Re: Symbol Grounding Problem (attn: Vend)

On 29 Apr 2007 09:11:51 -0700, Vend <vend82@xxxxxxxxxxx> wrote:

On 26 Apr, 23:04, HMSBeagle <jsb...@xxxxxxxxxxxxx> wrote:
I think many people go through a "stage" in their studies of AI where
they think that the meaning of a word is equal to its defn. So one
has to imagine writing some small program where a computer can read
and understand a story merely by giving it an english dictionary. The
deep assumption made here is that words are defined in terms of words.
And those words are defined in terms of other words and so on and so
on, ad infinitum.

After completing the entire vocabulary of your Story-Reading Program,
you have a network of symbols interconnected by defns; in essence, a
semantic network. When one eventually grows out of this "stage" in
their studies it's undoubtedly because one realizes that this semantic
network is not GROUNDED.

What do you mean by "understanding" ?

I suppose "understanding" will have to be defined by whatever the NLP
researcher requires of the computer. It could be making summaries, or
just translating into a different language.


This is how I initially learned about the Symbol Grounding Problem. I
came into it from the perspective of a person working on Natural
Language Processing (NLP). The metaphor I like to use is that
semantic networks made with dictionary defns are sort of "floating in
the air" disconnected from reality (the ground).

There have been sweeping and earnest research projects trying to get a
computer to read simple stories and give a summary. One of them were
people who claimed that if you give a computer enough common-sense
rules (stated in English) and give it advanced parsing ability, that
it will actually understand simple ***-and-Jane stories. (Anyone
know the name of this project? The PsyC project? I can't
remember...)

I have read doctoral dissertations about attempts to read english text
files by using very advanced parsing techniques that connect pronouns
to earlier words and other trickery.

I have met numerous people in chat rooms who think that they can
eventually get a computer to understand meaning if they have a large
enough "text corpus". One guy from Russia who was convinced that he
could find the meaning of a word by applying bayesian statistics on
large text corpi.

Actually, this is how automatic translation softwares are built. The
results are poor, but they work to some extent.

In theory, you could use bayesian inference (or some other kind of
statistical approach) to learn how to make summaries from an huge
corpora of texts with their summaries.
I'm skeptical about the practical feasibility of that, but it may not
be actually harder than learing to do summarize from the interaction
with the world and people.

However, the Symbol Grounding Problem never goes away. And I have
concluded, after so many years, that NLP is the biggest failure in the
history of AI. It failed because it was based on the false
assumptions about what Meaning is.

Meaning is enivitably linked to interacting with real objects in the
real world, with our real bodies. George Lakoff concurs with this
conclusion. As does Mark Johnson. Meaning is also related to use
of a symbol in a social context. (The social rules are the meaning.)
If you lack the social context, you lack the common meaning among your
peers who you are trying to communicate with, using said symbol.

Well, I could go on and on about this. I will stop here.

You are describing the symbol grounding problem as a practical problem
of natural language processing. This is ok, I agree that to be
succesful in NLP a system has to have some kind of model of the world
(or at least the domain of discussion) and of people psycology.

Anyway, I think that the link you provided talks about something
different (but I can't tell for sure since I find it a bit confusing).

It talks about a philosophical problem about "meaning",
"consciousness" and their relevance in computationalism (in fact, it
mentions the Turing Test and the Chinese Room). I think that
philosophical questions about StrongAI and Artificial Consciousness
are separate from issues of technical realization and attempts to
conflate them introduce confusion.


An interesting point for sure. One keeps seeing this same point made
about all philosophy in general. Sometimes I think all philosophy is
built around the confusion over the defn of words..but that would be
digressing. :)


I think that it's possible, in principle, to create human-level
intelligence with a computer. This derives from the assuption that all
the laws of physics are simulable in a computer to any arbitrary
degree of accuracy and that an human brain operates according to them,
thus it is simulable on a computer. This address the philosophical
questions AI, but by no means implies that we are able to create human-
level AI with current or near-future technology.


I agree.
And if this is what computationalism means, then I guess I am a
computationalist.


This is equivalent to claiming that it's theoretically possible to
build a living human by synthetizing all the biomolecules from raw
elements and assembling them in an appropriate configuration, but in
practice we can't do that.


.

Quantcast