Meaningful relations in a common-sense knowledge graph

Common-sense knowledge is the generally true but defeasible knowledge that most people know and use to reason about the world. We propose an ontology for knowledge graphs to store this information for use in reasoning systems. The relations are based on semantic roles in the style of FrameNet, and are strictly defined to facilitate consistent structuring of information. The different kinds of high-level relationship covered by a given relation are thus narrower than in an existing common-sense knowledge graph, ConceptNet. For example, the ConceptNet relation that an entity is ‘capable of’ some action is either an ‘agent’ or ‘instrument’ relation in the new ontology. We hypothesise that this will aid models’ generalisation ability. We annotate the facts within WorldTree, a science question answering dataset, with this ontology, to demonstrate its feasibility. This dataset then gives labelled facts in the graph in addition to the correct answer for use as supervision.