Information Technology Reference
In-Depth Information
Fig. 5.2 Architecture-Sketch
of an ADR System that
Answers TC. Input includes a
S tory and a Q uestion, along
with V isual content. Output
includes an A nswer and—in
the form of a proof or at least
a rigorous argument—a
J ustification
((S,Q),V)
NLP
META-R
(analogical
component)
VIVID
(representation /
re-representation)
Theorem
Provers
ATP
VATP
NLG
(A, J)
harnesses not only our own analogical system (META-R), and not only ATP technol-
ogy, but visual theorem-proving; for formal details see [ 2 ]. This architecture-sketch
is inspired by, but abstracts and extends beyond, AI systems able to solve analogy-
relevant problems. One such inspiring system is by Lovett et al. [ 33 ]; it can solve
items on the Raven's Progressive Matrices. But it cannot for example prove that its
answers are correct, which is part of what the architecture-sketch in Fig. 5.2 demands.
Given the context we have now set, we can articulate a new biconditional:
TCA 4 A computational system A for analogy generation answers the TC if and only if,
given as input no more than either
unstructured textual and/or visual data, or
a vast, pre-existing database not significantly pre-engineered ahead of time by humans
for any particular tests of A ,
is—in keeping with aforementioned Psychometric AI —able to consistently generate analo-
gies that enable A to perform provably well on precisely defined tests of cognitive ability
and skill.
To comment briefly on TCA 4 , first note that we remove the explicit requirement
that the ability to find useful analogies be stable across a variety of input forms and
domains. This is subsumed by the requirement of good performance on precisely
defined tests; it is assumed that a sufficiently difficult psychometric test would provide
questions that are both varied in their form (e.g., word problems, puzzle solving, story
comprehension) and in their domains. The implicit requirement of domain variety
rules out the possibility of an artificial reasoning agent that can only process, for
example, certain types of math problems, as an acceptable answer to the TC.
Some might also see TCA 4 and the use of PAI as too restrictive in that it relies
too heavily on problem-solving and not enough on either creative thinking or the
Search WWH ::




Custom Search