Correcting inconsistencies in knowledge graphs with correlated knowledge

Knowledge graphs (KGs) have been widely applied for semantic representation and intelligent decision-making. The usefulness and usability of KGs is often limited by quality of KGs. One common issue is the presence of inconsistent assertions in KGs. Inconsistencies in KGs are often caused by diverse data that are applied for automatically constructing large-scale KGs. To improve quality of KGs, in this paper, we investigate how to detect and correct inconsistent triples in KGs. We first identify entity-related inconsistency, relation-related inconsistency and type-related inconsistency. On the basis, we propose a framework of correcting the identified inconsistencies, which combines candidate generation, link prediction and constraint validation. We evaluate the proposed correction framework in the real-word dataset FB15k (from Freebase). The promising results confirm the capability of our framework in correcting the inconsistencies of knowledge graphs.