基于注意力机制的多尺度空洞卷积神经网络模型

现有的时间知识图谱表示方法不能较好地捕获四元组内的复杂关系,而基于神经网络的模型大都无法建模随时间变化的知识,不能捕获丰富的特征信息,实体和关系间的交互性也较差.因此,文中提出基于注意力机制的多尺度空洞卷积神经网络模型.首先利用长短期记忆网络获得时间感知的关系表示.再利用多尺度空洞卷积神经网络提高四元组的交互性.最后,使用多尺度注意力机制捕获关键特征,提高模型的补全能力.在多个公开时间数据集上的链路预测实验表明,文中模型性能较优.

Multi-scale Dilated Convolutional Neural Network Model Based on Attention Mechanism

The existing temporal knowledge graph representation methods cannot capture the complex relationships within quadruple well. Most of the neural network based models are unable to model time-varying knowledge and capture rich feature information. Moreover, the interaction between entities and relations in these models is poor. Therefore, a multi-scale dilated convolutional neural network model based on attention mechanism(MSDCA) is proposed. Firstly, a time-aware relation representation is obtained using long short-term memory. Secondly, a multi-scale dilated convolutional neural network is employed to improve the interactivity of the quadruple. Finally, a multi-scale attention mechanism is utilized to capture critical features to improve completion ability of MSDCA. Link prediction experiments on multiple public temporal datasets show the superiority of MSDCA.