PTrustE: A high-accuracy knowledge graph noise detection method based on path trustworthiness and triple embedding

Abstract

Many knowledge graphs (KGs) constructed automatically or manually contain noises inevitably when adding heterogeneous data nowadays. The existing methods either have difficulty in detecting noisy triples with conflicted relations or ignore to represent triples after noise elimination, which creates obstacles for the downstream tasks to manipulate KGs. To detect various types of noises in KGs and constitute noise-free triples representation, this paper proposes a high-accuracy KG noise detection method based on path trustworthiness and triple embedding (PTrustE). First, PTrustE constructs a correlation-based path trustworthiness network to learn the global and local features in the path from the head entity to the tail entity of the triple. Next, PTrustE integrates all features of the path into the Bi-directional Gated Recurrent Unit to learn the path score matrix and path trustworthiness, to keep the sequential nature of paths. Finally, PTrustE uses the path score matrix for triple representation learning and the path trustworthiness for judging whether the triple is correct or not. Extensive experiments validate the superior performance of PTrustE. Compared with the second-placed INDIGO, PTrustE offers a 7.1% increase in fMRR on FB15K.