[Top] [All Lists]

Re: [ontolog-forum] What is "understanding"

To: "[ontolog-forum] " <ontolog-forum@xxxxxxxxxxxxxxxx>
From: Sergei Nirenburg <sergei@xxxxxxxx>
Date: Sat, 5 Apr 2008 13:13:28 -0400
Message-id: <4EDE56C7-4A4B-4D77-A56E-B68FEB378978@xxxxxxxx>
On Apr 4, 2008, at 10:10 PM, John F. Sowa wrote:
> In their paper, Cimiano and Reyle criticized Nirenburg and Raskin for
> not having a "foundational ontology" or what many people call an
> "upper ontology":
>     In line with Nirenburg and Raskin's Ontological Semantics  
> framework,
>     we thus adhere to the basic tenet that natural language semantics
>     needs to be captured with respect to an explicitly formalized
>     ontology. Further, we argue for a novel direction in computational
>     semantics, i.e. what we will call foundational semantics.
>     Foundational semantics differs from ontological semantics in that
>     it is concerned with identifying that abstract meaning layer which
>     remains constant across domains and applications. In this respect
>     our approach differs crucially from the ontological semantics
>     framework of Nirenburg and Raskin, who are not concerned with
>     domain-independent aspects of meaning.
> My major objection is to people who use words like "novel" for
> techniques that people have been working on for many decades.
> Cyc, for example, has the largest formal ontology ever implemented,
> and it is thoroughly integrated with an upper (or foundational)
> ontology.  Lenat & Co. have had linguists working with them since
> Cyc was founded back in 1984, and their "integrated" approach has
> not enabled them to do significantly better than many other systems,
> including ones that Nirenburg & Raskin have been working on.
> Sergei is quite capable of defending himself    (01)

OK, I'll bite.    (02)

I read that paper. The problem with Cimiano and Reyle's statement
is that it's simply wrong. It doesn't seem that they read our stuff  
A more plausible source of difference between the approaches (which,
incidentally, are not that distant in attitude to describing the  
world) might
be our lack of interest in proving soundness of our knowledge base.
One of the reasons for that is the observation that people hold vast
amounts of contradicting beliefs and still manage to function (jokes
about how well they function aside). If our task is to approach human
levels of performance it is reasonable to try to operate with models
that have this feature. Not that we deliberately introduce  
contradictions...    (03)

> In the 40 years since Montague wrote his papers, there has never been
> any logic-based system that could translate a single page from any
> textbook to a formal logic of any kind.  When Kamp was a graduate
> student in the late 1960s, he was hired by Rand to translate (by
> hand) an article from the _Scientific American_ to logic.  That
> experience is what led him to move away from Montague's method to
> his own DRT.  But after working with many people who were using
> DRT, he realized that even DRT was insufficient.    (04)

So, Kamp came to realize what many people in AI have been long aware
of: that text understanding (and we should also say generation) depends
both on discourse context  and prior knowledge not recorded in text.
Logicians and theoretical linguists, being smart people, could not,  
of course,
deny this. However, they chose to ignore these issues. After all,  
until now there
are many linguists who think that semantics is "really not what we  
are about" --
one example is the insistence of Powerset people to do complete  
parse of the entire web before delving into hairy issues of meaning.    (05)

Issues such as vagueness, "unexpected" input, non-literal language,
reference and ellipsis and others require harnessing (at least)  
(or textual) context and prior knowledge. Building systems that
attempt to address these phenomena is not predicated on the syntax of
the metalanguage (which is -- I am somewhat simplifying here -- the  
main role
logic plays in such systems) but rather on its semantics, both  
(roughly, the real ontologies) and compositional (heuristic rules,  
from a variety of sources, including empirical ones -- a major point
promulgated by John Sowa, among others).
> So far, there is ZERO evidence that a logic-based method based
> on a uniform, formal foundational (or upper) ontology, by itself,
> would be adequate for NL semantics (or as a foundation for the
> Semantic Web or for computer system interoperability).
> I still believe that logic and ontology are important.  They are
> useful, but they're not magic.  And many people want magic.    (06)

A seemingly less pejorative word than magic could be "order"
(as opposed to the "chaos" of real-world semantics). But striving
for order using logics and formal ontologies only betrays a measure
of lack of respect for the complexity of our task... Note that I don't
exculpate myself from occasional lapses of this feeling of respect...
Especially when specific applications must be configured...    (07)

But I am trying to keep away from the widely accepted metaphor of
"low hanging fruit first". I prefer to be constantly reminded instead of
the cautionary metaphor of looking for the lost wallet under the  
street lamp.    (08)

Sergei    (09)

Message Archives: http://ontolog.cim3.net/forum/ontolog-forum/  
Subscribe/Config: 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 Post: mailto:ontolog-forum@xxxxxxxxxxxxxxxx    (010)

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