28 AI and Poetry
The main aim of a poem is to be pleasing rather than conveying a meaningful message.
—Pablo Gervás30
In 2010, literary critic Marjorie Perloff wrote, “Nothing quite prepared the poetry world for the claim, now being made by conceptual poets … that it is possible to write ‘poetry’ that is entirely ‘unoriginal’ and nevertheless qualifies as poetry.”31 One of those conceptual poets was Kenneth Goldsmith, an early proponent of electronically generated text. As Perloff writes, he pushed the concept of the unoriginal yet further when he “announced his advocacy of conceptual or ‘uncreative’ writing—a form of copying, recycling, or appropriation that ‘obstinately makes no claim on originality.’”32 His 2003 piece of conceptual poetry, “Work Day,” consists of an entire issue of the New York Times, from the headlines to the adverts, transformed into digital text.
In conceptual poetry, as in conceptual art, the concept—the idea behind the piece—is of the essence. This could mean that it is not necessary to read a text in the traditional sense to get the idea, akin to reading a text as it scrolls down a computer screen, without being able to go back if you missed anything.
Pablo Gervás and His Poetic Algorithms
There is something algorithmic in poetry, a metric, a syllable and rhythmic counting.
—Pablo Gervás33
“I can get a computer to write a poem by giving it models of how poetry is written and of how poetry works, but I can’t give a computer a model of how love works; that’s way beyond what machines can do at this stage,” says Pablo Gervás.34
One of Gervás’s principal interests is in developing an algorithm to generate Spanish verse. He calls it Wishful Automatic Spanish Poet (WASP).35 As he points out, poetry provides a way of circumventing some of the basic problems involved in using computers to tackle literary genres, starting with their lack of advanced linguistic skills and basic common sense. Poetry does not require high precision. The “main aim of a poem is to be pleasing rather than conveying a meaningful message,” he says, and it is often wildly open to interpretation.36 Gervás chose Spanish poetry because its phonetics are straightforward and it is governed by very specific rules, allowing him to maintain correct syntax while playing fast and loose with semantics.
He used classic Spanish poems as the initial data for WASP, then input words tagged according to the number of syllables, the position of certain syllables, and the rhyme. He then asked poetry lovers to assess the thousands of poems the algorithm produced, taking into account meaning and aesthetics. The judges rated them on a scale from 1 (nonsense, ugly) to 5 (a real poem, very pretty). None achieved a 5 for making total sense nor for aesthetics. Scores for syntax were close to 3, but scores for aesthetics were much lower.
WASP has gone through several manifestations in its logic and methods. Gervás has used evolutionary programming, selecting the best output and allowing it to evolve. “We need a way to separate good stuff made by machines even if it differs from what humans produce,” he says.
Besides computer-generated poetry, Gervás has also created the PropperWryter algorithm, which generates plots, among them the plot for the musical Beyond the Fence.37 He is particularly interested in what the process and products of computer-generated literature mean for human beings and computers, as well as for creativity.
He believes that just as young poets read the poetry of others and emulate it until they find their own style, something similar will happen with computers. First they will emulate and then broaden out to, perhaps, “modeling stuff like personal experience.”38 Personal experience? The problem is of course that, for the moment at least, computers have no way of comprehending, let alone feeling, emotions like love.
Gervás doesn’t believe poetry created by machines should be judged by the same criteria as regular poetry or that it should be judged according to whether it can be distinguished from poetry written by human poets. “If you only consider machine poetry to be good if it’s like human poetry, then what’s the point?” he asks. What we are looking for is for the computer to surprise us by coming up with something new and different. “It might be much better than what people do,” Gervás points out.
An essential element in creative writing is that “the magic doesn’t come from the writer, it comes from the reader.”39 It seems truly magical when “a string of characters on a page sparks memory and emotion,” like a work of visual art or of music can. Good writers take a minimalist approach, leaving it to the reader to figure things out rather than explaining everything. “That’s what we want to try to get computers to do,” Gervás explains.
Gervás is intrigued by artificial neural networks but for the moment doesn’t use them. There isn’t enough data in Spanish poetry, plus it is still unclear how neural networks reason, what goes on in the hidden layers. Gervás is uncomfortable with hidden layers because, he asserts, “a large portion of my motivation is to understand how people write poetry.”40 In other words, he is interested more in the process than the end product—and in neural nets this process is hidden away deep in the hidden layers.
He complains that poetry generation is given low priority. “People don’t like to invest in research that won’t make any money or improve the quality of life,” he says. Research funds tend to be limited to more urgent problems like curing disease. But creativity and how to model it are also problems worth working on—in particular, literary creativity: poetry and narrative. With all the data available, machines can easily churn out weather reports and newspaper articles, but literary creativity is in a different league. As Gervás puts it, “Literary creativity is to communication what Formula 1 driving is to transport.”41 The only way to make cars safe is to test them by driving them at high speed, even if this is something most people will never do. Similarly, rather than contenting ourselves with weather reports, we need to push machines to the next level, make them work harder, learn to write true literature. “We must bridge the complexity gap between how people write literary text and how machines produce news stories and weather reports,” says Gervás.
As to creativity, “the word itself is ambiguous. … But for me the phenomenon is clear and has nothing to do with labels. Do something surprising that has not been done yet!”42