基于不确定知识图谱嵌入的多关系近似推理模型

Multi-relation approximate reasoning model based on uncertain knowledge graph embedding

针对大规模知识图谱(KG)的不确定性嵌入模型中无法对多种逻辑关系进行近似推理的问题,提出一种基于不确定KG嵌入(UKGE)的多关系近似推理模型UDConEx(Uncertainty DistMult(Distance Multiplicative) and complex Convolution Embedding)。首先,UDConEx结合DistMult和ComplEx(Complex Embedding)模型的特点,使得UDConEx具有推理对称与非对称关系的能力;其次,UDConEx采用卷积神经网络(CNN)捕获不确定性KG中的交互信息,使它具有推理逆关系和传递关系的能力;最后,UDConEx利用神经网络对KG的不确定信息进行置信度学习,在UKGE空间中可以进行近似推理。在CN15k、NL27k和PPI5k这3个公开数据集上的实验结果表明,相较于MUKGE(Multiplex UKGE)模型,UDConEx在CN15k、NL27k和PPI5k的置信度预测任务中平均绝对误差(MAE)分别降低了6.3%,30.1%和44.9%;在关系事实排名任务中,基于线性的归一化折损累计增益(NDCG)在CN15k和NL27k数据集中分别提升了5.8%和2.6%;在多关系近似推理任务中验证了UDConEx具有多种逻辑关系的近似推理能力。UDConEx弥补了传统嵌入模型无法进行置信度预测的不足,实现了对多种逻辑关系的近似推理,具有更精确、具有可解释性的不确定性知识图谱推理能力。

Because the uncertain embedding model of large-scale Knowledge Graph(KG)can not perform approximate reasoning on multiple logical relationships,a multi-relation approximate reasoning model based on Uncertain KG Embedding(UKGE)named UDConEx(Uncertainty DistMult(Distance Multiplicative)and complex Convolution Embedding)was proposed.Firstly,the UDConEx combined the characteristics of DistMult and ComplEx(Complex Embedding),enabling it to infer symmetric and asymmetric relationships.Subsequently,Convolutional Neural Network(CNN)was employed by the UDConEx to capture the interactive information in the uncertain KG,thereby enabling it to reason inverse and transitive relationships.Lastly,the neural network was employed to carry out confidence learning of uncertain KG information,enabling the UDConEx to perform approximate reasoning within the UKGE space.The experimental results on three public data sets of CN15k,NL27k,and PPI5k show that,compared with MUKGE(Multiplex UKGE)model,the Mean Absolute Error(MAE)of confidence prediction is reduced by 6.3%,30.1%and 44.9%for CN15k,NL27k and PPI5k respectively;in the task of relation fact ranking,the linear-based Normalized Discounted Cumulative Gain(NDCG)is improved by 5.8%and 2.6%for CN15k and NL27k respectively;in the multi-relation approximate reasoning task,it is verified that the UDConEx has the approximate reasoning ability of multiple logical relationships.The inability of traditional embedding models to predict confidence is compensated for by the UDConEx,which achieves approximate reasoning for multiple logical relationships and offers enhanced accuracy and interpretability in uncertainty KG reasoning.