A generative adversarial network for single and multi-hop distributional knowledge base completion

Abstract

Knowledge bases (KBs) inherently lack reasoning ability, limiting their effectiveness for tasks such as question–answering and query expansion. Machine-learning is hence commonly employed for representation learning in order to learn semantic features useful for generalization. Most existing methods utilize discriminative models that require both positive and negative samples to learn a decision boundary. KBs, 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 frequently easily discriminable from positive samples, which can prevent learning of sufficiently robust classifiers.Generative models, however, do not require negative samples to learn the distribution of positive samples; stimulated by recent developments in Generative Adversarial Networks (GANs), we propose a novel framework, Knowledge Completion GANs (KCGANs), for competitively training generative link prediction models against discriminative belief prediction models. KCGAN thus invokes a game between generator-network G and discriminator-network D in which G aims to understand underlying KB structure by learning to perform link prediction while D tries to gain knowledge about the KB by learning predicate/triplet classification. Two key challenges are addressed: 1) Classical GAN architectures’ inability to easily generate samples over discrete entities; 2) the inefficiency of softmax for learning distributions over large sets of entities. As a step toward full first-order logical reasoning we further extend KCGAN to learn multi-hop logical entailment relations between entities by enabling G to compose a multi-hop relational path between entities and D to discriminate between real and fake paths.KCGAN is tested on benchmarks WordNet and FreeBase datasets and evaluated on link prediction and belief prediction tasks using MRR and HIT@ 10, achieving best-in-class performance.