The development of the current theory of Panlingua was a long process.  It
began as simply a method for encoding knowledge in such a way as to be
accessible to and usable by computers.

At first I experimented with templates.  A verb may have a subject, an
object, a time, a place, etc., so why not make a generic template for all
of these things?  The problem was and is that the number of constituents
(direct dependents of the verb) can be anything from 0 to some arbitrarily
large value.  I had to abandon the idea of templates, and so did Doug
Lenat, later, at Project Cyc.

But I saw that most sentences have a verb at their heart, and that this
verb could have a varying number of dependents.  I succumbed to the
temptation of overemphasizing this point and pushing it to become "ALL
sentences have a verb at their centers" instead of just "all sentences
have a top word in their dependency structures."  This turned out to be a
mistake.

This idea of the centrality of the verb and the notion that every sentence
has a verb at its center, either real or imagined, then led to the idea of
the "morphogen."  A verb can appear by itself, for example just "go."  But
sentences become better defined and more unique as we add "morphogens."
For example, "You go," then "You go to town," then "You go to town this
afternoon," then "You go to town with me this afternoon," etc.  I now
deprecate this idea, although it still seems to be valid.

And if templates weren't going to work, then I needed some other kind of
basic linguistic building block that could be used to build all of the
structures of language, so I came up with a "universal atom of meaning"
without seeing that what I was talking about was really just a "word."
John Soa did something similar.  He incorporated his idea into a system
that came to be known as "Predicate Calculus," which employs words or
atoms that are somewhat more complicated than mine were.  My work was ever
driven towards simplicity because of my profound belief that elegance is
simplicity.  Make a system that is as simple as it can possibly be and yet
produce the same result, and you have achieved maximum elegance, which is
also maximum efficiency.  So I thought that atoms of meaning were composed
of two and only two elements: the "use" concept and the "tag" concept, and
I was almost right, although I had still missed the mark.  What I failed
to realize was that my "universal atoms of meaning" were just WORDS.  My
problem was that, just like many who had pondered the mysteries of
language before me, I did not fully understand the meaning of "word."

So I had "universal atoms of meaning," and I found that these atoms
assembled themselves neatly into L-shaped Tinkertoy structures, but I was
still mistaken about the way these atoms bonded to one another.  I
believed that the direction of linkage was downwards and rightwards
instead of (as I later determined) upwards and leftwards.  Each atom had
three valence points (potential connecting points), etc.

For lack of a better name, and because it was an interlingua using English
names for its concepts, I called the knowledge representation system
"Interlinguish," and Interlinguish it has remained.

I was still wrong about many things, but I had the basic idea, and the
rest came slowly over many cups of coffee during many years.  The sequence
of the above developments in my thinking may be wrong--I can no longer
remember exactly--but the precise sequence is not as important as the
total process.  I made many mistakes, jumped to many false conclusions,
but always kept on going, kept an open mind, and remained ready to abandon
any idea that proved wrong.  And the good part was that during all of this
time I was working with computers, and unlike humans, who are emotionally
involved in everything and apt to defend the silliest of notions to the
bitter end, computers cannot lie.  So it was only _me_ that needed
checking at all times to make sure that I never became emotionally
involved in this effort to the point of clinging to false hopes or false
ideas, and thus it was also a process of introspection and the search for
real truth.

At one point, I came to believe that unlike what I had been taught in
college about computer programming projects, in this case there may not be
"more than one way to skin a cat."  Maybe language could only work in just
one and only one way.  Maybe instead of just trying this and trying that,
I really needed to understand the science of linguistics and human
intelligence.  Maybe unless I really got the theory right, I was just
beating my head against a stone wall.  Not good--what I knew of modern
linguistics didn't look a lot like science to me.  I needed scientific
rigor.  I needed to start building a new theory of language and human
intelligence that would guide my efforts with computers.  So I made a
conscious effort to stop trying to "hack" language and to start building
rigorous theory that could not fail, and this approach payed off
immensely.

and so this attempt to create a knowledge representation system for
artificial intelligence was transformed into a theoretical quest for
linguistic truth.  Again, I cannot remember the exact sequence of these
developments, but here is what I learned:

1.  The basic building blocks of language are not words, but something
called "linguistic links" from which words and other linguistic data
structures are created.  Each such link has a source, a type, and a
destination.

2.  Unless a word is meant to be ambiguous (have more than one meaning),
every word that is part of a coherent phrase in any language is simply two
linguistic links emanating from a common source.  These two links are
called a "semantic link" or "semlink," and a "syntactic link," or
"synlink," respectively.  The destination of the semlink is a "semantic
node," or "semnod," which exists in the "ontology," which is a collection
of meanings and the relationships between meanings.  The destination of
the synlink is another word in the same phrase or sentence except for the
synlink of the top word of the phrase or sentence, which has a type but no
destination.  Semlink type is semantic role (part of speech).  Synlink
type is syntactic role (case).

At first I called the links between semantic nodes (semnods, or word-sense
definitions, or meanings) semlinks.  This was a mistake, because that term
is needed for the links between words and their meanings, so I later
called the links between semnods "radlinks."

At first I thought that parsing was "lifting meaning from texts," but I
later realized that parsing is really what we call understanding, and that
a phrase or sentence has been understood when the right synlink and
semlink have been selected for each word.  So I realized that it would be
possible for machines to understand texts if only they could be made to
select the right semlink and synlink for every word in a sentence or
phrase.  The question was then, "But how can you do it?"  Like many other
researchers, I had been convinced that language is "rule based," so that
if we just knew the right rules and all their exceptions, then it would be
possible to build machines that could parse and generate sentences
correctly.  Again I was dead wrong.  Language is not rule based but
precedent based.  Once an individual has recognized that words put into a
certain order mean a certain thing, he/she can subsequently understand the
same sentence again and a lot of other similar sentences.  This is because
he/she has added the disambiguated form of the sentence to his/her corpus
of parsed sentences.  It will be noticed that once the correct synlink and
semlink have been selected for every word in a sentence, the word symbols
themselves, whether they be spoken or written, are unnecessary and
redundant.  Because of this, even though I do not know Vietnamese, all I
have to do is to examine the linkages between the parsed words of a
Vietnamese sentence to understand it, so long as I have the word-sense
definitions for the semnods the words link to in the Vietnamese ontology.
Each word in an ordinary sentence has just one and only one synlink and
just one and only one semlink, and it is by selecting the right pair of
links for every word that understanding occurs.

And by knowing these things I realized that my "universal atoms of
meaning" were just such disambiguated words in parsed sentences without
their external word symbols (the spoken or written words that exist
outside the automated system).  The old L-shaped Tinkertoy structures of
Interlinguish remained the same, and the way words linked to each other
remained the same except for link direction, but now I had it all down to
a theoretical science instead of just another computational model.
Remember?  Elegance is simplicity, and this is the simplest and most
elegant way possible to represent knowledge--not by creating any
artificial knowledge representation scheme, but simply by selecting the
right synlinks and semlinks for words and allowing the external symbols to
fall away.  The result is language inside automated systems in a form
accessible to the logic of automated systems, whether these systems be
computers or human brains.

Thus by searching for a right way to represent knowledge inside computers
for AI, I stumbled upon a whole new theory of language that had never been
known before.  Of course we all know these things internally, but not at a
conscious level, and it is this conscious awareness of the theory and its
workings that I am talking about.  Working night and day I had melted a
mixture of raw metals and dirt in the crucible of this witch's cauldron of
a noisy old tower computer on my floor, and out had flowed the pure gold
of an impeccable theory!  I called this theory "The Theory Of Panlingua,"
and I called the arrays that are left after selecting the right synlink
and semlink for each word and discarding the external symbols "Panlingua
arrays."

But again I was wrong.  The problem was that there was once a Brazilian
named Xul Solar who had already developed a "language" called Panlingua
more than half a century before.  I fear he had no rigorous theory, and
this "work" of his was bedeviled by superstition and the paranormal, but
there it was--he had called it Panlingua, and this put a blight on my work
because of the association.  But never mind, the cat was out of the bag. I
had named my theory "Panlingua," and google knew it, so it was too late to
change.  I have to live with the ignominy of appearing to have stolen this
fellow's word, which in fact I probably did.  Nowadays we read through so
much material on a daily basis that I must have seen the word somewhere
and liked it and used it without a careful check, so Panlingua it shall
remain.

And needless to say, there are still other errors and other errors, after
all I am human.  In the beginning, I did not realize that this would be a
completely linguistic affair.  All I wanted to do was to represent
knowledge on computers in such a way as to be readily accessible and
usable by AI.  With all due respect to Mr.  Levier, my old instructor, I
had no idea that in this case there wouldn't be the same old thousand ways
to skin a cat just like he said.  I didn't realize that it would turn out
that there is really only one right way of representing knowledge on
computers, that language is unique and cannot be made to work except in
just one way, etc.  I kinda thought it might be easy--you know, like the
one on Star Trek.  How wrong I was!

I thought we were going to "lift meaning" from sentences, represent this
meaning internally, and do things with it for AI.  Thus in the beginning I
was free and easy with my use of Interlinguish.  Sometimes I would leave a
word out here or add some other word there with the idea of making the
internal representation clearer.  But as I learned more and moved to a
more rigorous approach vis a vis language, I adopted a new rule: The
Interlinguish representation must never have a different number of words
than the original text.  And of course this tied in with the realization
that my Interlinguish representations were just words linked correctly and
with their external symbols removed.

So at one time I was putting an atom of my own at the head of each
subclause.  This was because, as I thought among other things, a word
cannot be a noun and a verb at the same time.  So, for example, if "She
said the man entered her room," and "she" was the subject and "said" was
the top verb and "the man enter her room" was the object (what she said),
then how could we connect "the man entered her room" to "she said?"  The
connecting point would have to be the top verb of the lower clause, but
how could "entered" be an object when it was a verb?  So I got creative
and created my own "head" atoms for such cases.  Such an artificial atom
would be the regent of "entered" and the object of "said."  I did this for
a long time, or rather made my poor machines all do it for me, until I
learned that a word with the same semantic role (semlink type) can have a
different syntactic role (synlink type) in different situations.  For
example, verbs can assume the roles of subject, object, etc., just like
nouns can.  When I learned this fact, I was able to do the following:

1.  Completely remove all artificial "head" atoms of the type described
above.

2.  Do away completely with the idea of "verb rotation," which you may
have read about in some of my other writings.

3.  Understand how a noun can assume an adjectival role in phrases like
"house cat."

4.  Understand how nouns can assume adverbial or adjectival roles as in
"He went bananas."

Etc.

In what I used to call "verb rotation," auxiliary verbs were simply
assuming adverbial syntactic roles.  Thus in the sentence, "I should go,"
"should" is playing the role of adverb and should be treated accordingly,
and not as a verb, the verb of the sentence being "go."

In the sentence "He was hitting hard," "hard" is still an adjective (its
semlink type is adjective), but its syntactic role is adverbial.

I have had to think and rethink link direction several times.  Every
linguistic link has a source, a type, and a destination, just as I said
above, but it is not always clear which of the linked nodes should be
sources and which destinations.  In some cases, as in synonymy, links seem
to have no direction at all, or maybe run in both directions.  In other
cases we might choose to make links run a certain way in order to
distinguish them from other kinds of links, as in semlinks and lexlinks.
 But as a general rule of thumb, I believe the following is the right way
to choose link direction: If several items link to a single item, then
link direction should run from the several to the one.  This works the
best way for computers, but may or may not be the way that things are done
within the human brain.  I have reminded you again of the human brain,
because I have become convinced that language can only work one way to
work right, and that right way is the way it works inside the human head.

I hope this explanation has helped you to understand Panlingua.  I call it
a theory, but it is actually a very good working hypothesis that is
continually being improved over time.  I suppose that some parts are
really theory, because they have stood up under many hard tests without
ever showing any signs of cracking.  At the same time, it would be
considerably less than honest to say that various parts of this theory
have not been dead wrong during the process of its formation, which has
been going on since 1994.  So with the Apostle Paul, I will say, "Try all
things and hold fast to that which is good."  The only thing that I might
add is this: If you really want to test anything, then test it on
computers, because unlike human beings, computers cannot lie.