Uncertain Knowledge Graph Embedding Using Probabilistic Logic Neural Networks

Knowledge graph (KG) embedding embeds components of a KG into low-dimensional continuous vector spaces, while preserving the inherent structure of the KG. Existing embedding methods (e.g. TransE, DistMult) have two main drawbacks: (1) They fail to leverage the underlying domain knowledge, which can be captured by the symbolic rule-based approach with first-order logic; and (2) they are mostly designed for deterministic KGs and thereby fail to model the inherent uncertainty present in real-world KGs. In this paper, we propose the uncertain probabilistic logic neural network (PKGE) which captures domain knowledge and uncertainty information, preserves entity-relation semantic information, preserves graph structure, and can be trained effectively. To achieve this, PKGE employs the Markov Logic Network (MLN) to learn firstorder logic and encodes uncertainty by leaning confidence scores using the novel Uncertain KG Embedding (UKGE) model. We conduct optimization using the variational EM algorithm.