Alleviating Sparsity of Open Knowledge Graphs with Ternary Contrastive Learning

Anonymous

Abstract: Sparsity of formal knowledge and roughness of non-ontological construction methods make sparsity problem particularly prominent in Open Knowledge Graphs (OpenKGs). Sparse links make few-shot entities unable to learn potential features. We hypothesize that negative samples could help sparse links highlight discriminative features. However, existing contrastive learning in Graphs model binary objects, none has studied contrastive learning to model ternary pattern in any KGs. In this paper, we propose a Ternary Contrastive Learning (TernaryCL) to alleviate the sparsity of OpenKGs. TernaryCL designs (1) Contrastive Entity and (2) Contrastive Relation to mine ternary discriminative features by both negative entities and relations. (3) Contrastive Self constructs a self positive sample to give zero-shot and few-shot entities chances to learn discriminative features. (4) Contrastive Fusion aggregates graph features by extending the pattern from 1-to-1 to 1-to-N. Extensive experiments on benchmarks show the superiority of TernaryCL over state-of-the-art models.

Paper Type: long