一种精细表示多值属性的知识图谱嵌入模型

A Knowledge Graph Embedding Model for Fine Representation of Multivalued Attributes

知识图谱嵌入模型KR-EAR用实体及其属性值的嵌入(向量)来定义属性三元组的评分函数,导致多值属性的不同属性值学得的嵌入很相似,即KR-EAR未能精细地表示多值属性,从而影响下游任务的准确度。论文通过改进KR-EAR的属性三元组表示来提出一种精细表示多值属性的知识图谱嵌入模型,称为KGE-EAV。在KGE-EAV的属性值空间中,每个实体都对应一个超平面,该实体的每个属性值嵌入都在该超平面上形成一个投影向量;KGE-EAV用这样的投影向量(而不是属性值嵌入)来定义属性三元组的评分函数,从而可以为多值属性的不同属性值学得不同的嵌入。实验表明,在实体预测和属性预测两项任务上,KGE-EAV的准确度均优于KR-EAR和三个基线模型。

The KR-EAR knowledge graph embedding model uses the embeddings(vectors)of both an entity and its attribute value to define the scoring function for the corresponding attribute triple,resulting in very similar embeddings learnt for different values of a multivalued attribute,that is,KR-EAR fails to finely represent multivalued attributes,thus affecting the accuracy of downstream tasks. In this paper,a knowledge graph embedding model,called KGE-EAV,for fine representation of multivalued attributes is proposed by improving the attribute triple representation of KR-EAR. In the attribute value space of KGE-EAV,each entity corresponds to a hyperplane,and each attribute value embedding of the entity forms a projection vector on the hyperplane.KGE-EAV then uses such a projection vector(instead of the attribute value embedding)to define the scoring function for the corresponding attribute triple,so that different embeddings can be learned for different values of a multi-valued attribute. Experiments show that the accuracy of KGE-EAV is better than that of KR-EAR and three baseline models in both entity prediction and attribute prediction tasks.