Generating Contextual Weighted Commonsense Knowledge Graphs

Abstract

There has been a renewed interest in commonsense knowledge and reasoning. To achieve artificial general intelligence, systems must exhibit not only the recognition abilities of humans but also other important aspects of being human, such as commonsense and causality. Recent literature has shown that external commonsense knowledge graphs are beneficial to a variety of systems in multiple ways, including improvements in the commonsense abilities of deep learning models. This paper investigates an auto-generation of weighted commonsense knowledge graphs representing general information, as well as graphs containing contextual information. The method leads to the construction of graphs equipped with frequency based weights associated with nodes and relations. The proposed construction methodology has the advantage of a never-ending learning paradigm. We evaluate the constructed contextual knowledge graphs qualitatively and quantitatively. The commonsense knowledge graphs are inherently explainable and can support commonsense reasoning. We analyze commonsense reasoning approaches using contextual graphs and discuss the results.