Greg Detre
14:30 late Tuesday, September 24,
2002
is reading
the opposite process of producing?
Wordnet is
the biggest relational lexical database
one of the most used tools in computational linguistics
70,000 synsets (nodes), each with a handful of relations
other
knowledge representation schemes which inherit from Quillian:
frames (slots and variable ranges)
scripts � more episodic
description logics (formal logic description of a node)
spreading activation networks (numerical weighted links)
can you see all models as just nodes + links?
book:
�Understanding language understanding� � contains a set of criteria for
evaluating a knowledge representation model
how do you
define symbol? does Quillian combine symbols? is a plane a new combined symbol,
or just an organisational device?
how would you put a word like �have� into
Quillian�s ontology? just the same way he includes �put�, I suppose�
what is a
relation?
which sort of relations do we want to be basic?
what sort of ontological commitments are you
making if you accept �and� and �or� as basic?
they�re kind of set-theory-like � they�re
discrete � individuable � despite his explicit assumptions, his system doesn�t
really do a good job of blurred, vague concepts
when don�t they feel natural? common-sense
reasoning, exceptions, context
do you believe in sets of discrete entities?
deb roy is all for softening and going
stochastic and spreading bets(???)
what�s the probability that there are 1, or 2,
or even 1.5 objects???
why don�t we all use self-organising/dynamic
systems?
create your own relations? doesn�t force you to
enumerate an exhaustive list
might or might not be discrete
you�re making little implicit assumptions in
your encoding, your architecture etc.
recursive � can you define �is a� in terms of
�is a�?
if your list of relations is really long +
arbitrary�?
you lose computational tractability
well??? how important is that if we�re not
interested in implementing it on a machine, but just in the science
computational tractability (i.e. turning into
an algorithm) as a way of abstracting/generalising/understanding scientifically
computational tractability is also about our
understanding what�s going on
doesn�t a NN break down the distinction between
algorithm + data structure???
what
happens when they update/add a synset in Wordnet? how do they maintain it, keep
it consistent? how do you achieve concept persistence over time?
NNs aren�t
symbol-systems � they aren�t semantically interpretable at different points
you can
connect things up to the real world in the wrong way too though, can�t
you???
he�s making
the assumption that we are a symbol system, the real work gets done by the
symbol system � he just wants a way to bootstrap into that system, i.e.
attaching his symbol-labels to the real world
questions
to ask: why is the chinese/chinese dictionary impossible???
we don�t
like his breakdown of human behaviour
what do we
think of the distinction between discrimination + identification? deb doesn�t
like treating them as separate behavioural abilities
which is higher-level? discrimination seems pretty low-level
at the just-noticeable-difference / finest-granularity, identification +
discrimination are the same?
but what do
you then do with the iconic representations???
you can do this in an image-processing kind of
way
cf Shepard, rotated shapes, recognition time is
proportional to angle of rotation
there is a continuum of processing from the
intiial unprocessed sensory representation through to the categorical
representation � how much processing constitutes the boundary that is iconic
representation�
are iconic representations 3D???
he says
Harnad has saved symbolic systems
even when you�re combining horse + stripe
symbols to get zebra � but you can do that in analogue just by projecting one
image onto another
you do need symbols to communicate the idea
it�s all very imagistic� but you can kind of
imagine it working for sound, or multi-modal stuff
does the grounding actually affect the
operation of the symbols? does it actually change your reasoning?
what more
is symbol grounding than isomorphism??
is wordnet bigger than cyc???
cyc contains 2m structures, but many of them are empty, and need cleaning up
also, the consistency and depth varies domain-by-domain
frames??? scripts??? description logics??? spreading activation networks???
frames vs scripts
accidental vs non-accidental edges (in visual perception)???