CTransE : Confidence-Based Translation Model for Uncertain Knowledge Graph Embedding

抄録

Knowledge graphs play an important role in many AI applications such as fact checking. Many studies focused on learning representations of a knowledge graph in a low-dimensional continuous vector space. However, most of the recent studies do not learn embedding representations on uncertain knowledge graphs. Uncertain knowledge graphs, e.g., NELL and Knowledge Vault, are valuable because they can automatically populate themselves with new facts. Nevertheless, the automatic process basically induces uncertainty to knowledge. In this study, we introduced knowledge graph embedding on uncertain knowledge graphs by using adapting confidence-margin-based loss function for translation-based models, namely CTransE, to deal with uncertainty on knowledge graphs. The results show that CTransE can robustly learn representations of uncertain knowledge graphs and outperforms the conventional method on knowledge graph completion task.