Contrastive knowledge embedding with discriminative self-weighted sampling.

Knowledge Graph (KG) embedding is to map components of a KG including entities and relations into a continuous low-dimensional space, with the goal of simplifying subsequent manipulation while preserving the inherent structure of the KG. Recent research efforts are directed towards the design of various types of scoring functions, yet the learning framework, which is essential for further improvement of KG embedding models, has received less attention. In this paper, we exploit Contrastive Learning (CL) for KG embedding, given its power in representation learning, which helps enhance the expressiveness of KG embedding models. However, traditional CL techniques uniformly sample negatives from the entire dataset, which can lead to inefficiency in KG embedding due to the existence of low-quality triplets. To alleviate this problem, we devise a flexible CL framework termed "Co‾ntrastive knowledge embedding with Di‾scriminative S‾elf-weighted S‾ampling" (CoDiSS) for KG embedding. Unlike conventional hard negative sampling techniques, which simply assume harder negatives are more informative, our CoDiSS employs an adaptive weighting mechanism that assigns importance to all negative triplets according to their corresponding contributions to model learning. Furthermore, we devise a Discriminative Weight Refinement (DWR) loss that reshapes the score distribution of negatives to enlarge the separation between informative and false negatives. As a result, the proposed CoDiSS encourages learning from the most informative negative triplets while suppressing the adverse effects of false negatives. Experimental results across multiple datasets have demonstrated that the proposed CoDiSS framework can improve the performance of various KG embedding models, such as TransE, ComplEx, and HousE, enabling them to produce expressive KG embeddings.