Computing and Informatics

Inference on Knowledge Bases (KBs) is an important way to construct more complete KBs and answer KB questions. Inference can be viewed as a process from evidence to conclusion following specific formulas. Traditional methods usually search on the KB to collect evidence, which cannot apply to large-scale KBs, because the running time of searching increases radically as the scale of KBs increases. What is worse, evidence cannot be found if one fact in it is missing, which may result in the failure of inference. To this end, we propose a fuzzy method of estimating evidence, which replaces searching by estimating the existence of evidence by constructing formula embeddings, and then we merge these estimations into a probabilistic model to infer conclusions. This method can apply to large-scale KBs, because estimating evidence is very fast and is irrelevant to the KB scale. Estimating evidence can also be viewed as fuzzy matching, so this method can handle the situation where facts are missing. We evaluate this method on the knowledge base completion task, and it achieves a better performance than state-of-the-art methods and has a shorter running time.