Learning Distributed Representations of High-Arity Relational Data with Non-linear Relational Embedding

摘要

We summarize Linear Relational Embedding (LRE), a method which has been recently proposed for generalizing over relational data. We show that LRE can represent any binary relations, but that there are relations of arity greater than 2 that it cannot represent. We then introduce Non-Linear Relational Embedding (NLRE) and show that it can learn any relation. Results of NLRE on the Family Tree Problem show that generalization is much better than the one obtained using backpropagation on the same problem.