Pat Hayes responded to a post on this thread.
OK, let's try to clarify the points that PatH either
misunderstood or just disagrees with.
[[1]] First, "definition"
[PC] > An issue that has occupied some of my
attention lately has been the
Ø question
of what basic ontological concepts are sufficient to support
Ø accurate
communication. I frame the issue as being analogous to the
Ø "defining
vocabulary" used by some dictionaries as a controlled vocabulary
Ø with
which they define all of their words. For the Longman's, it is around
Ø 2000
words. The analogous question is how many fundamental ontological
Ø elements
(types/classes, and relations/functions) are needed to logically
Ø specify
the meanings of all other terms used in a reasonably complex domain
Ø (having
perhaps 100,000+ terms), to some adequate level of detail?
[PH] > >
> Define "define". If you mean logically
define, then the number of defining terms is going to be comparable to the
total number of terms. All 'natural kind'
> terms, for example, are definitionally primitive.
When I used "define" in reference
to dictionary definitions, the meaning should be clear, and ample examples are
available. For the "Conceptual defining vocabulary" , the
foundation ontology nicknamed thus to emphasize its **analogy** to the
linguistic defining vocabulary used in dictionaries, I carefully used the term
"logically specify the meaning" because I am aware that using the
term "define" in an ontology context sets off fire alarms among
those who can only interpret that word as meaning "necessary and
sufficient definition". I suppose that calling it a
"Conceptual Defining Vocabulary" may cause a reaction in those who
are sensitive to the specialized meaning of "define" in logic, but
I thought that the distinction was clear. Nevertheless, I will continue
to use "conceptual defining vocabulary" because I think it is a
useful analogy, even though the "meanings" of the terms in the
ontologies created using the foundation ontology will rarely be specified as
necessary and sufficient logical "definitions".
To "specify the meaning" of a
subtype is to assert the necessary conditions for membership, for example:
(1) Asserting
it as a subtype
(2) Asserting
some property or relation not held by other subtypes – this could be as
simple as being a subtype of another type.
(3) If
there are necessary properties or relations not inherited from any of its
parent types, they must also be asserted, using terms already in the conceptual
defining vocabulary (aka foundation ontology)
(4) If a
subtype is precisely the intersection of two parent types, that would be a
necessary and sufficient definition. Of course, the parent types may be
primitives.
(5) Occasionally,
other means of specifying necessary and sufficient conditions will be used.
There are several criteria which I use to decide
whether a newly created concept representation is or is not primitive. I
have mentioned those before, and won't reiterate here.
[[2]] how much vocabulary do we have in common?
[pc] >> My own suspicion is that the similarity **in
the fundamental concepts**
>> has to be very close to 100%.
[PH]
Ø There
is quite convincing evidence that this is NOT the case. In particular, human
beings seem to
Ø communicate
adequately even while the terms they use in the communication are based on
Ø different
mental ontologies. This happens so often that it seems to be the norm rather
than
Ø the
exception; it is part of what makes it so hard for reasonable people to come to
agreement
Ø on
basic ontological questions (such as whether time is a dimension or not).
This is precisely the point that I dispute,
that the communication is based on different ontologies. This is
worth some discussion. The hypothesis I presented is that the part that
**is understood** is based on *exactly* the same ontology, and where the
ontologies differ there is misunderstanding. In fact, I can't
imagine how it can logically be otherwise. Different
different interpretation (at least when the parts that are different are part
of the interpretation process). But, **please** note that I said that
what is similar in people are the **fundamental concepts** - among which are
the naïve physics sort of notions that a weighty object will fall if not
supported, when near the earth's surface. The point is that there
are lots of these, including how people interact with each other, that we learn
from experience as we grow up. And when we put labels on those common
experiences, we have the most basic ontology, which is extended by analogy to
less concrete things. The exact process by which the analogical
extensions are created is a fascinating subject of research, but for the moment
I am only concerned with the end result – we all have a large (>
1000) body of common concepts for which we have terms that can be understood
precisely, when uttered in a familiar context (don't get started on that
– conversational and situational context, OK?). So it appears that
you disagree. But on what basis?
Funny thing you should mention time and
"dimension" because that is precisely one of the good illustrations
of this point. When I did an exercise, creating linguistic
(dictionary-style) definitions of 500 words not in the Longman's defining
vocabulary, the one word I decided that I needed that was not itself in the
original vocabulary was "dimension". They didn't
need it to define (the way they do their definitions) any of the 65,000 words
in the dictionary. Also to the point, 5-year olds who know thousands of
words and a lot of naïve physics and naïve sociology will typically not have a
clue as to what a "dimension" is. That is a word we
learn in school. At least in my case I am sure I didn't have a need
for it until I learned some geometry. Ergo, the reason that your
disputants couldn't agree on whether time is a "dimension" is
because "dimension" is not one of the fundamental concepts that are
lexicalized in the basic vocabulary. It's not part of the
linguistic defining vocabulary of words we all agree on. It's
needed for math, and in that context, describing visualizable two or three-dimensional
spaces, is probably uncontroversial. When used by analogy to refer to
other concepts, I am not surprised that terminology disputes can arise.
Of course, words differ in their meaning depending
on context, and that will screw up any but the most carefully planned
experiments. People who have no clue as to what an "ontology"
is communicate quite well. – in discussing accuracy of
linguistic communication, we are talking about the implied ontologies
people use in communicating, not the logic-based systems that computational
ontologists create. So with respect to your fascinating comment:
>> the terms they use in the
communication are based on different mental ontologies.
Just which terms are based on "different
ontologies"?? How did the researchers extract these
"ontologies"? How did they distinguish different *ontologies*
from different interpretations of *word sense*? This is
interesting. References??
If we can find the terms that differ, that could provide
us with clues as to what words are not in the common basic vocabulary.
[[3]] Relative speed of brain and computer
[PC] >> The reasoning is something like
this: if the
>> brain (or a simulation) does as much computation as one of our
laptops, then
>> it can run at least 1 million inferences per second.
[PH] > There is no evidence whatever that the brain is
capable of this. In fact, there near-conclusive evidence
> that the human brain is incapable of quite simple
logical inferences without external
> support (such as drawing a diagram or writing a
formula). In particular, people - even
> with logical training - consistently make
systematic errors when given
> simple modus tollens reasoning tasks, with confidences
in the 90% range.
First, note that I said ***if***
the brain does as much computation. It's true, I think it does
more, but of course not in a linear binary manner. Second, and more
importantly, I did *not* say that the brain does **conscious and
sequential** modus ponens inferencing of the type we do with our logic
formalisms. It does its inferencing by massively parallel signal
processing. With >10**11 neurons, each with > 1000 connections to
other neurons, firing at multiple times per second, it hardly seems like a bold
assertion to imagine that it will accomplish at least the FOL equivalent of 1
million inferences per second. The visual system processes images
multiple times per second to a level of detail that our fastest computers
cannot yet mimic. When a baseball batter sees a ball coming at him for a
few tenths of a second and starts swinging, just how inference-equivalents do
you think it would take to process the image, calculate the proper swing, and
send the signals to the proper muscles? Less than 300,000? Then you
should be able to write a program to do that really easily, at 2 gigahertz.
[[4]] How close are the ontologies of two different people
[PC] >> A similarity of 99.9% in two different
fundamental
>> ontologies may not be enough for any meaningful level of
communication.
[PH] > Look, this HAS to be nonsensical. If this really
were true, how could human
beings EVER succeed in communicating?
Because our basic ontologies are in fact closer
than that.
[PH} > There is no way you and I could come to a
logically secure agreement
> on 1000 axioms even if we had months to do it in and
had nothing else to do.
> But children learn natural language at a rate of around
20 new words per day around
> the ages of 4-6 years old, pretty much all by
themselves.
Talk about nonsense. If a child learned
language by itself, why don't the feral children speak the native
language of their area fluently? Yes, children learn first largely by
pointing to instances, but when a young child visits a zoo and points to a tiger
and says "pussy cat" we know that her internal definition of
Felidae has not yet reached the level of detail of an adult. The internal
meanings of words do get refined as people have more experience of the world
and learn more about the language.
First, in talking about the shared
common ontology expressed in language, I am discussing the kind of basic
notions people use in general communication, and in particular when trying to
explain things clearly to one another. These words are associated with
their meanings by experience in context, not by discussions among opinionated
academics. The intended meanings of the words people use when they are
trying to be clear are almost always understood properly – because the
words are chosen to minimize the potential for misinterpretation – the
speaker knows how the listener will interpret those words. The issue
remains, when I ask you to explain the meaning of something, how could I
possibly understand the answer unless you used words whose meanings we both
agree on? Yet that happens every day many times a day to almost everyone.
By age 5 or 6, children are capable of learning by being told. They
have a good part of the basic defining vocabulary by that point. I suspect that
the basic vocabulary isn't near its adult level until the teens.
Can you and me agree on an ontology?
Well, you are, let us say, contentious, but I still believe that we could
come to agreement on a lot more than 1000 axioms, provide that both of us had
the time to discuss it and neither of us insists on throwing out something that
the other proposes on any grounds other than a logical inconsistency with other
parts of the ontology, or a fatal vagueness that makes it impossible to relate
one to the other. The biggest reason people disagree on ontologies (when
they take the time to actually argue the points, and ignoring the most common
source of disagreement – different use of terms) are either due to some
insistence on only one means of expressing a concept, even though logically
consistent alternatives may be preferred by others – or simple personal
preference. Many differences revolve around different theories of certain
physical objects or processes. Those can be represented with logical
consistency by categorizing them as alternative theories (Newton, Einstein),
each useful in certain situations, neither necessarily expressing the
fundamental structure of (whatever). You yourself have expressed the
opinion that the differences among existing ontologies are smaller than they
are generally considered to be, and logical consistency can be demonstrated by
bridging axioms among different ways of expressing the same idea. If we
allow logically consistent ways to express the same concept, or include slight
variations on a concept, agreement will be much easier. In the SUMO
project, one of the problems was that Adam and Ian took a minimalist view of
the upper ontology. They wanted to include only one way of representing
things in the world. In part, they had to produce a deliverable in a
specified time frame, which forced them to avoid extended discussion. In
addition, their concern, as I interpreted it, was that any redundancy would
eventually result in an intractable complexity of reasoning. But that
doesn't mean that what they had in the ontology was in any sense *wrong*.
In some cases, it just didn't satisfy the *preferences* of
others. If they had adopted the tactic of putting in whatever anyone
wanted, properly related to the other parts, it may have had a wider
acceptance. The lack of time and funding for wider participation was, I
believe, also a major problem.
Do an experiment. Go to the Longman
dictionary site:
http://www.ldoceonline.com/
. . . and look at some of the definitions that
they give for words you select. (Don't worry too much about whether
the definitions actually give a good mental picture of the meaning of the word
in all its nuance and detail - they are designed to be simple, for
learners of English as a second language). The question is, do you think
that there are ambiguities in the defining words that make you uncertain as to
what they intend by those definitions? If you find any case where you
think that you don't properly understand the meanings of the words they
use, let us know which ones.
[[5]] how do native speakers learn to agree?
[[PC]] >> We all know that people differ
in assumptions and beliefs, and yet we do
manage to communicate reasonably well in most cases. How can that be?
[PH] > Because the assumptions and beliefs themselves are
irrelevant to the communication:
> all
that matters is that we come to agreement on the beliefs we EXPRESS to one
another.
(PC – I thought that's what the next
part of the paragraph was saying)
[PC] >> Well, it happens probably because we **know**
that we have different
assumptions and beliefs, and when communicating, only assume that there is a
certain fundamental set of knowledge in common, and only
rely on that basic
set of common assumptions and beliefs to express the ideas we want to
communicate.
[PH] No, that doesn't work. First, people don't know
this.
[PC] Oh, come on. We all learn at an early age that
people believe different religions, and even different politics. Everyone
knows that many other people are not experts in the field we are experts
in. We encounter differences of opinion as soon as we can talk.
When a specialist tries to explain something to a non-specialist, s/he
doesn't usually use technical jargon (unless s/he is oblivious to the fact
that it is jargon). I don't understand this objection. How
would you succinctly describe the Gricean conversational maxims?
[PH] Second, how is agreement on this vital common 'core'
achieved?
[PC] By growing up in an environment where the basic words
are used consistently to refer to the same common experiences that every person
has, regardless of where we grow up. We don't get together around a
table to agree on the basic words. We absorb them by living in an
environment where they are used consistently in the same sense, in a particular
situation. We need to see them used in a common linguistic environment to
begin to grasp the meaning.
===========================================
To understand the points above, recall that I am talking about the
basic vocabulary shared by most native speakers of a language, not the
technical words that we learn on special topics – even something so
fundamental-seeming as "dimension" or "identity" (also
not in LDOCE defining vocabulary). There are plenty of those common
words, and plenty more of the specialized ones that can be defined (dictionary
definition) using the common ones, though in some cases a picture makes
understanding a lot easier.
When building a foundation ontology to serve as a Conceptual Defining
Vocabulary, we will of course create abstractions that are not lexicalized in
the linguistic defining vocabulary, in order to analyze the more common
concepts into their more fundamental component parts. This
presents an opportunity/risk for different ways of analyzing the structure of
common basic concepts – like "event'. But when
the most fundamental components of meaning are extracted and represented, the
different ways that ontologists may prefer to represent the same thing can all
be represented with the most basic concepts, and the relation between those
different ways of viewing the same thing can be precisely specified. I am
aware that merely saying this will not convince anyone. It requires a
project to actually test the hypothesis. That was also a subject of a
previous posting. I wish that the IKRIS project had continued – I
think it would have produced some important data related to this hypothesis.
PatC
From: Pat Hayes
[mailto:phayes@xxxxxxx]
Sent: Thursday, March 06, 2008 4:15 PM
To: [ontolog-forum]
Cc: edbark@xxxxxxxx; Patrick Cassidy
Subject: Re: [ontolog-forum] Ontology similarity and accurate
communication
At 8:00 AM -0500 3/6/08, Patrick Cassidy wrote:
In the discussion on "orthogonal", Ed Barkmeyer
pointed out:
> My position is that two agents don't need to have non-overlapping
> ontologies to be unable to communicate effectively. Their ontologies
> can have a 90% overlap, but if there is one critical idea that one has
> and the other does not understand, they can't do business.
>
Ed focused on the problem that arises when one 'critical idea' differs
between the ontologies (or assumptions) of two different communicating
agents. I suspect that the problem can also arise when even minor
differences are present in the interpretations of communicated information,
because the interpretation of many concepts involve a very large number of
implications and associated inferences.
This question appears to me to be one that is worthy of a separate
field
of investigation: precisely how different can ontologies be while sustaining
an adequate level of accuracy in interpreting communications that rely on
the ontologies?
My own suspicion is that the similarity **in the fundamental concepts**
has to be very close to 100%.
There is quite convincing evidence that this is NOT the
case. In particular, human beings seem to communicate adequately even while the
terms they use in the communication are based on different mental ontologies.
This happens so often that it seems to be the norm rather than the exception;
it is part of what makes it so hard for reasonable people to come to agreement
on basic ontological questions (such as whether time is a dimension or not).
The reasoning is something like this: if the
brain (or a simulation) does as much computation as one of our laptops, then
it can run at least 1 million inferences per second.
There is no evidence whatever that the brain is capable of
this. In fact, there near-conclusive evidence that the human brain is incapable
of quite simple logical inferences without external support (such as drawing a
diagram or writing a formula). In particular, people - even with logical
training - consistently make systematic errors when given simple modus tollens
reasoning tasks, with confidences in the 90% range.
If (crudely
calculating) the inferences supported by the differing ontologies differ by
1 in 1000 then two different ontologies will generate 1000 differing
inferences per second from the same information. How much difference can
be
tolerated before something goes badly wrong - perhaps a direct logical
contradiction? My guess is that each serious "fact" that we
rely on to
support our everyday activities is supported by at least 1000 assumptions,
and getting one in a thousand wrong would invalidate the meaning of these
facts, making normal actions, expecting predictable results, effectively
impossible at any level. A similarity of 99.9% in two different fundamental
ontologies may not be enough for any meaningful level of communication.
Look, this HAS to be nonsensical. If this really were true,
how could human beings EVER succeed in communicating? There is no way you and I
could come to a logically secure agreement on 1000 axioms even if we had months
to do it in and had nothing else to do. But children learn natural language at
a rate of around 20 new words per day around the ages of 4-6 years old,
pretty much all by themselves.
But, as I said at the start, this is an issue that needs
investigation.
We all know that people differ in assumptions and beliefs, and yet
we do
manage to communicate reasonably well in most cases. How can that be?
Because the assumptions and beliefs themselves are
irrelevant to the communication: all that matters is that we come to agreement
on the beliefs we EXPRESS to one another.
Well, it happens probably because we **know** that we have
different
assumptions and beliefs, and when communicating, only assume that there is a
certain fundamental set of knowledge in common, and only
rely on that basic
set of common assumptions and beliefs to express the ideas we want to
communicate.
No, that doesn't work. First, people don't know this.
Second, how is agreement on this vital common 'core' achieved?
If we 'misunderestimate' what our fellow conversant
knows,
there can be and often is a miscommunication. The ability to communicate
effectively depends on the ability to guess correctly what facts,
assumptions, and beliefs are likely to be shared by those with whom we
communicate. Among specialists, of course, a lot more common technical
knowledge is assumed.
An issue that has occupied some of my attention lately has been
the
question of what basic ontological concepts are sufficient to support
accurate communication. I frame the issue as being analogous to the
"defining vocabulary" used by some dictionaries as a controlled
vocabulary
with which they define all of their words. For the Longman's, it is
around
2000 words. The analogous question is how many fundamental ontological
elements (types/classes, and relations/functions) are needed to logically
specify the meanings of all other terms used in a reasonably complex domain
(having perhaps 100,000+ terms), to some adequate level of detail?
Define "define". If you mean logically define,
then the number of defining terms is going to be comparable to the total number
of terms. All 'natural kind' terms, for example, are definitionally primitive.
Most ontology languages don't even support definitions. Why
are you so focused on definitions which usually don't, and often can't, exist?
And which if they did exist would be largely functionally irrelevant in any
case?
I don't
know, but I think that this is a question that is important enough to
warrant substantial effort. My guess is in the 6,000-10,000 concept
range,
and that many of those are fundamental enough to be common to many complex
domains.
Any other guesses?
See above. My guess is that most terms we use, both the
communicate with and to speak with, have no definitions and do not need them.
We eliminated definitions from KIF because they caused only harm and provided
no functionality.
Pat Hayes
Patrick Cassidy
MICRA, Inc.
908-561-3416
cell: 908-565-4053
cassidy@xxxxxxxxx
_________________________________________________________________
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
--
---------------------------------------------------------------------
IHMC
(850)434 8903 or (650)494
3973 home
40 South Alcaniz St. (850)202
4416 office
Pensacola
(850)202 4440 fax
FL 32502
(850)291 0667 cell
http://www.ihmc.us/users/phayes
phayesAT-SIGNihmc.us
http://www.flickr.com/pathayes/collections