Entity alignment in noisy knowledge graph

Abstract

Entity alignment is an important task in Knowledge Graph(KG), which aims to find identical entities in two different KGs. Existing methods include two steps, graph representation and alignment inference. The representation is learned based on the semantics and structure of KG. In applications, however, incorrect triples (which are also called structure noise) inevitably exist in KGs due to low-quality corpora and low-performance construction algorithms. The structure noise in KGs affects the representation of KGs and the alignment inference. To this end, we propose an entity alignment method in noisy knowledge graphs for the first time. Firstly, a noise-aware module is designed to recognize the noisy triples and exclude them from KG representation. Secondly, we design a more strict semi-supervised algorithm that combines local similarity and global alignment cost together to obtain high-quality pseudo-alignments in noisy environments. The experimental results demonstrate the effectiveness of our method in noisy KGs and the good compatibility with other baselines.