Adaptive Entity Alignment for Cross-Lingual Knowledge Graph

Abstract

Entity alignment is a key step in knowledge graph (KG) fusion, which aims to match the same entity from different KGs. Currently, embedding-based entity alignment is the mainstream. It embeds entities into low-dimensional vectors and transfers entities to the same vector space. However, the precision of entity alignment depends largely on the number of alignment seeds. Manually labelling alignment seeds is very expensive and inefficient. To address this problem, this paper proposes a novel adaptive entity alignment approach for cross-lingual KGs, namely AdaptiveEA. This approach adopts a new joint KG embedding network to accurately learn the embeddings of entities from both structural semantics and relational semantics. Then, it employs an adaptive entity alignment strategy to iteratively capture new aligned entity pairs that meet a dynamic alignment threshold. Then the new captured aligned entity pairs are utilized to expand alignment seeds to further guide the next embedding training. Experimental results on public datasets show that the proposed approach can achieve much better precision than the latest state-of-the-art models in a small number of labeled alignment seeds.