Previous: Embodiment, Symbol Grounding, and Natural
Language Semantics
Up: Wider Significance
Next: Some Pseudo-Philosophical Notes
The work presented in this
dissertation is relevant to the field of knowledge representation and
reasoning (KRR), although it is not what is typically understood by that
term. KRR involves a formal representation language with a set of inference rules
that operate on constructs of the language [B. Smith 1985]. Most KR
languages can be translated relatively easily into predicate logic of some
kind,
so I will assume a predicate-logic-like language
for the purpose of this discussion. The semantics of KR languages is
typically (if at all) defined in terms of Tarskian semantics or derivatives
thereof, with an interpretation function that maps terms and predicates
of the language (the syntactic domain) into objects or sets of objects (or
properties) in the domain of interpretation (the semantic domain), whether
that domain is taken to be (part of) the real world, a possible world, or
intensional objects in an agent's mind [Shapiro \& Rapaport 1987][Manin 1977]. I
see two problems with the practice of KRR as I have just outlined it (see also
critics of traditional AI, e.g., [Harnad 1990], [Lakoff 1987],
[Searle 1980], or recently [Angell 1993], to name just
a few). First, the semantic models of KRR systems are purely hypothetical,
in that they rarely, if ever, enter into the workings of an implemented KRR
system. Second, the semantic models used in model theory for logical
systems do not fit ``natural'' or perceptual categories at all.