融合多源信息的知识表示学习

Knowledge Representation Learning Based on Multi-source Information Combination

在知识图谱中,实体的文本描述信息、实体的层次类型信息和图的拓扑结构信息中隐藏着丰富的内容,它们可以形成对原始三元组的有效补充,帮助提高知识图谱各种任务的效果。为了充分利用这些多源异质信息,首先通过一维卷积神经网络嵌入文本描述信息,然后根据实体的层次类型信息构建投影矩阵,将三元组中的实体向量和实体的描述向量映射到特定的关系空间中来约束实体的语义信息,再基于图注意力机制融合图的拓扑结构信息,计算不同邻接点对实体的影响。在图注意力层中,计算了实体间的多跳关系来帮助改善数据稀疏的问题。最后,通过二维卷积神经网络来捕获不同维度间的全局信息,进一步提高模型的性能。链接预测实验结果表明,基于多源信息组合的知识表示学习模型(MCKRL)能够充分利用三元组以外的多源异质信息,因而相比于其他基线模型,该模型在链接预测任务上取得了更好的结果。

In knowledge graphs,there are rich contents hidden in the text description information of entity,the hierarchical type information of entity and the topological structure information of graph,and they can form an effective supplement to the triple information to get better performance.In order to make full use of these heterogeneous information,the convolutional neural networks are firstly used to encode entity description.Then a projection matrix is constructed according to hierarchical type information to project entity vectors and entity description vectors into specific relation space to constrain their semantic information.After that,the graph attention mechanism is introduced to fuse the topological structure information of graph and calculate the influence of different adjacency points on entities.Meanwhile,the multi-hop relationship information between entities is calculated to further solve the problem of data sparsity.Finally,a decoder is employed to capture the global information between different dimensions.Experimental results of link prediction show that the multi-source information combined knowledge representation learning(MCKRL)model can make good use of multi-source heterogeneous information beyond triples,so it obtains better results than other baseline models in link prediction.