KBGAN: Adversarial Learning for Knowledge Graph Embeddings

Proceedings of NAACL-HLT 2018, pages 1470–1480 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics KBGAN: Adversarial Learning for Knowledge Graph Embeddings Liwei Cai Department of Electronic Engineering Tsinghua University Beijing 100084 China cai.lw123@gmail.com William Yang Wang Department of Computer Science University of California, Santa Barbara Santa Barbara, CA 93106 USA william@cs.ucsb.edu Abstract We introduce KBGAN, an adversarial learning framework to improve the performances of a wide range of existing knowledge graph em- bedding models. Because knowledge graphs typically only contain positive facts, sampling useful negative training examples is a non- trivial task. Replacing the head or tail entity of a fact with a uniformly randomly selected entity is a conventional method for generat- ing negative facts, but the majority of the gen- erated negative facts can be easily discrimi- nated from positive facts, and will contribute little towards the training. Inspired by genera- tive adversarial networks (GANs), we use one knowledge graph embedding model as a neg- ative sample generator to assist the training of our desired model, which acts as the dis- criminator in GANs. This framework is inde- pendent of the concrete form of generator and discriminator, and therefore can utilize a wide variety of knowledge graph embedding mod- els as its building blocks. In experiments, we adversarially train two translation-based mod- els, TRANSE and TRANSD, each with assis- tance from one of the two probability-based models, DISTMULT and COMPLEX. We eval- uate the performances of KBGAN on the link prediction task, using three knowledge base completion datasets: FB15k-237, WN18 and WN18RR. Experimental results show that ad- versarial training substantially improves the performances of target embedding models un- der various settings. 1 Introduction Knowledge graph (Dong et al., 2014) is a pow- erful graph structure that can provide direct ac- cess of knowledge to users via various applica- tions such as structured search, question answer- ing, and intelligent virtual assistant. A common representation of knowledge graph beliefs is in the form of a discrete relational triple such as Locate- dIn(NewOrleans,Louisiana). A main challenge for using discrete represen- tation of knowledge graph is the lack of capa- bility of accessing the similarities among differ- ent entities and relations. Knowledge graph em- bedding (KGE) techniques (e.g., RESCAL (Nickel et al., 2011), TRANSE (Bordes et al., 2013), DIST- MULT (Yang et al., 2015), and COMPLEX (Trouil- lon et al., 2016)) have been proposed in recent years to deal with the issue. The main idea is to represent the entities and relations in a vec- tor space, and one can use machine learning tech- nique to learn the continuous representation of the knowledge graph in the latent space. However, even steady progress has been made in developing novel algorithms for knowledge graph embedding, there is still a common chal- lenge in this line of research. For space effi- ciency, common knowledge graphs such as Free- base (Bollacker et al., 2008), Yago (Suchanek et al., 2007), and NELL (Mitchell et al., 2015) by default only stores beliefs, rather …

Content truncated to 500 words