Meaningful relations in a common-sense knowledge graph

Abstract

We introduce a new structure for knowledge graphs that store commonsense information. This comprises a set of relations based on semantic roles that aim to be highly generalisable across multiple tasks and domains. • Common sense is the information that most people know and use to understand and reason about the world. It is generally true and widelyapplicable, but is defeasible (can be overruled in the presence of more specific information). Machine learning systems must have access to this to be able to complete reasoning tasks. • One way of storing information is in a knowledge graph. This has the benefit of being an explicit representation, rather than uninterpretable latent information stored in a neural network.