Doug, (01)
It is interesting that Cyc also felt the need for microtheories, but
to cut to the chase... (02)
On Tue, Feb 2, 2010 at 2:06 PM, doug foxvog <doug@xxxxxxxxxx> wrote:
> Rob Freeman wrote:
>
>> Think of it as a machine learning problem. I'm talking about learning
>> a whole ontology, theory, even logic, directly from raw patterns in
>> text, then using it to interpret a sentence in context.
>
> This seems like a far more complex problem than ontologizing all of
> "common sense knowledge" (Cyc's goal) by hand. (03)
It does, but computers don't mind doing the same thing over and over. (04)
And if the number of potential theories are infinite, a generation
process is surely the only way to find them all. (05)
Equally, if the theories are contradictory, you will need to do a lot
of processing to select between them at run time anyway. And we don't
even know how to do that. Another nice thing about generating theories
each time is that it automatically provides a means of theory
selection. (06)
>> The machine learning problem has been attempted before. The new thing
>> is that we agree anything you learn must be partial. Extremely partial
>> I believe. Specific to each sentence, or more. That is what makes this
>> machine learning effort different from all earlier efforts. But, given
>> that we've agreed whatever you learn must be partial, what will be the
>> limits on what can be learned?
>
> Learning from pattern analysis requires large data sets. Not learning
> from individual sentences (which I doubt is what you mean by extremely
> partial). To learn partial meanings for words and phrases in different
> contexts, the corpus would have to be divided up into different contexts.
> If this is not done manually, the problem of metaphorical usage of words
> would constantly arise. (07)
You are right. Learning from individual sentences is not what I mean
by "extremely partial". What I mean is that each sentence should
govern what is learned (from the whole corpus.) So a sentence just
selects. The "learning" takes place over the whole corpus. You end up
with a meaning (from the corpus) specific to each sentence. (08)
What is learned are the generalizations in the corpus which are
relevant to a given sentence. This gives you a different theory for
(=appropriate to) each sentence (or even more specific, you can apply
other selection criteria to the learning process if you wish.) (09)
-Rob (010)
_________________________________________________________________
Message Archives: http://ontolog.cim3.net/forum/ontolog-forum/
Config Subscr: http://ontolog.cim3.net/mailman/listinfo/ontolog-forum/
Unsubscribe: mailto:ontolog-forum-leave@xxxxxxxxxxxxxxxx
Shared Files: http://ontolog.cim3.net/file/
Community Wiki: http://ontolog.cim3.net/wiki/
To join: http://ontolog.cim3.net/cgi-bin/wiki.pl?WikiHomePage#nid1J
To Post: mailto:ontolog-forum@xxxxxxxxxxxxxxxx (011)
|