Modeling commonsense rules in an inference network

Abstract

In commonsense reasoning, conditional statements of the form “IF condition(s) THEN conclusion(s)” are the most common and important constructions. While material implication is generally used in classical logic based belief representation systems, its dual implication could be semantically too strong for expressing commonsense IF-THEN rules because not all contributing conditions of a rule can be expressed (the Qualification Problem [18]) and the negation of conclusions do not always imply the negation of the conditions. This paper studies a hybrid neural-symbolic belief representation system called Neural-Logic Belief Network (NLBN) [14] where IF-THEN rules can be more realistically captured for commonsense reasoning. Deduction of an IF-THEN rule in this formalism is considered as information flow from the condition(s) to the conclusion(s). In this system, the strength of conclusions can be modeled by using individual rule mapping functions.