Copyright 1989 Information Access Company;
                   Copyright Scientific American Inc. 1989
                              Scientific American

                                   June,  1989

SECTION: Vol. 260 ; No. 6 ; Pg. 122; ISSN: 0036-8733

LENGTH: 3420 words

HEADLINE: A potpourri of programmed prose and prosody; Computer Recreations;
computer-generated commentary

BYLINE:  Dewdney,  A.K.


    As the renowned Oxonian master of nonsense observed, semantics takes
precedence over syntax in creative writing.  Good literature is shaped by the
meaning contained in a writer's ideas.  Computers are not yet capable of ideas
and so cannot take care of a composition's sense.  Yet, as a number of
contemporary programs show, computers can certainly take care of the sounds.
But is it art?  That is for the reader to decide.

    Consider the writing of MARK V. SHANEY, a computer program created by Bruce
Ellis (now at the University of Sydney) that is based on an idea of Don P.
Mitchell of the AT&T Bell Laboratories.  MARK V. SHANEY is not exactly a
self-starter.  The program must first read and reflect on someone else's work.
It then produces a rambling and somewhat confused commentary on the work.  As an
example, I quote MARK V. SHANEY's reflections on reading an elementary chemistry

    "Wash the blackboard.  Watch it dry.  The water goes into the air.  When
water goes into the air it evaporates.  Tie a damp cloth to one end of a solid
or liquid.  Look around.  What are the solid things?  What are the only ones
that take part in the clouds themselves.  As clouds move air tosses the tiny
droplets of water. Sometimes the water which leaves the body in the leaves of
green plants.  It is quite porous, almost like a sponge.  But the walls of the
hollow tubes and cells are very strong.  Chemical changes take place when
something burns...."

    The commentary is hardly different when MARK V. SHANEY digests a book on
elementary mathematics.

    "Why do we count things in groups of five.  When people learned how to coun
tmany things, they matched them against their fingers. First they counted out
enough things to match the fingers of both hands.  Then they put these things
aside in one quart.  A giant-size bottle that will hold four quarts is a
three-digit number...."

    From here on, the program's output becomes increasingly confused, as readers
may see for themselves in the illustration on this page.

    Although sense is conspicuously absent from MARK V. SHANEY's writings, the
sounds are certainly there.  The overall impression is not unlike what remains
in the brain of an inattentive student after a late-night study session.
Indeed, after reading the output of MARK V. SHANEY, I find ordinary writing
almost equally strange and incomprehensible!

    How does MARK V. SHANEY produce its remarkable works?  The answer is rather
simple.  The program's name, a weak pun on "Markov chain," provides a clue.  In
abstract terms, a Markov chain is a sequence of symbols generated according to a
table of probabilities.  In the version relevant to MARK V. SHANEY's operation,
each row of the table corresponds to a pair of symbols.  The entries in each row
consist of individual symbols, each paired with an associated probability.  A
sequence of symbols is generated by an algorithm that begins with a "chain" of
two symbols and thereafter cycles through four simple steps.

    1. Find the last two symbols in the current chain.

    2. Go to the row of the table corresponding to the symbol pair.

    3. Select a symbol from the row according to its probability.

    4. Add the selected symbol to the end of the chain.

    For example, the first few entries of a Markov-chain table for the symbols
A, B, C and D might look like this: AB: B(.1) C(.1) D(.8) AC: A(.1) B(.2) C(.1)
D(.6) AD: B(.4) D(.6) BA: B(.3) C(.4) D(.3) BB: A(.5) C(.5)

    Given the symbol pair AB as the initial chain, the algorithm would have a 10
percent chance of selecting B, a 10 percent chance of selecting C and an 80
percent chance of selecting D as the next symbol.  How does the algorithm choose
a symbol according to the probabilities?  It divides the interval between 0 and
1 into numerical segments whose lengths correspond to the symbol probabilities.
It then chooses a random number between 0 and 1 and determines in which segment
the number has fallen.

    For row AB in the above table, the segments corresponding to the respective
 probabilities for B, C and D range between 0 and .1, between .1 and .2, and
between .2 and 1.  Suppose, then, that the computer's random-number generator
yields .0172.  Because that number lies in the first segment, B would be
selected as the next symbol in the chain.  The chain would then consist of the
symbols ABB.  The algorithm would next consult row BB in order to select a
fourth symbol for the chain.  Here again, a random number is generated.  If it
is less than or equal to .5, A is selected; otherwise the algorithm selects C.
Because of its dependence on chance, if the algorithm were restarted with the
same initial symbol pair, it might well produce an entirely different chain.

    Such an algorithm was actually applied in the 1940's by Claude E. Shannon of
Bell Laboratories to analyze the information content of human language.  He
constructed the algorithm's probability tables by scanning ordinary text and
counting how many times each individual character followed each pair of
characters (including blanks).  Once the character frequencies for a given text
were known, they could easily be changed into probabilities.  The Markov chains
of characters generated in this manner had statistical properties that resembled
the source text, although they rarely contained valid words.  How then does MARK
V. SHANEY apply Markov chains to produce understandable English words?

    The trick is to apply Shannon's algorithm for Markov chains but with entire
 words instead of characters as the concatenated symbols.  As MARK V. SHANEY
scans a text, it builds a frequency table for all words that follow all the word
pairs in the text.  The program then proceeds to babble probabilistically on the
basis of the word frequencies.

    A key feature of the program is that it regards any punctuation adjacent to
a word as part of the word.  That feature enables it to form sentences having a
beginning and an end.  Approximately half of them are even grammatical.  I
shudder to think what the program might produce after scanning this article!

    Indeed, others have already shuddered at MARK V. SHANEY's reflections, some
with rage and others with laughter.  Some years ago Ellis decided to go on-line
with his creation.  The victims of the program's analyses were the innocent
computer users who subscribed to an electronic bulletin board called  The bulletin board provides a place for male and female
scientists, engineers, programmers and graduate students from all over the
country to post their thoughts on subjects as diverse as dating, makeup and
personal relationships.  Why not have MARK V. SHANEY read the postings and
respond with its own "thoughts" on those subjects? Here are some of MARK V.
SHANEY's more thoughtful comments.

    "When I meet someone on a professional basis, I want them to shave their
arms.  While at a conference a few weeks back, I spent an interesting evening
with a grain of salt.  I wouldn't dare take them seriously!  This brings me back
to the brash people who dare others to do so or not.  I love a good flame
argument, probably more than anyone....

    "I am going to introduce a new topic: does anyone have any suggestions?
Anybody else have any comments experience on or about mixed race couples,
married or otherwise, discrimination forwards or reverse, and eye shadow?  This
probably the origin of makeup, though it is worth reading, let alone judge
another person for reading it or not?  Ye gods!"

    The opinions of the new correspondent drew mixed reviews.
Serious users of the bulletin board's services sensed satire.  Outraged, they
urged that someone "pull the plug" on MARK V.  SHANEY's monstrous rantings.
Others inquired almost admiringly whether the program was a secret artificial
intelligence project that was being tested in a human conversational
environment.  A few may even have thought that MARK V. SHANEY was a real person,
a tortured schizophrenic desperately seeking a like-minded companion.

    If the purpose of computer prose is to fool people into thinking that it was
written by a sane person, MARK V. SHANEY probably falls short.  MELL, the
brainchild of Bonnie V. Firner of Piscataway, N.J., probably comes a little
closer to that goal.  MELL writes weird science-fantasy stories with a peculiar
meditative quality.

    "The warrior scowls in the drought pulses.  He loves himself in the drought.
He molds himself in the drought.  He glowers at the warrior Dugaki in the
drought pulses.  He calls her in the drought. He snarls at her in the drought.
He sits beside her beside awareness.  He seizes her.

    "Oh I am mighty says Oban.  He smothers her in the drought.  He smashes her
of the turquoise....

    "Oban kills Dugaki.  He has it of turquoise.  He glares at it in the drought
pulses.  He calls it in the drought.  He snarls at it in the drought.  He sits
with it beside greed.  He seizes it.  He smothers it in the drought.  He smashes
it of the turquoise.  He needs it beneath the thunderbolt.  His body burns
beneath the thunderbolt...."

    MELL consists of some 1,500 lines of BASIC code; in contrast to MARK V.
SHANEY, Firner's program is complicated even at the conceptual level.  The
program's main loop generates two characters in terms of 16 randomly generated
"descriptors" whose values define qualities such as size, niceness,
occupation, age and health, smell, commitment (from indifferent to fanatical)
and even magical power.

    Having chosen the names and traits of the story's characters, the main loop
then determines what motivates a character by examining the values of its
various descriptors.  If one of the descriptors has a low value, MELL bases the
character's interaction with the other characters on that fact.  If no character
has a low descriptor value, MELL will decide on the nature of a character's
interaction based on his or her occupation.  For each cycle of the main loop,
the program then generates several sentences that describe the characters by
fitting names for their qualities into predetermined grammatical slots.  The
sentences thus generated also include what Firner calls "background" words.

    In the above story, for example, the background word "drought" sets the
scene.  Consequently that word creeps into many of the sentences.  After
generating a paragraph that describes an act by one of the characters in this
way, the sentence-generating process starts all over again.  Between iterations
the program can change a quality of one of the characters.  This helps to keep
the story (as much as there is of one, anyway) from stagnating.

    Is poetry any easier for a computer to generate than prose?  One doubts it,
because the meaning contained in prosody tends to be far more dense than in
prose.  Of the three versifying programs I shall discuss, only POETRY GENERATOR
,which was written by Rosemary West of Mission Hills, Calif., is fully automated
;the other two require human intervention to finish the product.  West describes
 her program as follows:

    "My approach ... was to supply a large vocabulary of words and phrases that
would be selected at random and combined according to a set of grammatical
rules.  For example, consider the following poem: 'The tree dips its bare
fingers/into the black ice-pond/just as three gray geese/slide down a nearby

    "Each line of the poem can be broken down into several parts.... 'The tree'
is a noun phrase; 'dips' is a verb; 'its bare fingers' is the object of that
verb.  Having categorized the parts, I can then come up with between 100 and 400
possible substitutions for each part, one of which is randomly selected by the
computer.  For example, using the same verse structure, I might get: 'A woman
hides five gray kittens/under the old jalopy/at the moment when the sad
clowns/enter your museum of pain.'"

    The poetic structures on which POETRY GENERATOR hangs its words may vary
considerably, lending variety to the syntax and to what seems to be the meaning
of the poem.  West has based some of POETRY GENERATOR's output on the
structures of her own poems, several of which have been published.

    Thomas A. Easton of Belfast, Me., thinks the best way to generate computer
poetry necessarily involves a symbiosis between human and machine.  He has
written a semiautomatic program called THUNDER THOUGHT that provides ideas for
poems.  Again, I quote the author of the program:

    "I conceive of the creative mind as having two components: the popcorn mind
and the critical mind.  The former generates random combinations of whatever
words, ideas and images happen to be in a sort of mental focus (along with
peripheral material, which is why the popcorn mind can surprise us).  The
critical mind then discards as garbage the vast bulk of what the popcorn mind
produces and edits, twists and elaborates the remainder to form poems."

    Relying on internal lists of nouns, verbs, adjectives and adverbs, THUNDER
THOUGHT operates roughly like West's POETRY GENERATOR, arranging words into
sentence frames to produce what Easton regards as raw poetic material for a
human mind to refine.

    "The intermediate result is ungrammatical, nonsensical, ridiculous garbage
... but not always.  Among the many lines of garbage there always lie a few
lines to which one responds.  They make sense--or seem to.  They beg one to
tweak them a little.  A pair of them insists that one make up a third line.
They stimulate one to think of other lines that can accompany them.  A little
editing, interpolation, elaboration and--voila!--a poem."

    Easton has written some 110 poems by this method and has published 32 of
them--some have even appeared in literary journals.  That ratio, he claims, is
one that would turn many real poets green with envy.

    The last word on computer poetry goes to ORPHEUS, a program designed by poet
Michael Newman of New York City.  ORPHEUS lays out strict frameworks, from haiku
to sonnets, into which human writers may insert their own chosen words.
Essentially ORPHEUS is a word processor ("poetry processor" in Newman's
parlance) that lays out the lines of a given poetic form.  The program allows a
human being to fill in the lines according to whim and then to end them with the
help of a rhyming dictionary.  Setting ORPHEUS in sonnet mode, for example, one
might write a pair of lines (in imitation of Shakespeare's 130th sonnet) as

    My Apple's screen is nothing like the sun;

    The Cray is faster far on problems big:

    Because the first line ends with the word "sun," ORPHEUS consults its
rhyming dictionary and displays a number of words that rhyme with sun: bun,
done, fun, gun and so forth.  Scanning them, one's eye might alight on the word
"fun." Is there a computer that is more fun than the Apple?

    An association with games brings the Atari computer to mind for the next

    If Apple pleasant be, th'Atari is more fun;

    The first quatrain is polished off by selecting a word that rhymes with big.

    If wires be hairs, her circuits are a wig:

    The rest of the sonnet can be read in the illustration at the bottom of the
opposite page.

    The poetry programs I have mentioned may be bought by readers who want to
sharpen their prosodic skills.  Newman will be happy to supply ordering
information for ORPHEUS NRD (Newman's Electronic Rhyming Dictionary) and related
products to those who write him at 12 West 68th Street, #2C, New York, N.Y.
10023.  Easton meets West in a software package containing THUNDER THOUGHT and
a program similar to it called versifier, which was written by West.  Readers
may inquire about the package by writing to Easton at Box 805, R.F.D.  2,
Belfast, Me.  04915.

    Readers may also be interested in what is perhaps MARK V. SHANEY'S magnum
opus: a 20-page commentary on the deconstructionist philosophy of Jean
Baudrillard.  That effort was directed by Rob Pike of the AT&T Bell
Laboratories, with assistance from Henry S. Baird and Catherine Richards.  The
commentary can be obtained by writing Pike at the AT&T Bell Laboratories, 600
Mountain Avenue, Murray Hill, N.J.  07974.

    Students in Italy, stockbrokers in Singapore and physicians in the U.S. have
joined the growing crowd of Mandelbrot-set devotees--all thanks to a ride on the
Mandelbus, which I described in the February column.  The effort to make the
set's basic iteration algorithm understandable paid off in ridership and perhaps
even readership. Yet, as a number of letters show, some confusion remains.  It
is only the Mandelbus' first stop that is being tested for membership in the
Mandelbrot set.  Subsequent stops may be either inside or outside the set, but
if any of them turns out to be more than two units away from the origin, the
first stop can be automatically excluded from the Mandelbrot set.

    Readers who attempted to visit the area of the Mandelbrot set that Andrew
LaMance calls Love Canal were disappointed to find empty space: they were given
a wrong coordinate address.  A minus sign placed in front of the first
coordinate given should put readers squarely on the site of the curiously
alluring pollution.

    I may have been overly impressed by the magnification of 54,000 that Ken
Philip of Fairbanks, Alaska, achieved in generating an image of a sea-horse
scepter.  Such magnification hardly marks the limit of a computer's acuity.
Indeed, magnifications of that order are nearly routine both for A. G. Davis
Philip of Schenectady, N.Y., and his brother Ken.  The Schenectady Philip
writes, "While I was in Fairbanks in November, my brother's Mac II produced a
picture of a Mandelbrot midget at a magnification of 2 x 10.sup.31.  I consider
that remarkable."

    Peter Garrison of Los Angeles coincidentally explains that the
double-precision mode available in most personal computers--in which the number
of bits in the computer's "words" are doubled--actually more than doubles the
computer's resolving power.  In fact, the power is doubled for each one-bit
increase in effective word size. Extremely high magnifications can therefore be
achieved by means of the even greater precision made possible by special
hardware and software.

    Near the end of the column I meantioned a fast new algorithm for computing
the Mandelbrot set that was described by Uval Fisher in the book The Science of
Fractal Images.  William S. Cleveland of the AT&T Bell Laboratories wrote to
explain that the algorithm was actually developed by William P. Thurston and
Allan R. Wilks. According to Cleveland, the new algorithm not only is faster but
also produces more accurate pictures of the set than the standard algorithm
does.  As Cleveland says: "If you get on board the bus with the Thurston-Wilks
algorithm painting the scene (in black and white), a wholly new and more
realistic world will open up to you."

    A new way to render beautiful Mandelbrot images using the traditional method
was communicated to me by Carl G. Nugent of Seattle.  It is now possible to see
the set in delicate bas-relief, looking like the compressed fossil of an alien
life-form.  Although shaded in a single color, the images are just as beatiful
as the full-color treatments because of their incredibly tactile nature: one can
"feel" those delicate tendrils.  The appearance is based on a trick that makes
the tiny, intricate details of the set look from afar as though they were
illuminated from one side.

    The underlying idea is to divide the display screen into diagonal rows of
pixels that run from the top left to the bottom right of the screen.  If the
iteration count generated for each pixel is taken to represent the pixel's
"height," then an imaginary light source in the top left corner of the screen
will cause a black "shadow" to be cast on certain pixels, depending on the
height of the neighboring pixel (above and to the left) in its diagonal row.  In
even diagonal rows a pixel is not to be colored black unless its neighbor's
height is strictly greater than that of the pixel; in odd diagonal rows,
however, a pixel is colored black even if its neighbor has an equal height.
Hence, in Mandelbrot "plateaus" (areas where all pixels have an equal iteration
count), the diagonal pixel rows will alternate in color.  Up close the displayed
plateaus have a checkerboard appearance.  "To make it look gray," says Nugent,
"just throw away your microscope!"