Uncertain Knowledge Graph Embedding: a Natural and Effective Approach

Uncertain Knowledge Graph (UKG) models uncertainty of the knowledge, which usually models the inherent uncertainty of relation facts in a knowledge base with a confidence value. Embedding such uncertain knowledge represents is still an unresolved challenge, although deterministic Knowledge Graph embedding has been extensively studied recently, which aims at representing entities and relations as vectors in a continuous vector space and preserving semantic information as much as possible. For capturing both structural and uncertainty information of relation facts in the continuous vector space, we propose a simple but effective two-steps approach to Uncertain Knowledge Graph Embedding (UKGE) based on the skip-gram/CBOW model and learning confidence score by Long Short-Term Memories (LSTM) neural network. We show that the embedding techniques achieve results comparing with state-of-the-art approaches on uncertain knowledge graph embedding. The approach achieves the best tradeoff between efficiency and accuracy of UKGE. Because of the simplicity, the method can also handle large size graphs in lower time consumption.