WARNING: This material was written using the old nomenclature in which a
semlink was a link between semnods and a lexlink was a link from a word to
a semnod.  Under the new system, a lexlink is a link from a semnod to a
lexnod (from a node in the ontology to a node in the lexicon), a radlink
(radical link) is a link between semnods, and a semlink is a link from a
word to a semnod.  Sorry.

Chapter 0, Why Panlingua?  A firm foundation upon simple observations.

By Chaumont Devin, Honolulu, November1, 2001.

Is there some "universal field theory" capable of pulling all of the
various
branches of linguistics together into a coherent and organized whole?  Is
there
some way of understanding language that will explain how every part of it
works
in comparatively simple terms?  Is there some universal theory of
linguistic
structure that can be imposed upon all areas of linguistic research and
endeavor so as to show what is relevant and important and what is not, and
to
provide a new core of linguistic knowledge to which fresh information can
be
fitted and attached in such a way as to be immediately usable not only by
humans but also by machines? We shall presently see.  Meantime I shall call
this theory and the universal subsurface language it describes PANLINGUA,
from
the Greek word, PAS, meaning ALL, and the latin word, LINGUA, meaning
LANGUAGE
or TONGUE.

In any natural language it will be found that Multiple meanings exist for
most
words.  Each such meaning is called a "word sense."  As a typical example
take
the ordinary English word, JUMP.  Here is an "American Heritage Dictionary"
definition with each word sense assigned an ordinal number:

jump, v.  jumped, jumping, jumps.  intr.  1 To spring off the ground or
other
base by a muscular effort of the legs and feet.  2 To move suddenly and in
one
motion: jumped out of bed.  3 To move involuntarily, as in surprise: jumped
at
the knock on the door.  4 To respond or act quickly: Jump when I give you
an
order.  5 To spring at with the intent to assail or censure: jumped at me
for
criticizing her.  6 To join with or show eagerness: jumped into the race
for
the nomination.  7 To grab at readily: jump at a bargain.  8 To arrive at
hastily or haphazardly: jump to conclusions.  9 To move randomly or
aimlessly:
jumps from job to job.  10 To undergo a sudden and pronounced increase:
Prices
jumped.  11 To rise suddenly in position orrank: jumped over two others
with
more seniority.  12 To skip over space or material, leaving a break in
continuity: The lecture kept jumping from one subject to another.  13 To be
displaced by a sudden jerk: The phonograph needle jumped.  14 To be
displaced
vertically or laterally due to improper alignment: The film jumped during
projection.  15 Computer Sci.  To move from one set of instructions in a
program to another farther ahead or behind rather than moving sequentially.
16
To move over an opponent's playing piece in checkers.  17 To make a jump
bid in
bridge.  18 To show enterprise and quickness.  19 To have a lively,
pulsating
quality: a disco that really jumps.  tr.  20 To leap over or across: jump a
fence.  21 To leap upon: jump a bus.  22 To spring upon in sudden attack:
Muggers jumped him in the park.  23 To move or start prematurely: jumped
the
starter's gun.  24 To cause to leap: jump ahorse over a fence.  25 To cause
to
increase suddenly and pronouncedly: Unexpected shortages jumped prices.  26
To
skip: The typewriter jumped a space.  27 To promote, esp.  by more than one
level: jumped him to head foreman.  28 To take (an opponent's piece) in
checkers by moving over it with one's own.  29 To raise (a partner's bid)
in
bridge by more than is necessary.  30 To leave (a course) through mishap:
The
train jumped the rail.  31 Slang.  a To leave hastily; skip: jumped town
before
I could catch up with him.  32 Slang.  To leave (a position) in violation
of a
contract: jumped the team.  n.  33 The act of jumping; leap.  34 The space
or
distance covered by a leap: a jump of seven feet.  35 A hurdle, barrier, or
span to be jumped.  36 A track sport featuring skill in jumping: the high
jump.
37.  a A sudden, pronounced rise, as in price or salary.  38 An impressive
promotion.  39 A step or level: managed to stay a jump ahead of the others.
40
A sudden or major transition, as from one career or subject to another.  41
A
short trip.  42 One in a series of moves and stopovers, as with a circus or
road show.  43 In checkers, a move made by jumping.  44 Computer Sci.  A
movement from one set of instructions to another.  45 An involuntary
nervous
movement, as when startled.  46 jumps.  The fidgets.

(Taken from the "American Heritage Dictionary," with the numbering system
altered so as to use ordinals only without regard for part of speech).

Linguists refer to this phenomenon as "polysemy."

Except in cases where double or multiple meaning is deliberately intended,
such
as in poetry, proverbs, puns, jokes, etc., every word of every spoken or
written sentence is generally linked to but one and only one intended
meaning
or word sense.  For example, in the sentence:

Prices jumped on the New York Stock Exchange.

the intended meaning for "jump" is obviously sense #10 (to undergo a sudden
increase) in the above definition, and we know this immediately by looking
at
the other words in the sentence.

When taken alone, the word, "jump," is ambiguous because it can mean any of
at
least 46 different things, and therefore has NO real meaning.  The process
of
selecting this one and only one meaning from the 46 possibilities given for
"jump" in the above definition is called "word sense disambiguation," and
its
purpose is to determine the semantic role of "jump" in the above example
sentence.  It follows from this that the semantic role of each word in a
sentence is determined by the word sense to which it is linked by some
invisible connection.

The syntactic role of a word is determined by a similarly invisible
connection--this time not between word and word sense, but between word and
word.  Except for the main verb of a sentence, each word in every sentence
has
some other word in the same sentence called its "regent," or "head."  This
relationship between a word and its "regent," unfortunately, is much harder
for
most people to understand than the relationship between a word and its
word-sense definition, and so syntax (the relationship between words in a
sentence) has often become the subject of heated discussion.

The word order of a sentence contributes to our ability to determine both
the
semantic and syntactic role for each word in the sentence, but word order
itself is inconclusive.  A high percentage of words are located close to
their
regents, but it is also true that a word and its regent are often located
many
word positions apart.  Thus, although the word, "syntax," has often been
equated with "word order," and although in simple and rigidly controlled
languages like computer programming codes these may be the same, in fact
the
true meaning of "syntax" is not "word order," but rather "the invisible
linkages that exist between every word in a sentence (except the main verb
of
the sentence) and its regent."

The study of how words link to their regents is called "dependency
grammar,"
and is too abstruse and complicated a subject to be adequately covered
here.
Also, at the time of my last interactions with people involved in
developing
this study, there was a great deal of confusion surrounding many issues,
and
many of the participants seemed to be ignorant of even the most basic
facts,
such as the fundamental theorem I have mentioned above, namely that for
each
word in a sentence except for its main verb, there exists one and only one
other word within the same sentence that is its regent.  And even now I am
convinced that if there are any dependency grammarians reading these
arguments,
at least 50% of them will immediately rise with righteous indignation in
their
breasts to challenge this simple and seemingly obvious theorem!  But I will
ignore them lest they stop me in my tracks.  And for those of you who have
been
convinced about the linkage between word and word sense but remain
unconvinced
about the linkage between word and regent, I will refer you, albeit with
considerable trepidation, to the literature that has been written on the
subject of dependency grammar.  The only thing I will say as a Parthian
shot in
my own defense is that simple people do dote upon complexity and loath
simple
theorems capable of collapsing the complex structures they devise.

So for every word in every coherent sentence that has ever been spoken,
thought, or written, we have a semantic role and a syntactic role, and
apparently nothing else.  And even the most cursory analysis will be
certain to
reveal that each of these roles is fully determined by a single link--a
link to
some word sense for the semantic role, and (except in the case of the main
verb
of the sentence) a link to some other (regent) word in the same sentence
for
the syntactic role.  Then, if we just barely nudge this line of reasoning a
little further, we will come face to face with the following astonishing
theorem:

Every word in every sentence is simply a pair of links emanating from a
common
point, or node.

The point from whence these two links spring has no shape, form, meaning,
or
other identifying characteristic of any kind.  Its only claim to existence
are
the two links locking it into the greater structure of language.  But these
links are invisible, as we have seen, and so also are all true points in
space
and time.  Thus every coherent written sentence, with all its words made up
of
alphabetical characters or other symbols arranged in a line, is simply the
visible representation of some invisible structure of links and nodes in
our
minds, and this invisible structure is Panlingua.

Many arguments have been propounded for and against a universal subsurface
language common to all of mankind.  But these have often been couched in
terms
more characteristic of abstruse religion than hard science.  They have
failed
or else never even attempted to work out the details of what such a
language
might look like or how it might work.  Such arguments convey an impression
of
overwhelming complexity, whereas the whole Panlingua idea is
underwhelmingly
simple, as I have just shown.  Yet its obvious ramifications for the fields
of
linguistics and computer science are enormous, so that it cannot safely be
ignored.  It is imperative, therefore, that this theory of Panlingua should
be
clearly understood, and this is why I am writing this manuscript for you.

Chapter 1, What the Theory covers.

By Chaumont Devin, Honolulu, May 7, 1998.


The Goal of This Manuscript.

My purpose here is to show how all the inner workings of the human
linguistic apparatus can be represented in terms of simple links and nodes
which can readily be modeled in binary computers.  My approach will be to
establish fundamental assumptions based upon observations of human speech
behavior, and to use these as tools to uncover those inner workings of the
brain heretofore hidden by natural processes of linguistic interpretation.

Although a great deal of the current terminology is appropriate and usable
for Panlingua, it will soon be apparent that my approach has not been to
build carefully within the framework of established linguistic traditions.
This work is not intended as an exercise in classical linguistics.
Instead, I have dealt with the whole problem of language as a computer
programming challenge, taking nothing for granted except what I am able to
prove by modeling upon automated systems.  It was by this means, and not
by linguistic training, that I came ultimately upon the theory (and I
believe there is really only one such theory) underlying all
manifestations of language and linguistic intelligence that can be run on
machines, be they biological, digital, or whatever.  To the underlying
principles of this theory, I believe, classical linguistics must
ultimately conform, and this conformance will require a good deal of
painstaking effort.  Fortunately, these principles are simple, and easily
verifiable by independant observation, so that once the linguist has
grasped them, it should be easy for him/her to see how they apply to
his/her special field.  My particular interest not being linguistics but
artificial intelligence, I would prefer to leave the finer linguistic
ramifications (of which there should be many) to be worked out by others
better qualified than myself.  The goal of this manuscript is therefore to
introduce this theory of Panlingua as simply and clearly as I can--not to
show off any superior knowledge of classical linguistics, but to use just
enough of the current terminology as to keep the linguist on track without
losing the non-linguist computer programmer or other interested lay
person.

Symbols and meaning.

It is said that a picture is worth a thousand words.  What is often
forgotten is that a single word can be worth many thousands of pictures.
As an example suppose you were to enter the world's most comprehensive
library with a single word in your mind, and that this word was "bird".
You might find literally thousands of books dealing with this subject, in
each of which were many images, paintings, and photographs.  And when you
stop to think about it, it may seem quite remarkable that so much
knowledge could be accessed through just that one word.

As it turns out, words are also worth many millions of computer pixels and
many hours of computer data processing time.  Just think, for example, of
how many pixels it might take to represent an identifiable cockroach.  Not
only would it be necessary to use enough pixels to create the general
outline of a cockroach, but it would also be necessary that the image of
the cockroach include enough unique characteristics to distinguish it from
other similar insects, such as certain beetles.  And yet the same entity
can be identified by the single written word, "cockroach", or just the
two-syllable sound, or even (as we shall see later) by a single integer
value.

Indeed, the sound of a word like "bird" or "cockroach" will start a
process of image building in our minds.  We may suddenly "see" a cockroach
in some vague manner in our minds.  Then we may remember that cockroaches
have six legs, and provide the image our minds have created with these
appendages, and so on.

At this level a human word is not much different, say, from a particular
kind of grunt or squeal made by a baboon.  Many animal species are known
to use a variety of such sounds to warn of the presence of particular
kinds of danger, using different sounds to indicate different predators,
etc.  The mechanism is presumably the same.  The sound evokes the image of
the associated predator in the brain of the hearer, and the hearer takes
the appropriate precautions.  Thus a single grunt, bark, squeal, etc., can
evoke a whole array of facts, images, memories, etc., for example about
lions.  The animal or human hears the sound, recognizes the sound as a
symbol, and activates the part of its brain that responds to the
identification of that symbol.

In summary then it can be said that certain sounds can activate certain
regions of certain brains.

Syntax.

So far we have discussed cases in which the direction of activation is
either straight in, as when a single identifiable sound is heard, or else
straight out, as when an animal sounds a warning to its comrades.  No
lateral activation is involved.  No links are activated in any sideways
direction perpendicular to the incoming or outgoing signal.  No syntactic
element is involved, only the semantic association of a sound with some
region of the brain.

But in higher animals it can be seen that it is possible for certain
symbols to be clustered together in meaningful ways.  For example the
human sounds for "eat",  "monkeys"  and "bananas"  might be arranged in
such a way as to indicate that "Monkeys eat bananas."

To better understand what is going on, let us imagine the human brain as
consisting of horizontal planes, or layers, with gaps between them.  We
might then have the following:

1.  Top layer, the lexical or phonological plane.  Here complex sensory
patterns such as sights and sounds are identified and used to activate
single nodes.  Each node in this plane corresponds to some unique pattern
that can be perceived in the real world.  The perception of a pattern
activates its appropriate node, after which the actual pattern that
activated the node can be forgotten, and all further processing can be
carried out in terms of nothing else but links and nodes.  The nodes of
this plane are called lexical nodes, or LEXNODS.

2.  Second layer, the syntactic plane.  The nodes of this layer are words.

3.  Third layer, the semantic plane.  The nodes of this layer are called
semnods, or "semantic nodes".  Each semnod can be defined by a single
word-sense definition such as might be found in any standard dictionary.

4.  Fourth layer, the lower brain.  Here reside the various regions of the
brain that handle stimuli associated with various symbols.  Please note
that our use of the term, "lower brain" may not coincide with its use in
brain physiology.

The particular word order required to represent a particular meaning will
differ from language to language, but the meaning will remain the same.
Let us call this word order "surface syntax".

As an example to illustrate how word order may vary from language to
language to represent the meaning, "Monkeys eat bananas", let 'S' stand
for "subject", 'V' stand for "verb", and 'O' stand for "object".  Then we
might have the following different languages and syntaxes:

Language A: Monkeys eat bananas.  (SVO)

Language B: Bananas eat monkeys.  (OVS)

Language C: Eat monkeys bananas.  (VSO)

Language D: Eat bananas monkeys.  (VOS)

Language E: Monkeys bananas eat.  (SOV)

Language F: Bananas monkeys eat.  (OSV)

Human language consists of symbols that translate to sounds, images, and
other sensory patterns, and all the links that can exist between these
symbols and the regions of the brain that handle them.  Needless to say,
in human language the patterns created by such symbols and the links
between them can quickly become very complex, and this is why we need some
simplifying model like that given above in order to understand them.

In summary we might observe that in higher animals, and especially in man,
besides the pathways that go straight into and straight out of the brain,
such as (in the inward direction):

sound -> word -> semnod -> associated lower brain function

and (in the outward direction):

lower brain function -> semnod -> word -> sound,

there also exist lateral links between words that manifest themselves in
surface syntax.  Furthermore we might observe that words can stand in
various relationships to one another at some deeper level, but that these
deep relationships are not completely determined by the linear order in
which the symbols occur (or in other words by surface syntax) since this
surface word order may vary from language to language.  And we might
deduce that there exists some kind of subsurface syntax that holds words
together without regard for the word order of surface syntaxes.  This
subsurface syntax is commonly called "word dependency" by linguists, and
is in fact the substratum upon which all surface word order is built.

Word Dependency.

As we have seen, words are the nodes of the syntactic plane.  In fact all
texts (grammatical strings of words) should be thought of as existing in
the syntactic plane.

Then word dependency is the way in which words relate to one another
within the syntactic plane.  Words are said to "depend" (hang) from other
words higher in GRAMMATICAL IMPORTANCE, or RANK.  For example, in the noun
phrase, "blue sky", the thing that is being talked about is "sky", and so
"sky" is said to be higher in rank than "blue".  The word, "blue" is then
said to depend from "sky".  The word, "sky" is then called the "regent"
and the word "blue" is called its "dependent".

As we have noted, word order may vary from language to language.  Thus,
for example, in Malay, the word order for "blue sky" would be "sky blue",
and for "big house" the correct Malay syntax would be "house big".
However the dependency between these words remains the same for any
surface language.  "Big" depends from "house" and "blue" depends from
"sky".  This fact, commonly recognized among most linguists, is strong
evidence for the existence of a universal subsurface language whose syntax
is not word order but word dependency.

Now let us imagine two African warriors walking across the savannah.  One
of them makes out a lion lurking in the grass.  The sight of this predator
triggers a response from the region of his brain that deals with lions.
Part of this response is a warning to his comrade.  "Lion!" he says, and
both men suddenly come to a standstill, their spears poised for attack.
What has happened from a linguistic point of view?

1.  Warrior #1 recognizes "lion", which activates the region of his lower
brain that deals with lions.

2.  The "lion" region of the lower brain of warrior #1 sends an impulse
out to the phonological plane of his brain where it is translated into the
audible word that is the symbol for "lion" in his language.

3.  Warrior #2 hears the word and recognizes it as the symbol for "lion",
which means that the "lion symbol" region of the phonological plane of his
brain becomes activated.

4.  The activation of this region of the phonological plane of warrior #2
sends an impulse into the region of his lower brain that deals with lions,
and suddenly he is also able to see the lion camouflaged in the grass.

In the above explanation we have deliberately ignored the semantic and
syntactic planes for purposes of simplification.  There is no syntactic
component to the utterance of the single word "lion".  A signal passes
straight up out of the lower brain of warrior #1 and straight down into
the lower brain of warrior #2.

But what if our two warriors were to happen upon the freshly killed
carcass of a zebra, and warrior #1 were to say to warrior #2, "This zebra
was killed by a lion"?  What major linguistic linkages would be involved
then?

Before we tackle this problem let us establish some definitions:

A link is a linguistic relationship between two linguistic nodes.

A synlink (syntactic link) is a dependency link between two words.

A lexlink (lexical link) is a link between a word and a semnod.

A semlink is a link between two semnods.

In general terms the processes involved are similar to those in the
preceding example.  Signals pass up out of the lower brain of warrior #1,
get converted into symbolic sounds, get heard and interpreted as symbols
by warrior #2, and pass into the lower brain of warrior #2.  But this time
more than one symbol is involved, and so we have the phenomenon of syntax.
The symbols occur in a certain linear order (or sequence) determined by
surface syntax, and the words relate to one another in certain ways
determined by subsurface syntax, or word dependency.

It is easy to determine the word order of surface syntax just by listening
to speech or by reading printed texts.  Regarding the functions of the
lower brain, on the other hand, although we know there must be various
regions that do various things, at this point very little is understood
about them, and no adequate means has been devised for their study, so
they remain a sealed book.  But by subjecting surface syntax to various
processes of deduction it is comparatively easy to identify and map out
linkages in the syntactic and semantic planes.  So instead of remaining
engaged in the seemingly endless analysis of surface syntax phenomena for
surface syntax' sake, or of attempting to delve directly into the
intricacies of the lower brain without understanding anything about the
way brains work, let us now proceed by focusing our energies upon this
more immediately productive and promising research.  Our goal will be to
map out ALL the linkages that occur between nodes in the syntactic and
semantic planes, and to investigate the things that can be done by
manipulating these links  in various ways.

"This zebra was killed by a lion."

In the following explanations, the words in capitals represent Panlingua
atoms that reflect the same words in the sentence just given above:

THIS is linked to ZEBRA by a synlink of type "determiner".  ZEBRA is
linked to KILLED by a synlink of type "patient".  A is linked to LION by a
synlink of type "determiner".  LION is linked to KILLED by a synlink of
type "agent".

All of the above links are between Panlingua atoms, or words, and lie
completely within the syntactic plane.  The following links connect words
to their semnods.  Only lexlink TYPE is given, since the semnods
themselves are already identified by the words that link to them.

THIS has a lexlink of type "default".  ZEBRA has a lexlink of type
"singular".  KILLED has a lexlink of type "past-tense declarative
transitional".  A has a lexlink of type "default".  LION has a lexlink of
type "singular".

In the above lines, "default" simply means "no special type in
particular".

The synlinks and lexlinks given above define the Panlingua representation
of the sentence, "This zebra was killed by a lion".  The fact that the
surface syntax was in the passive voice was not included in this Panlingua
representation because the meaning is exactly the same as "A lion killed
this zebra" where no pronominal references are involved.

In summary, ALL thoughts are constructed by setting up synlinks and
lexlinks between nodes in the syntactic and semantic planes.  In this work
we will examine the details of these linkages, how they can be modeled
using present-day computers, and what a better understanding of them will
mean for the future of linguistics, computer science, and other fields of
research.

Chapter 2, The Ontology: Ultimate Foundation of ALL Knowledge.

By Chaumont Devin, Honolulu, May 7, 1998.

The most comprehensive traditional grammars and dictionaries only skim the
surface. --Noam Chomsky.

In the last chapter we had a detailed look at how "A lion killed this
zebra" might be stored as a Panlingua representation using nothing more
than a collection of elementary links and nodes.  In fact ALL Panlingua
structures are built from nothing but various kinds of synlinks and
lexlinks, and there is no thought that cannot be represented in Panlingua.

But is Panlingua all there is?  Is there any other way in which knowledge
can be stored in the three-dimensional model we have discovered? The
answer is YES.  Knowledge can also be stored in SEMLINKS, which we have
not implemented thus far.  Semlinks link semnods to each other, and lie
completely within the semantic plane.

Another name for the collection of all semlinks and semnods is "the
ontology".  The word "ontology" comes from the Greek words "ontos" and
"logos", and means roughly, "the study of what IS", or "words about what
IS".  As it turns out this is a very good name, because the major link
types of the ontology (semlink types) correspond closely to English
auxiliary verbs (recall that the main such verb is "to be" which in Greek
is "ontos").

Now imagine with me that an African woman has walked down to the stream at
the edge of the village with her child.  The child sees a crab scurry away
and hide among the rocks.  "What's that?" she asks her mother, pointing.

"A crab," her mother responds.

"What is a crab?" her daughter persists.

"A kind of animal."

"Do crabs bite?"

"Yes."

"Do they hurt?"

"A little bit."

"Can we eat them?"

"Yes."

Etc.

What is happening here is the fundamental learning of fundamental
knowledge by a child.  Almost all of this information can be stored in an
ontology, where it is evidently more permanent and can be accessed more
rapidly than if it were stored in Panlingua arrays.

I have created a computer program that serves as a lexicon and ontology.
Using such a program one might enter the following:

crab sng crabs plr * (asterisk means "of same") crab isa animal crab can
bite crab can not injure

crabs are edible etc.

In this mechanized example the word in the middle is a mnemonic.  In my
program it is always three letters chosen to help humans remember.  When
referring to the ontology such link types are also sometimes called
operators, and the two words they link are sometimes called operands.

***** Note that the class of semnods called operators can be expanded to
include all verbs and prepositions, but whether this is the case in real
human ontologies remains unclear.

Whenever such ontological entries are made, either (1) a word is defined
and assigned a semnod, as in:

crab sng

or else (2) a semlink is forged between two previously-established
semnods, as in:

crabs can bite

The properties of such ontologies have only begun to be explored, but have
already proven to be quite amazing.  One of the greatest strengths of
ontological structures is their ability to "grasp" the ramifications of
new information.  For example, suppose an ontology is up and running, and
the following entry is made:

William is a man

Then the following interactions should immediately be possible:

Can William talk Yes

can William menstruate No

can William give_birth No

has William testicles YES

has William vagina No

and many other such things.

This power to subsume and know many things about new items without ever
being told enables massive amounts of information to be stored using
minimal computational resources.  In fact an ontology can ideally return
relational information on all semlink types for n(n-1) word pairs, where n
is the number of words in the lexicon.  Thus for only 10,000 words and 10
semlink types, an ontology can return information about 99,990,000 * 10 =
999,900,000 combinations.  The ontology is thus by far the most powerful
knowledge storage medium ever known to man.

But ontological knowledge, although very fast and very massive, is limited
strictly to knowledge about binary relations.  Thus it is possible for an
ontology to know that "Roses are red" but it is not possible for an
ontology to know that "John can dance the jitterbug".  In the ontology,
knowledge is organized into a sort of rough-and-ready instant reflection
of the real world, but an ontology cannot be used to store specific
knowledge about people, places, and events.

***** Note that if any verb is permitted as an operator, then "John dance
jitterbug" can be recorded, indicating that John can dance the jitterbug.

So ontological knowledge might be viewed as the foundation of all
knowledge, because it exists within the linguistic planes but at a
sublinguistic level.  In other words, Panlingua is a language--indeed, the
mother of all languages, but ontological information is not.

Do animals possess ontologies?  Almost certainly.  In fact it may be that
ALL animals possess minimal ontologies of some sort or another.  Otherwise
how could animals know that:

item_a is edible item_b is not edible item_c is edible etc.

And many animals can probably construct and manipulate Panlingua
structures as well.  As a matter of fact it has been shown that predatory
insects such as spiders can carefully plan and execute their attacks.  It
may be that they employ something other than Panlingua, but this would not
seem likely for reasons we shall later explore.

But it is important to notice that not all ontological operators
(semlinks) are created equal.  Certain semlink types are of crucial
importance, and must be present in nearly all ontologies.  The first of
these is probably the hypernym link (something is a kind of something).
The second is probably the holonym link (something is a part of
something).  Then antonym, synonym, and others.  And some semlinks must be
handled in special ways.  For example, X is a hypernym of A if A is a B
and B is a C and ...  is an X.  And A is a holonym of B if A has a B, A is
something that has a B, or A has something that is a B.

Among other indispensable things, the hypernym link enables a great deal
to be said about many words using only a relatively few entries.  The
holonym link also enables many questions to be answered using minimal
data.

To understand how ontologies can return such powerful results, let us
examine briefly how they work.  Suppose we have the entries:

man isa human human isa ape ape isa simian simian isa primate primate isa
mammal mammal isa vertebrate vertebrate isa animal animal isa organism
organism isa material_object material_object isa material_thing
material_thing isa thing man isa male_mammal male_mammal has testicle
woman isa human woman isa female_mammal female_mammal can give_birth
animal can eat William isa man

Such an ontology would "know" that William can eat because:

William isa human human isa ape ape isa simian simian isa primate primate
isa mammal mammal isa vertebrate vertebrate isa animal animal can eat

And it would know that William can't give birth because there is nothing
above William in the hypernym hierarchy that can give birth.  But it would
know that William has testicle because:

William isa man man isa male_mammal male_mammal has testicle

>From these examples it can be seen that if A is a hypernym of B, and A
can do something, then B can do that same thing also unless specifically
stated otherwise.  And if A is a hypernym of B, and an A has something,
then a B has that something also.  But if A is a holonym of B, and B is a
holonym of C, then an A has a C and a B has a C, and so on.

The hypernymy hierarchy serves mainly to scramble up and down within an
ontology looking for other relations.  A is a hypernym of B if B is a kind
of an A, and B is a hyponym of A if b is a kind of an A.

For holonymy and meronymy, on the other hand, A is a holonym of B if an A
has a B, and A is a meronym of B if a B has an A.  Thus "brake" is a
meronym of "automobile", and "automobile" is a holonym of "brake" (cars
have brakes).

So a hypernym is the opposite of a hyponym, and a holonym is the opposite
of a meronym.

An important use of the holonym/meronym relation is to discover the topic
of discourse for a segment of text.  Thus for a sentence like:

He adjusted the brakes and set the timing.

a check of the ontology might reveal the following semlinks:

a brake is part of a wheel a wheel is part of an automobile timing is part
of an engine an engine is part of an automobile

Thus from the ontology we might find that automobiles have both brakes and
timing, and surmise correctly that the topic of discourse of this sentence
is an automobile.

Besides its role as a repository of world knowledge, the ontology plays a
crucial role in parsing without which certain kinds of disambiguation
would be virtually impossible.  As an example, take the word "had" in
"Martha had a baby in the hospital".  Does this mean that Martha possessed
or that Martha gave birth?  Well, if the ontology "knows" that Martha is a
woman, and if in the general Panlingua reference there exists an entry
like "Women give birth in hospitals", then by traversing the general
reference using the ontology the system should be able to figure out that
"Martha had a baby in the hospital" probably means that Martha gave birth.

The ontology also plays an important role in search-match operations in
which identical words are not used.  For example, if a Panlingua reference
says that "Automobiles travel on roads" and it is searched to learn
whether a Chevrolet can travel on Highway 99, such a search should return
"Yes" because the ontology says that a Chevrolet is a kind of automobile
and Highway 99 is a particular road, etc.

Ontologies tend to differ from person to person, and multiple ontologies
may exist for certain people.  To illustrate what I mean, I was once
fascinated by a Philippino woman who lived in California.  She is surely
one of the most brilliant people I have ever known.  But I observed three
basic personalities in her.  She was a smart Philippino with a perfect
command of Tagalog, and when she spoke Tagalog, her whole being went into
Philippino gear, and she became a talkative Philippino lady.  But in
English she had two personalities.  One was the very shrewd, very
professional businesswoman, and another was plain Miss Everyday American.
If I had to guess, I would quickly say that she must have had a separate
ontology for each of these three modes.

As another example, suppose that one man is a criminal.  He believes that
stealing is good because it brings him satisfaction.  The semnod linked to
"steal" in his ontology is linked by hypernymy to "good thing".  There is
another man who is honest, and believes that stealing is a crime.  In his
ontology, the semnod that links to "steal" is linked by hypernymy to the
semnod linked to "bad thing".  The two ontologies are not the same
primarily because their link patterns differ.

Ontologies also differ from language to language.  For example, in some
Southeast Asian languages, green things and blue things employ the same
semnod.  Thus for a native speaker of such a language learning English,
he/she may be crossing the street when the light is "blue" for many years.

But once reliable ontologies have been built for two languages, bridging
relations can be forged between the semnods of the two ontologies for use
in machine translation.

Lastly ontologies can be made to self economize by the following rules:

If A has a C and B has a C and a B is a kind of an A, then break the "B
has a C" link.

If an A has a C and an A has a B and a B has a C, then break the "A has a
C" link.

Etc.

So to summarize, ontologies are built from nothing but semlinks--that is,
links between semnods, or semantic nodes.  They contain massive amounts of
the most basic kinds of knowledge, and no truly human language or
intelligence is possible without them.  They differ largely from culture
to culture, and to a lesser degree also from individual to individual
within the same culture.  Each semnod in an ontology represents an
unambiguous meaning, and can be defined precisely by a single definition.
The definition for any semnod is simply a word-sense definition for some
word linked to that semnod.  Ontologies are much more effective than
dictionaries in that besides simply defining word senses or semantic
nodes, they also define the relationships between them, thus generating
powerful dynamic models of the real world.

Chapter 3, The Lexicon: Interface with the Outside World.

By Chaumont Devin, Honolulu, May 8, 1998.


What is a lexicon?  There will be many answers, but in the context of
Panlingua we have the following definition:

A lexicon is a collection of lexical nodes, or LEXNODS.  Each lexnod is
linked
by lexlink to one or more semnods.  Furthermore, each lexnod is linked to
one
or more physical patterns, or symbols, in the real world.

***** Note that here we have used the same word (lexlink) to specify the
link
between a lexnod and a semnod as we have used to specify the link between a
Panlingua atom and a semnod.  This probably should not have been done,
because
these two kinds of linkage are really quite different from each other.
Although it is rather late in the game to do this now, I am thinking of
redefining the names of the links of Panlingua as follows:

Lexlink: link from lexnod to semnod.

Semlink: link from Panlingua atom to semnod.

Synlink: link from one panlingua atom to another.

Sublink: link from one semnod to another.

At present, however, I feel great reluctance to do so, because the old
terms
(LEXLINK, SYNLINK, and SEMLINK) have worked so well.

By some means or other, the animal brain is capable of reducing physical
patterns to lexnods at incredible speeds.  As an example, the sound of the
word, "bird," the written word, "bird," a Chinese pictograph representing a
bird, a tiny picture of a bird, and the sight of a real bird all seem to be
capable of activating the lexnod associated with "bird."  As partial proof
of
this statement, just write the sentence, "A bird flew over the house."
Then
erase the word, "bird," and replace it with a small drawing representing a
bird, and show the altered sentence to a child.  The child will probably be
able to read the altered sentence without any trouble.  Furthermore, the
human
brain is capable of interpreting what seem to be an infinity of variations
on
the perceived patterns for "bird" without the slightest difficulty.  Just
how
such feats are accomplished remains an unsolved mystery.

But since computers are fundamentally stupid, and hardly capable of
anything at
all where pattern recognition is concerned, for purposes of computational
linguistics, we must limit ourselves for the present to direct links from
lexnods to clusters of characters forming word symbols in plain text files.
Some speech recognition systems have been developed, but at this writing
they
are still too unreliable for most serious applications.

The simplest--and hence the best--way of linking each lexnod to an English
word
symbol is simply to put the word symbol on a line by itself in an English
word
list, and use the ordinal value representing the position of the word
symbol in
the list as the integer representing its lexnod.  Thus for the following
sample
list:

this
here
I
bird
through

the lexnod for "this" would be lexnod #1, the lexnod for "here" would be
lexnod
#2, etc.

Another form of lexicon simplification is to use part of speech for lexlink
type.  Just as lexlink type can limit synlink type (what kinds of syntactic
linkages a word can have), so also part of speech can define potential
lexlink
types for a language.  As an example, for English, that class of words
ending
in "-ing," generally known as "present participles," can potentially have
lexlink types of noun, adjective, present participle, etc., depending on
syntax.  By taking advantage of this kind of knowledge for a particular
language, a great deal of lexicon space can be saved by using part of
speech as
lexlink type instead of providing a separate lexlink for each potential
lexlink
type (a three to one saving in our example).  The tradeoff is that more ad
hoc
programming will be necessary to process such part-of-speech entries in the
lexicon during parsing and text generation.

***** Notice that in the above paragraph the word, LEXLINK, is being used
loosely to mean either the link between a Panlingua atom and a semnod or
the
link between a lexnod and a semnod.

Of course lexlink type (the type of a link from a lexnod to a semnod) is a
kind
of part of speech, but of much finer grain, and thus of much greater
precision,
than say the "standard" eight parts of speech often given for the English
language.

But notice that each lexnod is tied to one or more complete physical
patterns
or symbols in the real world.  This implies a distinct and separate lexnod
for
every "form of a word," without regard for morphology.  Thus no lexicon
entry
can ever be a prefix, suffix, or other part of a word, and the lexicon must
contain separate entries for all of the following: eat, eats, eating,
eaten,
ate, eatable, etc.  Except in certain rigidly structured languages,
therefore,
morphology is useless as anything but a mnemonic hint for deducing the
probable
linkages for new words.  A missionary friend of mine once demonstrated this
to
his own dismay by attempting to make "shyness" out of "shy" using the
standard
morphological transformations.  What he ended up with was "genitalia,"
which
turned out to be a most unforgetable mistake!  He was leading the
"testimony
service" during an evangelistic crusade on the island of Ternate.  When he
failed to get the people to jump up and "testify," he asked them if their
genitalia (not shyness) had them bound to their benches!

The lexicon is the gateway through which internal linguistic systems are
linked
to the outside world.  In other words, on a Panlingua-based system, all
internal functions could hum along just fine without any lexicon, but it
would
be impossible for the system to communicate with the outside world in
either
direction.  So Panlingua cannot function without an ontology, but it CAN
function just fine without a lexicon as long as no output is required.
Thus a
person might think to himself, "I wonder what he is doing now?" and
actually
set this thought up internally in Panlingua, but never need to use the
lexicon
in his own brain unless he wishes to vocalize the words.

A symbol is a sound, image, or other physical pattern that represents
something else.  Spoken and written words are symbols, but the atoms of
Panlingua are NOT.  Nor are the semnods of the ontology or the lexnods of
the
lexicon.  In the following paragraphs I will generally use "symbol" instead
of
"word" for spoken or written words.  This is because I have also called the
atoms of Panlingua "words", and I wish to distinguish spoken and written
words
from them.

Parts of speech are word classes that seem to have evolved naturally in
spoken
languages over time.  Most languages have such classes as nouns, verbs,
adjectives, and adverbs, but there is no guarantee that all the parts of
speech found in one language will be found in another.  For our purposes,
we
will adhere to the following definition:

Part of speech is a classification that limits a symbol to a particular set
of
synlinks and lexlinks.

As an example, let us take "noun".  If we know that a symbol is a noun,
then
we know that this symbol can be linked to its regent (dependency grammar)
by a
synlink of type agent, patient, prepositional object, etc.  Part of speech
doesn't tell us which of these synlink types a noun will have in a
particular
sentence.  It only tells us what kinds of synlink a noun CAN have and what
kinds of synlink it CANNOT have.  Lexlink type (the type of a link from a
lexnod to a semnod) is really just part of speech brought to greater
precision.

A word is a linguistic node having one and only one synlink to a regent and
one and only one lexlink to a semnod.

When words are defined in this way, the invisible atoms of Panlingua are
true
words, whereas the physical word symbols of the external world are only
symbols
representing words.  As an example let us take the sentence:

Rust eats iron.

A quick check of the lexicon will tell us a multitude of things.  Among
others
that:

RUST is a noun linked to a semnod linked by synonymy to another semnod
linked
to CORROSION.

RUST is a noun linked to a semnod linked by holonymy to a semnod linked to
FUNGUS.

RUST is a verb linked to a semnod linked to RUSTS, RUSTED, and RUSTING.

EAT is linked to a semnod to which another semnod linked to ANIMAL is
linked
by a link of type CAN (Agency.  Animals CAN eat.).

EAT is linked to a semnod linked to CORRODE.

Etc.

But as if by magic we immediately know that this kind of rust is not a
fungus,
and that this kind of eating is not something that animals do.  How our
brains
do this at all, and how they can do this with such great accuracy and
speed,
is probably the greatest mystery in linguistics today. We know part of how
this is done, but only part, and there are almost certainly some very
important things we are missing.

Whatever the case, as speakers of English we immediately know that for
"Rust
eats iron", we have the following two links for each word:

A lexlink links RUST to the semnod linked to CORROSION, and a synlink of
type
"agent" links RUST to EATS.

A lexlink of type "transitional repetitive" linksEATS to a semnod linked to
CORRODES, and a synlink of type "declarative" links EATS to some regent
word.

***** Note that EATS, being top verb, may not actually link to any real
word,
but to a dummy word whose only function is to provide the synlink from EATS
with somewhere to go.  In a sentence like "He says that rust eats iron,"
EATS
would link to THAT.  But for top verbs, there is no real regent word to
link
to.

After disambiguation, a lexlink links IRON to a semnod linked by hypernymy
to
METAL, and a synlink of type "patient" links IRON to EATS.

Because we are intuitively aware of the existence of these linkages, we
call
these written and spoken symbols "words," even though they are really not
words
at all but only the physical symbols representing words.  As we have seen,
Panlingua has no symbols, and consists only of synlinks, lexlinks, and
nodes.
Yet the words or atoms of Panlingua, also called universal atoms of
meaning,
have precisely the same synlinks and lexlinks as the ones we can recognize
in
the spoken and written symbols we call words.  Thus Panlingua may be
thought of
as a "distilled" form of language without the symbols and linear word order
characteristic of texts.  Or again Panlingua may be seen as the structural
framework upon which all surface languages are formed.  The fortunate thing
for
US is that Panlingua can easily be modeled on computers using structures
representing links and nodes.

Because each word symbol in a sentence like "Rust eats iron" may have more
than
one potential lexlink and more than one potential synlink, such sentences
are
said to contain "ambiguity", and to require "disambiguation".  The process
of
disambiguation, or selecting just one and only one synlink and just one and
only one lexlink for each word, is also known as "parsing".  Thus parsing
is
the conversion of written or spoken texts to Panlingua representations by
means
of disambiguation.

The need for a lexicon in parsing is obvious.  Without the lexicon there
would
be no linkage between the external symbols of written and spoken language
and
the semnods of the ontology, and thus no means of converting these symbols
to
Panlingua representations.  Through the lexicon each external symbol is
linked
to one or more semnods.  During the parsing process, all extraneous
lexlinks
are culled leaving just one and only one lexlink, and just one and only one
of
the possible synlinks is selected based on the part of speech returned for
the
symbol by the lexicon.

The same is true of the process of creating strings of written or spoken
words
from Panlingua representations.  For each Panlingua atom a search is made
of
the lexicon in order to find an external symbol linked to the same semnod
as
the Panlingua atom by the same kind of lexlink and having a part of speech
that would allow the symbol to be used in a way reflecting its position and
role in the Panlingua structure.  This process, known as "text generation",
is
much simpler than parsing because Panlingua structures are essentially
unambiguous.

For some collocations, such as "house cat," it may at first seem reasonable
to
create lexicon entries such as "house_cat" or "house-cat."  But further
examination of the problem will show this kind of solution to be a kludge,
and
not the universal answer that will be required to handle compound words for
the
entire English language.  The right approach seems to be to create lexlink
types of "left coupler" and "right coupler."  Then both "house" and "cat"
will
lexlink to the same "house cat" semnod, but "house" will have a lexlink
type of
"left coupler" instead of "noun."

Besides just handling simple compound words like "house cat" and "grey
matter,"
which never admit any intervening words between their two parts, this
approach
can also be made to handle examples such as:

come about
comes about
coming about
came about

For each of these we might also have constructions like:

come slowly and carefully about
comes quickly about
coming awkwardly about
came quietly about
etc.

In each case, ABOUT is just a right coupler linked to the same semnod used
(linked to) by this sense of COME.

A similar problem arises for compound names such as "The United States of
America."  The solution seems to be to create a semnod for the USA, and
lexlink
types of name1, name2, ..., name9.  Then all of the parts of such a name
can be
lexlinked to the same semnod, as follows:

the name1
United name2
States name3
of name4
America name5

Now suppose we are using a computer Panlingua-based lexicon/ontology, and
we
make the following entry at the keyboard:

roses are red

What happens?

***** Note that for the following we are using our the lexicon-ontology
(SEMLEX), and not the English parser.

First of all, for the computer even to accept this input it will be
necessary
that "roses", "are" and "red" already be defined in our lexicon and
ontology.
"Are" is linked to a semnod used as an an operator (see Chapter 2) in our
ontology.  Roses and red have been defined as follows:

rose sng (ROSE is a singular noun.  Create a semnod for it.)
roses plr * (ROSES is a plural noun.  Use same semnod as last entry.)
red adj (Red is an adjective.  Create a new semnod for it.)

The entry:

roses are red

then causes the following operations (in simplified form):

Find a noun lexlink (link from lexnodto semnod) for ROSES in the lexicon.
Find the semnod to which this lexlink connects.
Find an adjective lexlink for RED.
Find the semnod to which this lexlink connects.
Forge a semlink of type ARE from the first semnod to this one.

Then when the following query is entered:

are roses red

for each semnod to which ROSES is linked a search is made to see if a
semlink
of type ARE exists linking it to another semnod linked to red.  If such a
semlink is found, the system returns "Yes".  If no such semlink is found,
the
system returns "No".

As you may have gathered from the foregoing, the implementation of such
systems of links and nodes on a binary computer can be less than
straightforward.  It may be that someday we will have computers better
adapted
to this kind of modeling.  And yet, even with current computer technology
there is no appreciable delay when all these components are linked
correctly.

As far as I can understand them at this point, these are the basic
components
of systems using Panlingua:

1. The ontology.

2. Various arrays of Panlingua itself.

3. The lexicon.
4. Another kind of data structure used in parsing (still classified).

To these must be added a parser and text generator for interaction with the
outside world.

All further components of the human linguistic apparatus, as best I can
understand them, are encoded as long Panlingua arrays.  In other words,
after
all of the information that can be stored in the lexicon and ontology have
been
stored there, then whatever else is left over must be encoded in the form
of
Panlingua arrays.

Although nothing can compare with the elegance of the ontology, large
amounts
of information stored as Panlingua arrays can also be accessed at very high
speeds, as we shall soon se.

Chapter 4, Basic Definitions, Assumptions, and Theorems.

By Chaumont Devin, Honolulu, December 22, 2002.

As regards the whole structure of language, of course it collapses
immediately
without WORDS, which are its basic building blocks, or atoms.  Yet very few
of
us have ever really given much thought to what a word is, how it is
actually
fitted into the framework of language, etc.  In modern nations, grammar is
taught in schools, hence most educated children know immediately how to
recognize separate words.  Surprisingly, this is NOT the case with people
who
have never been exposed to modern education.  When compiling a dictionary
of
the language of Buru Island, for example,  I found that I first had to
train
most of my informants to distinguish separate words in their own language
before they were able to be of much help to me.  This skill often came
readily
enough, since teaching it was only a matter of defining and externalizing
an
internal process with which they were already familiar at a subconscious
level.
Still, I remember one individual, probably the best-educated of them all,
and
proud of the fact, who could never quite discern word boundaries with 100%
accuracy.  This kind of observation goes to show just how seemingly
impenetrable a barrier exists between our conscious and subconscious
minds--a
barrier through which we must forge inroads in order to understand the real
meaning of language along with much else.

But even tenured professors will not be able to tell you what a word is.
The
truth is that although we may use thousands of words every day, and even be
capable of discoursing at length on linguistics and entymology, when pushed
to
the wall, most of us are powerless to define what a word really is.  Here,
for
example, is the best that the American Heritage Dictionary of a few years
ago
was able to do:

"A sound or a combination of sounds, or its representation in writing or
printing, that symbolizes and communicates a meaning and may consist of a
single morpheme or of a combination of morphemes."

Such a definition may be okay for the American Heritage Dictionary, but in
reality it is a pretty shaky foundation upon which to build our knowledge
of
language.  Let us take just a moment to examine:

"A sound or combination of sounds or its representation in writing."  So
does
this mean that a word is not a word until it has been spoken or written?

"That symbolizes and communicates a meaning."  But here AHD is overly
simplistic and even misleading, because most such sounds and written words
can
symbolize not "a," or just ONE meaning, but several or even many meanings,
no
particular one of which can be determined without linkages to other words
and
meanings.  In fact, I think the slippery meaning of "word" has just
slithered
off the page!

"May consist of a single morpheme or of a combination of morphemes."  Well,
I
suppose.  But although endless pages have been written on morphology, in
fact
morphemes cannot be relied upon without embarrassment, even in the most
regular
of natural languages.  AHD might almost as well have written, "May consist
of a
single character of the Roman alphabet or of a combination of such
characters."

So a major goal of this chapter will have to be to attempt some less warm
and
fuzzy definition of the word, "word," else we are lost forever!

WARNING: In standard English, "word" can mean either the thing symbolized
by a
symbol in a text or else the symbol itself--hence a great deal of
confusion.
In the following sentences I will be using "word" to mean the elusive thing
symbolized by a symbol in an array of text and not the symbol itself.

Definition: A symbol is a sound, image, or other physical pattern
(four-dimensional pattern) representing something else.

Definition: Texts are understandable (meaningful) linear arrangements of
symbols.

Definition: A surface language is a language constructed using symbols
arranged
into texts.

Definition: Words are the entities represented by the symbols in
understandable
texts.

Definition: A subsurface language is a language whose structures are built
only
of words with no symbols involved.  It can have no symbols.

Definition: Syntax is the study of the ways in which words relate to one
another.

Definition: Semantics is the study of the way words relate to meanings.

Definition: A link is a vector (directed binary relation) connecting two
nodes.

Definition: A node is the point at which two or more links are joined.

Assumption: Nodes contain no information in and of themselves.

Assumption: Since nothing is encoded in nodes, then it naturally follows
that
all grammatical information must be encoded in the links between them, or,
more
specifically, in link type.

Assumption: Every word has a syntactic and a semantic role.

Assumption: Besides its syntactic and semantic role, no word ever has any
other
kind of role.

Theorem: Every word is linked to a meaning.
Proof: Every word has a semantic role.

Theorem: Every word is a node.
proof: The things that are linked to other things are called nodes.

Theorem: Every meaning is a node.
proof: The things that are linked to other things are called nodes.

Definition: A semnod (semantic node) is a node to which one or more words
are
linked--in other words, a meaning.

Assumption: Many classes of word can link to a single semnod.  For example,
redden, reddens, reddened, and reddening all link to the same semnod
through
various types of lexlink when they appear in a sentence.

Assumption: There exists a finite and predictable set of word types that
can be
linked to each semnod.  For example, the words, BIG, BIGGER, and BIGGEST
all
link to the same semnod.

Definition: A semlink (semantic link) is a link between two semnods.

Definition: The semantic plane is the collection of all semlinks and
semnods.
All semlinks and semnods are said to lie within the semantic plane.  The
nodes
of the semantic plane are semnods, and semlinks are the links between then.

Definition: An ontology is a representation of the semantic plane.  It is a
netlike pattern of semlinks defining hierarchical and non-hierarchical
relations between semnods.

Assumption: The ontology contains certain kinds of knowledge.

Definition: A lexlink (lexical link) is the link connecting a word to its
semnod, or meaning.

Theorem: In general, all words are connected to other words.
Proof: Every word has a syntactic role.

Definition: A synlink is the link between two words.

Definition: The syntactic plane is the collection of all synlinks and
words.
All synlinks and words are said to exist within the syntactic plane.  Words
are
its nodes, and synlinks are the links between them.

Theorem: Words have rank.
Proof: The links between words are directional.  Thus words to which other
words are linked can be thought of as having higher rank and vice versa.

Definition: A regent is a word to which one or more other words are linked.

Definition: A dependent is a word linked to a regent.

Assumption: Dependents serve to modify the final meanings of their regents.

Definition: A regent is one rank level above its dependents.

Assumption: A regent can have more than one dependent.

Definition: A sibling is a co-dependent of the same regent.

Theorem: Siblings maintain a certain order of proximity to their common
regent
in relationship to one another.
Proof1: Noun modifiers naturally assume various proximity positions
relative to
their regent nouns according to their semantic classifications, and this is
true across many if not all languages.
Proof2: Texts of narratives are maintained in chronological sequence.

Theorem: Siblings cannot maintain order of proximity if all are connected
directly to their common regents.
Proof: A regent is a node, which is just a connecting point containing no
information of any kind whatsoever.

Definition: A left sibling is a co-dependent one position closer to the
common
regent.

Theorem: Only first dependents are linked directly to regents.  Other
dependents are linked to their left siblings.
Proof: The human linguistic apparatus maintains information about the
relationships of siblings to each other--specifically about their proximity
to
their regent--and this information cannot be syntactically encoded in any
other
way.

Theorem: The links between words naturally assume L-shaped Tinkertoy
structures.
Proof: Every sibling is linked horizontally to its left sibling except for
the
first, which is linked vertically to the common regent.

Theorem: A word is a node connected to its regent or left sibling by a
synlink
and to its semnod by a lexlink.
Proof: Every word has a syntactic and semantic role.
Note: This is my final definition for the nonsymbolic meaning of "word."

Theorem: The meaning of every word is completely determined by (1) a
synlink to
a regent or left sibling, (2) a lexlink to a semnod, and (3) any dependents
that might link to it.
Proof: Every word in every language always has a syntactic and semantic
role
with meaning potentially modified by dependents.

The following theorem works for a subset of languages including English:

Theorem: Once the synlink and lexlink have been determined for every word
in a phrase and its external symbols have been removed, what is left is the
Panlingua representation of the phrase which consists of nothing but links
and nodes.

This theorem does not work for Austronesian languages in which the
auxiliary verb, "be" need not be expressed in surface syntax.  For example,
in a sentence transliterating to "I [was] a child then," it is clear that
"I" and "child" and "then" are all words of the same rank and require a
common regent.

Theorem: In languages like Malay it is often the case that the subsurface
representation of a sentence differs from its surface representation by the
presence of an "invisible 'be' verb."

Corollary: There can be more words in the subsurface representation of a
thought than in its surface representation.

Definition: A state is the way something is.

Assumption: All language encodes nothing but the maintenance or assumption
of
various states by various things and the agents responsible for these same
maintenances and changes.

Definition: A verb is a word representing the assumption or maintenance of
a
state.

Definition: A telic verb is a verb representing the assumption of a state.

Definition: A stative verb is a word representing the maintenance of a
state.

Definition: An adjective is a word representing a maintained state.

Definition: A noun is a word representing a thing.

Assumption: All words encode either things (linguists like to call these
"entities") or states, but never both.

Assumption: No word linked to the same semnod can encode more than one kind
of
state or thing.

Assumption: All verbs always encode a single state, either assumed or
maintained by the action of the verb.

Assumption: The state encoded by a verb always passes to its patient (if it
has
a patient).

Theorem: State flow is independent of dependency.
Proof: The state of a verb flows down to its dependent noun, while the
state of
an adjective flows up to its regent noun.

Assumption: A noun is the patient of a verb if and only if it assumes or
maintains the state encoded by that verb.

Definition: A frame is a regent and its optional dependents.

Assumption: All language consists of repeated instantiations of similar
frames.

Assumption: All language is made up of two basic frame types: (1) noun
frames,
and (2) non-noun frames.

Assumption: In non-noun frames states are passed downwards to nouns from
their
regents.

Assumption: In noun frames states are passed upwards from modifiers to
their
regent nouns.

Assumption: Except for the root atom of a frame, all words within a frame
have
their regents within the same frame.

Theorem: There exists but one and only one subsurface language.
Proof1: All of the above definitions and observations apply to every known
language of every kind.  Even artificial languages, such as computer
instructions, obey all the same rules, their only difference from natural
languages being greater simplicity and zero ambiguity.
Proof2: Any healthy human infant can readily acquire any natural language
known
to man.

Definition: This universal subsurface language is Panlingua.

Assumption: Knowledge is represented in the structures of Panlingua.

Definition: A person is a human being or other entity possessing human
linguistic capability.

Definition: A speaker is a person generating text.

Definition: A hearer is a person interpreting text.

Theorem: Texts (surface language) are ambiguous.
Proof: Their very symbols are usually ambiguous.  In general, each such
"word"
in a sentence can potentially synlink to any other "word" in the same
sentence
and lexlink to several semnods.

Assumption: Efficient processing requires information to be in unambiguous
form.  Neither computers nor humans can do anything with natural language
until
it has been interpreted or parsed according to the grammatical rules which
generated it in the first place.

Definition: Parsing is the process of disambiguating texts and encoding
their
meanings in Panlingua.

Theorem: Panlingua is unambiguous.
Proof: Parsing is the process of disambiguating texts and Panlingua is the
result of this process.

Definition: Understanding is what occurs when a hearer successfully parses
a
full sentence.

Definition: A situation model is a model of the world, either real or
imagined,
temporarily held by a person.  It consists of the visualization of various
things causing, assuming, and maintaining various states.

Definition: Visualization is the generation of situation models.

Theorem: A translational equivalence exists between situation models and
their
Panlingua representations.
Proof: Panlingua encodes the maintenances or assumptions of various states
by
various things and the agents of these maintenances or assumptions, and
situation models consist of the same.

Theorem: Panlingua representations can be converted to situation models and
vice versa.
Proof: A translational equivalence exists between situation models and
their
Panlingua representations.

Definition: An explicit state is a single state explicitly indicated by a
word
such as an adjective or verb.

Definition: An implicit state is a state not explicit in any word, but
implied
by a situation model.

Theorem: Panlingua representations can be built from situation models.
Proof: A translational equivalence exists between situation models and
their
Panlingua representations.

Theorem: Thoughts encoded in the treelike structures of Panlingua must be
converted to linear texts before they can be transmitted from one person to
another.
Proof: During transmission, language must pass through the physical world,
where it is encoded in symbols.  Panlingua has no symbols and therefore is
not
text, therefore text must be generated based on Panlingua representation
before
transmission can occur.

Definition: Text generation is the process of generating texts from
Panlingua
representations.

Definition: A lexicon is a collection of symbols and their potential
synlink-lexlink combinations.

Lexlinks constitute bridges between the syntactic and semantic planes.
Proof: Words are the nodes of the syntactic plane, semnods are the nodes of
the
semantic plane, and lexlinks link words to semnods.

Definition: A template is a Panlingua representation employing semantic
nodes
deliberately taken as high as is known to be possible in the hypernym
hierarchy
of the ontology.

Definition: A wildcard value is a value that will match anything.

Theorem: Aspect is coded in lexlink type.
Proof: Grammatical information is encoded in link type.  Aspect is not part
of
syntax and so cannot be encoded in synlinks, and lexlinks are the only
other
links available for all words.

Theorem: Other information mutually exclusive with aspect is also encoded
as
lexlink type.
Proof: Grammatical information is encoded in link type.  Part of speech is
not
part of syntax and so cannot be encoded in synlinks, and lexlinks are the
only
other links available for all words.

Theorem: Panlingua exists nowhere outside the syntactic plane exceptin the
space between the syntactic and semantic planes.
Proof:  Panlingua is constructed exclusively from synlinks and lexlinks.

Assumption: The only places where knowledge can be "written" in our minds
are
in the ontology and various Panlingua arrays.

Theorem: No knowledge exists anywhere except in the syntactic and semantic
planes of our minds.
Proof: All written "words" are only mnemonic symbols used to jog our
memories
and to communicate the thoughts of others to our minds.  Knowledge exists
ONLY
in our minds, and the only structures that can hold it are the ontology and
Panlingua arrays.

Theorem: A particular surface language may not take advantage of all the
predictable word types that can be linked to a particular semnod.
Proof: In English there is no stative verb for the semnods to which many
adjectives are linked, while in other languages these verbs are abundant.

Theorem: Although Panlingua is essentially the same for all surface
languages,
each will leave a slightly different signature in its underlying structure.
Proof: Disambiguate sentences having identical meaning in two surface
languages.  Resulting Panlingua structures may differ significantly.

Theorem: Besides subsurface Panlingua, which bears the signature of the
surface
language, there exists a sub-subsurface, or deep Panlingua.
Evidence1: Many people can translate very quickly between natural
languages.
Evidence2: While thinking certain kinds of thoughts, people are often heard
talking to themselves.  During other kinds of thinking people are NOT heard
talking to themselves.

Theorem: All evolving linguistic systems converge upon Panlingua over time.
Proof1: Researchers encounter identical phenomena from language to
language,
hence the great corpus of common linguistic terminology.
Proof2: The similarity of computational systems increases and never
decreases
as linguistic performance increases over time.
Proof3: Panlingua is the same for every language and for all of mankind.

Theorem: The basic structure of Panlingua itself is not evolving over time.
Proof: Humans have had time to evolve various physical differences,
therefore
if Panlingua were evolving it would not now be the same for all mankind.

Theorem: No linguistic system can function at a truly human level without
Panlingua.
Proof: Panlingua underlies ALL languages, therefore anything not based on
Panlingua is not even a real language at all, much less a human or
human-like
language.

Conclusion: Panlingua is obviously the solution, as nature has abundantly
proven, so lets figure it out and apply it to all the great machines of the
future.

Chapter 5, Deductions.

By Chaumont Devin, Honolulu, December 23, 2002.

If the words of languages all depend on other words, and some word order
can be
changed in surface languages without altering meaning, and Panlingua
underlies
this surface language, then Panlingua must not be strictly linear in form.

If Panlingua is nonlinear, and any form of linearity must be retained in
surface languages, then Panlingua is somehow encoding this linearity within
its
structure.

If Panlingua encodes some kinds of linearity, and deems other kinds of
linearity irrelevant, and a regent can have many dependents, and it is
found
that only the linearity of dependents is retained, and this only for
certain
classes of regent, and that this linearity actually reflects proximity to
the
regent, then siblings must point to their leftsiblings until the first
sibling
is reached, and this first sibling must point upwards to the common regent,
so
that Panlingua assumes L-shaped Tinkertoy structures in the syntactic
plane.

If the sticks of this Tinkertoy structure are synlinks, and atoms connected
by
horizontal synlinks are siblings, and siblings have the same parent node,
and
synlink type encodes word dependency type, then the types of these
horizontal
synlinks encode the relationship of the dependent to its regent, and not
the
relationship of the dependent to its left sibling at the terminus of its
synlink.

If all synlinks only ever point to left siblings and regents, then the
syntactic structure of Panlingua can be visualized as a collection of
arrows
chasing one another in a leftward and upward direction.

If a noun is a word representing a thing, and thoughts are things, and
entire
thoughts can replace nouns in all surface languages, then entire thoughts
must
be treated as nouns in Panlingua as well.

If a verb is a word representing the maintenance or assumption of some
state,
and every thought has but one and only one verb and its dependents, and
this is
the same definition as that for "clause", and all subordinate clauses are
treated as nouns, and the heads of all subordinate clauses are verbs, then
the
subordinate verb must link upwards as if it were a noun, while its
dependents
link to it just as to any other verb.

If the purpose of part-of-speech classification is to determine what
synlinks
and lexlinks are possible for a certain symbol if it represents a word, and
if
this information is contained in the lexicon, and the purpose of the
lexicon is
only for communication with the outside world, and the lexicon is not
necessary
to the internal functions of language, then part-of-speech is not related
to
the ontology or Panlingua but only to external symbols.

If part of speech is employed as a semnod-use classification, then it is
part
of Panlingua.

If part of speech is used for semnod classification, then part of speech is
part of the ontology.

If each word in a sentence has but one and only one meaning, and each word
in a
sentence has but one and only one syntactic role, then each word in a
sentence
must have but one and only one lexlink and synlink.

If the only components of a word are a synlink and a lexlink and a node,
and
the node is really nothing but the origin where these two links are joined,
then a word is really nothing but a synlink and a lexlink emanating from
the
same node.

If a word is really only such a synlink and a lexlink, then a real word
requires no symbol or sequential position in a line of text.

If a node must originate both a synlink and a lexlink in order to be a
word,
then meaning can never be divorced from syntax in any word.

If Panlingua is a purely logical entity, then it can be represented in any
medium--for example as a schema on this page or as a structure made of
Tinkertoys.

If Panlingua Is purely logical, and can be represented in any medium, then
the
reason humans with internal representations of Panlingua can speak and
Tinkertoys cannot must lie in the difference between mediums and their
abilities to fully implement all of the features of Panlingua.

If the human linguistic apparatus uses Panlingua, and the human linguistic
apparatus works, and no other linguistic apparatus works, then the only
linguistic devices that work use Panlingua.

If the only linguistic apparatuses that work use Panlingua, and computers
are
linguistic devices, and linguistic ability improves as Panlingua is
approached,
then the better a computer can perform linguistically, the closer it has
come
to implementing Panlingua.

If better linguistic performance means approaching Panlingua, then
computational linguistic systems approach Panlingua as they improve.

If computational systems improve as they approach Panlingua, then all
refinements in computational linguistics converge on Panlingua.

If computational linguistic systems improve steadily, and this improvement
means approaching Panlingua, then it is not necessary to conserve surface
languages to deduce Panlingua, and people can learn Panlingua by studying
machines.

If many surface languages are available, and traits common to many must
belong
to Panlingua, and no machine has yet learned a language, and it doesn't
look
like any machine will do so very soon, then to learn Panlingua from surface
languages may be faster than to learn it from machines.

If a state is the way something is, and brains can model the way various
things
are in various states, then a brain can model a thing in some state,  just
by
knowing what thing and what state it is in.

If a brain can model various things in various states, and language can
represent various things and the states they are in, and such a
representation
can reflect reality, then a brain can use language to model reality.

If each set of things and states can be called a frame, and moving through
a
set of frames can reflect reality over time, and changing the states of
some
things makes a next frame, and language can represent these changes of
state,
then a brain or computer can use language to update its model through time.

If without exception all languages encode the maintenance or assumption of
various states by various things, and the agents or causes of such
phenomena,
and nothing less or more, then Panlingua must also encode the maintenance
or
assumption of various states by various things and nothing less or more.

If a brain can use such information to form its model of reality, then this
must be what brains do with Panlingua all the time.

If Panlingua is the universal subsurface language common to all of mankind,
and
the human species has been evolving over time, and races have been
separated
long enough to change physically as they have, yet infants of any race can
readily acquire the language of any other, then Panlingua itself cannot be
changing over time.

If Panlingua never changes over time, and surface languages express ideas
generated in Panlingua, and surface languages are free to assume various
forms,
and what CAN happen WILL happen, then it is reasonable to assume that any
feature of Panlingua will probably evidence itself in some surface language
over time.

If Panlingua is stable and not evolving, and parsing is translation from
surface to subsurface language, and generation is translation from
subsurface
to surface language, and these translational processes are continually at
work
in the human mind, and translation systems work more efficiently as source
and
destination languages converge, and all biological organisms evolve towards
maximum efficiency, then there will be a selectional pressure against any
surface language ever straying very far from Panlingua.

If the structures of surface languages are constrained from straying very
far
from those of Panlingua, and some feature is found to be common among
surface
languages, then that feature probably reflects some feature of Panlingua.

If Panlingua's features can be deduced by examining many surface languages,
and
surface languages are free to take any form, and only a few surface
languages
survive, then the probability of successfully mastering Panlingua will be
small.

If Panlingua's features can be deduced by examining many surface languages,
and
a detailed knowledge of Panlingua turns out to be essential for the
machines of
tomorrow, then it may be critical that all surface languages be preserved
without exception now.

If Panlingua underlies all languages, and languages are to be preserved,
then
it is silly to think of systematically preserving languages without some
knowledge of Panlingua.  If Panlingua itself is immutable, and yet
subsurface
Panlingua has a slightly different structure for every surface tongue, then
Panlingua must be very flexible, and capable of taking many forms.

If Panlingua is highly flexible, and takes on different forms for different
surface languages, then while Panlingua constrains surface languages,
surface
languages also influence Panlingua within immutable bounds.

If surface languages influence Panlingua, and Panlingua is the substance of
our
thoughts, then the natural language we are using influences our minds and
personalities.

If words are already words before they are spoken, and there exists a deep
Panlingua without spoken words, then animals may also use Panlingua in
another
form.

If Panlingua has never changed in man, and Panlingua can assume various
forms,
then maybe it has never changed in animals either, and it permeates all
life.

If there is no intelligence without life, nor life without intelligence of
some
kind, and all knowledge exists only in ontologies and Panlingua structures,
then maybe Panlingua is the intelligence that drives all life.

If Panlingua is immutable, and Panlingua permeats all living things, and
panlingua is really the same thing in animals and man, and the more animals
evolve the more they utilize all the features of Panlingua, then the
evolutionary process may be converging on forms best adapted to make full
use
of all the features of Panlingua.

If mankind is still evolving, and evolution is converging upon forms that
best
utilize Panlingua, then our progeny will probably exhibit better language
and
intellectual skills than we do, and be able to think in new and better
ways.

Has this been happening? You bet!

Chapter 6, Searching and Matching on the Panlingua-Based Machine.

By Chaumont Devin, Honolulu, May 10, 1998.

Basic Matching in Panlingua.

Recall that a thought is a verb and its dependents.  In Panlingua, rather
than
matching texts character-by-character and word-by-word, entire phrases and
thoughts can be matched to see if they mean the same thing.

Suppose we have a knowledge base represented entirely in Panlingua, and
that
it runs something like this:

Eggs cost $1.50 per dozen.
Coffee costs $1.50 per pound.
TV dinners cost $2.00 each.
Ice cream costs $2.50 per half gallon.
Fresh bread costs $1.50 per loaf.
Etc.

What would this look like in Panlingua?

Recall that every word in Panlingua has a synlink and a lexlink.  Let us
disregard the lexlinks temporarily and focus upon the synlinks.  Recall
that
in the syntactic plane the nodes and links of Panlingua have a Tinkertoy
appearance.  The Tinkertoy sticks correspond to synlinks, and the wheels
correspond to words, or nodes.  Every one of the sentences given above is a
thought -- namely a verb and its dependents.  The sibling dependents all
link
together horizontally and the first dependent links vertically to the verb,
creating the typical Panlingua down-and-right-branching effect.  The
general
pattern is a series of L shapes, one L for each thought, and with the verb
of
the thought at the top of each L.

But what about the verbs at the tops of the L formations.  They have to
have
synlinks too.  Their regent may be a real word or it may be a dummy word
placed to fill this requirement at the head of the array.  And because all
the
verbs of the structure are of the same rank (all of them are dependents of
the
same regent), they are all siblings, which means that they will be joined
by
horizontal synlinks.  As a result, Panlingua representations of long
strings
of sentences like the above would appear to have a verbal backbone with the
dependents of these verbs hanging beneath them.

Now suppose that this Panlingua representation is held in an artificial
intelligence that can interact with a user, and that the user asks the AI,
"How much does coffee cost?"

First the AI will parse the English entry to obtain yet another Panlingua
representation.  Then the AI will examine the Panlingua results and "see"
that
the newly-parsed structure represents a query.  Then the AI will change the
synlink type of the parsed query from "complex interrogative" to
"declarative", so that no mismatch will occur because the sentences
represented in the knowledge base are declarative.  It will also change the
lexlink of the word representing "what" so that it links to a wildcard
semnod
that will match anything.  Once this has happened, both the query and the
knowledge base are in appropriate Panlingua representation, and Panlingua
matching can begin.  The AI will take the verb of the query and compare it
to
the first verb of the knowledge base to see if these two atoms match.  If
they
do not, the AI will skip ahead to the next verb (which is linked to the
first
by a synlink), and see if that one will match, and so on until a match is
found or the AI reaches the end of the Panlingua representation.  But if
they
do match (as they must, since both will be "cost"), it will then compare
the
dependents of both verbs to see if a match can be found among the
dependents
of the verb in the knowledge base for every dependent of the verb of the
query.  A Panlingua match occurs whenever a regent and all its dependents
can
be matched to another regent and its dependents without regard for the
order
in which the dependents occur.

If the search fails, the AI will respond with an "I don't know".  If it
succeeds, the AI will pass the subtree that matched the wildcard semnod,
which
will be the price, to the text generator, and the system will say, "$1.50
per
dozen".

Needless to say, this kind of searching is exceedingly fast because instead
of
having to muddle through texts one character at a time, the AI can skip
along
from verb to verb to verb.  Furthermore, instead of matching the characters
of
each word, it need only check to see if the semnod to which two words are
linked is the same and whether the synlink types of the two words are the
same.

Of course this is only a simple example designed to provide a basic idea of
how Panlingua matching works.  In a real system special algorithms would be
in place to interpret queries like, "How much are eggs?" in such a way that
the
system would make sure a price was returned (since "how much" in this case
would be assumed to mean a price), etc.

But it will be seen that even at the most basic level, Panlingua has killed
two old birds with one stone.  Not only can Panlingua search much faster,
but
it is also able to match on a thought rather than a word level, so that
thoughts expressed differently will still match as long as their meanings
are
the same.

Fuzzy matching.

But what if instead of just saying, "I don't know", every time a less than
perfect match were found, you might like the AI to give you more
information?
Panlingua provides several ways for determining fuzzy values.  Suppose you
asked the same AI, "What does day-old bread cost?" and the AI found that
there
was no entry for "day-old bread", but only one for "fresh bread".  Using a
simple matching function, instead of returning 0, meaning "perfect match",
the
function might return 1, meaning "one miss".  And when compared with the
returns on all the other matches, this value of 1 might turn out to be the
lowest.  The AI might then respond, "I don't know, but fresh bread costs
$1.50
per loaf".  This is a simple example of fuzzy matching.

To convert the fuzzy value obtained in the above to a percentage match
value,
just divide the 1 by the number of words in the query, which happens to
parse
to four atoms, and multiply by 100 for a result of 25%.

More refined matching algorithms can be written to weight these mismatches
according to rank.  In other words, make each dependent missed count as 2,
but
make each dependent of a dependent missed count only for 1, etc.  This
exact
approach is not recommended, but is provided only as an example.

Yet another kind of fuzzy matching might take into account and weight
ontological relationships.  For example, suppose biscuits and bread both
have
hypernym links to "bakery products" in the ontology.  Then if the user were
to
ask, "What do biscuits cost?" instead of just answering, "I don't know", it
might answer, "I don't know, but fresh bread costs $1.50 per loaf", and so
on.

Template Matching.

Another important kind of Panlingua matching is what I will here call
"template matching".  Using Panlingua it is possible to create templates
against which many sentences can match.  This is done again by matching in
conjunction with the ontology.  As an example, suppose we have a string of
Panlingua templates, one of which is:

Truck drivers drive trucks.

And suppose further that in the ontology we have an entry:

Sam is a truck driver

and we want "Sam drives trucks" to match "Truck drivers drive trucks".  A
matching function can do this if it is designed for template matching,
because
not only does it check to see if the semnods to which two atoms link are
the
same, but if they are not the same then it uses the ontology to see if the
semnod to which the atom in the template is linked is a hypernym of the
semnod
to which the query word is linked.  In this kind of matching, any word in
structure A matches the corresponding word in structure B if the word in
structure A is linked to the same semnod as the word in structure B, or
else
if the word in structure A is a hyponym of the corresponding word in
structure
B.

This kind of matching is very useful for searches into generic references
in
order to see if a sentence is of a recognized type or pattern because a few
template representations will be able to match many test sentences.

Matching Direction.

It is important to notice that Panlingua matching is directional.  Thus the
fact that sentence A matches sentence B does not necessarily mean that
sentence B will match sentence A.  As an example, "She parked the car" will
match "She parked the car in the garage", but "She parked the car in the
garage" will not match "She parked the car".  This is because the
fundamental
matching rule for Panlingua reads as follows:

Subtree A matches subtree B if the regent of subtree A matches the regent
of
subtree B and the dependents of subtree A can all be matched among the
dependents of subtree B.

And this rule must be applied recursively whenever dependents of dependents
occur.

Ignoring Markers.

Many Panlingua structures may require marker atoms to indicate things like
tone of voice, emphasis, what to do in case of a match, etc.  These can all
be
safely ignored where simple queries are being matched to knowledge bases
because the queries will be free of these markers though the knowledge base
may have them.  Recall that in Panlingua matching, if structure B has all
the
dependents of structure A and more, then A will match B, because all that
is
required is that what is in A be found with the same syntax in B.

But care must be taken in cases where structure A may contain special
dependents not present in structure B.  For applications involving such
cases
special code will have to be written so that these additional dependents
will
be ignored when they are not part of the comparison to be performed.

Further Confirmation.

These matching properties, so humanlike in character, provide striking
confirmation for the validity of our deductions about the existence and
structural features of Panlingua.

Further research.

Needless to say, if even these cursory examinations of the matching
properties
of Panlingua can yield so much, then there may be astonishing new
discoveries
to be made in the areas of Panlingua-based matching and inference
processing.

Chapter 7, The Various Kinds of Panlingua Data Structures Needed by
Machines.

By Chaumont Devin, Honolulu, May 10, 1998.

Any Panlingua-based system must have an ontology because Panlingua cannot
work
without one.  Every Panlingua atom is linked by a lexlink to a semnod in
the
ontology, and many Panlingua operations require the information contained
in
the semlinks of the ontology.  And if a Panlingua-based system is to
communicate with the outside world at all, then it must have a lexicon,
because it is only by means of the lexicon that the atoms of Panlingua can
be
converted to the symbols of the outside world.  These two structures will
be
basic, then, to almost any system involving Panlingua.

But besides these two critical data structures which have no portions
encoded
in Panlingua, there are many more that are of importance to Panlingua-based
systems, all of them represented in Panlingua itself.  In this chapter I
will
attempt an examination of some of them.

The General Reference.

There are many things a fully functional linguistic apparatus needs to know
that require more than a binary relation to be defined.  An ontology can
tell
us simple things, like, "Roses are red", because only two words are
involved,
and the verb is a common ontological operator (semlink type).  But for
sentences like, "Soldiers shoot guns", "The sun rises in the east" etc.,
several binary relations will be required.  These trivial facts about the
world that every good computer must know should be stored in a Panlingua
reference of the template variety.  Recall that in Chapter 6 I described
such
Panlingua references under the heading, "Template Matching".  Furthermore,
this general reference should be handled dynamically in such a way that the
system can learn.  Thus, for example, if the general reference contains the
sentence:

The sun rises in the east.

and later the system receives from a reliable source the information that:

Heavenly bodies rise in the east.

and a check of the ontology reveals that the sun is a heavenly body, then
the
original sentence should be upgraded to:

Heavenly bodies rise in the east.

and all versions of this sentence in the general ontology that match the
upgraded entry should be destroyed.  Then supposing there was no original
entry stating that:

The moon rises in the east.

now, providing the ontology has an entry for

moon isa heavenly body

the system will automatically know that the moon also rises in the east
without being told.

**** Note: An ontology CAN record information such as "soldiers shoot guns"
if
ALL verbs (including SHOOT) are allowed as operators.

Discourse Logs.

A Panlingua log should be kept of the last few sentences that have been
communicated in discourse with each user and for each text or Panlingua
structure being output or read.  The system will use these to resolve
problems
of discourse such as what the conversation is about, who the last male
subject
was, etc.

Cyclopedic References.

In addition to the general reference, the system may require specific
encyclopedic information about places, people, events, and other things.
These can take various forms, including large files on disk.  Not all the
information the system can get at through these need be represented in
Panlingua.  For example, it may be desirable to display copies of original
texts, graphics, sound and video clips, etc.  Or when certain topics are
referenced it may be desirable to have the system invoke some dedicated
interactive process.  All these things can easily be done in Panlingua by
adding various marker atoms to the Panlingua structures of the reference.
Instead of pointing to a semnod, the semnod identifier field of such atoms
could be used to identify other data files, programs, etc.

Idiom Reference.

This reference would contain such idioms as, "Kick the bucket", "Turn
green",
etc.  The parsing process would then go like this:

1. Parse into Panlingua.
2. Check to see whether parsed subtree is an idiom.
3. Skip ahead if it is not.
4. Replace the idiom with its true meaning if it is, and mark the subtree
to
show that in the original text an Idiom was employed.

And why can't this idiom check be made before parsing?  Simple.  "Raining
cats
and dogs", "Raining bloody cats and dogs", and "Raining cats and
blankety-blank dogs" all contain the same idiom, which cannot be matched at
a
word level but only at a thought level, which requires that the match be
made
in Panlingua.

Most people agree that something -- namely alternative meanings when
ambiguities occur -- is lost in translation.  However anything can be
marked.
For example, "It was raining (idiom originally used for next word) hard".
This can be accomplished using a marker dependent in Panlingua which
retains
the information that the original version used an idiom instead of a single
word.

To make a marker in a Panlingua array, simply employ the same structure
used
for an atom, or Panlingua word, with a type of "Just a marker, please
ignore"
indicated by some mnemonic.  Then make your code ignore such atoms wherever
appropriate.

Other References.

In fact anything that can be said can be represented in Panlingua, and once
in
Panlingua this information can be scanned at high speeds, and searching can
be
done with high selectivity and accuracy.  Not only this, but for the
computers
of the future, many processors can be assigned, each one to a different
Panlingua structure, so that split-second searches can be done over truly
vast
amounts of information.

But in the meantime, even with the personal computers of today, Panlingua
implementations whose atoms are only seven bytes long can handle many
billions
of different semnods and words.  And as an example, to record a simple fact
such as, "Soldiers shoot guns" requires only 21 bytes of computer memory,
which means that nearly fifty thousand such facts could be stored in just
one
megabyte of RAM.  And because many such items will be stored as templates,
each one of such items might match many sentences.  For example, if the
general reference says that "Soldiers shoot guns", and the ontology tells
us
that a corporal is a soldier, then it will be known by the system that
corporals shoot guns.  So the potential of Panlingua-based systems is
indeed
very large, and it is doubtful whether other types of systems could emulate
this kind of performance.

Further Confirmation.

I must point out that all these facts about Panlingua-based systems,
versatility, speed, highly selective matching, and sheer amount of
information
that can be stored, all serve only to confirm what I have already asserted
before, namely that:

1. Panlingua is the unique basis upon which all truly functional automated
linguistic systems must be built.

2. All refinements of automated linguistic systems must converge upon
Panlingua.

3. No automated linguistic system can ever really be made to work without
Panlingua.

Etc.

And these are not just idle claims.

Chapter 8, How to Translate Between human Languages with Machines.

By Chaumont Devin, Honolulu, May 10, 1998.

But how, you may ask, can Panlingua be used to translate between languages?
Here are the steps required to translate from Language A to Language B:

1. Create separate lexicon-ontologies for Language A and Language B.

2. Create a translation table linking the semnods of Language A to the
semnods
of Language B.

3. Create any ad hoc algorithms required to convert Panlingua
representations
using the semnods of Language A to Panlingua representations using the
semnods
of Language B.

4. Create a parser for Language A and a text generator for Language B.

It may be easy to understand why separate lexicons will be required for
Language A and Language B.  After all, every language employs a different
set
of symbols for communication.  But you may wonder why the two ontologies as
well.  Wouldn't it be better to create a single ontology to handle both
languages?  After all, the external symbols of language may differ, but
aren't
things still the same things, and aren't the states these things can be in
still the same states?  It may well be possible to do this, at least for a
limited number of languages at a time, especially if those languages belong
to
people of similar cultures.  But cultures tend to differ dramatically, so
that
what is considered a good thing by the speakers of one language may be
considered to be a terrible thing by the speakers of another language, etc.
For example, in the Moluccan Islands of Indonesia, the idea of rain falling
while the sun is shining may strike terror to the heart.  When this
happens,
the devil comes out looking for victims to curse with vile afflictions.  So
"hujan panas" is a very bad thing.  But in the Hawaiian Islands many people
love the misty showers of manoa because they are thought of as sweet and
refreshing, and often come in the company of brilliant rainbows, so in the
Hawaiian islands such a thing would be a pleasant, sweet, and good thing.
And
in the Moluccan Islands people might say, "Let's cross the street.  The
light
is blue". because to many people in the Moluccas blue and green mean the
same
thing.  But if you said the street light was blue in Hawaii people would
think
you were complaining about the color of the light.  And while in Hawaii a
basket is just a basket, I have never been able to find a word for just
"basket" in Maluku.  A basket is never just a basket, but always some kind
of a
basket, perhaps a "bakul", or a "keranjang", or a "kamboti", etc.

For some of these difficulties the solutions would seem simple enough.  For
example if in the Moluccas green lights are blue lights, then all we would
have to do is make the lexlink from the word, "green", link to the same
semnod
as the lexlink from "blue", and all would be well.  Well, that is, until we
tried to translate from Moluccan Malay into English.  Then we would be
faced
with a problem, because "blue" and "green" link to separate semnods in
English, so that a choice would have to be made, and this would require
special processing.  If Both Moluccan Malay and English were made to share
the
same ontology, then besides a semnod linked to the English word, "green",
and
another linked to "blue", there would have to be another one that linked to
the equivalents of both these words in Moluccan Malay.  But why do both of
these words even exist in Moluccan Malay?  Why not just some word like
"greenblue?"  I cannot answer this question, but for some strange reason
they
do, and this is the reality with which we must deal.  As you can see from
this
example, having a common ontology, even for two languages, will necessitate
additional nodes.  This means that if an ontology were to be developed to
include all the semnods of all the languages known to man, then this
ontology
might have to be very large.  I am not saying that it would have to be
larger
than the capacity of a modern computer, but it would have to be large.

And for the basket case, if both languages shared the same ontology, then
there would simply be no lexlink to the semnod for English "basket" from
Malay.  So far so good.  But what happens to the hypernym links of the
common
ontology?  To the American fellow, a "bakul" is a basket, and a basket is a
container, but to the Moluccan guy a "bakul" is just a container.  Can we
mark
semlinks for language?  If we could, then the Moluccan semlink from the
thing
called a "bakul" would jump directly to the semnod for "container", but the
English semlinks would go from "bakul" to "basket" to "container" (notice
that
I have gotten lazy and stopped writing "the semnod linked to" before each
word).  Now things are getting uncomfortable, because it suddenly becomes
necessary to change all our software to make it handle different semlinks
for
multiple languages, and this would seem to be an instant nightmare.

I will not even bother with the problem of sunny rain, because in both
English
and Moluccan Malay this requires more than just a single semlink because in
both languages "sunny rain" requires two words.  In short, it can be seen
that
common ontologies may not be the right way to go.

And in fact if we return to the human model, it is probable that common
ontologies are not being used even there.  People who speak many languages
fluently often seem to adopt whole different personalities for each
language
they speak.  They have found that they fit into a certain personality type
for
English, another for French, another for German, etc.  The link between
personality and ontology is clear.  The ontology holds all the key
perceptions
of the individual about his/her world.

It may well be that people have a common ontology for many "core"
relations,
such as "houses are big", "kittens are tiny", etc., and then use limited
language-specific ontologies for certain special things.  This might seem
to
be a logical approach, but I have failed to recognize any evidence for it
thus
far.

So the safest and easiest approach must be to create two ontologies and a
translation table.  Each entry in this table would then be a link from a
semnod of Language A to a semnod of Language B.  Ideally these links would
be
bidirectional, so that translation would be possible in both directions,
but
for simplicity in this chapter I will assume that they are directed from
Language A to Language B.  Each such link would then have the following
components:

source semnod of Language A
destination semnod of language B
special function identifier

Because we are translating from Language A to Language B, the identity of
the
source semnod, which is of Language A, will always be known.    In most
cases,
the destination semnod will also be known, because the two languages will
have
more semnods that map cleanly than those that do not.  A special function
may
or may not be given in cases where the destination semnod is known, but
definitely must be provided where no destination semnod is known.  In cases
where a special function is provided, this special function will be part of
the ad hoc code (item #3).  The function will be called with the address of
the current translation link and the current word in the source Panlingua
representation and its regent as arguments.  It will be designed
specifically
to deal with the problems known to arise when trying to translate this
semnod,
and will generate a circumlocution in the target Panlingua representation,
or
do anything else that may be required to make the translation work well.
The
farther Language A is from Language B, obviously, the greater the number of
such ad hoc functions that will be required.  Thus the number of ad hoc
functions required to translate from English to Japanese would probably be
very large, whereas the number of such functions required to translate from
English to Dutch or German would be very small.

But although the ad hoc functions required for translation will not be easy
to
write, those required for parsing will be much worse.  The parser is by far
the most difficult component to develop for any such system.  The problem,
of
course, is that texts are ambiguous.  The texts of surface languages employ
strings of symbols to represent Panlingua atoms for communication.  Each
such
symbol is linked to an entry of a lexicon for that language, and for each
such
symbol the lexicon may provide many potential synlink-lexlink pairs.  In
order
to parse sentences, all but one of the potential synlink-lexlink pairs for
each word must be discarded, and the destination of the remaining synlink
must
be determined.  After this has been accomplished the original symbol for
the
word as well as its linear position in the word sequence can also be
discarded.  At this point the word has been parsed, and  is represented in
Panlingua.

For text generation, a grammar of the target language is used to rearrange
the
Panlingua atoms in a linear sequence.  Then the lexicon is consulted to
find a
symbol with which to replace each Panlingua atom, and the process is
complete.

Traditional researchers identify two main kinds of parsing: TRANSFER and
INTERLINGUA.  The transfer method is basically one of manipulating words
around
using ad hoc code with little regard for what they mean.  The program
recognizes some pattern, and generates an appropriate output if it can.
The
interlingua method employs some kind of interlingua, such as an
implementation
of Panlingua, to hold the meaning of parsed sentences, then generates text
in
the target language from these internal structures.  The
machine-translation
method I have outlined above is, of course, an interlingua one.  But it
resembles the transfer method in that only subsurface Panlingua, and not
sub-subsurface, or deep Panlingua, is used.  A better and fuller
implementation
of the true interlingua method would be to first parse from the source
language
to subsurface Panlingua, and then to convert this subsurface Panlingua to
deep
Panlingua before outputting into the target language.  Because text in the
target language will be generated from deep Panlingua instead from the more
language-dependent subsurface variety, the text generator will work to
output
from ANY source language instead of just one.

Of course these are oversimplified descriptions of parsing and text
generation, but they do tell the basic steps that must be taken.  There
must
be some gimmick in the parsing process that we are missing completely, but
so
far no one has ever been able to see it.  Thus for a single programmer to
write even a moderately reliable parser using current knowledge may take
years.  No one has ever been able to write a parser that can even come
close
to human performance, and thus it is very difficult to get materials
converted
from surface languages to Panlingua.  Clearly this issue of parsing remains
the major impediment to progress in the field of artificial intelligence
today, because without getting materials into Panlingua it is virtually
impossible to do anything really meaningful with them automatically using a
computer.  But a vast new horizon awaits us once this difficulty has been
overcome.

Chapter 9, How to Program Panlingua Applications using Standard ANSI C.

By Chaumont Devin, Honolulu, May 10, 1998.

Many of you may have been wondering just how one might implement a system
of
links and nodes on a computer, and this is not a trivial question.  One
might
begin by analyzing the information that is needed in order to create a
standardized link for the system.  Since nodes generally carry less
information, even in the computational implementation, they can generally
use
exactly the same number of bytes as the links and have room to spare.  In
this
way all the links and nodes of a link-node system can be thought of as the
elements of a single array.  So the first step is to determine the sizes
and
identities of the elements required for each link.    Once this has been
done,
it will be a simple matter to define the data elements needed for the nodes
within the same number of bytes.

As an example, suppose we intend to create a simpleminded ontology capable
of
holding 65535 semnods.  Recall that the links of the ontology are semlinks,
which link one semnod to another within the semantic plane.  To create a
system that can learn, the semlinks should be weighted, meaning that each
link
should hold statistical information that tells the system how much it is
used.
Each link will also have a type and a destination semnod identifier so the
system can know what kind of a link it is and where it links to.  Thus for
each link we will need a weight byte, a type byte, and at least two bytes
for
a semnod identifier.  So the type of structure we will need, in C, should
look
something like this:

Typedef struct { /* semlink structure */
unsigned char wgt; /* weight byte */
unsigned char typ; /* semlink type */
unsigned      tag; /* target semnod identifier */
} sem_t;

With this type definition we might then create an array containing space
for
10,000 links and  nodes as follows:

sem_t ont[10000];

Now recall that although we have defined the semlink structure with its
various elements, or fields, we have still to consider what the semnod
structure should be like.  For this we will need a value that tells the
system
how many elements of the array are taken up by the semnod and its pointers
and
an identifying integer value for the semnod.  For the number of elements
used
by the semnod and its pointers a byte value should do, so we can just use
the
first byte field of the semlink array, which was "wgt".  Then for the
integer
identifier we already have the "tag" field which serves the same purpose in
the semlink structure, so we can use IT.  Then for a particular semnod
entry
in the array we might have:

5 elements, semnod #21935

The "5 elements" part tells us that this semnod element will be followed
immediately by four semlink elements in the array.  Then for these we might
have:

wgt=255  typ=isa tag=20029
wgt=33 typ=has tag=15293
wgt=55 typ=has tag=995
wgt=92 typ=can tag=11504

And these four semlinks would then be followed immediately by the next
semnod
in the array, and so on until a semnod was reached containing all zeros,
meaning that we had gotten past the last semnod and the last semlink in the
array.

Since semnod tag values should increment consecutively, a check of the
integrity of the data in the array can easily be made by skipping from
semnod
to semnod in the array to see if the numbers follow one another
consecutively
until the zero at the end of the data has been found.

To add a semlink to a semnod, the next semnod and everything after it will
have to be shifted one position to the right so the semlink can be put in
its
place.  Then the "wgt" byte of the semnod to which the new semlink has been
added will need to be incremented to show one element more.

To remove a semlink from a semnod, simply shift everything right of the
semlink one position leftward and decrement the "wgt" value of the semnod
to reflect one less element in use.

An arbitrary decision will have to be made regarding semlink direction.  In
my
systems the semlinks point from the current semnod to the semnods
identified
in their tag fields.  But it might be just as valid to choose the opposite
direction and assume that all the semlinks were pointing to the semnod to
which they belong.  Then their tag values would not indicate where they
were
pointing TO, but instead where they were pointing FROM.

So using a digital computer it is a very easy matter to model systems of
links
and nodes.  It is easy to traverse all the semnods in the example system
above
just by adding the "wgt" value to the current index at every semnod and
jumping to the next semnod.  It is easy to visit all the semlinks
pertaining
to one semnod just by finding the semnod, then using the "wgt" value of the
semnod to increment the index until the next semnod has been reached.  Etc.

But what about data structures like Panlingua, which are not simple systems
of
links and nodes?  Yes, it would be possible to model Panlingua using linked
lists, but this would not ve very easy to do, and it would be slow and
expensive in terms of computational resources.  A simpleminded structure
for
the Panlingua atom would be as follows:

typedef struct { /* the Panlingua atom */
unsigned char sib; /* number of atoms in subtree */
unsigned char syn; /* synlink type */
unsigned char lex; /* lexlink type */
unsigned short tag; /* semnod identifier */
} atm_t;

This structure employs only five bytes of memory per atom--two bytes for
the
tag value and one byte for the three other values.

A Panlingua system using this atom structure could handle up to 65535
unique
words and 65535 unique semnods, yet every atom would only be five bytes
long.
Thus an array containing a single megabyte of these structures could hold
69,905 three-word assertions or 104,857 two-word assertions in a form that
could be traversed at very high speeds.

But how can Panlingua representations, which are known to be
three-dimensional
structures of links and nodes, be held in such an array?  In my
implementation
of Panlingua I use a trick.  It works like this:

Every atom that has a sib value of 0 is the terminator atom of an array.
Every atom that has no dependents will have a sib value of 1.  Every atom
that
has one or more dependents will have a sib value greater than 1.  For every
atom that has one or more dependents, its first dependent will be the next
atom in the array.  If an atom has a sibling to the right, that sibling
will
be located at the address of the current atom plus sib.  Thus if the 5th
atom
of an array has a sib value of 4, and if that 5th atom has a sibling, then
its
sibling will be the 9th atom in the array.  Every atom located at the
location
of the current atom plus its sib value is either the sibling of the current
atom or else an atom of higher rank that is not the regent of the current
atom.

Thus, ignoring synlink and lexlink types, we might represent the following
sentence in the following way:

John loves Mary
3 loves (LOVES heads a subtree containing three atoms)
1 John (JOHN is an atom with no dependents and IS a dependent of loves)
1 Mary (MARY is an atom with no dependents and a right sibling of JOHN)

Notice that even though MARY has no siblings to the right, it still has a
sib
value of 1.  Because sib values indicate the number of atoms in the subtree
and not the location of the next sibling, it is necessary to take
appropriate
programming precautions to make sure that, say, the next atom to the right
of
Mary will not be construed as a right sibling.  This is done using an
"upper
limit index", which always points to the next atom past the current
subtree.
Thus if a process is traversing a subtree at a lower level than its head
atom,
and the index of any atom plus its sib value equals the upper limit index,
then the atom at that calculated position is not a sibling to the right of
the
current atom.

All this may seem rather awkward, but by arranging Panlingua atoms
according
to dependency in this way great savings can be achieved in terms of
computational resources because it is not necessary to store an identifying
integer for the regent or left sibling of  the current atom in the
Panlingua
atom structure.  Recall that each Panlingua atom, or word, employs two
links.
The target of the lexlink is identified by the tag field.  Lexlink type is
given by the lex field.  And synlink type is given by the syn field, but no
synlink target need be identified because that target is just the regent or
left sibling of the current atom.

Thus each atom in a Panlingua array points just beyond the remaining atoms
in
its subtree by means of its sib value.

This arrangement of computer data structures representing Panlingua atoms
into
arrays as described above is called "Interlinguish."

Interlinguish arrays lend themselves readily to fast search, match, swap,
move,
and other operations, for which many of the algorithms have been made the
property of the public domain.  At first the programmer may be bewildered
by
the seeming complexity involved, but after some familiarization it is not
difficult to gain a reliable feel for where things should be when atoms
have
been arranged in this manner.  Mastering this system is crucial for those
wishing to write parsers, text generators, and other algorithms for
Panlingua-based systems.

Chapter 10, Panlingua and Robotics.

By Chaumont Devin, Honolulu, May 12, 1998.

What does Panlingua mean for robotics?  A very great deal.  With a
knowledge of
Panlingua it will be possible for people building robotic devices to think
in
important new ways.  Why?  Because once Panlingua is known it is possible
to
integrate all the various parts of a robot so that they will harmonize with
each other in modular ways.  What Panlingua offers is a new kind of target
or
goal for robot subsystem development.

Let us study, for example, the problem of electronic image pattern
recognition.  What would be the output of such a system without Panlingua?
Probably something very limited in scope, very esoteric, and very difficult
to
use except with a very specific system.  But what if an image processing
system were being designed to interface with a system using Panlingua?  The
goal of the image processing system would then be to set up reports in
Panlingua.  Such reports might go something like:

I see a car at center moving right at high speed.
I see another car passing behind it moving left.

Etc.

Of course such electronic perceptions may be impossible at this stage of
pattern recognition and image processing development, but they serve to
illustrate what I want to say.  At least at this time no image processing
system will be able to accurately identify everything it sees.  But
supposing
the developer manages to create an image processor that can even return
information about two kinds of things.  If even this seemingly trivial
amount
of information can be represented using Panlingua it will be of great
value.
In the first place, in an integrated system that included all the major
linguistic components, the system would immediately be able to say what it
saw
in the surface languages it used.  For example, "I see a car", etc.  Even
this
small achievement would be of immediate and possibly incalculable value to
a
blind man.  But besides being able to say what it saw in English, the
system
would also have this information available for immediate use by any other
subsystem.  For example, supposing the system were designed to drive cars.
The Panlingua representation would immediately be available to be passed to
a
"car driving" module for use.  This car driving module would then keep
analyzing all the incoming reports from the image processor and telling its
subsystems what to do based on its driving decisions, also in Panlingua.
Each
such subsystem would only be required to understand that subset of
Panlingua
for which it was designed.  For example, the throttle mechanism might only
understand such things as, "faster", "slower", "full throttle" and "cut
power".  These Panlingua representations would be all that this particular
subsystem would ever need to know.  But because they were coded in
Panlingua,
a person testing the system would still be able to analyze or listen to
each
command as it was produced in plain English if necessary just by asking the
system to monitor the signals sent to this subsystem.  In fact no matter
what
the robotic subsystem, this would be the case.  This would constitute a
vast
improvement over all kinds of systems and subsystems in use today, each one
of
them using proprietary protocols that may have to be studied for days
before
they can be used.

Many marvelous things would be possible with systems held to an interface
standard that used only Panlingua.  For example you might go down to your
local appliance store and buy an electric range, bring it home, plug it
into
your computer, assign it the name, "Charley", and say, "Charley, I want
half
heat on your front left burner for fifteen minutes". Or you may have told
Charley to heat his oven to 350 degrees, but you are not quite sure whether
or
not it is time to put in the cake, and ask, "Charley, how hot is your oven
now?"  Or you might call your vacuum cleaner "Mr. Clean", and say to your
computer, which is linked to Mr. Clean by Panlingua interface using infra
red,
"Mr. Clean, go clean up that mess in the corner.  No, not there.  Turn
left.
Now go straight".  Etc.

The point is that once everything is running on Panlingua and can interpret
Panlingua in the performance of its particular tasks, then any new device
can
be voice-controlled just by plugging it into some master system capable of
interpreting speech, converting it to text, parsing text, and determining
what
to do with the results.  And this is true for a self contained robot of the
traditional anthropoid variety with two arms and two legs as well as for a
distributed system capable of running all the new appliances in a
21st-century
home.

Incredibly, nothing much more is required for the command and control unit
of
such a system than just the same old Panlingua-based linguistic apparatus I
have already described.  Here are its components again:

English-to-Panlingua parser
Panlingua command processor
Panlingua-to-English text generator

And in finer detail, the shared components to make these larger systems
work?

lexicon
ontology
Panlingua template general reference
Panlingua idiom reference
Panlingua discourse log
Panlingua cyclopedic reference
Panlingua scenario reference
Panlingua command reference

Much work is still needed to know how to make a robot perform all the tasks
regarded by humans as "thinking".  What appears clear is that all these
thought processes will involve one or another of the data structures listed
above in some way.  Notice that the lexicon and ontology are the only two
such
references not represented in Panlingua.  The implication is clear.  Most
thought processes will be carried out using Panlingua.

Of particular interest to robot builders will be the Panlingua command
reference array.  Suppose you told your faithful robot, who had just
finished
digging in your garden, "Go wash your hands".  From this input the parsing
apparatus would set up two thought representations in Panlingua.  The first
would be just, "Go".  The second would be, "Wash your hands".  Because
sentence
type is encoded in Panlingua, the robot would perceive that these two
sentences
were both imperative ones, so instead of just standing there with an
unreadable
expression while it checked to see if it already had this information
stored
somewhere before storing it, the computer would search its Panlingua
command
processor for a match to the command, "Go".  When the Panlingua command
reference found the match, it would return an event code to the command
processor of the robot, and guided by this event code the robot's command
processor would pass this "go" command to its mobility subsystem, and the
robot
would start moving away.  Then it would process "wash your hands" in the
same
way, and pass this second command to a subsystem handling hand things.
This
subsystem would in turn pass this command to a special hand washing
subsystem,
which would know that this operation requires water and a sink, and pass a
"find the sink" command to the mobility system, which would send the robot
in
the direction of a sink, etc., etc.  Thus by simply hanging an event code
onto
a template in a Panlingua command reference array, it is possible to select
any
action the robot can perform using standard Panlingua search-match
operations.

Or what if you asked your robot a question, like, "What is the chemical
composition of fiberglass resin?"  Once again, sentence type is also
represented in Panlingua (as verb synlink type), so your faithful robot
would
know that you had asked it a question and start looking for the answer.  If
you had asked it whether roses were red it would have realized that the
assertion, if it existed in its memory, would b a binary one, and look for
it
in the ontology.  But because this question involves a complex structure,
it
searches the cyclopedic reference, from which it retrieves the answer, uses
it
to generate text, converts this to speech, and tells you in plain English.

Many questions remain unanswered about the general problems of robot
design.
Some of these are comparatively straightforward, for example:  How should a
robot determine what should be transferred from short to long term memory,
when, and how?  In a Panlingua-based implementation, of course, this would
involve moving knowledge from the discourse log to the ontology, the
general
template reference, the cyclopedic, and other Panlingua references in the
system.  More subtle are questions about how robots learn.  I have
explained
one or two of these processes in previous chapters, but there are doubtless
many more.  As far as I can tell, robot learning will involve the creation,
destruction, and migration of various links among various nodes.  And then
there are those most difficult questions about inferences drawn from long
acquired facts, and the synthesis of new ideas.  I am confident that
through a
knowledge of Panlingua these things are now coming within reach, and this
fact
makes our time a very exciting time to be alive.

But one of the most important problems that must be overcome before
reliable
distributed Panlingua-based systems can be developed is that of the
ontology.
Recall that ontologies differ from language to language, and even from
individual to individual.  Before, say, English can be used to drive any of
the distributed systems I have described above, it will be necessary that
compatibility be ensured by setting up a standard English ontology to be
used
by everyone.  This will be possible because although the semlink patterns
of
ontologies may vary, the set of semnods remains more-or-less the same.
Almost
everywhere English is spoken, therefore, a rose is a rose is a rose, and a
dog
is a dog is a dog, etc.  What needs to be done is to agree upon the
specific
semnod identifier to be used (for example 15329) for the semnod linked to
"rose" (the flower), and some other integer identifier for the semnod
linked to
"dog" (the animal, etc.  Otherwise when you tell your faithful robot to go
get
a screwdriver he may return with a tree!  Such semantic standardization
would
not be difficult to achieve, but knowing my American countrymen, I suppose
many
thousands of mutually incompatible ontologies will be developed before this
happens.  Semnod identifiers are only integers, and no special sequence or
order of any kind is required.  The only major difficulty will be getting
any
two red-blooded Americans to agree upon which of the billions of integer
values
available to use for things like "dog" and "rose".  Hopefully some
automated
means will be devised to clean up the mess and integrate them all in the
end,
so that, say, in another 100 years or so we will have worked out a common
ontology for English, and modular systems will be able to work as I have
described.

The possibilities I have mentioned may all sound quite interesting, but by
far
the most important of all is the potential for making machines that really
think.  We already know a few of the processes involved, but there remain
many
more of which we know nothing at all. The exciting thing is that Panlingua
theory may give us a key to unlock the doors.  As everyone knows, computer
processors are getting faster and faster every year.  What this means to
robotics is that if we can someday learn to make computers think, then it
may
also be possible to make them think very fast.  To build such a machine
would
not seem impossible for a creature who can barely run 15mph but can fly
aircraft at many times the speed of sound.  It might be possible, say, to
compress a thousand man-hours of human thought into a mere ten minutes or
so.
Thus it might be possible to create an entity capable of transcending the
limits imposed by a lifespan of 70-80 years in order to bring long-term
thought processes to maturity.  And not only bring them to maturity, but do
so
almost immediately.  Albert Einstein spent the last decades of his life
searching for a unified field theory for physics in vain.  Before he had
time
to discover this secret his life was cut short.  A machine with the same
intellectual powers might hit upon it in less time than it takes to boil
an egg!

Simply extrapolate the curve of technological development and it should
become
clear that the things I am saying are not mere fantasies but the very stuff
of
our future.

Chapter 11, Some Tough Implications for Modern Science to Ponder.

By Chaumont Devin, Honolulu, May 13, 1998.

(I think this chapter may need some rather tough editing and revision)

If by this time you have failed to sympathize with the theoretical
framework
behind Panlingua, then I bid you read no further because this chapter
assumes
that Panlingua is fact.

The discovery of Panlingua has far-reaching implications.  On the one hand
it
promises amazing new technological breakthroughs during Century 21.  On the
other hand it tells us some amazing facts about our universe and ourselves.

One of the most important deductions implied by the existence of Panlingua
is
that from the time men first spoke on this earth Panlingua has never
evolved.
The reason for this deduction is obvious.  The separation of humans into
isolated groups has consistently resulted in physical variations.  Observe
that we are not one but many races, yet Panlingua has remained exactly the
same.  All language works exactly the same way.  We know this because of
conformity to certain principles across many languages, and because there
exists no spoken language that a healthy infant of any human origin cannot
readily acquire.

Some of the major principles that are never violated across many languages
are
these:

1. No human language ever does ought else but to encode the assumption or
maintenance of various states by various things and the agents that cause
these assumings and maintenances.

2. All human languages are made up of repeating frames of two basic types:
noun frames and non-noun frames.

Let us now take a few moments to study these phenomena of state and framing
more carefully.  A noun frame consists of a noun and its dependents.  The
dependents of nouns are various kinds of noun modifiers, including
adjectives,
determiners, nouns that modify nouns, prepositional phrases and even whole
clauses that modify nouns.  These modifiers serve to tell us more about the
nouns they modify -- in other words to better distinguish or define their
noun regents.  Thus noun modifiers always indicate some state the things
indicated by their noun regents are in.  I use the term "state", in a broad
sense to include such states as, for example, being "the one who put out
the fire" in noun frames such as "the man who put out the fire".

Notice that in all such noun frames the direction of "state flow" is upward
to
the noun regent, or head, and never downward from the noun.  Recall that in
the Tinkertoy structures of syntax, regents are always parent nodes while
dependents are their daughters.  Thus the regent of a noun frame is always
a
single noun at the root of the subtree, or the "head" of the subtree, and
this
noun receives various states from its modifiers/daughters/dependents, which
are
positioned further down.

In non-noun frames, on the other hand, state flow may occur in either
direction.  In verb frames, the explicit state of the verb passes to a noun
called its "patient", which often heads a noun frame in its own right.  For
example, in "The monkey ate the banana", the specific state of "ate", which
is
"eaten" gets passed to "banana", because it is the banana that gets eaten.
In
sentences like "John loves Mary", this state flow may be less clear because
John is engaged in the act of loving, and maintains a state that seems to
be that of the verb.  But in fact John is only the agent, and the explicit
state of "loves" is "loved", and it is Mary that gets loved whether she
likes it or knows it or not.

It is important to notice that the patient of a verb is not always its
object.
For example, in a passive sentence like, "The city was covered by fog",
"city"
is subject instead of object, and yet "city" is also patient.  In English
the
object is usually also the patient of the verb, as in, "Fog covered the
city",
but this is not always the case.  The thing to remember is that the subject
is
what becomes the antecedent of the next pronoun, whereas the patient is the
thing that receives the explicit state of the verb.  Thus the term,
"object",
should be thought of as pertaining to discourse, whereas "patient" should
be
thought of as pertaining to state flow.

Besides patient, of course, verb frames hold adverbs and prepositional
phrases, from both of which state flows upward to the regent verb.

The role of agent appears to lie outside the domain of state flow, the
agent
of the verb frame acting as the cause of state flow rather than engaging in
any state flow itself.  In English, of course, the agent is usually the
subject.

Underlying every prepositional phrase is a prepositional frame in
Panlingua.
In such frames, the explicit state of the preposition always passes
downwards
to its object.  For example in the phrase, "through the water", the water
gets
penetrated in a "through" manner. This is hard to see because in English
there
is no verb series for "through", such as: THROUGH, THROUGHS, THROUGHING,
THROUGHED.  But if there was, it might readily be seen by most native
English
speakers that in "through the water" the water would be the thing that got
"throughed".

Thus prepositions appear to be just a special class of degenerate verbs,
and
so it is natural that prepositional frames should belong to the same class
as
verb frames.  But in prepositional frames patient is always object without
exception, whereas in verb frames patient may be either subject or object.

But just as in verb frames, state flow in prepositional frames may move
either
up OR down.  For example, in the phrase, "right up the hill", "right"
serves
as an adverb modifying "up" in exactly the same way as an adverb might
modify
a regent verb, with state flowing upward from "right" to "up", and down
from
"up" to "hill".

Another feature common to all human languages is the presence of
subordinate
clauses.  As an example, in the sentence, "To steal other people's property
is
wrong", the clause, "steal other people's property", is a whole separate
thought unto itself which acts as the subject of the sentence--in other
words
a noun.  As it turns out, all thoughts or clauses are nouns, and for these
to
fit into the syntactic structure of Panlingua it is necessary to treat them
as
such.

Biologists tell us that our DNA is 96% identical to that of a chimpanzee.
What is the most significant difference between humans and chimpanzees?
Granted chimpanzees are more physically robust, the lengths of the limbs
don't
match, details of the feet are different, etc., and these physical
differences
must account for at least part of the difference in DNA.  But what does
this
tell us about language?  It tells us that those parts of the human
linguistic
apparatus that differ from those of a chimpanzee must be accounted for by
less
than 4% of the genome.  And it tells us that all the sequences of human DNA
that differ from chimpanzee DNA should be suspected of encoding human
linguistic capability.  Furthermore, for animals like apes, dogs, parrots,
and
dolphins--namely those animals that have near-human linguistic abilities --
any DNA sequences found in common among them should be suspected of
encoding
language if they are not also found in other animals.

But if human language is not evolving, and if conscious intelligence and
awareness is coupled to language, then anything modern man can do at this
time
could have been done by the first human beings.  In other words, if it were
possible to go back in time and snatch an infant from Homo sapiens sapiens
of
120,000 years ago and bring him/her up in an American home, then his/her
chances of making it through college would probably be just as good as
those
of anyone else in the modern world.  It may well be that the only
differences
between us and Homo sapiens sapiens of 120,000 years ago are cultural.

Unfortunately it is not yet possible for us to analyze the DNA of those
first
beings who spoke words as men.  But if we could, and if we found that even
they showed this 4% difference with chimpanzees, as the fact that language
has not changed would indicate, then we might be forced to live with
another
hypothesis about ourselves, namely that  because of the large difference of
4% between earliest humans and modern apes, it would appear that man has
not
evolved slowly and steadily from speechless subhuman into modern man as we
once supposed, but that some kind of quantum evolutionary leap has
occurred.
The question would then be what caused this leap, and whether it would
still
be possible to believe in the Darwinian model, at least for the case of
man.

But if we are willing to carry the correspondence between language and
intelligence even further, then yet other mysteries will unfold.  This is
because if language can be linked to intelligence, then life itself can be
linked to intelligence even more.  Show me any living thing and I will show
you intelligence.  Even the most humble bacterium is in some way a master
of
complex chemical and electrical processes, and perhaps even other
manifestations of intelligence of which we are not yet aware.  But although
there can be no doubt that these processes require an intelligence as yet
beyond the bounds of modern science (none of us has ever been able to
design
and build a working bacterium), most of us would probably not attribute
consciousness to such intelligence.  But if language and intelligence and
life
are all inextricably one, then there must exist linguistic processes at
work
even within the most primitive of life forms.  We can see this because
linguistic equivalents exist for all the manifestations we are able to
observe.  For example, "If the environment is too salty, do process x.  If
the
temperature is too high, then do process y". Etc.  For some of us it would
seem difficult to understand how such a consistent equivalence could exist
without there being in fact some real correspondence.

And we will observe that although the "lowest" forms of life exhibit an
intelligence greater than our own conscious intelligence, this intelligence
is
internal rather than external, and subconscious rather than conscious.  But
as
animals express greater and greater physical sophistication their
coefficient
of external intelligence keeps rising higher and higher as well until at
last
it culminates in the highly linguistic intelligence of man.  And yet
despite
this external expression of linguistic intelligence, it is impossible to
determine whether in fact the cells of the human body are in any way really
more intelligent than so many bacteria.  It would seem that they are, but
how
can we really know?

Chapter 12, Early Beginnings, Progress, goals, and the Riddles that Remain.

By Chaumont Devin, Honolulu, December 24, 2002 (upgraded four years to the
day
from the date of the original manuscript).

Breaking the "Text Barrier."

The original problem was to store human thoughts in a computer without
using
texts.  during nine years of research between 1985 and 1994, I had learned
that
some very interesting results could be obtained using computers to
manipulate
textual data alone.  But I also learned that there must exist some
exponential
graph of improvement in results versus programming difficulty as one goes
on.
In other words, I found that in order to make even marginal improvements in
system performance, it may become necessary to invest many hours of
programming
time.  In fact I was up against a wall!

In 1994 this led me to develop a computational thought representation
system I
called "Interlinguish."  It was a computer interlingua employing
English-based
concepts.

At that time I knew nothing of a three-dimensional model of language, or
how
the innards of the human linguistic apparatus might work.  All I wanted to
do
was to make a computer capable of representing thoughts in some way a
machine
could really understand and use.  I had taken text-based language
processing
far enough to clearly see the brick wall of intractability, and I needed a
way
through.  The "text barrier" had to be broken.

In 1994 I was aware of the frame-based system employed by Project Cyc, and
so
for awhile I pondered the possibility of using frames.  But I quickly saw
the
futility of this approach because frames, by their very definition, are
limited
to a set number of elements, whereas words have unlimited numbers of
dependents.  What I needed was a universal atom of meaning--a basic
building
block using which ANY possible thought might be constructed.  I solved this
problem by creating the Interlinguish atom.

But this first universal atom of meaning was similar to Newton's equations
in
that it worked, but in that it failed to account for every detail.  It
wasn't
until some years later that I discovered the three-dimensional model of
language I an using today.  Now, finally, I think I can tell the whole
story,
and it is this:

Every word is a universal atom of meaning consisting of a node and two
links,
one to another word, and the other to a semantic node.

In the external environment of course words also consist in graphic symbols
and
sounds, but no word is ever anything more or less than this inside our
heads.
It was upon realizing this fact that I also discovered that the
Interlinguish
universal atoms of meaning I had labored so intently to create were in fact
merely words.  At that point my computational effort and my theoretical
understanding of language had merged.  This did not happen at once.  At
first I
held back from admitting that my universal atoms of meaning were words.
But
after much thought and many tests to find counter examples, I knew that
they
WERE the same things, and my "unified field theory" for language was born.

I have later come to realize that in English the meanings of "word" include
a
sense not written in most dictionaries.  Most people think of a word as a
sound
or a written symbol, yet native English speakers often say things like, "No
word was spoken," which is closer to the sense I am after.  What is a word
before it is spoken?  And here is the crucial question: Before a word is
spoken, does it really exist at all?  If your answer is YES, then you agree
with me, and you have glimpsed the real meaning of "word" which is not
written
in most dictionaries, as I have said.  A word is not really a symbol in our
4-dimensional universe at all, but a purely nonmaterial entity of some
kind.
The thing we hear or see written on a page is not really a word, though we
may
call it that, but rather a symbol REPRESENTING a word.  Real words exist
nowhere except in our minds.

At first I saw Panlingua words as the nodes of a binary tree.
Syntactically,
each word was the daughter of some other word, and might have two daughters
of
its own: a dependent and/or a sibling.  This seemed to make sense to my
computer programmer's mind.  I had read somewhere of tree structures with
nodes
having multiple daughters being converted logically to binary trees with
nodes
having only as many as  two daughters in order to effect standardization of
node size (force all nodes to be exactly the same).  I also perceived the
links
between these binary nodes as pointing downwards and to the right.  I had
found
what I wanted--a rock-solid representation that would be identical for
every
word in every language known to man--my universal building block and
universal
atom of meaning!

At last I had what I was looking for, and I was able to prove that it
worked by
using it to build my first commercially marketed version of Brainchild,
which I
called Brainchild1.  Anyone now using BC1 can see that my perception of
words
at the time was as I have described by running BC1 with the /d switch
turned
on.  BC1 then displays a diagram for every sentence it parses with little
arrows pointing down and to the right.

But although my original binary-tree model worked, as I labored to develop
Brainchild2, I noticed that instead of pointing downwards and to the right
as I
perceived them in my theoretical model, my synlinks were actually pointing
upwards and to the left.  The reason I had taken so long to see this was
because it only happened during parsing, after which my results were stored
in
the kind of Panlingua arrays I have described in a previous chapter.
Another
reason was that my original attempt at an interlingua, the Interlinguish of
1994, had been more syntactically oriented.  I had been unaware of the
three-dimensional model consisting of syntactic and semantic planes at the
time.

So in my final analysis, I have thrown out any mention of binary trees,
which
are really of zero relevance to Panlingua.  Instead, every word is
described
simply as a node from which emanate a single synlink and a single lexlink
and
nothing more.  The synlinks are of two basic varieties: synlinks that link
directly to a regent and synlinks that link to a left sibling.  So far I
have
never had to formally encode this basic synlink type distinction as part of
overall synlink type.  Instead it is always indicated by synlink
orientation.
When drawn on paper, vertical synlinks (synlinks pointing upwards) always
link
directly to regents, whereas horizontal synlinks (synlinks pointing
leftwards)
always link to siblings.  In Panlingua arrays, the first dependent after
the
regent is always assumed to link directly to the regent while subsequent
dependents link to their left siblings.

The historical overview given above is a general one, and does not include
a
host of false starts and details.  For those who would like to work out an
exact chronology for the discovery of Panlingua, I suggest an examination
of
the archives of comp.ai.nat-lang, dg-list@UGA.CC.UGA.EDU, and other
archives.

Although we have been able to make rapid progress in the analysis of
Panlingua,
in the NLP world there are questions that remain unresolved.  The important
thing with parsing is to know what one is parsing FROM and what one is
parsing
TO.  The matter of what we are parsing from is simple.  It is English,
French,
or some other spoken language, or maybe some dead language, or even some
artificial language like Esperanto or a language of computer commands.  But
the
issue of what we are parsing TO is more problematic.  Many linguists
believe
they are parsing by producing annotated texts.  For example, they may write
taggers which can tell you the part of speech for which every word in a
sentence is used with an accuracy of better than 95%.  Or they may even be
able
to cluster constituent phrases and mark them for case role, etc.  But none
of
these exercises produces much of anything useful to a machine.  They are
mere
academic exercises for academic consumption.

So we first need a very clear idea of what we are parsing TO, and what the
results of our parsings will be.  As you already know from reading this
work,
my idea of parsing is to end up with some implementation of Panlingua that
a
computer can employ without difficulty in higher level searching, matching,
and
analysis.  But in order to do so, we first need to know precisely how every
kind of thought is represented in Panlingua, otherwise we will have no idea
where to begin.

Subsurface Panlingua has a one-to-one word correspondence with the symbols
of
surface language, as I have already shown.  For each surface symbol
representing a word, merely select the right lexlink to the right semnod
and
the right synlink to some other word within the same sentence, and you end
up
with subsurface Panlingua.  Thus one might think of parsing as "lifting the
meaning from texts."  But subsurface Panlingua bears the marks of the
surface
language from which it has been lifted (parsed or derived), as I have said.
Therefore although it is unambiguous and machine-ready, it is
language-specific
and nonstandard.

So the big question arises: Is there such a thing as a STANDARDIZED deep
version of Panlingua in which the representation of any meaning conforms to
a
set of rigidly-enforced universal rules?  We have already assumed that a
deep
version of Panlingua exists, so the question is only whether or not deep
Panlingua is universally the same.  This idea is TERRIBLY tempting.  Its
ramifications include subliminal machine control, telepathy, the
downloading of
minds, mind extension, mind control, instant education, communication with
animals, animal speech, and more.  My personal suspicion is that no such
rigorous standardization exists.  Is this not why some of us are brilliant
while others remain idiots?  Still its possibilities make it well worth
looking
for.  But even if no rigorous standardization exists for deep Panlingua, it
is
clear that deep Panlingua MUST obey the basic rules of Panlingua, and it
would
seem probable that a deep Panlingua would have more grammatical constraints
than the subsurface variety.  In either case, in order to fully implement
all
the features of Panlingua, it is imperative that we ultimately come to know
EXACTLY what rules apply.

At the same time, there is no apparent reason why we might not artificially
standardize deep Panlingua for our machines.  Could this be the next great
evolutionary leap toward full implementation of Panlingua--one not to be
made
by humans but by machines?  Only time will tell.  I assume that we wish to
make
our machines ever faster and better than we are, and yet to keep them under
our
control.  Machines using a standardized deep Panlingua would certainly be
better than we are.  Ideally they would be able to communicate with each
other
telepathically (probably using radio waves) at a subliminal level.  They
would
be able to master any field of knowledge almost instantaneously.  They
would be
able to extend there minds ad infinitum, and more.

But most importantly for us, with a standardized deep Panlingua we could
build
better machines right now--machines that could be made to communicate
efficiently between robotic components and acquire new knowledge bases
immediately as I have already shown.  This is why I have been pushing so
hard
to get people to start using Panlingua, and then to agree upon some common
ontology.  Without this, we will be groping blindly forward into the 21st
century when we ought to be flying.

We ought now to be taking very seriously how to solve the problems of our
species and this planet, because time is against us, and we may not have
all
that much longer to do so.  Every day that we lose simply provides another
opportunity for some fool like Osama bin Laden to blow us all back into the
middle ages and reduce Planet Earth to a mud ball.  Living in the
comparative
sterility of academic and business environments, people often forget that
the
real world is positively teeming with such men.  Our only hope lies in the
overwhelming superiority of our technology, and our best hope of
maintaining
that superiority is by developing the kinds of systems I have been
discussing
here.

Of course it will be necessary to ascertain precisely what is happening
with
Panlingua in every case before it will be possible to realize its full
potential and build machines to take full advantage of its features in the
ways
I have described.  I feel comfortable with the general structure of
Panlingua
as I have described it.  The theory is rock-solid, and I know that this is
no
idle claim or delusion because I have often published and argued its
details in
appropriate Internet fora and in personal exchanges with acquaintances and
friends.  I have yet to read or hear any argument capable of tearing it
down.
Furthermore (believe it or not) my computational implementations of
Panlingua
really work.  And yet human language being nothing if not interpretative, I
am
necessarily limited to deduction rather than direct observation.  Almost
everything known about Panlingua has to be deduced by examining texts that
some
human brain or other has generated from Panlingua, and text generation is
fundamentally interpretation.  Because of this it is never easy to be sure
which one is right when two or more possible explanations might work
equally
well within the three-dimensional model.

Verb Rotation.

To give some idea of the kinds of problem that come up when trying to
discover
the structure of deep Panlingua, let us examine that of the auxiliary verb.
I
believe that in going from subsurface to deep Panlingua, a process takes
place
which I call "verb rotation."  Recall that one of the assumptions upon
which
our theory is predicated is that:

At the heart of every thought exists a single verb.

But in English we have a minor constellation of auxiliary verbs which occur
in
simple sentences along with main verbs.  Yet our assumption tells us that
each
thought must contain only a single verb--not two, not zero, but always one.
And a closer look at the structure of Panlingua will confirm that this
truly
must be so, because we know that every word  must be connected to another
word
in order to have meaning, and that sentences must be connected to each
other,
and that the only consistent place that sentences can be reliably joined is
at
their central verbs.  Allow each thought in a Panlingua structure to have
more
than one central verb, and the very structure of Panlingua falls apart.

Observation across many languages shows that many natural languages do not
use
auxiliary verbs as in English.  Why not?  Probably because processing
auxiliary
verbs is computationally inefficient, and deep Panlingua may not have them.
Consider the Malay sentences:

1.  Ia marah: He/she/it is angry).

2.  Itu manusia: That is a human being.

3.  Bisulnya bengkak: His/her/its boil swells or is swollen.

English demands that instead of "He angry" (#1), the thought be represented
as
"He IS angry."  English also demands that instead of "That a human being"
(#2),
the sentence should read "That IS a human being."  And as for the ambiguous
sentence "His boil swell", English demands either "His boil is swollen" or
else
"His boil swelled", etc.

So which is right--English or Malay?  Upon first reaching southeastern Asia
our
forebears tended to look upon Malays as human but something less than
Europeans.  They might be bright, but linguistically they might be excused
for
getting some things wrong.  But today we have a very different view of the
realities of human evolution, and so we are forced to ask this question: If
malay has no auxiliary verb, and yet English demands one, then how are
these
thoughts really represented inside our heads?  There are at least three
distinct possibilities as follows:

1.  The auxiliary verb is present in deep Panlingua but malay speakers have
agreed to let it drop from surface texts whenever ambiguity will not
result,
and English auxiliary verbs are always represented as verbs in deep
Panlingua.

2.  In deep Panlingua, the Malay predicate noun or adjective is actually a
verb.  Thus "angry", in "he angry", is a verb instead of an adjective, and
"a
human being", in "That a human being", is somehow a verb or some kind of
verb
phrase as well.

3.  Some English auxiliary verbs are also verbs in deep Panlingua, whereas
others are not.

If #1 is the correct hypothesis, then for the "one and only one verb"
assumption to hold it will be necessary that many simple English sentences
be
broken up into two thoughts.  For example, take:

1.  He can | fly airplanes.

2.  She might | like a kiss.

3.  He was | walking in the park.

4.  She did | eat the apple.

In each of the above cases the second clause (thought with a central verb)
behaves exactly as a noun object or predicate adjective should.  "Fly
airplanes" is the thing (noun object) that he can do in sentence #1.  "Like
a
kiss" is what she might, and "walking in the park" is the way (adjective)
he
was.

So what really happens to a sentence like #1 ("He can fly airplanes.")? In
subsurface Panlingua, of course, the "can" belongs to an upper and the
"fly" to
a lower thought, or clause.  This is because disambiguation is the only
process
that has occurred.  But making "can" the top, or central verb, and "fly" a
lower verb is counterintuitive, and does not happen except in certain
natural
languages.  In searching for a template telling whether or not someone can
fly,
why would one first try to match "can" when the topic is flight? This
suggests
that in deep Panlingua the real main verb must therefore be "fly," and not
"can," which would seem rather to be some kind of modifier.  A kind of
rotation
must be taking place in which "fly" moves up into the main verb position
while
"can" drops down into a modifier position.

And what if predicate adjectives linked to their subjects by copulas are
actually represented as verbs in Panlingua?  This possibility (See sentence
#3
above) has a lot of merit because after millions of years of evolutionary
refinement we would expect to see maximum efficiency, and one word would
use
less computational resources than two.  Thus for "He was walking in the
park."
we might expect "walking" to rotate up into top verb position and "was" to
rotate down into a modifier position and finally be subsumed by the lexlink
type code for "walking," where verb aspect should properly be encoded--in
other
words, to drop out completely from the structure of deep Panlingua.

Now let us consider sentence #4.  Does the "do" verb ever really appear
like
this in deep Panlingua?  Probably not.  Although it appears to be
consistent
with the same underlying form, what "She did eat the apple" is actually
saying
is simply that, in spite of everything to the contrary, she really ate the
apple.  Since "do" as used here is a more or less "empty" word, this must
really be just one thought instead of two, with a main verb of "eat."  I
suspect that in deep Panlingua, the English "do" is replaced by a sort of
dummy
word indicating stress or emphasis.  Another alternative would be to encode
the
declarative stress in the synlink type of the central verb.  Both methods
seem
equally possible, and I see no way of determining which is actually
employed,
if not both of them.  A deep Panlingua capable of using either method would
be
of the unstandardized variety, but a deep Panlingua allowing but one and
only
one method might belong to the rigidly standardized alternative I have
described.  Once again, the interpretative processes involved in text
generation do not allow us to glimpse the original deep Panlingua
structure.

So far so good, but my verb-rotation hypothesis runs into trouble with
copula
constructions whose objects cannot be converted to verbs.  As an example,
consider the sentence, "He is a human being."  How can "a human being" be
rotated up into the position of main verb?  Clearly this kind of
construction
must be handled quite differently, but how?  In languages like Malay, the
surface construction runs something like "He a human being," but what does
this
show us?  Where is the main verb?

It appears that some English auxiliary verbs do also appear as verbs in
deep
Panlingua while others do not.  The problem is to untangle the question of
which do and which don't.  Personally I tend to favor the following,
although I
am in no wise convinced, nor do I claim to have covered all possibilities:

1.  Verbs that mean to "be" or to "become" must be retained in deep
Panlingua
when their objects represent things, otherwise they are rotated.  In:

She was washing her clothes.

the "was" disappears and the "washing" remains but with past tense encoded
in
its synlink and incomplete aspect encoded in its lexlink.

The state-flow idea makes the retention of copula verbs in deep Panlingua
look
unlikely for predicate adjectives.  In all other Panlingua frames, state
flows
either from regent to dependent or else vice versa, but in the copula
arrangement state flows not from regent to dependent or from dependent to
regent, but from sibling to sibling.  At the same time, in Malay we see
adjectives used as stative verbs everywhere.  Thus I would presume that
predicate adjectives are rotated up to become stative verbs.

But in a sentence like, "John is a man," it would seem clear that "man"
could
not be somehow rotated into top verb position!

2.  Verbs that mean "do" are dropped or replaced by adverbs of emphasis
where
necessary.  Thus for sentences like:

Did they die?

the "did" disappears completely, and the synlink type for "die" becomes
past-tense simple interrogative (I perceive verb synlink type as also being
sentence type).

3.  For the English auxiliary verbs, "shall" and "will," the only thing
that
need be retained in deep Panlingua is the subtle distinction between them
that
is not assiduously maintained even in many human examples.  e.g.  what is
the
real difference between "will go" and "shall go"?  Many school teachers can
tell you, but many other native speakers cannot.  This distinction can be
indicated by replacing the auxiliary verb with an adverb when the promise
is
indicated and by dropping it completely when all that is indicated is verb
tense.  In either case, the synlink of the main verb must then be changed
to
indicate future tense.

4.  The remaining English auxiliary verbs are all rotated down into a
modifier
position and converted to adverbs.  The conversion of auxiliary verbs to
adverbs is clearly reflected in Malay, where "can" or "bisa" is in fact
primarily thought of as an adverb.  The same kind of rotation will work for
each of the following sentences:

He could | fly airplanes.  He might | fly airplanes.  He has | flown
airplanes
(with appropriate aspect and tense modifications to "fly").  He would | fly
airplanes.

In these latter examples, of course, it would certainly be possible to
retain
the upper-lower clause representation we observe in English.  But the
question
is not whether this CAN be done, but rather if in fact this IS done in deep
Panlingua.  In any case it is necessary to  decide upon these details for
certain before moving ahead with the development of Panlingua-based
machines.

My personal feeling is that these sentences ARE rotated, because of the
great
improvement in search efficiency such verb rotations provide.  Suppose, for
example, that you had a search engine searching for an answer to the
question:

Can the lion escape?

and that the only information in the database on lion escapes was:

The lion might escape if you forget to lock his door.

And suppose that these two sentences were not rotated, so that the main
verb of
the first was "can" and that of the second was "might."  The search engine
would find nothing by doing a verb-to-verb-to-verb search, because "can"
cannot
be matched to "might."  Only a word-by-word search of the entire database
would
reveal the answer.  But if these sentences were rotated, then "escape"
would be
the main verb in both sentences, and the match would be both quick and
sure.

For machine matching, at least, rotation clearly offers an improvement, and
it
is hard to see how brains could work much differently in this respect.  So
if
Panlingua is truly universal, and must work the same way wherever it is
found,
and if improvements in computational linguistics are indeed converging on
Panlingua as I have said, then almost certainly this verb rotation must
occur.
And if verb rotation can occur, then there would seem to be no reason to
believe that other kinds of translation (also called "transformation") do
not
take place between subsurface and deep Panlingua as well.

Trinary Links.

At one time I believed I had made a major breakthrough.  I thought I saw
that
instead of a single link having type, the links that I used in my
linguistic
models could be broken down further into trinary links consisting simply of
three diodes.  One of these three diodes would emanate from the original
source
node and terminate at a node from which the remaining two diodes emanated.
One
of this pair would then terminate at the semnod for the type while the
other
terminated at the original destination semnod.  The simplicity of this
design,
I thought, would make linguistic links and nodes easier to model in
electronic
hardware.  Unfortunately, the links I have defined carry too much
information
for this to work, and apparently cannot be modeled using anything less than
a
transistor or two-input AND gate--in other words a unidirectional valve.
As a
two-input AND gate, a link then has one input tied to the source node, the
other input to the type node, and the output to the destination node.

Lexlinks.

Recall that lexlinks (lexical links) are the links that link words to their
semnods.  In fact a single word in the lexicon may link through several
lexlinks to several semnods, each such link defining a different sense of
the
word.  But in Panlingua representations, which are not lexicon entries,
every
Panlingua word has just one and only one lexlink to a single semantic node.

At this point, the question may naturally arise, "But how do we know that
each
word has just two kinds of links?  What if there were more?"  And our
answer
must be that we DON'T know, but we can only guess.  If there were more,
then of
course our present three-dimensional model of language would no longer
suffice,
and a whole new model of language would be required.  This is quite
possible,
but difficult to conceive since so many things might have to be changed.
And
in fact we have never been able to discover anything that would indicate
the
presence or necessity of such an additional link.  So we simply assume that
it
is safe to assume that every word has just two and only two links, and that
the
absolute and entire meaning of every word is determined by the three
components
that make up its integral whole, namely (1) a connecting point, or node,
(2) a
link to another word, and (3) a link to a semantic node.  Furthermore our
theoretical understanding indicates that nodes, just like the connecting
points
in an electronic circuit, have no real meaning in and of themselves.  And
then
we know that all the syntactic information about a word (how it relates to
its
regent) is encoded in its synlink (syntactic link) type and destination.
And
so it follows that the lexlink and its destination MUST encode all that is
left
of the meaning of any word except for that which might be encoded by its
dependents.

So what, precisely, IS encoded in lexlink type, anyway?  There are several
semantic phenomena that appear to have no relationship to syntax.  Among
these
is a phenomenon commonly called "aspect."  Aspect is the "state of being" a
word calls for by the type of its link to its semnod.  For example:

durative: Her heart was breaking.  Perfective: Her heard was broken.

So for verbs, lexlink type encodes that particular kind of state of being
(that
of something breaking, broken, etc.) commonly referred to as aspect.

At first I was overly ambitious, and presumed that lexlink type also
encoded
things like potentiality (the state of being breakable), the state of being
the
action itself (breakage), the state of being possessor (his), the state of
being equipped with (winged), the state of being abundant (nutty), the
state of
being composed of (gaseous), the state of pertaining to (girlish), etc.  I
now
see that this was an error.  In fact most of this information is properly
encoded in semlink type.  The problem was that I was trying to force too
many
different kinds of words to link to a single semnod.  Everything now works
okay
after further breaking down semnod classification in order to create more
resolution (breaking one semnod up into two or more to give each a more
definite meaning).

For those of you who have access to my old manuscripts, you will see that
some
of the unanswered questions that bothered me about lexlink type before were
these:

1.  Precisely how many such states of being occur in Panlingua?

2.  What are they?

3.  Does Panlingua use different lexlink type values to encode each of
them,
or, because mutually exclusive sets of these states exist for nouns, for
verbs,
and for adjective/adverb modifiers, does Panlingua use the same set to mean
different things depending on synlink type?

I now see that these questions were just red herrings.  What I needed was
(and
still is) the cooperation of a dedicated team of scientists.  What I got
was
ridicule.  Yet by working tirelessly in the face of great loneliness, some
poverty, and unending personal loss, I have still managed to uncover this
veritable mountain of gold, and I am not finished yet!

Some of the beautiful things about studying Panlingua are that the raw
materials of natural language are ubiquitous, a great deal can be learned
just
by trying to crawl back into one's own mind, and all that is needed to
model
and test one's results is a simple computer.  Yet no other field of
research
offers greater promise for the very future of mankind.

Chapter 13, Subsurface and Deep Panlingua Structures.

By Chaumont Devin, Honolulu, December 28, 1998.


Subsurface and deep structures in Panlingua.

Recall that subsurface panlingua is defined simply as disambiguated text,
whereas the structure of deep Panlingua is not fully known.  Recall
further
that both are made up of words, and that each word is nothing but a node
and a
synlink and a lexlink emanating from that node.

Idiom.

Now that we are familiar with the idea of verb rotation, let us also
consider
yet another linguistic phenomenon--that of idiom.  What happens, for
example,
when we hear a sentence like:

There she was, in the nude!

We might first parse this as a Panlingua structure like this:

was: Past tense declarative verb, 6 atoms.  there: Adverb of location, 1
atom.
she: subject, 1 atom.  in: preposition meaning 'wearing', 3 atoms.  nude:
Prepositional object, 2 atoms.  the: determiner, 1 atom.

But what would such a structure mean?  That she was at a location distant
from
the speaker (there) in something called "the nude"?  Obviously not.  What
the
automated system really needs in order to understand this sentence is:

was: Past tense declarative verb, 4 atoms.  she: subject, 1 atom.  there:
Adverb of location, 1 atom.  naked: Predicate adjective, 1 atom.

In other words, she was there naked.

In English we have many such idiomatic expressions, and it is clear that
we
cannot derive their meanings by simply blocking off segments in a linear
string
of words.  In other words, creating an automated system that can simply
replace
every occurrance of "in the nude" with "naked" will not be enough.  This
is
because it turns out that phrases like "in the blinking nude", "in the
absolute
nude", and "in nothing but the nude" mean almost the same thing to any
native
speaker.  The additional words serve only to slightly modify the "naked"
meaning.

There are many other examples of this phenomenon in English.  For example,
"kick the bucket", "kick the old bucket", "blinkin kicked the blinkin
bucket",
"quickly kicked the bucket", etc.

Words that may disappear in deep Panlingua.

Besides the case of idioms, in which a whole phrase may collapse into a
single
word in the deep structure, there are also English words that apparently
disappear from the deep structure altogether.  I am writing of words like
"There," in sentences like "There is a man in the kitchen," and "it" in
sentences like "It is raining outside."  Recall that all texts only ever
record
things and the states things may be in.  Such words as these do NOT
represent
THINGS or STATES or anything else, but only fulfill some syntactic
requirement
of English.

Of course such words also link to semnods, but these semnods are dummy
semnods
that exist only for these words and represent no meaning whatsoever.  We
are
able to claim with some measure of confidence that they do not exist in
the
deep structure because of the following:

1. These words are not likely to occur in languages distant from English.

2. We would expect a deep internal representation to be maximally simple
and
efficient, and it would not be simple or efficient to process a
representation
cluttered with meaningless words.

Words That May or May Not Appear in Surface Languages.

On the other hand, many words in the deep structure may disappear
altogether
from surface representations.  As an example, consider the two sentences:

He thought she worked in an office.

and

He thought THAT she worked in an office.

Both sentences mean the same thing, but the second sentence has the word,
"that," while the first one does not.  What is happening, and if the
meanings
are identical, then is there a word corresponding to "that" popping in and
out
of existence at the deep level as well?  It may be that deep Panlingua is
flexible, and capable of representing the same meaning both ways--both
with or
without a word corresponding to English "that."  On the other hand, it may
be
that there really IS some universal deep grammar that is rigid and always
the
same.  Then again, the deep structure may be mostly rigid but only
slightly
flexible within certain predefined bounds.

My feeling is that a word corresponding to English "that" must be required
in
deep Panlingua, and this is the only reason the English "that" ever
appears at
all.  The subsurface language is quite flexible, as we have seen, but
probably
not so the deep structure.  My reason for assuming that some word like the
English "that" must be required in the deep structure begins with the fact
that
the ontology is clearly divided between noun and non-noun nodes, as we
have
seen.  The meaning of a word is always some THING or else some STATE a
thing is
currently in or can assume, but never both.  Next "thought" is being used
transitively, which means that it must have some patient, which in English
usually means some object, and this object is "she worked in an office,"
which
is a subordinate clause, and as we have seen, all clauses are essentially
THINGS.  But as we have also seen, all clauses have verbs as their
regents, so
that the only way a clause can link to anything else is through its regent
verb.  Yet the whole lower clause, as seen from above (from the upper
clause)
is a noun, while no verb can ever really be a noun or lexlink to the
semnod for
a THING.  Thus deep Panlingua must require some noun or other, if even a
dummy
noun, to link to the verb of the upper clause, and to which the verb of
the
lower clause can link so as to avoid being both a verb and a noun.  So my
assumption is that a word corresponding to the English "that" is ALWAYS
present
in deep Panlingua, but that because we know it is there at some
subconscious
level, and because saying what everybody knows is tedious, we often drop
it
from the surface structure for convenience whenever dropping it will not
lead
to unresolvable ambiguity.

Another such dummy word that disappears and reappears in English
subsurface
structure, but this time with strict regularity, is "to."  Consider the
sentences:

To steal chocolate is bad.

and

Stealing chocolate is bad.

In both cases, the meaning of "steal" is being used in a noun role.  In
the
first case, the "to" forms a nice, dummy-noun connecting point through
which to
link as subject to "is," but in the second sentence, which means exactly
the
same thing, there is no such convenience.  Instead, English tells
something
unusual is going on by changing the verb form from "steal" to "stealing,"
and
we guess the rest without even knowing we have done so at a conscious
level.

I believe that here the English "to" is the disappearing surface
representation
of a word that must always be present in deep Panlingua, and that each
English
gerund represents not one but two words in the simpler, more rigid deep
structure.

Another example is the sentence:

She gave me a cup of coffee.

and

She gave a cup of coffee to me.

The meanings are exactly the same, so I would assume that the deep
structure
for the two sentences is identical, the only difference between them being
a
trick of surface syntax.  So is the "to" always present in the deep
structure,
always absent from the deep structure, or neither?  Once again, the answer
must
lie in the fact that something cannot mean a noun and a non-noun at the
same
time.  Here the English "to" is a preposition of destination, which might
easily be replaced by any adverb of destination, such as in the following:

She gave a cup of coffee upstairs.

Now "me" represents a THING, whereas "to" represents motion towards and
arrival
at some destination.  I would assume therefore that in order for the
destination, "me," to link to the verb, "gave," as an adverb of
destination
would require an actual word representing a non-thing in a simpler, more
grammatically rigorous deep structure.  I would therefore assume that for:

She gave me a cup of coffee.

the English "me" is ALWAYS represented by two words in the deep
structure--one
representing a person and the other representing the state of being in
motion
towards some destination.

Another interesting example is the use of the English word, "tomorrow."
In
fact tomorrow is a day, which is a thing, and yet we use it as an adverb
in
sentences like:

I will see you tomorrow.

Here "I" is the subject-agent and "you" is the object patient, so
"tomorrow"
cannot be agent or patient.  It is clearly being used as an adverb of time
(temporal adverb).  How can this be?  Because "tomorrow" has a built-in
adverb,
and you can see it for yourself by pulling the word apart after "to-."  Of
course in modern English we would not write:

I will see you to morrow.

because "to" no longer has this temporal meaning, which it must once have
had.
So we have apparently merged the preposition with its object and left it
stuck
there while the stand-alone version has changed to "on" ("on the morrow,
etc.)
over time.  It would seem obvious that the original preposition has been
somehow retained in the deep structure while being allowed to disappear
from
the surface representation by mutual agreement over time, and this is how
"tomorrow," which represents a thing, is allowed to play an adverbial case
role
in English.

There are probably many more ways of verifying if not actually proving the
existence of a deeper, more universal Panlingua, but I think I have gone
far
enough to show that there is good reason to believe it exists.

But before going on, I must point out once again that the only way we can
"see"
Panlingua at this time is through the filter of interpretation.  We are
only
seeing the surface representations which have been translated from deep
Panlingua and not deep Panlingua itself.  It is therefore mostly true that
our
best assumptions are nothing more than mere guesses.  I have proposed that
words like "to" and "that" are ALWAYS present in the deep structure.  I
have
also said that we would expect the deep structure to be maximally
efficient.
So it might be argued that words like "that" and "to" would only clutter
the
deep structure, and that the deep structure does not need them.  In fact
it may
well be that even prepositions are entirely absent from the deep structure
because their meanings can be encoded in synlink type.  This is a very
reasonable argument, and a very elegant solution.  The reasons I believe
it is
suspect are as follows:

1. If a word sometimes appears in surface structure and at other times
does
not, then it is not necessary for surface syntax, and must therefore
reflect a
word in the deep structure.

2. If such step-down words as I have mentioned do not exist as regents for
verbs, then prepositions would probably not exist as regents for nouns.
But we
can prove that besides serving as regents for nouns, prepositions do exist
in
and of themselves in the form of adverbs without noun modifiers and as
regents
for adverbs.  As examples, consider:

She climbed up.
She climbed up the hill.
She climbed RIGHT up.

In the last example, "right" does not modify "climbed" or "hill."  It
modifies
"up" as a sibling of "hill."  I believe this more or less proves the
actual
existence of prepositions within the deep structure, else how could the
meaning
of "right" be encoded?

3. In order to be simple and elegant, the deep structure must almost
certainly
be all one way or the other--either all of these non-dummy words that
appear
from time to time in surface syntax must ALWAYS be present or else ALWAYS
absent, but not both, and since we have shown that they must be present at
least for prepositions, then they are probably ALWAYS present in ALL
cases.

Two Levels of Panlingua.

So it appears that Panlingua exists on at least two levels, which we might
call
"subsurface" and "deep", respectively.  At the first, or subsurface level,
the
thought is represented in Panlingua, but its meaning might be all wrong if
it
were taken to be deep Panlingua, or Panlingua in its purest form.  In
parsing
from English, therefore, there exists an intermediate Panlingua
representation
specific to English, that must be transformed further in order to yield
Pure
Panlingua representation, and this is probably true for ALL natural
languages,
although it would not be true for simplified and rigidly structured
languages,
such as computer programming languages and other codes.

Thus in parsing, it is necessary first to set up a literal subsurface
Panlingua
representation which maps word-for-word to the original.  This subsurface
representation will contain idioms, unrotated lower clauses, word
omissions,
etc., just as they appear in surface language.  Then to this must be
applied a
second analytical process in order to convert idioms into their true
meanings,
rotate badly formed upper-lower clause pairs, add in omitted words, etc.,
to
yield the final product, which is deep Panlingua.  And for text generation
two
distinct processes are also required, which are the exact opposites of
those
employed in parsing.  Deep Panlingua structures must first be converted to
idioms, rotated structures containing the lower clauses common in English,
etc., in order to yield subsurface Panlingua, and then this subsurface
Panlingua must be used to generate texts.

So how many levels of Panlingua are actually used within the human brain?
We
cannot be sure, but there must be at least these two.

It is sad that at this point in time there is no parser in the world that
can
even reliably get English through the door into subsurface Panlingua when
what
we really need to know are the grammatical rules for deep Panlingua and
how we
can use these to build the machines of tomorrow.

Warning: Such sentence constructions as the Hawaiian, "Aia ke kixi o
Kamehameha makai o ka hale alixi," meaning roughly, "There [is] the statue
of Kamehameha seaward of the palace," would indicate that even to obtain
the subsurface form it is necessary in many cases to insert an extra word,
namely the "be" verb.  This would seem clear because if every word but the
top verb must have a regent, then all three words, "aia," "kixi," and
"makai" require a regent from which to depend but that this regent is not
present in the original text.  It must therefore be present in the
subsurface representation although absent in the text, which means that
simple disambiguation by working out the two links for each word cannot
work unless this invisible word is recognized and used as regent.

Chapter 14, The CLOCKLESS Ontology.

By Chaumont Devin, Honolulu, December 20, 2002.


The Problem.

Until now, the only kind of automated system we have been able to apply to
linguistics has been the clocked binary computer--mostly something more or
less resembling an IBM PC.  While it appears that such machines are
ultimately capable of modeling anything that can be reduced to a series of
logical operations (IF, THEN, AND, OR, NOT, and the mathematical
expressions that can be represented using such logic), they are quickly
overwhelmed when applied to certain kinds of natural phenomena.  This is
because they are always limited by their speeds and their sizes, so we
respond by building ever bigger and faster computers.  And for our most
difficult problems, such as the modeling of environmental phenomena, we
build supercomputers having many times the speed and capacity of ordinary
computers.  We also find that we can further enhance the speed of
supercomputers by providing them with multiple processors, so that certain
operations can be broken down and distributed into thousands of individual
components, each one of which can be handled by a separate processor at
the same time.

In linguistics we are confronted by challenges that cannot be easily
overcome using any kind of ordinary computer or supercomputer.  The basic
problem is one of permutations and speed.  Each word in a sentence can
have many meanings.  You can easily verify this by looking up words in a
dictionary.  Each word usually has several senses, and each such sense has
a different meaning.  In order to exhaustively and reliably parse any
sentence, EVERY sense of EVERY word must be considered.  In other words,
every possible PERMUTATION of meaning must be examined and either accepted
or rejected as a reasonable meaning for the sentence as a whole.  This is
not difficult for sentences of two or three words, but quickly gets out of
hand as more words are added.  As an example, suppose that we have a
sentence that is five words long, and the senses for each word are as
follows: 5, 3, 22, 2, and 6.  5 * 3 * 22 * 2 * 6 = 3,960 permutations to
be considered.  Now suppose that instead of five words, as described
above, we have a sentence with six words--the five words already mentioned
plus a word that can have ten meanings.  Now the number of possible
permutations to be examined has risen from 3,960 to 39,600 for only six
words.  Yet many sentences run on past 20 and even 30 words!  How many
permutations would be possible for a 20-word sentence each word of which
averaged five senses?  5 * 5 * 5 * 5 * 5 * 5 * 5 * 5 * 5 * 5 * 5 * 5 * 5 *
5 * 5 * 5 * 5 * 5 * 5 * 5 = 95,367,431,640,625--or something over 95
trillion!  Now suppose our binary computer has a clock speed of 2ghz.
Even if each evaluation took only one clock tick, which by current
technology is impossible, this sentence would take at least 95 * 10^12 / 2
* 10^9 = 47,500 seconds, or about 132 hours to parse.

The example I have given above is only intended to provide the reader with
an intuitive feel for the kinds of numbers involved in natural-language
processing.  In reality, of course, a good programmer would write many
kludges into his software capable of selecting out inappropriate word
senses to keep them from further evaluation, and this would save a good
deal of time.  But on the other hand I only gave one clock cycle per
evaluation, which is ridiculously conservative and entirely out of the
question at this time.  My point is simply that we cannot hope to parse
natural languages in real time without a dramatic improvement in computer
design.

So let us try to determine precisely what part of natural-language
processing takes up most of our time.  It is clearly our use of the
ontology, which must be consulted at every step during parsing as well as
in various other operations.  Among other things, the ontology must be
able to answer the question of what can do what, so let us consider the
question of whether dogs can cry.  What we need is a YES or NO answer, and
we need it right NOW.  But in a clocked ontology (an ontology set up as
software and data in an ordinary computer) this cannot be.  The ontology
software begins by looking at the entry for DOG to see if there is any
direct agency link to CRY, and it finds there is none.  So it searches the
DOG entry to see if a dog is also something else, and finds that a dog is
a mammal.  Then it searches for an agency link from MAMMAL to CRY, and
finds that there is none.  By this time the ontology has probably consumed
some thousands of clock cycles, and yet we have only begun.  Now the
ontology tries to determine whether a mammal is anything else, and finds
that it is a vertebrate, and the whole cycle is repeated for VERTEBRATE.
This process continues up through CORDATE and ANIMAL and
OBJECT-HAVING-PHYSICAL-FORM and MATERIAL-THING and finally as far as
THING, which has no hypernyms, before finally being able to return an
answer of NO, dogs cannot cry.  Thus the time it takes the ontology to say
NO may be so long that a computer user can actualy perceive the delay,
even at a clock speed of 1 or 2 gigahertz.

The Solution.

What is needed is a kind of black box that can answer any such query
IMMEDIATELY--in other words, a CLOCKLESS ONTOLOGY.  Such a device would
result in a quantum leap along the path toward 100% reliable parsing,
machine understanding, and the Holy Grail of artificial intelligence.

What we need is a "black box" which can be accessed by our software.  It
must have a minimum of 16k nodes (connecting points) to which various
links will be attached.  These nodes should be internally addressable by
the device.    We should be able to send it a message such as "forge 12519
agt 16019," meaning to create a link of type "agent" from node #12519 to
node #16019.  We should also be able to send it a message such as "verify
12519 agt 16019," meaning to return YES if such a link has been defined,
and NO if it has not.  And finally we should be able to send it a message
such as "break 12519 agt 16019," meaning to break the existing link.
Since the device will be essentially unclocked, it should be able to
respond without any delay except for the clock ticks that may be required
internally to set up the query and response with no clock ticks lost in
actually determining whether a link exists or not.

It may well be possible to design a basic hardware ontology that would
work across many languages.  After all, dogs must still be mammals and
mammals animals in Eskimo and Swahili.  And the design of such a device
would remain fairly stable over time, after all, dogs are not about to
change into frogs during our lifetimes.  So the same design, once
implemented, could be used again and again for a very long time.  A great
deal of work would be required to select the 16k semnods that are most
common in everyday use in languages around the world, and to come up with
noun groupings that would not be susceptible to much alteration later on.
But once these 16k basic semnods have been established, we should be able
to use them to build a fast core hardware ontology capable of serving for
any natural language.  Upon this solid core foundation, then, might be
built slower software to handle less common semnods and language-specific
exceptions of various kinds.

Reality or Dreams?

As long ago as August of 1996, I conceived the idea of a DIODE ONTOLOGY.
In its simplest form, it might be visualized as a network of pins on a
board, all of which are connected to diodes.  Each diode would then
represent a hypernym link, and each pin would represent a semantic node.
Thus, for example, from the pin representing MONKEY, there might be a
diode leading to PRIMATE.  In this way, whenever a voltage were applied to
the pin for MONKEY, the pin for PRIMATE would also go high, indicating
that a monkey is a primate, and if PRIMATE were in turn connected by
another diode to MAMMAL, then the pin for MAMMAL would also go high, etc.
If there were then a means whereby each pin might be selected and tested
by number, even such a simple diode ontology would be of vast benefit to
us at this time, since a great deal of the queries made to the ontology
are about hypernymy.  Whenever our software needed to test for hypernymy,
it could just set the pin representing some hyponym to high, and then
check to see if the pin representing the possible hypernym went high as
well.  Additional diode ontologies might then be created for such things
as patiency, agency, inalienable possession, etc.  Then to find out
whether dogs can bark, we could set the pin for DOG to high and check to
see whether the pin for BARK went high also.

But this particular kind of hardware ontology would be cumbersome for
various reasons, some obvious and some not so obvious.  The first problem
would be that if we were to create a separate diode ontology for each kind
of semantic link (one for hypernymy, another for synonymy, another for
agency, another for patiency, etc.), then instead of having only one pin
to represent each semantic node, for each such diode ontology we would
have to have a whole separate array of pins representing the same semantic
nodes.  Much time would have to be wasted going back and forth between
ontologies, and any change to the general ontology might mean having to
make corresponding changes in a whole series of ontologies dedicated to
specific semantic link types.

Better results might be achieved using transistors or two-input AND-gates
to represent semlinks.  Then we might have one input of an AND gate
connected to DOG, another input to ISA, and its output to MAMMAL.  To test
whether a dog were a mammal, then we would apply a YES to the ISA pin and
the DOG pin and test to see whether the MAMMAL pin were a YES or a NO.  Or
we might have another AND gate with inputs DOG and AGT and output BARK,
and still another with inputs BULLDOG and ISA and output DOG. Then by
applying YES to BULLDOG and ISA and AGT, and checking to see whether BARK
were then YES as well, we might be able to determine whether bulldogs can
bark.  This solution is very elegant because it can automatically ascend
the hypernym network and traverse the agency link in a single operation.
But it DOES seem to imply the necessity of being able to set more than
two nodes to YES at a time, as in the case of BULLDOG, ISA, and AGT in the
above example.  We could try to use diodes instead of AND gates for
hypernymy, which would be the same thing as tying ISA permanently to YES,
but doing so might interfere with other operations.  At the time of this
writing, I have not yet succeeded in thinking this through, but I realize
that whether it is possible or not may depend largely upon the fact that
the semnods of the ontology fall into two broad classes--semnods
representing things and semnods that do not represent things--and that
neither can be a hypernym of the other.  Furthermore, all the operators in
the ontology are simply semnods of the non-thing classification.

When designing a hardware ontology, there are two major problems that
must be considered.  These I have called BACKFLOW and METAFLOW.  In
backflow, fluid pressure is allowed to propagate both forwards and
backwards where it should be allowed to propagate in a forwards direction
only.  In METAFLOW, fluid pressure is allowed to propagate forwards beyond
some point at which it should stop.  In the diode hypernym box, backflow
cannot happen because diodes are small check valves that allow the
electric fluid to flow in a forwards direction only.  Thus it is possible
to get a YES reading for such propositions as "house is a building," and
still get a NO reading for "building is a house." Electricity can flow
from the semnod for "house" to the semnod for "building," but there is no
way it can get back from "building" to "house," so the hypernym function
will work.  But in hardware ontologies employing multiple links, possibly
through both diodes and transistors, both backflow and metaflow might be
a problem.

From these two examples (that of the diode ontology and that of the
transistor or AND-gate ontology) it should be clear that I am not
dreaming.  These things ARE possible, and possible with current
technology.

Basic Underlying Theory.

In clockless ontologies the operational principle is not the recursive
traversal of semlinks through time, but just changes in the pressure of a
fluid.  This fluid might be water, oil, or anything else, but using
current technology, of course, our fluid of choice would be the same old
electricity we use for all other computer parts.  In such a system it is
no longer necessary to traverse links in order to determine continuity.
All one need determine is whether pressure applied at the source reaches
the destination.  Fluid flow requires time, but fluid pressure change is
almost instantaneous throughout a continuous hydraulic system.  Because of
this, a clockless ontology would virtually eliminate the problem of
processing time--not for the entire linguistic processing system, but for
this major component we call the ontology.

The path towards incalculably faster, human-like computers may not lie in
the development of ever-faster CPU chips and more kb of RAM.  Most of us
already have more hardware capability than we know what to do with, and
this could well be our problem.  Biological systems operate at far slower
speeds than do modern digital devices, and yet every healthy cockroach can
outperform all the best supercomputers of our time.  It would seem that
our deficiency lies not in any lack of CPU power, clock speed, or memory
capacity, but rather in the matter of internal arrangement and design.  We
are often informed that the reason animals can do what they do is because
of distributed processing over untold millions of CPUs.  I don't think I
can agree.  Give a man a million CPUs and he will probably end up with
nothing more than a million copies of his same old unsolved problems.  The
plain truth is that we suffer from a great ignorance of how we should put
together what we have.  But as we have seen, besides actual dissection,
the study of language provides our only detailed insight into the logical
workings of our minds, and hence into the workings of other automated
systems.  Even my own simple observations above have shown us how to build
a better, faster computer--one capable of doing things our best current
machines simply cannot be made to do in any way, shape, or form, and that
at very minimal cost.  You may smile and say, "But that's only for
linguistics."  Yet it would appear that linguistic phenomena lie at the
heart of all intelligence, and that there can be no life without
intelligence nor intelligence without life.  We may have blinded ourselves
by staring too long at the current computer paradigm.  We may not need
anything bigger or faster or more "powerful," but only a smarter
arrangement of the things we already have, and the way to learn this
smarter arrangement is by carefully observing nature and deducing the way
she gets things done.  And this is precisely what the theory of Panlingua
is all about--namely learning to do things better and smarter by a
careful and painstaking analysis of what is really going on inside our
heads.

Chapter 15, Religion.

By Chaumont Devin, Honolulu, December 23, 2002.


After looking over my notes about Panlingua, I see that here we have some
pretty potent stuff, in fact the very stuff of God!

1. Without beginning.
2. Indestructible.
3. Without physical form.
4. Omnipresent.
5. Immutable over eons of time.
6. Representing all the knowledge that ever was.
7. Able to synthesize new knowledge from old.
8. Invisible.
9. Utterly hidden to unbelievers.
10. Shrouded in mystery.
11. Underlying all things.
12. Holding the promise of unlimited power.

This is the very essence of Jehovah, and all within our grasp!  What more
could
one ask?

I shudder.

The last thing I would EVER want to do would be to create a new religion.
But
just for fun, let's see how Panlingua might relate to one of the old
religions
we already know:

The Gospel According to St. John.
The King James Version!

Chapter 1 (with minor alterations and annotations).

1.  In the beginning was Panlingua, and Panlingua was with God, and
Panlingua
was God.
2.  Panlingua was in the beginning with God (With God and actually God as
well
because Panlingua is the invisible essence encoding ALL intelligence
including
that of God).
3.  All things were made by Panlingua; and without Panlingua was not any
thing
made that was made (because there IS no life wihout embedded intelligence
of
some kind, and ALL intelligence is Panlingua).
4.  In Panlingua was life (because no atoms can join to form living tissue
without intelligence to guide them); and the life was the light of men
(because
life has resulted in human awareness).
5.  And the light shineth in darkness (on comp.ai.nat-lang); and the
darkness
comprehended it not (and you can say THAT again!).
6.  There was a man sent from God, whose name was Joe.
7.  The same came for a witness, to bear witness of the light, that all men
through Joe might believe (Ha, ha!).
8.  Joe was not that light, but was sent to bear witness of that light (and
oh,
how he did!).
9.  That was the true light, which lighteth every man that cometh into the
world (because ne-ery a one of them but what was driven by Panlingua from
the
womb).
10.  Panlingua was in the world, and the world was made by him, and the
world
knew him not (So sad, too bad!).
11.  Panlingua came unto his own (tenured professors), and his own (tenured
professors) received him not (but made as much fun of him as they had
time).
12.  But as many as received Panlingua, to them gave God power to become
his
very own sons, even to them that believe on Panlingua's name:
13.  Which were born, not of blood, nor of the will of the flesh, nor of
the
will of man, but of God (who really IS this invisible Panlingua).
14.  and Panlingua was made flesh (in all of us, because we are all the
hosts
of this very same old immutable Panlingua), and dwelt among us (and in us),
(and we beheld his glory, the glory as of the only begotten of the father,)
full of grace and truth.
15.  Joe bare witness of Panlingua and cried, saying, this was he of whom i
spake, he that cometh after me is preferred before me: for he was before me
(Yep, quite a bit before me, as a matter of fact).
16.  and of his fulness have all we received (because nobody is big enough
or
evolved enough to implement it all), and grace for grace.
17.  for the law was given by moses, but grace and truth came by jesus
christ
(ultimate implementation of Panlingua for all time).

Kai ho Logos in to skotia fainei!

Kinda funny, but actually pretty deep.  As a matter of fact, I know the
original words of John were burning in my de-Christianized subconscious all
the
time that I was working on the theory.  Did John really know things?  But
HOW,
and at what level?

Chapter 16, Universal Grammar.

By Chaumont Devin, Honolulu, February 12, 2003.

So does this thing, the universal grammar, really exist?

The answer is a definite YES, but its rules are very simple, and here they
are:

1. All linguistic structures are built from nothing but linguistic links
and nodes.  A node is either the source or destination of a link.  Every
linguistic link has three elements: a source node, a type semnod, and a
destination node.  Thus every linguistic link consists of three simpler
direct links having no type but only direction.  These are, (1) a link
from the source node to a central node, (2) a link from the central node
to the destination node, and (3) a link from the central node to the type
semnod.  The central node is completely isolated from everything outside
the linguistic link of which it is a part--nothing else can ever link
directly to it, and it can link to nothing else.

2. Every word of every language is nothing but two links and a node from
which they both emanate.  One of these links, called a synlink, links the
node of the current word to the node of another word called its regent.
The other link, called a lexlink, links the node of the current word to a
semnod.

3. Each word can have multiple dependents, but only one regent.  A
dependent is a word whose regent is the current word.

4. Besides the regent-dependent relationship, there is also a proximity
relationship between each word and its regent.  Thus each regent has a
first dependent, a second dependent, a third dependent, and so on.  The
first dependent is nearest to the regent, the second dependent is one
dependent away from the regent, and so on.

These are the cardinal rules of universal grammar upon which all
linguistic phenomena are constructed.  The rest is only detail.  For
example, there are proximity rules for dependents that remain to be worked
out by means of studies across many languages, etc.

Chapter 17, What Is Grammar?

By Chaumont Devin, Honolulu, February 13, 2003.

If you look up the word, GRAMMAR, in a dictionary, you will almost
certainly find something to do with rules, although different dictionaries
may have differing views as to just what these rules may be about.  For
Panlingua, the definition of GRAMMAR is as follows:

The set of facts that needs to be known in order to (1) generate linear
texts in some language from Panlingua arrays, and (2) convert linear texts
written or spoken in the same language to Panlingua arrays.

The latter process, of course, is known as parsing.  This can be a simple
process using simple rules, as in the parsing of rigidly proscribed
computer instructions.  But in the case of natural-language parsing, it is
the most difficult problem known in computer science today.

Syntactic rules.

The first set of grammatical facts or rules to be considered is usually
syntax (what word symbols can go with what other ones).  For example,
every native speaker of English knows without even thinking about it that
"I" can have a subject relationship with "am," as in "I am."  Likewise
he/she will immediately know that "he am," "she am," and "it am" are
wrong.  Why?  Over many generations, native speakers of English have
gradually come to associate "I" with "am;" "he," "she," and "it" with
"is," etc., so that these have come to be seen as "rules" of English
grammar.  These rules of surface syntax have no deeper significance
whatsoever, but we rely upon them for UNDERSTANDING, which is just another
word for PARSING.  When we see an "I" before an "am," we know immediately
that "I" is the subject of the verb, "am."  As might be expected, these
syntactic rules are more developed for the most commonly used words of our
language, such as groupings involving personal pronouns and the verb, to
"be."  They help us understand what is being said or written quickly and
with minimal effort.

But the same syntactic rules used in parsing or understanding have an
inverse role in the opposite direction.  Thus when we think of putting the
1st person singular pronoun into a subject role, we immediately think,
"I," and not "me."  And if this is followed by the "be" verb, we
immediately think, "am," and not "is."

All of these things come naturally with the human linguistic apparatus,
and all seem to flow seemlessly and effortlessly together in the swift
stream of human expression and thought.  No one need tell us these myriad
rules, because they are all "set in stone" within our minds, and have been
ever since we were three or four years old.  But to a foreigner attempting
to express him/herself in English, this is a very different matter.  Even
intelligent human beings who were equally endowed with all the linguistic
machinery possessed by native speakers of English in the beginning may
find it quite impossible ever to master our native language completely.
The foreigner may say or write everything correctly for a long time, then
suddenly say something like, "When my parents arrived to America from
Russia," and we have nailed him/her as a foreigner beyond the shadow of a
doubt, because every native speaker of English would have said "arrived IN
America" and not "arrived TO America."  So these grammatical "rules" do
exist, and there may be literally millions of them.

But now we are trying to teach language to dumb computers, who have never
had any kind of linguistic apparatus whatsoever.  How can we possibly
teach these many rules, which seem endless when we think about them, to a
stupid machine?  We shall see.

Morphological Rules.

The morphology of a word is the shape it assumes.  Please notice that here
I am using "word" in its most precise meaning, which is literally a
Panlingua atom, and not a written or spoken symbol.

The word, to "be," can take various forms, each one of which is recognized
immediately and with unquestionable reliability by every native English
speaker, as we have already seen.  AM, IS, WAS, ARE, WERE, BEEN,
BEING--all these are different surface forms assumed by word symbols
generated from Panlingua atoms linked to the same semnod, and we know them
very well.  But the "be" verb is one of the most common words in the
English language, and this is why it has these special unique forms that
conform to no morphological rules.  Other, less frequently used words,
employ standardized forms for similar nuances of meaning.  For example,
KILL, KILLS, KILLED, KILLING; JUMP, JUMPS, JUMPED, JUMPING; etc.  We can
remember AM, IS, WAS, ARE, WERE, BE, BEEN, and BEING with no trouble, but
our minds would boggle were we to be forced to memorize unique and special
words covering all of these forms for all English verbs.  So Mother
English has provided a way around this by giving us different word endings
for different shades of meaning.  These are, of course, -s for 3rd-person
singular, -ed for past participle and past tense, and -ing for present
participle or else something that sounds similar to these endings.

For some languages, this kind of standardization is applied very rigidly
to just about every word symbol in the entire language without exception.
For example, consider the regular verb endings for Latin and Greek.  Such
languages are so thorough in this matter that they even provide special
case forms for every noun, and perhaps even gender.  Thus traditional
grammarians have made many seemingly exhaustive inquiries into this matter
of morphology.

The unfortunate fact is, however, that morphology provides only a kind of
rough-and-ready rule of thumb by which humans can sometimes figure out the
probable meanings of unfamiliar words, and then imposes the very tedious
requirement of having to select the precise form required by consensus
during text generation.  Examples of mistakes in this latter process
abound among children."I go-ed to the park," etc.  Fortunately for us,
English speakers are not required to memorize hundreds of verb
conjugations, as in Latin.  Instead, we have been saddled with the worse
curse of having to learn to spell every single word in the language by
heart, which would seem infinitely more brutal.

But having unique and precise forms for every word meaning is a great boon
to computers, since it makes parsing much easier because there are fewer
potential word senses to disambiguate.  Also, computers can easily be
programmed to accurately output any number of inflexions or morphemes.

Semantic Rules.

At last we come to the most essential and yet most elusive set of facts
about language.  For manyd ecades, and perhaps even centuries, linguists
have pondered sentences such as "Fruit flies like a banana," "Little
Johnny found his blocks in the corner of his pen," etc.  How can humans
possibly know that the "flies" in "Fruit flies like a banana" is not a
verb?  How can humans know immediately that Little Johnny's pen is not a
writing instrument? etc.  There seem to be some kind of grammatical
"rules" underlying all of this, but what are they, and how do they work?
These are the questions that have brought computational linguistics
straight up against a brick wall.

Consider the sentences, "She sang with her fingers," and "She thought with
her toes."  Are these sentences grammatically correct?  Some may say so,
while others will disagree.  Syntactically, they seem to break no rules;
but grammar is not just syntax.  Our idea of correct grammar is not
usually confined to the question of whether words having the right
part-of-speech designations end up in the right places.  It also demands
that sentences make good sense.  Thus "Fruit flies like a banana" cannot
make very good sense unless we know that a fruit fly is a kind of
herbivorous insect.

People try to collect all the basic rules required to understand and
communicate in languages into little books known as "grammars."  But can
we really write a book that would make it possible for us to understand
and generate every conceivable sentence even in a single human language?
The answer must clearly be NO, because the number of nonsensical sentences
that spring immediately to mind would seem to be infinite.

The reason for this apparent grammatical intractability is the lack of an
essential ingredient, namely the ontology.  People have compiled word
lists and dictionaries for centuries, but until recently, no one knew the
importance of various kinds of linkages between words.  I will not attempt
to reiterate my description of the ontology here, since it has been more
completely covered in another section.  It is only with the help of an
ontology that it is finally even conceivable that one might really be able
to write some kind of grammar capable of covering every possible
construction in a particular natural language.  After all, besides
beingworking computer modules, ontologies can also be written out
tediously upon paper.

So how can we teach computers this seemingly infinite set of rules?  Again
I will fudge and say a little bit about the ontology here.  Suppose we
have an ontology in which every kind of bird islinked to the semnod for
BIRD.  For example:

goose is a bird
duck is a bird
oriol is a bird
sparrow is a bird
finch is a bird
...

Now suppose that we have forgotten to tell the computer that birds can
fly.  Then the sentence, "The hen flew over the chicken coup" will be
meaningless unless we have linked in airplane flight, in which case the
hearer might assume that, oh yes, the hen was aboard a jet aircraft.  And
if the sentence is meaningless, then for all intents and purposes it MUST
be grammatically incorrect.

But then suppose that we merely add ONE SINGLE link to our growing
ontology, namely:

birds can fly

Then, through the rules of hypernymy, our ontology immediately knows that
any bird can fly unless specifically linked otherwise (linked to FLY by a
negative link).  Thus by adding just one single link, we have immediately
increased the number of grammatical "rules" our system knows about by
potentially thousands, since there are thousands of species of birds in
the world, not to mention the proper names of individual birds.

This, then, is my answer to the question I proposed above, namely, "How
can we possibly teach these many rules to a stupid machine?"  My
predecessors were oblivious to the existence of any ontology in the modern
sense of the word, and this rendered them incapable of accounting for the
knowledge of children.  Where did this amazing knowledge come from?  Plato
thought it must have been inherited from past lives by reincarnation.
Noam Chomsky suspected that children were somehow born with it already
present inside them.  "How do we come to have such rich and specific
knowledge," he wrote, "or such intricate systems of belief and
understanding, when the evidence available to us is so meager?" (Chomsky,
1987).  And of course this same belief in the existance of an innate body
of knowledge passed down from previous generations also led to his
particular theory of universal grammar and to statements like the
following:

"Abstractly, and under a radical but quite useful idealization, we may
then think of language-learning as a process of selecting a grammar of the
appropriate form that relates sound and meaning in a way consistent with
the available data and that is valued as highly, in terms of the
evaluation measure, as any grammar meeting these empirical conditions."
--"Studies on Semantics in Generative Grammar," 1972.

It seemed impossible that any child could master the intracacies of
language in just two or three years of unsystematic exposure to a chaotic
world, therefore each child must be born with many "forms" of universal
grammar already inside his/her own head to choose from.  With our present
understanding of the ontology of course such a hypothesis is no longer
necessary, and yet it is clear that we are all born with massive
intelligence inside our heads.  But instead of being static, as in a
limited array of preset forms of universal grammar from which to make a
selection, this intelligence is dynamic and able to learn things quickly
in ways we had never dreamed.  In fact the number of possible grammars it
can handle is for all intents and purposes infinite.

So we conclude that what we call grammar must have two essential elements:
the syntactic and the semantic.  Of these two, of course, syntax is the
simpler, and can be fairly thoroughly described for any particular
language within the pages of a single volume.  But it is impossible to
understand the semantic component without an ontology, and the modern
ontology is a very recent arrival on the scene.  Syntactic rules can be
hard-coded into computer programs, but semantic rules cannot.  This is
essentially the mysterious reason why natural-language computer processing
(machine understanding, machine translation, etc.) has never seemed to
work.

As an example of the power of the ontology, let us consider the problem of
the native speaker and the foreign speaker of English mentioned above.
One would say, "My parents arrived in America," while the other might
unhesitatingly assert that his parents "arrived TO America."  The ontology
can immediately tell us which of the two versions is grammatically correct
because it will have a link such as "arrive in place," but the "arrive to
place" link will be absent.

Is this matter of the "to" and the "in" a grammatical rule?  Many would be
quick to say it IS, but I would disagree.  This is precisely the kind of
"rule" which is really no rule at all, but only a fact buried in the guts
of a correctly formed ontology.  It exists only because a semlink has been
forged from "arrive" to "place" through "in."  Every ontology contains
thousands of such linkages from which the presence or absence of literally
billions of others can be infered.  Were we to attempt to perceive these
as "rules," then the whole world and half the universe itself would be
filled with only rules, more rules, and nothing else but rules!  Such
things must not be thought of as rules, but rather as facts that can be
determined by consulting a correctly formed ontology.

So I have summarized the essentials of grammar, hopefully in a way no one
else has done before.  I hope that this brief description will serve to
open your eyes to the true nature of language, which defies and transcends
all the bounds of traditional "grammar" because its very nature and
existence assume the presence of the ontology, which has never been
developed outside the human mind until now.  But before closing, I would
like to leave you with MY definition of grammar, which is this:

The GRAMMAR of any language is correctly defined as the relationship of
that language to Panlingua.

For this reason, all grammatical works written in ignorance of Panlingua
are something like descriptions of biology written in ignorance of
animals.  They have their place, but they lack any definitive goal, and
they fail to fit the natural languages they describe into the larger
framework of language in general.  It is my ernest hope that instead of
being ridiculed and ignored, the work I have done will now be taken
seriously and applied to the many disappearing languages of the world
before it is everlastingly too late, because we will have no second
chance and we don't have much time.

Chapter 18, Natural Language Acquisition.

From the earliest stages, the child knows vastly more than experience has
provided. --Noam Chomsky.

Chapter 19, Machine Learning.

How Ontologies Learn.

In order to make this section easier to understand, let us assume that the
electronic version of the ontology under discussion here has been created
using a rigidly modular approach, so that when we speak of "the ontology,"
we speak not only of the data in the ontology, but also about all of the
algorithms used to access and manipulate that data.

First let us consider the kind of learning that results from a direct
instruction from an external, controlling intelligence.  For example,
there is a computer user at the computer console who has instructed the
ontology to create a semlink such as:

sparrow isa bird

indicating that a sparrow is a kind of bird.

The ontology has no choice but to insert the link, "sparrow isa bird."

But here, already, and without further adoo, the ontology has learned a
bewildering myriad of facts.  Now it knows that sparrows eat, sparrows
drink, sparrows fly, sparrows live, sparrows procreate, sparrows have
eggs, sparrows die, sparrows breathe, sparrows have legs, sparrows have
wings, sparrows have beaks and feathers, etc.  In other words, the
ontology now knows that a sparrow has all the generic attributes of birds.

But what if the user instructs the ontology to forge the link:

ostrich isa bird

Then the user decides: "No, that may be true, but its not quite right."
So he types:

flightless_bird isa bird

and

ostrich isa flightless_bird
flightless_birds cnt fly (flightless birds can't fly)

What happens now?  The ontology already knows that an ostrich is a bird,
but now it learns that an ostrich is a flightless bird, which is also a
kind of bird.  At this point it becomes clear that certain decisions will
have to be made by the ontology itself, and that certain automated
processes must occur.  The ontology "slides" the terminus of the hypernym
link down from bird to flightless_bird.  In other words the ontology
itself has learned that there can be no direct hypernym link from ostrich
to bird, but that the linkage must be ostrich isa flightless_bird isa
bird.  So the ontology has learned something automatically without any
direct outside intervention.

A similar kind of learning occurs in the case of agency.  For example, the
ontology may be instructed to forge the semlink:

sun can set

Then, at some later time, the ontology receives another instruction, this
time telling it to forge the semlinks:

Sun isa heavenly_body
heavenly_body can set

What must happen?  In this case, instead of the terminus of the semlink,
"sun can set" sliding downwards from a higher node in the hypernymy
hierarchy, it is the source or origin of the semlink that must move up
from sun to heavenly_body.  So the ontology itself has learned
automatically that it doesn't need the link, "sun can set," because the
sun is a heavenly body, and heavenly bodies can set, therefor it is
already clear that the sun can set unless specified otherwise.

But what if the ontology has never been instructed to forge the link:

birds can fly

Yet it has the links:

duck isa bird
ducks can fly
sparrow isa bird
sparros can fly
goose isa bird
geese can fly
etc., etc., etc.

There are two options: either the ontology is programmed to ask a human
operator whether birds can fly, or else the ontology may be programmed to
automatically assume that birds can fly because all the birds that it
knows of can fly.  In either case, some automated learning will occur.

For if the ontology determines that birds can fly, then it will have to
break all of the following:

ducks can fly
geese can fly
sparrows can fly
etc., etc., etc.

which is also a kind of learning, because the ontology has automatically
learned that these links are unnecessary.  But this process is more often
refered to as ECONOMIZATION or KNOWLEDGE CONSOLIDATION.  So automated
learning and economization occur simultaneously in the ontology.

A further refinement based upon observations made by the Brazilian
philosopher, Sergio Navega, runs as follows.  Suppose during the creation
of an ontology, the human operator enters the following:

duck isa bird
duck can fly
sparrow isa bird

and suppose that the operator has not yet entered the following:

bird can fly

Then suppose the operator enters:

sparrow can fly

An ontology equipped with this refinement would then check to see whether
sparrow might have a hypernym in common with anything else that can fly and
find out that ducks are also birds and that ducks can also fly.  The
ontology would then create a semnod linked to no english word, which we
shall represent by a dot (.), as follows:

. isa bird
. can fly
duck isa .
sparrow isa .

The ontology would automatically break the "duck can fly" link as
previously explained.

Then if the operator entered:

bird can fly

the link:

. can fly

would disappear, and . would have no outgoing links except:

. isa bird

so the ontology would simply combine . and bird.

However this approach becomes complicated in cases where, after creating
the nameless semnod, which can fly, an operator enters something else that
is a bird, because it will be impossible for the ontology to know whether
or not the new entry can also fly.  One way of keeping this from happening
might be to wait until there are x number of birds that can fly before
creating the nameless semnod, where x is some arbitrary threshold.

Still another kind of machine learning associated with the ontology works
as follows.  Suppose the human operator enters the following:

apple sng
apples plr *
apples are red

but then, at some later time enters:

apples cbe green

A good ontology should know that nothing can be red and green at the same
time, and should therefore automatically change the entry, "apples are
red," to "apples cbe red," denoting potential only.

Upgrading the Ontology from Panlingua Arrays.

At a higher level, systems that employ Panlingua can learn by examining
Panlingua arrays.

For example, suppose the system inspects a Panlingua representation of the
sentence:

The sparrow flew swiftly through the air.

From this sentence alone the following semlinks can be gathered for the
ontology:

sparrow cbe the
sparrows can fly
fly cbe swiftly
fly cbe through
air cbj through

So the controlling program can automatically analyse all Panlingua arrays,
pull out the semlinks it finds in them, and instruct the ontology to add
them to its data.

Furthermore, parsers can be written to query human operators when they
cannot find enough information in the ontology and other linguistic
structures to parse certain phrases.  Then, once the sentence has been
correctly parsed, again the system can analyse the resulting Panlingua
array as above so that the parser will know how to handle similar phrases
in the future.

The Problem of False Information.

But how can a computer be kept from learning bogus information?  What if
the operator steps out for a cup of coffee and a mischievous child starts
typing information at the console.  For example:

Bird isa fluid
fluid isa solid
etc., etc.

Computer security is always an issue, and this kind of problem can be
controlled by means of passwords, old-fashioned vigilance, and other
conventional means.

But no matter what, errors will occur.  The operator may get confused and
type in something bogus.  Parsed information may be bogus.  But more often
an imperfect parser will set up a bad Panlingua array.

One way of limiting the incorporation of bogus information into the
ontology is by providing it with a mode switch which can be toggled
between "learn mode" and "do not learn mode."  Then unless the ontology is
in learn mode, no matter what happens during a session, nothing will be
saved and no changes will be made to the original data.

Weighting.

But the most sophisticated way to ensure that bogus information will not
become part of an ontology is through the use of WEIGHTING.

In computer jargon, a weight byte is a byte containing information about
how important, accurate, etc., some snippit of information may be--in
other words, how much weight should be attached to it.

The simplest approach to weighting for the semlinks of an ontology is
probably using "frequency of use" as criterion.  Using this method, each
time a semlink is used, the weight byte gets bumped up (a one is added to
the weight byte) unless a fixed maximum has been reached.  Then a periodic
purge is performed upon the ontology during which all semlinks with weight
byte values below a certain threshold are discarded, and all weight bytes
are reset to zero.  In this manner, any bogus semlink will ultimately
disappear.  The problem is that valid semlinks may also disappear, making
it necessary for the system to establish them all over again.  Then the
shorter the period between purgings, the less bogus information will lurk
in the ontology but the more valid semlinks will be lost; and the longer
the period between purgings, the less good semlinks will disappear but the
more bogus semlinks will remain.  Etc.

Is this also a form of machine learning?  Yes, because the ontology
automatically learns that certain semlinks are bogus and destroys them.

Learning the Contents of Panlingua Arrays.

And finally, after comprehensive analysis for ontological information, the
Panlingua arrays themselves can be stored as fresh knowledge and used to
respond to queries and execute tasks.

Summary.

From the above it can be seen that an ontology can learn very easily and
quickly using even the simple algorithms I have just given, and that this
learning is of two types: (1) learning directly using inputs from the
outside world, and (2) learning through self examination.  Are there other
machine-learning techniques associated with the ontology that remain to be
discovered?  Only time will tell.  And as for the analysis of Panlingua
arrays, I have only brushed the surface in this essay.  So this is a very
exciting field, and a wide-open one in which to make new discoveries.