基于链接预测模型的融合研究

Fusion research based on link prediction model

知识图谱技术在信息检索、智能问答领域有着重要作用.为了解决大型知识图谱中的有效实体缺失问题,需要使用链接预测手段自动预测实体之间缺失的链接,完善知识图谱中缺失的实体.现在有许多链接预测的方法,其中基于神经网络的Grail模型侧重于对局部子图进行推理,具有很强的归纳倾向,但在语义层次建模方面存在不足.因此,提出了Grail&HAKE融合模型,通过将Grail模型与HAKE模型进行融合,解决了Grail模型在语义层次建模方面存在的不足.实验结果表明,Grail&HAKE的融合模型比单独使用Grail和HAKE模型的MRR值分别高出0.1005和0.3063,并且在部分数据集上优于其他融合模型.说明Grail&HAKE融合模型在知识图谱的有效实体链接预测方面是有效可用的.

The knowledge graph technology had played an important role in information retrieval and intelligent question answering.In order to solve the problem of missing entities in large-scale knowledge graphs,the automatic link prediction of knowledge graph was needed.There were many methods of link prediction,among which the neural network-based Grail model focused on reasoning on local sub graphs,which had a strong inductive tendency,but was insufficient for semantic-level modeling.Therefore,the fusion model of Grail&HAKE was proposed to solve the problem of insufficient semantic-level modeling about Grail model.The experimental results showed that the fusion model of Grail&HAKE was at least 0.1005 and 0.3063 higher in MRR than Grail and HAKE alone,and it outperformed other fusion models on some datasets.It was proved that the fusion model of Grail&HAKE was valid and useful to predict missing entities in knowledge graph.