Title: Commonsense Knowledge Relation Classification

Abstract: We present an analysis of different approaches to relation classification applied to ConceptNet, a commonsense knowledge repository. We approach this as a multi-way classification problem and assess performance for each relation type. ConceptNet relations have particular characteristics – they often have multi-word arguments, are tightly connected and they have been crowd-sourced, so the resource may contain noise or distributional biases. We explore underlying factors of the difficulty of common sense relation classification based on ConceptNet. Our experimental settings and models are chosen to highlight potential influencing factors. Results and data analysis show that (i) ConceptNet relations can be classified with high performance with models that are well adapted to their specific –possibly multi-word– argument properties; (ii) a considerable amount of misclassifications seems due to relation ambiguity; (iii) the current structure of ConceptNet is not well suited for relation learning with graph-based models.