Negative Sampling Method For Translational Model In Knowledge Representation Learning

Knowledge representation learning(KRL)aims to embed entities and relations in knowledge graph(KG)into dense and low-dimensional vector spaces.As a representative model of KRL,TransE and its extensions adopt a pairwise ranking loss function to separate positives and negatives.This optimization goal determines that the training process of these models can not be carried out without negative triples.Therefore,the quality of negative triples plays a significant role in model training.However,to the best of our knowledge,the study on negative sampling method has not drawn much attention yet.Most translation-based models adopt uniformly random negative sampling method by corrupting either the head or tail of the facts with a random entity.Random methods will incur lots of low equality negatives,resulting in slower convergence to the model.Furthermore,existing negative sampling methods ignored the potential noises and community structure within a KG.In order to tackle these problems,the following works are studied:(1)confidence-aware negative sampling method for noisy knowledge graph embedding.The concept of negative triple confidence is first introduced in nosiy knowledge graph.Based on this concept,a negative generating method considering confidence is designed,which is then applied to noise detection in KG and achieved a higher accuracy on standard benchmarks with different ratios of noises.(2)incorporating community structure for negative sampling.A KG is partitioned into different communities,and then triples are corrupted by different strategies based on the entity density in a community.Experiments demonstrate that the proposed strategy is effective.