Fast Confidence Prediction of Uncertainty based on Knowledge Graph Embedding

Abstract

The uncertainty is an inherent feature of Knowledge Graph (KG), which is often modelled as confidence scores of relation facts. Although Knowledge Graph Embedding (KGE) has been a great success recently, it is still a big challenge to predict confidence of unseen facts in KG in the continuous vector space. There are several reasons for this situation. First, the current KGE is often concerned with the deterministic knowledge, in which unseen facts’ confidence are treated as zero, otherwise as one. Second, in the embedding space, uncertainty features are not well preserved. Third, approximate reasoning in embedding spaces is often unexplainable and not intuitive. Furthermore, the time and space cost of obtaining embedding spaces with uncertainty preserved are always very high. To address these issues, considering Uncertain Knowledge Graph (UKG), we propose a fast and effective embedding method, UKGsE, in which approximate reasoning and calculation can be quickly performed after generating an Uncertain Knowledge Graph Embedding (UKGE) space in a high speed and reasonable accuracy. The idea is that treating relation facts as short sentences and pre-handling are benefit to the learning and training confidence scores of them. The experiment shows that the method is suitable for the downstream task, confidence prediction of relation facts, whether they are seen in UKG or not. It achieves the best tradeoff between efficiency and accuracy of predicting uncertain confidence of knowledge. Further, we found that the model outperforms state-of-the-art uncertain link prediction baselines on CN15k dataset.