gmane.comp.ai.conceptual-graphs

Re: [] Deep question answering: Watson vs. Aristotle

John F Sowa — 2013-06-20T04:38:28

After I sent my note about Adam Lally's talk about Watson, I came across
an article that goes into more detail about the machine learning methods
that Watson used to estimate its confidence in a candidate answer:

http://researcher.watson.ibm.com/researcher/files/us-heq/W(16)%20ANSWERS%20MERGING_RANKING%2006177810.pdf

During the Q/A period of his talk, Lally discussed the complexity of
the evaluation process, which makes it quite different from typical
machine learning tasks.  They used various "models" for different
question types, and the "features" that were significant for one
model might be absent or irrelevant for another model.

As examples, Lally mentioned two Jeopardy! questions for which Watson
produced unusually bad answers:

   Q: When sending packages to soldiers in Iraq, don't send meat
      products from this animal.

   A: What is reindeer?

If religious taboos has been considered, the answer should have been
pig, hog, or swine.  It's possible that the feature that dominated the
evaluation

Re: [] Deep question answering: Watson vs. Aristotle

John F Sowa — 2013-06-19T20:00:54

Last week, Adam Lally presented an ACM webinar about Watson.
Registration is required, but it's free:

http://event.on24.com/eventRegistration/console/EventConsoleNG.jsp?uimode=nextgeneration&eventid=615568&sessionid=1&username=&partnerref=&format=fhaudio&mobile=false&flashsupportedmobiledevice=false&helpcenter=false&key=BD4EF1A7EC24CBCDEAA15DC3A4DCB83A&text_language_id=en&playerwidth=1000&playerheight=650&overwritelobby=y&eventuserid=82820970&contenttype=A&mediametricsessionid=65934872&mediametricid=1057031&usercd=82820970&mode=launch#

For anybody who saw the original Jeopardy show or read any articles
about Watson, skip the first 15 minutes of the talk.  At minute 19,
Lally mentioned a point that should be emphasized: Watson uses multiple
reasoning algorithms or paradigms for deriving possible answers.  The
first version, which used only six algorithms performed rather poorly,
but as they added more, it improved rapidly.  When they continued to
add more algorithms, the incremental improvement slowed. The

Re: [] Deep question answering: Watson vs. Aristotle

John F Sowa — 2013-06-20T04:38:28

After I sent my note about Adam Lally's talk about Watson, I came across
an article that goes into more detail about the machine learning methods
that Watson used to estimate its confidence in a candidate answer:

http://researcher.watson.ibm.com/researcher/files/us-heq/W(16)%20ANSWERS%20MERGING_RANKING%2006177810.pdf

During the Q/A period of his talk, Lally discussed the complexity of
the evaluation process, which makes it quite different from typical
machine learning tasks.  They used various "models" for different
question types, and the "features" that were significant for one
model might be absent or irrelevant for another model.

As examples, Lally mentioned two Jeopardy! questions for which Watson
produced unusually bad answers:

   Q: When sending packages to soldiers in Iraq, don't send meat
      products from this animal.

   A: What is reindeer?

If religious taboos has been considered, the answer should have been
pig, hog, or swine.  It's possible that the feature that dominated the
evaluation

Re: [] Deep question answering: Watson vs. Aristotle

John F Sowa — 2013-06-19T20:00:54

Last week, Adam Lally presented an ACM webinar about Watson.
Registration is required, but it's free:

http://event.on24.com/eventRegistration/console/EventConsoleNG.jsp?uimode=nextgeneration&eventid=615568&sessionid=1&username=&partnerref=&format=fhaudio&mobile=false&flashsupportedmobiledevice=false&helpcenter=false&key=BD4EF1A7EC24CBCDEAA15DC3A4DCB83A&text_language_id=en&playerwidth=1000&playerheight=650&overwritelobby=y&eventuserid=82820970&contenttype=A&mediametricsessionid=65934872&mediametricid=1057031&usercd=82820970&mode=launch#

For anybody who saw the original Jeopardy show or read any articles
about Watson, skip the first 15 minutes of the talk.  At minute 19,
Lally mentioned a point that should be emphasized: Watson uses multiple
reasoning algorithms or paradigms for deriving possible answers.  The
first version, which used only six algorithms performed rather poorly,
but as they added more, it improved rapidly.  When they continued to
add more algorithms, the incremental improvement slowed. The

[] Deep question answering: Watson vs. Aristotle

John F Sowa — 2013-06-15T17:19:16

I came across the URL of a recent posting that compares IBM's Watson
to the work on the Digital Aristotle project at Vulcan:

http://haleyai.com/wordpress/2013/06/15/deep-question-answering-watson-vs-aristotle/

That web page has many URLs that point to slides and articles about
the Vulcan work and to articles from the IBM Journal of R & D
about Watson.  The web page itself doesn't go into much detail,
but the references it cites are very useful.

Among the references is a pointer to slides on "Acquiring Rich
Knowledge from Text," which were presented by Benjamin Grosof
and Paul Haley at the recent Semantic Technology Conference.
Following is a web page with many citations for "background"
on that presentation:

http://haleyai.com/wordpress/2013/05/28/background-for-our-semantic-technology-2013-presentation/

Following is the URL for the slides themselves:

haleyai.com/wordpress/wp-content/uploads/2013/06/Acquiring-Rich-Logical-Knowledge-from-Text-SemTech2013.pptx

And following is a pointer to a web page (w

Re: [] [OpenDeepQA] Interest in major CG/Topic Map project?

John F Sowa — 2013-06-14T18:22:05

Jack,


They do seem to have some very useful mappings to and from many
different representations.  Has anyone else on CG list done anything
with HyperGraphDB?  Any comments about its performance, tool set, or
ease of use?

The mapping from HyperGraphDB to CGIF should be fairly straightforward,
but as with anything, the devil is in the details.  Writing such a
mapping might be useful for many applications.

In my previous note, I cited a 69-page TR on slot grammar by
Michael McCord.  That TR contains a lot of useful information
that could be adapted to other systems.  But it may be overkill
for anyone who is not planning to use that system.

Following is a 14-page report that goes into more detail about how
the ESG (English Slot Grammar) parser is used in conjunction with
other tools and notations in order to analyze Jeopardy! questions:

http://www.cs.cornell.edu/courses/cs4740/2013sp/papers/lally-et-al-2012.pdf

As Michael said in the SG report, the original system was implemented
in Prolog, but he later m

Re: [] [OpenDeepQA] Interest in major CG/Topic Map project?

Jack Park — 2013-06-12T00:43:13

Thank you very much John.  I see your comments as confirmation, of
sorts, on the literature search I've been conducting for this context.
I went so far, late last year, to get a copy of _Knowledge Systems and
Prolog_.

I am having a pretty good time studying and taking notes from Adil
Kabbaj's copious documentation on Amine; it is not entirely clear that
I have the chops necessary to understand all of it, but worming
through the code in Eclipse is being enlightening. Adapting it to a
database should be possible.

I've spent some time looking at http://www.hypergraphdb.org/  which
happens to have plug in applications ranging from Prolog, to WordNet,
to importing OWL documents, to a TMDM-style topic map and more. One of
the projects for it is a natural language system which uses OWL,
Prolog, WordNet, and the CMU link grammar parser. Quite a bit to study
there. The platform uses a key-value store, presently BerkeleyDB, but,
it appears that Apache Accumulo (courtesy the NSA) could be plugged
in.

In any case, th

[] [WoMO13] 3rd CfP: 7th Int'l Workshop on Modular Ontologies (WoMO-13)

Chiara Del Vescovo — 2013-06-06T00:01:44

========================================================
     7th Int'l Workshop on Modular Ontologies (WoMO)
            Corunna, Spain, September, 2013
           held in conjunction with LPNMR 2013

             --- Third Call for Papers ---

      --- Student Travel Grants available ---
========================================================
           Submission deadline: July 5, 2013
========================================================

http://www.iaoa.org/womo/2013.html

MODULARITY, studied for years in software engineering, allows mechanisms
for easy and flexible reuse, generalization, structuring, maintenance,
design patterns, and comprehension. In formal and applied ontology,
modularity is central to reducing the complexity of designing and
understanding ontologies, and to facilitating ontology verification,
reasoning, development, maintenance and integration.

Recent research on ontology modularity shows substantial progress in
foundations of modularity, techniques of modularization and modul

Re: [] [OpenDeepQA] Interest in major CG/Topic Map project?

John F Sowa — 2013-06-05T16:20:24

Jack,

I basically agree with your observations, and I'd like to add some
comments and suggestions.


Yes.  I would also point out that Watson uses a wide range of different
technologies, some of which had been developed over a long period.

Their NLP component, for example, is based on a Prolog system that
Michael McCord had developed during his entire career at IBM.  He
published an early article on _slot grammar_ in 1980, when he was
teaching at the University of Kentucky:

    http://acl.ldc.upenn.edu/J/J80/J80-1003.pdf

Michael was one of the "early adopters" of Prolog, and he joined IBM
Research shortly after he published this article.  He continued to
develop his system over the next 30 years.  Following is a 69-page
IBM Technical Report from 2010:

http://domino.research.ibm.com/library/cyberdig.nsf/papers/FB5445D25B7E3932852576F10047E1C2/$File/rc23978revised.pdf

Around 1995, I was consulting with a conceptual graph project at IBM
Santa Teresa that was using Michael's parser as a front end.  The slo

[] Re: [OpenDeepQA] Interest in major CG/Topic Map project?

Jack Park — 2013-06-02T18:56:08

An update to this:

The Amine project has been recommended many times. I downloaded all
nearly 700 pages of its documentation into Word then to a pdf so I
could spend time studying that on my tablet. I've created an Eclipse
project to study the code. There may be an approach which starts with
that code.

Cheers
Jack

On Mon, May 27, 2013 at 9:24 PM, Jack Park <jackpark-Re5JQEeQqe8AvxtiuMwx3w< at >public.gmane.org> wrote:

---------------------------------------------------------------------
To unsubscribe, e-mail: cg-unsubscribe-ndqbGWLCXyQvU1jxblBdo2D2FQJk+8+b< at >public.gmane.org
For additional commands, e-mail: cg-help-ndqbGWLCXyQvU1jxblBdo2D2FQJk+8+b< at >public.gmane.org

[] [OpenDeepQA] Interest in major CG/Topic Map project?

Jack Park — 2013-05-28T04:24:57

Just an abstract for now:

Strong claim: it is a reasonable concept that IBM's DeepQA (Watson) can be
replicated in general capabilities today. After all, they open sourced
UIMA, and published a large body of descriptive material with which skilled
developers can begin experiments.

Hypothesis: a framework based on Conceptual Graphs, certain harvesting
concepts, and topic maps to keep all concepts and relations well organized
can satisfy at least one variant of such a platform.

The CG platform in question here, it seems to me, will necessarily be more
complex than any of the present open source implementations, even though
there are very good ones available. My present thinking calls for starting
fresh with the framework of CG/TM in mind, complete with scalable
persistence/indexes.  My own code is starting using Apache Solr for
indexing.

I am seeking interest, conversation, whatever it takes to move forward on a
"movement" I think of as Open Deep QA.

Consider one instance of a reason for doing so. We know

[] What is the role of an upper level ontology?

John F Sowa — 2013-05-19T18:51:52

The following slightly edited note to Ontolog Forum summarizes
an ongoing thread about the topic in the subject line.

Anyone who may be interested in the topic can extract any phrase
quoted below, enclose it in quotes, and use any search engine
to find the original note (and thread) in the Ontolog archive.

Another useful resource is the proceedings of the KI2003 Workshop
on Reference Ontologies and Application Ontologies.  The articles
are now 10 years old, but they are still good position papers
on the various issues.  Most of the topics they discussed are
still ongoing R & D issues today:

http://sunsite.informatik.rwth-aachen.de/Publications/CEUR-WS/Vol-94/

John Sowa

-------- Original Message --------
Date: Sun, 19 May 2013
To: ontolog-forum-TYD3xnkLNKlCLzx1llviGw< at >public.gmane.org

My main point is that any upper-level ontology that claims to be
broadly applicable should avoid detailed axioms.  It should be
as neutral as possible with respect to any or all ontologies that
have proved to be useful for

[] GKR< at >IJCAI 2013 Second Call for Papers

Madalina Croitoru — 2013-03-04T13:35:48

==============================GKR< at >IJCAI 2013 SECOND CALL FOR 
PAPERS===================================

THE THIRD IJCAI INTERNATIONAL WORKSHOP ON GRAPH STRUCTURES FOR KNOWLEDGE 
REPRESENTATION AND REASONING

======================================================================================================

Graph-based knowledge representation and reasoning is a growing area of 
research, with more and more important contributions appearing over the 
last few years.

The workshop welcomes contributions that:

[] Archives for the Philosophy and History of Soft Computing

John F Sowa — 2013-02-27T15:37:46

The attachment announces a new online journal devoted to Archives
for the Philosophy and History of Soft Computing.

Most of the founders belong to the fuzzy systems community, but
they are broadening their scope to interdisciplinary studies
that focus on issues that relate technology and the humanities.

John Sowa

---------------------------------------------------------------------
To unsubscribe, e-mail: cg-unsubscribe-ndqbGWLCXyQvU1jxblBdo2D2FQJk+8+b< at >public.gmane.org
For additional commands, e-mail: cg-help-ndqbGWLCXyQvU1jxblBdo2D2FQJk+8+b< at >public.gmane.org

[] Sapir-Whorf Hypothesis

John F Sowa — 2013-02-26T13:10:52

Following is a slightly edited note to Ontolog Forum about the way
the structures of different languages influence thought.  The first
URL in this note points to a TED talk on the subject.

At the end is a reference to the research article written by
the speaker.  The full paper is 56 pages long, but that includes
a lot of data and the list of 126 different languages in the study.

-------- Original Message --------
From: John F Sowa
To: ontolog-forum-TYD3xnkLNKlCLzx1llviGw< at >public.gmane.org

On 2/26/2013 5:22 AM, Hassan Aït-Kaci wrote:

Thanks for the pointer. Keith Chen is a good speaker, and he presented
the hypothesis and the results very well.  I recommend it.

Relationship to ontology:  Many of us have found that a 4D ontology
for spacetime is more useful for representing and reasoning about
certain kinds of relationships than a 3+1 D ontology.

In effect, English is a strongly 3+1 D language that forces every
statement to be situated on a time line.  But Chinese is more
of a 4D language that ignores t

Re: [] Prolog + CHR (Constraint Handling Rules)

John F Sowa — 2013-02-22T18:33:04

I received an offline note about the ClioPatria Semantic Web server,
which is written in Prolog and enables Prolog to do reasoning with
and about Semantic Web data:

    http://cliopatria.swi-prolog.org/help/whitepaper.html

This software is freely available, and it could be combined with other
Prolog-based tools.

John

---------------------------------------------------------------------
To unsubscribe, e-mail: cg-unsubscribe-ndqbGWLCXyQvU1jxblBdo2D2FQJk+8+b< at >public.gmane.org
For additional commands, e-mail: cg-help-ndqbGWLCXyQvU1jxblBdo2D2FQJk+8+b< at >public.gmane.org

[] Ubuntu for tablets

John F Sowa — 2013-02-21T14:35:15

According to various market estimates, Smartphones based on the Android
OS from Google had a 68% market share in 2012 vs 20% for Apple.

Since Android is based on the Linux kernel, it was easy for the
developers of the popular Ubuntu version of Linux to make Ubuntu
run on top of Android.

But now they have taken the next logical step -- provide an Android-like
front end to Ubuntu that will run on any form factor:  Smartphones,
tablets, PCs, and TVs.  They also provide software that makes it easy
for vendors to convert Android apps to Ubuntu apps.

Since a tablet or PC has a larger screen than a phone, Ubuntu enables
multiple Smartphone apps to run concurrently in separate windows on
a larger screen.  See

    http://www.ubuntu.com/devices/tablet

Commercial developers are beginning to get into the Ubuntu market for
Smartphones.  That should help promote competition.   For more info,
type "ubuntu for tablets" to your favorite search engine.

John

--------------------------------------------------------------

[] Prolog + CHR (Constraint Handling Rules)

John F Sowa — 2013-02-20T20:50:15

Following is a set of 90 slides about using Prolog + CHR to support
large-scale applications in finance:

    http://dtai.cs.kuleuven.be/CHR/files/Elston_SecuritEase.pdf

Prolog is a powerful platform for developing intelligent systems,
but it has fallen out of fashion as newer fads have sprung up.
But as this presentation shows, it is a highly scalable platform
that can support major applications.

The Experian credit bureau is a very large corporation, which uses
Prolog to check everybody's credit worthiness.  They use it so heavily
that they bought Prologia, the company founded by Alain Colmerauer,
who implemented the first version of Prolog.

But Experian is also a secretive company that doesn't tell anybody
what they do or how they do it.  Nevertheless, their size is a strong 
argument for the power of Prolog (and related technology).

The CHR notation compiles into Prolog:

    http://www.swi-prolog.org/man/chr.html

Acknowledgment: Thanks to Adrian Walker (in cc list above) for
sending me a copy of th

[] References on neuroscience and cognitive linguistics

John F Sowa — 2013-02-18T14:21:05

The Human Connectome Project is carrying out a detailed mapping
of the connections in the brain as a basis for further study and
for various attempts to simulate the human brain.

A BBC science correspondent volunteered to have the full 45 minute
scan, of which only 50 have been carried out so far.  Following is
his description of the experience with some colored pictures of the
wiring in his brain:

    http://www.bbc.co.uk/news/science-environment-21487016

For more pictures and more detail, see the project home page:

    http://www.humanconnectomeproject.org/

The discussion is keyed to the various parts of the brain. I presented
a quick overview with some pictures and terminology in Section 1 of
my talk for ICCS in January (Slides 3 to 19):

    http://www.jfsowa.com/talks/relating.pdf

Section 3 of that talk is based on publications by René Thom,
Wolfgang Wildgen, and Jean Petitot.  See slide 20 for URLs.
Following are the slides from a talk by Wildgen in 2006:

http://www.slideserve.com/mariah/emerge

[] Architectural considerations in ontology development

John F Sowa — 2013-02-13T16:10:36

As my publications show, I often discuss philosophical issues
and relate them to software design and development.  But it's
important to recognize when and whether a distinction is relevant.

Following is a slightly edited version of a note to Ontolog Forum.

John Sowa

-------- Original Message --------
Subject: Re: Architectural considerations in ontology development
Date: Wed, 13 Feb 2013 10:56:17 -0500

...

There is a serious question of which philosophical distinctions
are relevant to an ontology and whether they belong at the *top*
level or the *bottom* levels.


That distinction can be critical for some applications and irrelevant
for others.  Many people (and computer applications) can get along very
well without ever thinking about or using those distinctions.


Those are also important philosophical issues.  But applications
that make different choices should be able to communicate with and
interoperate with other applications that ignore those distinctions.

In particular, legacy systems that hav

[] Research directions

John F Sowa — 2013-02-12T16:30:10

In my books and papers, I like to show how the latest and greatest R & D
is related to historical developments over the past few millennia.

In the following note to Corpora list, I quote some examples and
summarize the issues in the last two paragraphs.

John

-------- Original Message --------
Subject: Re: [Corpora-List] Broader linguistic resources
Date: Tue, 12 Feb 2013 11:05:57 -0500
To: corpora-afly1vuFV30< at >public.gmane.org

On 2/12/2013 8:55 AM, Dominic P Rout wrote:

That question reminds me of a thread from last week (Feb 5).
The subject line was "New techniques in text processing":

Amac Herdagdelen asked:

Phil Gooch replied:

Adam Kilgarriff replied:

An excerpt from Adam's paper:

An excerpt from the Chambers & Jurafsky paper:

Adam's paper describes important methods for analyzing corpora.
They belong in the toolkit of anyone who processes large volumes
of NL texts.

But the paper by Chambers & Jurafsky shows how issues that were
popular 30 years ago can be revived as "cutting edge" research to