View article

Knowledge graphs (KGs) inherently lack reasoning ability which limits their effectiveness for tasks such as question–answering and query expansion. KG embedding (KGE) is a predominant approach where proximity between relations and entities in the embedding space is used for reasoning over KGs. Most existing KGE approaches use structural information of triplets and disregard contextual information which could be crucial to learning long-term relations between entities. Moreover, KGE approaches mostly use discriminative models which require both positive and negative samples to learn a decision boundary. KGs, by contrast, contain only positive samples, necessitating that negative samples are generated by replacing the head/tail of predicates with randomly chosen entities. They are thus usually irrational and easily discriminable from positive samples, which can prevent the learning of sufficiently robust …