He held out a hand

On hers without speaking, and

Felt a devotion.

The girl exclaimed, with an air

Of injured pride she started

Up. Their glances met

In a mist of bargaining

And hyperbole.

–Edith Wharton and the machine

In the Richard Powers novel Galatea 2.2, a cognitive neurologist and a novelist endeavor to create a neural net sentient enough to pass a PhD comprehensive in English literature. The machine, dubbed Helen, ultimately develops a mind of its own. Eric Elshtain and Jon Trowbridge aren’t quite that far along with their Gnoetry poetry-writing program yet–but then they’re only on version 0.2.

The concept for Gnoetry (that’s pronounced with a hard G) grew out of a conversation Elshtain and Trowbridge had in a Hyde Park diner a little over four years ago. Elshtain, who’s 38 and a doctoral candidate in the U. of C.’s Committee on the History of Culture, had been investigating early attempts at machine-generated poetry. Trowbridge, a 36-year-old software programmer, found the idea intriguing.

“Haiku machines have been around for a while,” Elshtain says. “As far as we were concerned they would just spew out these strings of gobbledygook.” Early poetry programs were hampered by primitive computing power, Trowbridge explains. And a lot were no better than Mad-Libs, “massive computer-generated fill-in-the-blank. Give me a verb, give me a noun, then plug them in and get a sentence. There’s no real structure there. You get stuff correct in some grammatical sense, but it turns out sounding like bad stream of consciousness.”

Gnoetry, by contrast, “requires substantially high processing power,” Trowbridge says. The program’s approach is purely statistical: it has no knowledge of grammar, and it can’t tell a noun from a verb. All it does is scan for patterns–word frequencies, common and uncommon word combinations. From there it randomly generates new sequences of words subject to the patterns it’s observed.

That’s where the human element comes in. To actually use Gnoetry you first choose a poetic form (haiku, renga, tanka, or blank verse) and a source text or texts from the program’s storehouse (most are copyright-free downloads from Project Gutenberg). Then, if you’re using multiple texts, you assign weights for the analysis–say, 30 percent Moby-Dick, 50 percent Origin of Species, and 20 percent Heart of Darkness. Hit go and voila, you have a gnoem. From there you can alter it in any fashion–change a word, a line, a verse. Only the user decides when it’s a poem.

Trowbridge claims this is beyond his ken. “I don’t know anything about poetry. I can write an algorithm,” he says. Elshtain, on the other hand, is a former poetry editor for the Chicago Review who’s had his work published in journals like McSweeney’s and Ploughshares (his mother, Jean Bethke Elshtain, is a professor at the University of Chicago Divinity School).

“Within the parameters of Gnoetry, I have a very conservative view of poetry,” he says. “I think that I want people to immediately see the machine’s output as poetry, and syllabic verse forms like haiku and accentual-syllabic forms like blank verse mark the poems with the banner ‘See–this is poetry!’ If the computer spat out free verse I think that too many people would just dismiss it as a fluke and say, ‘Well, anyone can do that.'”

In addition to literary works Elshtain and Trowbridge have downloaded rap lyrics, Dylan songs, and newswire stories for use as source texts. “The first thing that blew my mind that Gnoetry spit out was like this gangsta rap haiku,” Elshtain says. “Mostly because it is coherent and grammatically correct and truly as if Ice-T had written a haiku: ‘But fuck the whole cake / Bitch, you crazy, you crazy! / Make a move and see.'”

Her Social Frame is a chapbook of 35 Gnoetry-generated renga (a Japanese collaborative verse form with a 5-7-5, 7-7, 5-7-5 syllable count) made using the statistical properties of Edith Wharton’s novel The Custom of the Country. Elshtain was amazed at these because they seemed so perfect, even without human manipulation. “The Wharton poems are like distillations of her novels,” he says. “Lines like ‘Their glances met / in a mist of bargaining / and hyperbole’ completely capture the feel of reading her. . . . We even checked the text to make sure Gnoetry wasn’t just lifting phrases. And it wasn’t. There was very little use of regeneration. All I did in terms of editing was just change some punctuation.”

My first experiment with the machine didn’t turn out so bad, I thought. My formula was 50 percent P.G. Wodehouse’s Right Ho, Jeeves, 25 percent Edgar Allan Poe’s “The Fall of the House of Usher,” and 25 percent Poe’s “The Masque of the Red Death” in two-line, five-verse blank verse. The result:

A very matey scheme. The storm, the thing.

The butler came in. There were aching hearts

In Brinkley Court? About a fellow, but

A bit. A very matey greeting. I

Remember years ago? The fourth? The thing

Again. The thought amuses you? The thing?

The truth. The storm, the matter of the same

Unpleasant tone, the brow a bit. The thing

Again. The prizes. I anticipate

A great relief. In short, the thing in this.

Sure, it could stand a little tweaking, but I’ve read worse.

Gnoetry 0.3 is still in an early stage of development, but Elshtain and Trowbridge plan to make it downloadable for both PC and Mac users. (Version 0.1 is available online, but it’s not much use unless you’re a Linux geek.) For more on the program, and for selected human/machine collaborations–including Her Social Frame–see beardofbees.com.

Art accompanying story in printed newspaper (not available in this archive): photo/Jim Newberry.