COVERAGE AND COMPETENCY

A central problem in artificial intelligence (AI) is the development of large-scale formalizations of commonsense knowledge (McCarthy, 1959). Progress toward this goal has been slow because (1) sufficiently powerful and expressive formal languages need to be developed, (2) the millions of facts that people know and use must be articulated, (3) these facts need to be encoded in a formal language, and (4) reasoning systems that efficiently use this knowledge need to be developed (Davis and Morgenstern, 2004). Put another way, the challenges of formalizing commonsense knowledge are barriers to both breadth and depth. In breadth, researchers are aiming for formalizations of commonsense knowledge that have broad coverage over the concepts in the domains in which commonsense reasoning problems exist. In depth, researchers require that these formalisms have the competency to solve commonsense problems in an automated manner.

Achieving both coverage and competency in the formalization of commonsense knowledge has been an elusive prize. The tradition in this field has been to focus on depth (competency) first, and to leave the problem of breadth (coverage) as somebody else's problem. Today, the field could best be described as a system of interactions between a community of authors and a handful of aggregators. In the general case, the set of authors is made up of academic researchers who simultaneously work on (1) a formal language that is slightly more expressive or powerful than previously published languages, (2) capturing a small handful of facts that people know and reason with, and (3) encoding these facts into the formal language that they have developed. These researchers publish their results in commonsense reasoning symposiums, knowledge representation conferences, and selected academic journals. The aggregators take the best products of these efforts and integrate them into large-scale knowledge bases, typically by rewriting them into their own formal languages. An important example of an aggregator is Cycorp Inc., which maintains the CYC knowledge base (Lenat, 1995), a confederation of thousands of micro theories authored by Cycorp knowledge engineers who base their work, when possible, on the best theories from the commonsense reasoning research community.