Hi Pat --
You wrote...
As for killer apps, I
actually think that a natural Language interpreter that can handle basic
English at the level of a 6-year old would represent a remarkable advance in computational
linguistics and artificial intelligence, and might even have applications in
that form. In any case it should serve to demonstrate that an ontology is
useful for supporting such a complex application. And if it could develop to
the linguistic level of a 16-year old, it would form the basis for any number
of powerful applications. Like most human 16-year-olds, it would not be
particularly capable in any field, but would be capable of *learning*
quickly by being taught in language, and once any expert skill is learned, the
application can be very cheaply reproduced, unlike human experts. That
kind of learning capability is one of Cyc’s goals, but what I have seen
of their NL program does not suggest to me that they are yet close to that
capability.
Well yes, but this would appear to be just a slightly modified version of the "AI-complete" problem. That is, once you have a system that can read and understand English, it can take off and improve itself by being taught, or even autonomously -- these days perhaps by browsing the Web.
But, as the label "AI-complete" suggests, this is a hard problem. Very talented people have worked on it for over 25 years. Doug Lenat, and the work of Mooney's group at U Texas comes to mind. What also comes to mind is John McCarthy's advice about AI -- to sit back and think for a long time, rather than to keep building similar experimental systems.
I like to think of the situation in terms of thesis-antithesis-synthesis. Roughly, Thesis: "we can build a natural language understanding system based on Chomsky grammars, ontologies, and a few other nice familiar things". Antithesis: "experience is teaching us that ain't so". Synthesis: "Some new insight into the nature of human languages".
Just some after-lunch musings.
Cheers, -- Adrian
Internet Business Logic
A Wiki and SOA Endpoint for Executable Open Vocabulary English over SQL and RDF
Online at www.reengineeringllc.com Shared use is free
Adrian Walker
Reengineering
On Tue, Apr 21, 2009 at 1:06 PM, Patrick Cassidy <pat@xxxxxxxxx> wrote:
Adrian:
Re:
[AW] > Hi Pat --
> Ambitious project.
> Is your aim (1) academic recognition that COSMO is correct, or (2) a
killer application ?
> If (2), what is the killer app?
>
Cheers, -- Adrian
The goal is to encourage
adoption of *some* common foundation ontology by some large community .
COSMO will not be *it*, but it may be one of the resources used to
create that community ontology. My own goals are to provide evidence that
a good FO can be used for NL, but the heavy lifting has to be done by a large
group, because the task is very complex. The FO has to be built by a consortium
including a large group of users with real inference-dependent applications –it
will serve as a common ‘language’ for the computers, and any useful
language (one including a basic vocabulary) has to be built and maintained by
the users. One of the main problems with the existing foundation ontologies
is not in their structure, but in the fact that they are built and maintained by
small groups that present them to the world, with little or very slow
mechanisms for feedback. I am also getting increasingly convinced
that no FO will gain wide usage unless it has a very good NL interface, which
will allow users to interrogate the ontology to determine if it already has a
concept in it that the user needs; and as part of that capability, can take a concept
description in ordinary language and convert it into the logical description that
will fit into the FO (or into some extension). Having spent time trying
to understand the structures of existing FOs, it seems unlikely to me that any
of them will be easily usable without such an interface. (try
it – you won’t like it) The need for an NL interface is one
reason I want to make as many as possible of the COSMO elements as close
as possible to linguistic usage of concepts. A good NL conversational
capability is also essential not just for usability, but because without it, any
ontology applications will be hard for non-ontologists to distinguish from just
another database or just another program.
As for killer apps, I
actually think that a natural Language interpreter that can handle basic
English at the level of a 6-year old would represent a remarkable advance in computational
linguistics and artificial intelligence, and might even have applications in
that form. In any case it should serve to demonstrate that an ontology is
useful for supporting such a complex application. And if it could develop to
the linguistic level of a 16-year old, it would form the basis for any number
of powerful applications. Like most human 16-year-olds, it would not be
particularly capable in any field, but would be capable of *learning*
quickly by being taught in language, and once any expert skill is learned, the
application can be very cheaply reproduced, unlike human experts. That
kind of learning capability is one of Cyc’s goals, but what I have seen
of their NL program does not suggest to me that they are yet close to that
capability.
John Sowa has explained why he thinks a good modular
architecture (e.g. a Minsky-esque ‘society of mind’) is needed
to build powerful AI applications. I agree. In addition to the control
architecture, I believe that a powerful foundation ontology is needed to
represent the meanings of the information passed among the modules. I
leave the system architecture to others, and am focusing on the FO that will
have that capability, and the NL interface that is needed to make the FO
usable.
Hi Pat --
Ambitious project.
Is your aim (1) academic recognition that COSMO is correct, or (2) a killer
application ?
If (2), what is the killer app?
Cheers, -- Adrian
On Mon, Apr 20, 2009 at 10:17 PM, Patrick Cassidy <pat@xxxxxxxxx> wrote:
John, Richard:
> PC> But WordNet still represents a tremendous and useful effort,
> > and is useful for NL at a shallow semantic level.
>
> I agree with most of what you said about WordNet, including this
> sentence. However, the following sentence is asking for something
> totally different -- not just a revised WordNet.
>
> PC> It is a good start, but something similar with a more precise
> > semantics is needed.
>
> The synsets of WordNet are at the same level as the word senses of
> a typical English dictionary. The process of deriving a dictionary
> such as the OED begins with dozens or even hundreds of highly
> trained lexicographers who take millions of citations gathered by
> thousands of people (many of them volunteers) who extract those
> citations from a truly immense volume of English.
>
> The old shoe boxes full of paper slips have been computerized,
> but the amount of human effort is measured in person-centuries.
> What you find in the dictionary (or in WordNet) is a boiled-down
> or *condensed* extract of the "average" meaning over many, many
> different occurrences of each word sense.
>
> If you want precision, you won't get it by averaging from raw data.
> You can only get precision by examining the precise *microsenses*
> of each word as it is used in each and every citation in the total
> mass of raw data.
>
> This implies that the precise semantics will be truly immense.
> And instead of being listed in alphabetic order, the precise
> meanings will be grouped in something like the microtheories
> of Cyc. But there will be an enormous number of them. In 2004,
> Lenat & Co. estimated that they had about 6000 microtheories,
> and they may have many more by now. But every time they get
> a new application, they need at least one new microtheory,
> and often quite a few more microtheories.
>
> Remember the line that Amanda mentioned and I highlighted:
>
> Ontology is fractal.
>
> That means that the amount of detail that is necessary at each
> level is the same at every level you examine. That implies that
> we will need something of the size of WordNet for every topic
> of every branch of human knowledge and activity. The completely
> precise version you are asking for will dwarf the current WWW.
>
> Yet a child at the age of 3 has a command of language that is
> far better than any computer system today. And that child
> doesn't need Cyc or WordNet or formal logic. I believe we
> should focus on what makes a child smart -- and it's not Cyc
> or anything remotely like it.
>
> RHM> You [Pat C] consistently said mapping from WordNet to xxx.
> > Do you realize that OpenCyc is mapping from its concepts to
WordNet?
>
> Lots of people have been mapping their ontologies to and from WordNet.
> But no computer can understand language as well as a 3-year-old child.
>
[PC] I agree with most of what John says, except for the
'fractal' part.
Ontology is certainly *capable* of being extended to indefinite levels of
detail, but practical applications do not require indefinite levels of
detail. One needs to describe all the detail that is important for the
application at hand, and that is enough. If more detail is needed, more
detail should be added.
What do we need next?
>> PC> But WordNet still represents a tremendous
and useful effort,
> > and is useful for NL at a shallow semantic level.
>
> [JS] I agree with most of what you said about
WordNet, including this
> sentence. However, the following sentence is asking
for something
> totally different -- not just a revised WordNet.
>
Yes, that is the point - what an ontology with pointers to
words that label
the concept is, is not just another WordNet, though the 'synsets' derivable
from it would *look* like WordNet synsets. It would have support for much
greater levels of semantic detail and semantic precision, and would include
important rules and functional representations that WordNet cannot
represent. The point of referencing WordNet in the ontology is to reassure
NL researchers that the ontology can be used in the same way as WordNet (if
an appropriate tagged corpus becomes available), but that the 'synsets' can
also be used for logical inferencing. This will hopefully encourage NL
researchers to try it out. Eventually users will need to do things that
are
not and cannot be done with WordNet - logical inference on the information
derived from the NL processing.
I am well aware of the problems one encounters in identifying 'word senses'
for dictionary purposes with coherence ontological concepts, having been
concerned about precisely that for the past twenty years. What I am
trying
to do now is to test the use of an ontology, with the ontology elements
serving as 'senses' and specifying which linguistic labels (English words or
phrases) are used to refer to those concepts in ordinary language. But
this
is a large task, and I have to confine my own efforts to the basic
vocabulary of 2000-5000 words.
As for existing mappings:
> RHM> You [Pat C] consistently said mapping from
WordNet to xxx.
> > Do you realize that OpenCyc is mapping from its concepts to
WordNet?
>
> Lots of people have been mapping their ontologies to and from WordNet.
> But no computer can understand language as well as a 3-year-old child.
>
As I said in my post:
[PC] >> the WordNet structure is not based on principles of inheritance;
so
a simple 'mapping' of WordNet
to an ontology like Cyc or SUMO is of limited
usefulness, and does not correct the problems.
The existing 'mappings' do not serve the same function as
the kind of
'mappings' I am doing in COSMO, because the synsets are not atomic concepts,
and not all words in a synset are in fact conventional labels for the
ontology concepts. The Cyc 'mappings' don't serve the purpose that a
complete rethinking of the synsets themselves will serve - and a rethinking
of the synsets is what is necessary, and what is part of my efforts with the
COSMO. The existing WordNet synsets are excellent references and
resources,
but one cannot take them as coherent ontological concepts, though that is
the way that some people attempt to use them.
_________________________________________________________________
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
_________________________________________________________________
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 (01)
|