Contexts: a formalization and some applications

Traditionally, logical formulas (and other structures) used in representation are supposed to be objective, decontextualized truths. In reality, however, they have many contextual aspects. Representations often assume a context of use, the assumptions they make depend on the context in which they were created, etc. The notion of "context" appears in many guises across a wide spectrum of problems in AI. These contextual subtleties are usually ignored or explicitly put aside to keep matters simple. This thesis attempts to remedy this situation by formalizing and developing several applications for contexts. We incorporate contexts as rich objects in a first-order framework, extending the logic (semantics and proof theory) as required. The basic change to first order logic is that formulas are not just true or false, they are true or false in some context. The most common context-related operation is to lift a formula from one context into another. Doing this requires relative (partial) decontextualization. We present a formal definition of this operation and develop a set of general axioms to handle most of the common cases of relative decontextualization. We introduce the notions of entering a context and problem solving within a context. We show that the mechanism of contexts can be used to simplify representation and to integrate theories/data bases that make different assumptions and simplifications. We also briefly discuss the use of contexts for viewing natural language processing as a constraint resolution process. The discussions of the use of contexts employ a number of examples from Cyc, a large multi-domain common sense knowledge base.