高级检索

    基于镜像空间的平移嵌入模型

    Mirrored Translation Embedding Model Based on Mirrored Space

    • 摘要: 知识图广泛应用于许多人工智能(AI)任务。然而,现有知识图通常是不完整的,需对知识图进行补全或链接预测。本文通过对知识图中的实体和关系进行嵌入来预测知识图的缺失环节:首先,引入镜像空间的概念,使得模型具有学习对称和反对称模式的能力;其次,在新的空间模型中,关系仍然被建模为平移,而实体被建模为具有镜像点的点;最后,提出了MTransE模型将镜像空间的概念应用到TransE上,并在4个广泛使用的数据集上进行实验。实验结果表明,该方法能减少参数的规模,并提高了在4个广泛使用的知识补全数据集上的性能。

       

      Abstract: Knowledge graphs have been widely used in many artificial intelligence (AI) tasks. However, the existing knowledge map is usually incomplete and needs to be supplemented or linked for prediction knowledge graphs. In this paper, the problem of missing links in the knowledge map is predicted by embedding entities and relationships into the knowledge map. A mirror space translation method is introduced to learn the symmetric/antisymmetric patterns. In the new space model, relations are still modelled as translations, while entities are modelled as points with mirrors. In this space, the translation-based models gain the ability to model symmetry/anti-symmetry relations. Finally, the proposed model MTransE applies the concept of mirrored space to TransE. By the experiments on four well-known datasets, it is verified that the proposed mode can attain better performance than other baseline models.

       

    /

    返回文章
    返回