Re: Interoperability and Vagueness
On Monday 17 January 2005 09:20, John F. Sowa wrote:
> Rob,
>
> I believe that theories about emergent syntax and
> related topics, such as your discussions of the ad-hoc
> nature of syntax, are on the right track. But there
> is an enormous amount of work necessary to fill in
> the details about how that ad-hoc syntax relates to
> whatever it is that happens when people "understand"
> a sentence, paragraph, or any discourse of any kind.
>
> ...
>
> I would first ask how your theory or anything similar
> would handle the "Grand Challenge" proposal that I
> submitted to DARPA, along with a couple of colleagues:
>
> http://www.jfsowa.com/ai/gcprop.pdf
>
> Following is the short summary of the problem:
>
> The task we suggested is one that nearly every two-year-old
> child solves: the problem of learning to integrate visual,
> tactile, and motor information with language. To evaluate
> progress on this task, we proposed that any AI research
> group that wished to respond to the challenge be given a
> collection of binocular pictures, still or moving, together
> with some natural-language questions about those pictures.
> Any AI system they develop would be asked to determine which
> pictures could answer any of the questions and to state those
> answers.
You're right, that's a big challenge.
The cognition of images is a problem in its own right. While our model still
can't drive a car down a road, it can be expected to say little of interest
about the road.
Personally I think the underlying solution is the same, ad-hoc
generalizations. But as you say, such ideas would have to be elaborated and
tested. Even the language bit which primarily interests me needs to be
elaborated and tested, though I am much more sure of e.g. the parameters of
generalization for language.
What bothers me is not that people are still unconvinced, but that I see no
work along these lines at all (or very little, and uncoordinated.)
I'll take the emergent handle, but it is dangerous. Most people don't know
what emergent structure is. If they think about it at all they confuse it
with evolving structure. Even of those working on emergent structure, they
largely ignore the possibility of instantaneous reinterpretation, the
subjective classes which I see as central. There is some work on distributed
structure, vector information retrieval, memory-based reasoning, some of
which can be seen as emergent, and with reference to which emergence is often
dismissed. But all of that work is trapped within a global classification
paradigm. Neural networks train to global classes, HMM's find global classes,
MBR, KNN select global classes. Really, who is working on subjective classes
in language (and cognition)?
Adam Kilgarriff rejected word senses, but the driving force in his research
domain remained, and remains SENSEVAL, a competition to classify senses. I
can't avoid the conclusion that the field remains focused on the
classification challenge because that is the challenge the field is focused
on, not because it continues to make "sense" to the researchers.
Similarly it seems clear to me that the majority of subscribers to this list
will not change their goals from a catalog to a search engine, even if they
accept the number of enumerable senses is infinite. Presumably because they
are paid, or trained, to make catalogs.
Some complexity is unavoidable, but I think there is great value in painting
the broad strokes large. It doesn't require much maths to say "emergent
structure". I prefer "ad-hoc" as a stronger statement emphasizing that the
emergent classes are subjective. If your model does not assume ad-hoc
meaning, then it is useful to say that too.
Let's at least make the distinction, so people know _what_ they are
"elaborating and testing".
So in the spirit of your "Grand Challenge", let me make a first step "petit"
challenge to the members of SUO (DARPA can fund it if they wish :-) Who here
is working on meaning as a process for finding ad-hoc classes, rather than as
an enumerable set of classes, and if not, why not?
-Rob