Knowledge Base Error Detection with Relation Sensitive Embedding

Abstract

Recently, knowledge bases (KBs) have become more and more essential and helpful data source for various applications and researches. Although modern KBs have included thousands of millions of facts, they still suffer from incompleteness compared with the total amount of facts in real world. Furthermore, a lot of inaccurate and outdated facts may be contained in the KBs. Although there have been many studies dealing with incompleteness of the KBs, very few of works have taken into account detecting the errors in the KBs. Broadly speaking, there are three main challenges in detecting errors in the KBs. (1) Symbolic and logical form of the knowledge representations cannot detect the inconsistencies very well on large scale KBs. (2) It is hard to capture the correlations between relations. (3) There is no golden standard to learn or observe the patterns of inaccurate facts. In this work, we propose a Relation Sensitive Embedding Approach (RSEA) to detect the inconsistencies from KBs. We first design two correlation functions to measure the relatedness between two relations. Then, a dynamic cluster algorithm is presented to aggregate highly correlated relations into the same clusters. Finally, we encode discrete knowledge facts with effects of correlated relations into continuous vector space, which can effectively detect errors in KBs. We perform extensive experiments on two benchmark datasets, and the results show that our approach achieves high performance in detecting incorrect knowledge facts in these KBs.