Arts & Culture Language
What Kind of Sonnets Will Computers Write?
If poetry is an act of discovery for a writer, then even a computational poem has to uncover something new.
Computers can outperform people. I have never beaten a calculator at long division or memorized the entire catalog of a library. But soon after I started a doctorate in computer science, I began to realize just how much computers were inching their way into new and wild domains. When I heard of a presentation on a computer program that would purportedly help you write a short story, I had to attend.
A mix of grad students and faculty had crowded into the department’s small conference room to hear the talk. The professor described a new line of research on collaborative storytelling. In his project, a person and a computer took turns writing sentences for a short story inspired by a New Yorker cartoon, sans caption. The difficulty seemed to be getting the computer to write something that made sense given what the person had written. Computers are notoriously bad at following an extended line of reasoning. They are like shoddy improv partners, ignoring whatever the other person is trying to set up. So an audience member asked if poetry—where unexpected juxtapositions from a computer might be less frustrating—could be a better domain for this work.
“Poetry is too easy,” the professor told us. “It doesn’t make any sense anyway.”
As a poet, this disturbed me. Feeling young but indignant, I tried to defend poetry as an endeavor of meaning-making. Poetry is where we turn when prose is not enough, when we can’t communicate in any other way. It’s not nonsensical or random; if anything, it heightens our attention to what words can mean. The professor was unconvinced.
“Whatever we see in poetry is constructed,” he argued. “I could give you nonsense and you might find it meaningful. That doesn’t make it interesting, computationally.”
Such was my first interaction with poetry in my new world of computer science. Welcome to graduate school.
*
Soon afterward, a well-known research group published a paper about a computer program that could write a sonnet on any topic. Again, I was thrilled that this kind of work would be accepted by my academic community. I read the paper deeply. I brought it to a reading group so I could discuss it with older PhD students, and they helped me properly understand the work.
First, a person inputs a topic for the poem, and the program finds a set of related words that can serve as potential line endings. Trained on 90,000 English song lyrics, the program then generates a sonnet where each line ends on one of those topical words. As a computer scientist, I appreciated the challenge that the sonnets represented, as well as their achievement. The method was fascinating: generating the text backwards, starting at the last word of the last line, helped the program achieve the rhyming constraints. But as a writer, I knew that the results read like a poem in which you had cut out all the lines that scared you.
A key component of a poem is the location of the volta, the emotional or conceptual “turn” in the poem when the poet changes their mind or takes the idea of the poem in a new direction. Italian, or Petrarchan, sonnets turn on the ninth line; Shakespearean sonnets turn on the thirteenth. These computer-generated sonnets had no volta. Instead, they blandly circled their topics with no particular direction or intent. A poem on ‘Bipolar Disorder’ ends with, “A journey through the long and winding roads.” A poem about ‘Love at First Sight’ reads, “Relax and make the other people happy.” Another line from the same sonnet reads, “And wander down the streets of Cincinnati.” It’s a nice rhyme with “happy,” but there’s no particular reason the poem uses that town other than to make the rhyming work. It doesn’t make me feel or recognize anything about falling in love (or about Cincinnati), and it doesn’t show me anything unexpected or new.
The computer-generated sonnets had no volta. They blandly circled their topics with no particular direction or intent. Did the scientists working on the sonnet generator know about voltas? Or did they simply not care for this particular feature of the form? Regardless, the poems didn’t feel much like sonnets at all.
What is a sonnet? In this computer science paper, a sonnet is defined formally: fourteen lines in loose iambic pentameter, an ABAB CDCD EFEF GG rhyming scheme, and lines ending with a period or comma. Were the sonnets the program wrote on-topic? Yes. Were they rhythmically correct? Sure. Yet this definition seems superficial. Sonnets have changed in the four hundred years since Shakespeare wrote snarky love poems. The seventh sonnet of Terrance Hayes’s American Sonnets for my Future and Past Assassins , for example, begins:
I lock you in an American sonnet that is part prison,
Hayes asks me to deal with the American lineage in which I live. The sonnet becomes a metaphor for a country in flux. Hayes’s sonnets don’t follow the ‘correct’ rhyming scheme. They don’t rhyme at all, which is typical of contemporary American sonnets. Instead, Hayes’s sonnets push up against the form as a kind of confinement, as well as a means of liberation. In contrast, the lines from the computer program held no heat. Poetry, like so much art, often tries to upend our ideas of what the art form does. It changes over time, and it challenges the viewer. The best poems don’t give us answers, but open us up to more questions. I’ve always been taught that even the poet should be surprised by what they write, because poetry is as much an act of discovery as it is creation.
It’s not that a computer should (or even could) be writing like Terrance Hayes. But the richness of the sonnet as a form goes far beyond a rhyming and metrical schema. If the goal of computer scientists is to take on a random rhyming challenge, then they should go ahead. But of course they’re not really interested in rhymes. The computer scientists wanted the program to generate “interesting, creative text” that would be indistinguishable from a human sonnet. They were searching for humanness. Maybe they hadn’t gotten there quite yet, but presumably this was their eager first step. Yet to me, it seemed they were on the wrong path. If they wanted to create human-like sonnets, they didn’t seem to know much about why we read these ‘human’ sonnets in the first place, or what we get from them.
*
I had entered the doctorate program to study what computer programs might learn from reading poetry, perhaps something they couldn’t learn from reading newspaper articles. This seemed like a path to making computation more human, as well as a way to unite my disparate interests. I have always felt pulled in opposite directions: on one end, to the intense joy of making art; on the other, to the inherent beauty in mathematics and computation. I thought I was finally finding a way to bring these two loves together. But the shitty computer-generated sonnets made me mad, though at first I couldn’t quite explain why. Perhaps bringing poetry and computation together was a kind of sacrilege, and I should have kept my art separate from my academics.
It’s strange to me that someone might want to create a poem without caring much about what poetry sets out to do. Then I was introduced to the work of poet and programmer Allison Parrish . Parrish also worked in this liminal space between computation and poetry, but she was able to articulate all the problems she saw there—and in doing so, help me understand my own unease. In a bold talk to computer scientists, Parrish once commented that academic work on poetry generation tends to think of poetry as whatever the researchers were forced to read in high school. Hence a desire for rhyming sonnets that are vaguely on topic but devoid of any intensity.
Parrish said that much of what computer scientists consider poetic simply isn’t interesting to her: She doesn’t want to read poetry that’s fluent, or likable, or touching. She’s looking for something else, and has made a career out of asking how we can make words a bit more like paint: something we can squish or stretch or smear together. Her poetry often eschews meaning in favor of sound. She spent a long time thinking about how to turn lines of poetry into long lists of numbers representing how a line of poetry sounds. Once you can convert the poetry into numbers, you can do all kinds of unthinkable things, like measure the sonic distance between poems, add lines together, or find words that sound “in between.”
Her book Articulations was created by finding lines of public domain poetry that are most sonically alike, and then stitching them together in sequence to produce a long prose poem that moves through the sound-space of poetry. It tends toward a shifting repetition, filled with internal rhyme. One part reads:
Whispered to it, Westward. Westward. To eastward and westward to eastward and to westward reaching westward to its source, for straight to each nest they flew, in wild quest which to itself and by itself is true.
Parrish isn’t interested in creating poetry that can “pass” for human poetry. Instead, she asks how computers let us do things with words we wouldn’t otherwise be able to do. I was mesmerized by Parrish’s work—its musicality, how it managed to create something with language I hadn’t seen anywhere else—instead of reducing poetry to some standardized form. Her poems were her own, human endeavor, not some hopeful duplicate of something else. She found a way to unite computation and poetry in a way that didn’t denigrate either.
*
Lillian-Yvonne Bertram , another poet and programmer, also has different ideas about what poetry should do. In an early computational piece by Bertram, titled ‘ Forever Gwendolyn Brooks ,’ the reader can generate a new Gwendolyn Brooks-esque poem indefinitely. Each generated poem echoes Brooks’s voice, and the echo can go on forever.
At the time, Bertram was reading Brooks’s book RIOT , written in response to the riots in Chicago after the assassination of Dr. Martin Luther King, Jr. RIOT is a short and intense book, carefully crafted from the design and the colors to the language. Bertram was coding combinatorial poetry generators that relied on templates and so was thinking about Brooks’s grammar and syntax, her unique sentence style. But Bertram also was thinking about RIOT as a path forward for Black communities and Black liberation, a way to understand the events of its day, the events of our current day. Bertram says that ‘Forever Gwendolyn Brooks’ is an homage to Brooks: to her writing, to her legacy, but also to her social commitments.
In 2019, Bertram published Travesty Generator , a book in which they use computer programs from the 1960s to create poems about brutality against Black people, and the algorithmic nature of this brutality—if algorithms are sets of instructions people can follow again and again to get a desired result, white supremacy and brutality against Black people seems similarly encoded in our society. Bertram’s poems contain traces of the 1960s computer programs as well as the poems she uses as source material. As in ‘Forever Gwendolyn Brooks,’ Bertram draws from a very different canon than most of computer science, which privileges white, Western literature. Reading their work, I finally found computational poetry that was subversive. Instead of sonnets with no substance, Bertram’s poems challenge me just like Hayes’s sonnets do. Bertram’s work was hard for me. It still is. Which is why I keep coming back to it. Others agree— Travesty Generator was longlisted for the National Book Award for Poetry
In some ways, Bertram’s computational poetry cannot help but be subversive. People typically think of coders as white men. In 2019, only thirteen out of the 1,649 PhDs in computer science awarded in the US were awarded to Black people, according to the Computing Resource Association. The field of computer science has historically been, and in most practical ways continues to be, inhospitable, if not downright terrorizing, to anyone considered an ‘other.’ As a result, Bertram often discusses their work using the language of poet Harryette Mullen, who wrote about being an “unimagined reader.” As a literate Black woman with the rights of citizenship, Mullen was never an imagined or intended subject of, say, the Declaration of Independence. Similarly, Bertram is the “unimagined coder,” using algorithms never intended for them as a Black woman. Their work tends to center Black authors and this is deliberate because, as they say, “The times feel too urgent to do otherwise.” In Bertram’s poems, I found work that foregrounds political choices, because all coding is political, even if many computer scientists choose to ignore this.
*
My early experiences of the intersection of poetry and computation seemed to strip poetry of everything I loved; Parrish and Bertram showed me computational poetry that moved and challenged me. I needed that sign that these things could be brought together, not just in a meaningful way, but in a way that retained a critical care for each.
As far as I can tell, most computer scientists don’t read poetry, and certainly not contemporary poetry. No shade: It seems like most people don’t read poetry. But it’s strange to me that someone might want to create a poem without caring much about what poetry sets out to do. The computer-generated sonnets revealed the problem with so much work in technology, and with so many of the people in its midst: a lack of respect for the medium in which it works. When that professor told me that poetry doesn’t make any sense, he was betraying his disinterest in poetry as an art form. He saw it as meaningless blather. Of course, he was wrong. A sonnet isn’t simply a series of rhyming lines. It’s a lineage. It’s a history. It’s a future. Even the most abstract poetry, like some melodic lines from Parrish’s Articulations , can make us feel something, in the way music makes us feel something without relying on the meanings of words.
A sonnet isn’t simply a series of rhyming lines. It’s a lineage. It’s a history. It’s a future. When Parrish reprimanded computer scientists for thinking about poetry as their high school English canon, she didn’t say they needed to widen their view of poetry. She said they needed to own their ideas of poetry. Implicit in her critique was the question of who gets to decide what poetry is. She was reprimanding their stance that poetry might be universally defined; that, in their search for ‘humanness,’ they might get to define it. But this search for some generalized human behavior will always be fraught, and Bertram’s work, as an unimagined coder, makes the fractures clear.
By the time Bertram’s Travesty Generator came out, I had shifted my research away from trying to help computers better understand people and towards helping people better use computers. My advisor had convinced me that centering people, rather than programs, was more exciting, maybe even more challenging. Instead of trying to figure out what computers can do on their own, I began to spend more time thinking about what computers should be doing to help us.
I started making programs that would give writers ideas. I made a series of computer-generated thesauruses that reflected different styles and domains. In this shift, I had finally managed to align computation and poetry in my work. Still, this doesn’t make me any happier with the typical computer science approach.
I don’t believe every computer scientist interested in poetry generation needs to be a poetry expert, or even a poetry lover—though the latter might help. Instead, I think we need to acknowledge that we’re working in domains with rich histories and diverse voices. The poetry scene of today has much to give. It tries to grapple with its past. So often computer scientists ignore the worlds we parachute into, and our ability to gain success without first learning and finding respect is a privilege we need to lose.
If poetry is an act of discovery for the writer, then even a computational poem has to uncover something new for the programmer. This is what Parrish and Bertram do with computer programs, though anything can be a tool to help us with discovery, whether it be a prompt, a letter, a friend, a thesaurus, a butterfly. Computers are just another tool, another way to access what we’re trying to get at with poetry: expression, communication, an unearthing, something sublime or maybe even physical. But like any tool, it matters who wields it, and what they want to do.
