"Reducing" CLASSIC to Practice: Knowledge Representation Theory Meets Reality

Most recent key developments in research on knowledge representation (KR) have been of the more theoretical sort, involving worstcase complexity results, solutions to technical challenge problems, etc. While some of this work has influenced practice in Artificial Intelligence, it is rarely—if ever—made clear what is compromised when the transition is made from relatively abstract theory to the real world. classic is a description logic with an ancestry of extensive theoretical work (tracing back over twenty years to kl-one), and several novel contributions to KR theory. Basic research on classic paved the way for an implementation that has been used significantly in practice, including by users not versed in KR theory. In moving from a pure logic to a practical tool, many compromises and changes of perspective were necessary. We report on this transition and articulate some of the profound influences practice can have on relatively idealistic theoretical work. We have found that classic has been quite useful in practice, yet still strongly retains most of its original spirit, but much of our thinking and many details had to change along the way. FORWARD February 2009 The practical instantiation of theories is an important, but often neglected, component of knowledge representation. The work on classic, described in this paper, was undertaken at AT&T Bell Labs around 1990, and fits into this component, taking the formal theories of description logic and crafting a useful system for representing information. The state of the art in description logics has changed considerably since classic was first envisioned in the late 1980s. Advances in algorithms and computers have made for effective reasoners for theoretically intractable description ? A slightly different version published in Artificial Intelligence, 114, October 1999, pages 203–237. ?? Supported in part by the AT&T Foundation and NSF IRI 9119310 & 9619979. logics, leading up to modern reasoners for description logics like shoiq, where reasoning is NExpTime complete. These advances in reasoning have led to the adaptation of description logics to the Semantic Web. The changes needed were theoretically quite small, amounting mostly to using IRIs as names and adding the datatypes of XML Schema. This lead, through a number of intermediate steps, to the W3C Web Ontology Language OWL. However, quite a bit of practical adaptation was needed as well, including allowing non-logical decorations, adding a method of combining documents on the web into one knowledge base, building user interface systems like Protégé (http://www.co-ode.org/downloads/protege-x/), and providing APIs for popular programming languages. There are a now multiple reasoning systems for OWL, including Pellet (http://clarkparsia.com/pellet/), Racer (http://www.racersystems.com/), and FaCT++ (http://code.google.com/p/factplusplus/), which are based on this combination of theoretical underpinnings and attention to practical details, now prevalent in the description logic community.