1. Kate Zhiming Zeng
Daniel Nelson
WRT 105
02 May 2014
Is Computer Poetry Poetry? Behind the Machine of Machine-Generated Poetry
Machine-generated poetry is essentially a string of text strung together
through a step-by-step process to resemble poetry. Automated poetry immediately
raises the question of whether such text can be deemed poetry at all. Although
automation can generate poems that fulfill the measurable stylistic demands of a
poetic form, such as the iambic pentameter of a sonnet, it seems impossible that it
can generate text that makes sense, let alone text that moves us. However, the
poetry produced with Charles Hartman’s programs challenges the assumption that
there is nothing poetic about machine-generated poetry. The way he uses the
machine underlines that the extent in which a machine-generated poem sounds
poetic depends largely on how a person collaborates with the machine.
“The object of poetry is confessedly to act upon the emotions,” John Stuart
Mill said. Good poetry evokes an emotional response in its readers. Since a machine
cannot think and feel on its own and thus cannot assess the emotional content of its
work, it is understandable that literary critics would call machine-generated poetry
mere imitations of the original. The main decisions a machine makes in creating
poetry are choosing words and arranging the word order based on a set of random
or arbitrary rules. Such random or arbitrary decisions do not guarantee that the
resulting text will have any emotional impact. A machine can only try out various
2. combinations of these decisions in the hopes that one of the resulting stanzas has
poetic potential. Unsurprisingly, such a process does not guarantee a high
proportion of good stanzas, out of all the stanzas produced. Hartman described how
he had to comb through “piles of computer paper, searching in vain for oracular
truths” even after he has done various fine-tuning to the program to increase the
chances of getting good poetry. In comparison, when a person writes poetry, he or
she has access to his or her personal experiences and is able to ponder on their
nuances in such a way that a set of instructions and data fails to encapsulate. Since
we have yet to program a machine that can perfectly imitate the way we process
information (and it’s arguable if that is even possible) and thus cannot produce a
reasonable proportion of poetry that actually moves us, it seems conclusive to argue
that any machine-generated text is not poetry.
While it would be exceedingly difficult to program the complexity of our
emotions and thoughts, this does not imply that machine-generated text has no
poetic quality. Even though a machine does not have a mind like ours but a limited
set of input and instructions (e.g. word banks and syntax structures), based on those
input alone, it can create thousands of word combinations that we normally do not
use. The meaning of these words is ambiguous - we cannot at first glance discern a
single, clear meaning. Through this form of ambiguity, machine-generated poetry
can act on our emotions and engage us. In Seven Types of Ambiguity, William
Empson described ambiguity as “any verbal nuance, however slight, which gives
room for alternative reactions to the same piece of language.” Ambiguity is seen as
an indispensable quality of good poetry because it leaves a poem open to a rich
3. variety of interpretations. Ambiguity gives poetry what Jane Hirshfield calls, the
“mysterious surplus”. Through the juxtaposition of images and words that are
seemingly unrelated by the standards of everyday language, a poem creates an area
of grayness that is hard to grasp, at least at first glance. The reader is invited to
ruminate on the possible meanings that can be constructed based on not just the
word itself, but also the connection between each word. In other words, the reader
fills in the gaps between disjointed words. This gap, or disjunction, is crucial in good
poetry because it engages the reader to supplement the gaps with their own
experiences, thus transforming the words into a poem that resonates personally. It
keeps the poem fresh as each reader brings a different interpretation to the same
poem.
So how effective are programmable processes at creating ambiguity? When
the program uses a randomness function, it becomes especially effective at finding
uncommon combination of words and images because it makes decisions
unpredictably, uninfluenced by human preferences, will or circumstances. In the
case of Hartman’s program, the program arranges parts of speech (e.g. adjective,
noun) randomly to create a syntax template, before proceeding to fill up the
template with randomly selected words that fit the syntax demanded. Without any
references to how people actually combine words beyond grammatical rules, the
resulting text is largely an incongruous arrangement of words - syntactically correct
yet rarely used in everyday life. If we measure a good poem by its degree of
disjunction alone, we find that an automated process with a randomness function
generates good poetry.
4. The question then is, when does the gap become so large that it loses the
audience? A poem that is too straightforward becomes clichéd stale poetry. A poem
that is too hard to comprehend cannot resonate with readers. The trade-off between
creating enough disjunction and leaving behind enough clues is a balancing act that
poets perform. “In literature,” Robert Frost wrote in a letter to John Bartlett, “it is
our business to give people the thing that will make them say, “Oh yes I know what
you mean.” It is never to tell them something they don’t know, but something they
know and hadn’t thought of saying.” In other words, a good poem involves arranging
words in such a way that is, at once, familiar and unfamiliar to the reader. Coming
up with unusual word combination is not enough to create good poetry. Good
poetry also needs to echo what has been said to resonate with readers. While it is
very simple to design programs to find unfamiliar combinations of words, it is an
extremely hard task to design programs that can find and store all the common
ways people use language, especially when people find new ways of using language.
Since it is difficult, if not impossible, to reduce the balancing act to a set of machine-
executable procedures, getting a machine to generate text that resonates with
people would be like finding a needle in the haystack of infinity.
However, just like how the poet does not require all the knowledge in the
world to create good poetry, a machine does not need to model all the ways people
use words in order to generate good poetry. Ferdinand de Saussure, who viewed
language as a closed or self-contained system, would probably defend the notion
that a machine can generate text that makes sense to people as long as it is
programmed to obey syntax structures. If language is indeed a self-contained
5. system, any text (including machine-generated ones) with an identifiable syntax
structure is enough to be comprehensible in terms of the way we understand
poetry.
A case in point would be Lewis Carroll’s famous poem, “Jabberwocky.”
’Twas brillig, and the slithy toves
Did gyre and gimble in the wabe:
All mimsy were the borogoves,
And the mome raths outgrabe.
Even though we have no prior knowledge of what most of these words mean,
it is not hard to guess what they mean because the underlying grammatical
structure suggests the relationship between those words. “Brillig” is an adjective
that seems to describe the weather. We can tell “toves” is a noun that probably
refers to a pack of animals because of the active verbs “gyre” and “gimble”. Other
kinds of poetry also work similarly by obscuring the reference of the word. In terms
of balancing between the familiar and unfamiliar, it seems that good grammar
already provides enough familiarity to any bizarre arrangement of words. Going
down this road, it doesn’t seem far-fetched to claim that a machine with the ability
to generate syntactically sound sentences can produce poetry.
Yet, the claim that “Jabberwocky” engages us merely because of its syntax is
simplistic. Even though we do not use those words in daily life, they bear semblance
to words we do use in such a way that we can build a context around them. “Brillig”
sounds like “brilliant” and “slithy” sounds like “slithering”. It seems that these words
are deliberately chosen to sound like it might refer to an actual object or concept.
Moreover, the way the words are grouped together also inform their meanings, such
6. as how “toves”, “gyre” and “gimble” create a notion of animals moving. Having
correct syntax alone is not sufficient for readers to find familiarity with the text. The
text also needs to suggest a context. Again, it would be extremely complicated to
reduce all the ways in which we pick and arrange words to express a context into a
set of programmable functions. It seems that machine-generated text could never
become comprehensible enough to resonate with readers.
Hartman ended up modifying not just the program but also its output, many
times, in fact. Hartman admitted that most of the machine’s original output did not
contain the “flashes […] of ordinary or extraordinary lucidity” he had hoped they
contained. Granted, he did find “endless tempting sentences, perhaps one in five or
ten” but most contained so much disjunction that they did not resonate with him. In
the excerpt below, Hartman commented on why he manually modified the output to
make it sound more poetic.
"The court of color (radiation of the center) is stress above any building."
Nonsense, yes, but with the subliminal promise of an image: open air, surrounding
white buildings, uncanny color. Take out "stress," which is abstract in this context.
Notice that "color" makes "radiation" unnecessary (though the connection between
them may have first called my attention to "color"). And "court" (as in "courtyard")
might contain the implications of both "center" and "building" and made those
words unnecessary. So "The court of color is . . ." what? Air, really, or all the air
considered as a whole: "atmosphere." "Atmosphere" might also be the courtroom of
colors, judiciously discriminating near from far (as in aerial perspective), bright
from dim.
But "atmosphere" could never have been produced by the program, not being
in its dictionary. So, came the subversive voice, add it to the dictionary.
The decision to change the output does indicate the limited ability of
programmable processes to generate full-fledged poetry. Hartman could have added
improvements to his program but instead he chose to modify the output directly.
8. Merril used an Ouija board to dictate verses to him: the end result is an award-
winning poem “The Changing Light at Sandover”.
Should we disdain the use of automation in the creation of a poem just
because of its arbitrariness? We are eager to forget how randomness plays a part in
the creative process. When we look at say, the sonnet, it seems arbitrary that each
line should contain five iambs. How does this number help to evoke an emotional
response? Much like how poets have used various poetic forms to write poetry,
Hartman uses the randomness function and the syntax template to find new ways of
combining images. A machine simply applies these templates to a great amount of
input at a speed faster than what a human can do. Could Shakespeare have written
his poems without the sonnet? Could Hartman have imagined his edited poems
without the machine? Even though there is no definitive explanation on how these
templates act on our emotions, they have curiously been found in poems that
resonate with readers. So long as the poet uses the templates, or programmable
processes, in such a way that the final form of the poem hits the sweet spot between
the familiar and the unfamiliar, computer poetry text can indeed be good poetry.
Bibliography
1. Carroll, Lewis. Alice's Adventures in Wonderland. New York: MacMillan, 1865.
Print.
2. De Saussure, Ferdinand. Course in General Linguistics. Illinois: Open Court
Publishing, 1983. Print.
9. 3. Devitt, Michael, and Kim Sterelny. Language and Reality: An Introduction to
the Philosophy of Language. 2nd ed. Cambridge: A Bradford Book, 1999. Print.
4. Empson, William. Seven Types of Ambiguity. New York: Meridian Books, 1960.
5. Frost, Robert. The Letters of Robert Frost. Volume 1. Ed. Sheehy, Donald
Gerard; Richardson, Mark; Faggen, Robert. Cambridge: The Belknap Press of
Harvard University Press, 2014. Print.
6. Hartman, Charles O. Virtual Muse: Experiments in Computer Poetry.
Connecticut: Wesleyan University Press, 1996. Print.
7. Hirshfield, Jane. “Poetry, Transformation, and the Column of Tears.” The
American Poetry Review. 42.6 (Nov/Dec 2013): 37. Web. 17 Apr 2014.
8. McHale, Brian. “Poetry as Prosthesis.” Poetics Today 21.1 (Spring 2000): 1-32.
Web. 17 Apr 2014.
9. Mill, John Stuart. “What is Poetry?” Essays on Poetry. South Carolina:
University of South Carolina Press, 1976. Print.
10. Nemerov, Howard. "Poetry (literature)." Encyclopedia Britannica Online.
Encyclopedia Britannica, n.d. Web. 17 Apr 2014.