BRUCE EDMONDS and VAROL AKMAN

The development of models and techniques that utilize contextrelated ideas have been driven, to a large extent, by the understandable failure of generic approaches. When people act in a certain way or say a particular thing, they do so in a context. As a result, in all the things that they do or say there are embedded background assumptions available only through the context. In the past, the importance of this background has not been sufficiently emphasized – it has been thought that going for the most general model is desirable and that context-dependency is something to be avoided as much as possible. However, recently it has been appreciated that the explicit recognition of context and its use in the manipulation of knowledge can be useful. While recognizing that context is important is surely something in the right direction, this does not immediately give one clues as to how it must be represented. In natural language, meaning is definitely context-bound, but context is not bounded. To quote Derrida (1979, p. 81), “[N]o meaning can be determined out of context, but no context permits saturation” [Note 1]. This discouraging aspect of language meaning gives rise to interesting puzzles in interpretation. Thus, Culler (1983, p. 124) observes that “any attempt to codify context can always be grafted onto the context it sought to describe, yielding a new context which escapes the previous formulation” [Note 2]. Is it possible to take a more practical approach? Can we decline to answer the question “What is in a context?” and still do useful work? At least some people think so. In Artificial Intelligence (AI), context was introduced by McCarthy (1987) specifically with the aim of solving the riddle of generality, viz. finding solutions to assorted