[Top] [All Lists]

Re: [ontolog-forum] Foundations for Ontology

To: "[ontolog-forum]" <ontolog-forum@xxxxxxxxxxxxxxxx>
From: "John F. Sowa" <sowa@xxxxxxxxxxx>
Date: Thu, 29 Sep 2011 13:38:09 -0400
Message-id: <4E84AD01.6020708@xxxxxxxxxxx>
On 9/29/2011 4:00 AM, Rob Freeman wrote:
> A vector compositional model of this kind is exactly what I've been
> promoting on this list and elsewhere for some years now.
> Just recently I found the papers Daoud Clarke cites by Mitchell and
> Lapata, 2008, Stephen Clark, and more recently Grefenstette. I was
> delighted to see that a small community seems to have sprung up
> pursuing the idea.    (01)

Vector composition is computationally useful.  But you have to look
at the *meaning* of what is being composed, the *meaning* of the
compositional operators, and the *meaning* of the results.    (02)

As I said in my previous note, any meaning representation must be
translatable to ordinary language, as a word, phrase, or sentence.
Logic-based representations (including conceptual graphs) meet
those requirements, but statistical vectors don't.  We use vectors
in VivoMind, but our vectors can be translated back to CGs that
are logically equivalent to the ones from which they were derived.    (03)

In the ontofound.pdf slides, I surveyed studies from psycholinguistics
and neuroscience for evidence of how people process language, and I
compared those results to computational methods.  One thing is very
clear:  people *understand* language.  They relate the words and
sentences to mental models that have very accurate mappings to
perception and action.    (04)

Even our cousins the apes have very accurate mental models.  They spend
a lifetime swinging from branch to branch through the forest. Any lapse
in their hand-eye coordination would eliminate them from the gene pool.    (05)

Another example:  Consider an orchestra performing a symphony.
Each member controls an instrument while viewing a musical score,
relating it to the sounds of neighboring instruments, and coordinating
it with the gestures of the conductor.  Each individual has extremely
precise mental models of multiple sensory inputs:  tactile feel of
the instrument, auditory input from multiple instruments, visual input
from the musical score, other instruments in the orchestra, and the
gestures of the conductor.  The results are extremely precise.    (06)

Similar examples of precision come from every skilled activity:
sewing, carpentry, surgery, arts of all kinds, sports, gymnastics,
dancing, driving a car, flying a plane, climbing a mountain...    (07)

In every one of those activities, experts in the field communicate
precisely with other experts.  They don't depend on any kind of
probabilistic mush.  What they say is precise, and they can correct
any mistaken interpretations to an arbitrarily precise level of
accuracy.    (08)

In my ontofound.pdf slides, I surveyed the early hopes and limited
successes of logic-based methods.  But I want to emphasize that I
have much more sympathy for the logic-based methods than I have for
most of the currently popular statistical methods in computational 
linguistics.  That was the basis for my negative remarks about
Clarke's paper.    (09)

The statistical people talk about "lexical entailment" for question
answering in which they consider 70% accuracy good.  Our VivoMind
company, competed for a contract with a dozen other groups, which
included universities and multinational companies. The test involved
processing a number of documents, extracting information from them,
and organizing the results in tables.  The scores were the percentage
of items extracted from the documents and correctly placed in tables.    (010)

For that test, the VivoMind score was 96%.  Second place was 73%.
Third place was 53%, and all the rest were below 50%.  We had a
runoff against the second-place group on documents from a very
different domain, and we also scored better on that test.    (011)

We do use statistics, but our meaning representation is conceptual
graphs.  The techniques are along the lines of the examples in
Section 6 of the ontofound.pdf slides and the articles cited in
the final slide.    (012)

Fundamental principle:  Statistics are valuable for many purposes,
but *not* as a meaning representation.    (013)

John    (014)

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    (015)

<Prev in Thread] Current Thread [Next in Thread>