Common Sense Reasoning for Knowledge Integration

Abstract

Recently, the field of Artificial Intelligence has made significant advances in building deep models of complex tasks. One area, however, that seems to have lagged behind is the domain of "common sense reasoning." Much common sense reasoning research has defined the term in a way that requires deep models of everyday phenomena, making progress difficult. Common sense reasoning does not in itself need to be a deep process. It can be limited to a single task - integrating the knowledge provided by other models and information sources, focusing only on the dominant conditions identified by each model. To remedy these shortcomings, we propose an approach based on an architecture that we refer to as CS+CM, for "Common Sense Plus Constituent Models". In a CS+CM architecture, the common sense (CS) layer plays an integration role that only requires limited inferential capabilities. In addition, we propose concrete limits to representational and inferential capabilities, and identify the structural requirements for deeper models to support integration at the common sense layer. We illustrate this approach with examples in diverse areas of capital investment, team selection, and behavioral finance.