Cyc: toward programs with common sense

Reviewer: Fred J. Damerau

In early AI research, attention focused on general problem solving and search methods. When it became clear that intelligent machine behavior was unlikely to be achieved if every action had to be derived from first principles, the field moved toward a “knowledge is power” paradigm, exemplified by most of the work on expert systems and their deep but narrow knowledge bases (KBs). These systems are not suitable for tasks such as text understanding, which require access to considerable commonsense knowledge of the world. The Cyc project at MCC appears to be the first serious attempt to actually build a broad KB (although the issues have been explored before, by the Common Sense Summer project at SRI, for example). The authors give an overview of the Cyc project, its representation language, inferencing system, and classification system. Discussion of the special inferencing rules used provides an introduction to the detailed problems that must be faced in this kind of endeavor. Similarly, discussion of the Cyc ontology illustrates the difficulty of classifying experience, particularly notions of time, persistence, and causality. The authors hope that manual update of the Cyc KB can be replaced in the next few years by feeding text through an associated natural language understanding program, updating Cyc directly from this program's output. This well-written, balanced paper, most of which is accessible to anyone, describes an interesting and potentially important project. Even if Cyc does not achieve its ambitious goals, the attempt alone should greatly benefit the AI community.