Scholar's discussion on HIP as an alternative to AI/SW research
4/5/2004
2:35 PM
Comment from an AI scholar:
There are a several points that should be crystal clear:
1. RDF is arbitrarily expressive.
It allows you to encode ternary relations, and these are arbitrarily
expressive. [18]
2. Reduced expressivity is a "design feature" of description
logics and first-order predicate calculus people intended to buy better
computational and proof-theoretic properties.
3. Turing machines are universal. They can perform any computation, provided we don't concern ourselves with how long it takes or whether it terminates. [19]
Any discussions therefore should center on how we represent our problem and support efficient algorithms (space & time computational complexity).
Critiques of AI in general are meaningless because of points 1 and 3. [18] [19]
Critiques of the semantic web and specific AI approaches must center on point 2, ie can we adequately represent the problem? Can we compute interesting results with reasonable resources? [20]
At first approximation, the AI and NL communities have no idea about informal knowledge (e.g, the latter Wittgenstein), including how to represent it, and how to work with it. This leads to well-known problems:
* Brittleness in AI systems when concepts are ungrounded in the
phenomena from which they were derived;
* Pre-paradigmatic weakness (or just a void) in semantic perception, reference, and representation in NL systems.
I'm not sure how many funding decisions for research in these areas are informed by an awareness of these basic issues.