数据集上预测主三元组实体时相较于基线方法中最优结果，HRNF的MRR、Hit@1、Hit@10 分别提升了 0.6、1.1、1.8 个百分点。在 JF17K 数据集上的MRR、Hit@1、Hit@10 分别提升了 0.5、0.5、2.9 个百分点。这一显著提升证明了HRNF方法在处理超关系知识图谱补全任务中可以有效地减少限定符带来的噪声问题。

关键词: 噪声过滤, 限定符, 超关系事实, 超关系知识图谱补全, 普通三元组