Knowledge Graph Contrastive Learning Based on Relation-Symmetrical Structure

Knowledge graph embedding (KGE) aims at learning powerful representations to benefit various artificial intelligence applications. Meanwhile, contrastive has been widely leveraged in as an effective mechanism enhance the discriminative capacity of learned representations. However, complex structures KG make it hard construct appropriate pairs. Only a few attempts have integrated strategies with KGE. But, most them rely on language models ( e.g., Bert) for pair construction instead fully mining information underlying structure, hindering expressive ability. Surprisingly, we find that entities within relational symmetrical structure are usually similar and correlated. To this end, propose knowledge framework based relation-symmetrical KGE-SymCL, which mines KGs ability KGE models. Concretely, plug-and-play approach is proposed by taking positions positive Besides, self-supervised alignment loss designed pull together Experimental results link prediction entity classification datasets demonstrate our KGE-SymCL can be easily adopted performance improvements. Moreover, extensive experiments show model could outperform other state-of-the-art baselines.