Knowledge Representation Learning with Dynamic Path

Yiqin Luo ； Liang Chang ； Long Li ； Wei Chen ； Tianlong Gu

《電腦學刊》 30卷6期 (2019/12) Pp. 207-215

https://doi.org/10.3966/199115992019123006016

摘要

Representation learning aims to embed knowledge graphs into a low-dimensional, dense, and real-valued vector space, in which entities and relations in a knowledge graph are represented as vectors. Many models have been proposed in the literature for the embedding. A model will perform better if it can capture more information than other models during the embedding process. Compared with the classical model TransE, the model PTransE takes into account not only direct relations but also multi-step relations (i.e, paths) between each pair of entities, and thereby achieves significant improvement in the tasks of entity prediction and relation prediction. However, in the case that there are many multi-step relations between a pair of entities, PTransE doesn't make any distinguish between them. In this paper, by introducing dynamic factors into the path embedding process of the PTransE model, we propose a dynamic path translation (DPT) method to capture different paths between each pair of entities. Experimental results show that the DPT method has a significant improvement in the entity prediction task and the relation prediction task.

關鍵字

dynamic translation ； knowledge graph ； multi-step relations ； representation learning