中国科学院自动化研究所机构知识库

知识图谱嵌入将离散的符号化实体和关系映射到连续的低维空间中, 使得后续的任务(如推理、图谱补全等) 可以通过代数运算的方式进行. 大体而言, 知识图谱嵌入可以分成两类主流方法: 平移距离模型和语义匹配模型. 本文从知识图谱的图结构出发, 分析平移距离模型的学习瓶颈, 指出其最大的缺点是没有区别对待实体在头部和尾部两个位置所应有的不同语义. 据此, 本文提出了位置敏感的嵌入表示模型(Location-Sensitive Embedding, LSE). 与现有模型不同的是, LSE 仅对头部实体进行由不同关系决定的映射, 并且将关系建模为一般化的线性变换而不仅是平移. 本文从理论上分析了LSE 的表示能力, 及与其他模型之间的联系. 同时, 为了提高模型的训练和推理效率, 本文提出了简化模型LSEd, 将线性变换限制为对角阵. 本文在4个大规模通用知识图谱数据集上使用链接预测任务进行测试. 实验结果表明, 本文提出的模型达到了最高或者与当前最先进模型持平的性能.

Knowledge graph embedding maps the discrete symbolic entities and relations into continuous low-dimensional space, enabling the downstream task (e.g., inference, knowledge graph completion and etc.) to be carried on in an algebraic manner. Roughly speaking, there are two main-streams towards knowledge graph embedding, translational distance models and semantic matching models. In this paper, we analyse the learning bottleneck of the former caused by special graph structures, and nd out that, translational distance models are to be blamed for treating head and tail entities from a same viewpoint. So, we propose a Location-Sensitive Embedding model (LSE). Unlike previous models, it just transforms the head entity with relational-specific mapping. And it models relation as a general linear transformation instead of a translational operator. We theoretically analyse its representation capacity and the relationship to current models. Also, to make it ecient, we simplify the model by assuming a diagonal matrix for transformation, forming a simplified model LSEd. We evaluate the proposed model on four large-scale knowledge graph datasets for link prediction task. Experiments have shown that the proposed model achieves highest, or at least competitive, performance compared with the state-of-the-art models.