Inference-Friendly Translation of ResearchCyc to First Order Logic

We are concerned with reasoning efficiently over large-scale commonsense knowledge bases (KBs). In this paper we report on a case study of reasoning with ResearchCyc, a large KB of everyday-world facts represented in a high-order logic. The contributions of this paper are two-fold. First, we present a translation of a large part (over 90%) of the ResearchCyc ontology into First-Order Logic (FOL). Our translation avoids translations that use axiomatic Set Theory as a basis since it is known to be difficult for theorem-proving. We also verify the correctness of the translation empirically. Second, we present an experimental comparison of current theorem provers on the original ResearchCyc KB and the resulting FOL KB. Our experiments show that the higher-order inference procedure of ResearchCyc consistently beats current FOL theorem provers in terms of speed. It is particularly telling that all FOL theorem provers have trouble proving a valid query even after 90% of the irrelevant clauses are removed. This shows that reasoning with high-order logics is a valid alternative to FOL reasoning in large-scale commonsense domains. On the other hand, our results suggest that an axiomatic set theory (inherently difficult for theorem provers) is not needed for most of the translation. Finally, the result of our translation offers a new challenge to the FOL theorem proving community.