×

Dartmouth Contest Asks Whether a Software Program Can Write



Friday, July 31, 2015
Hanover — What breeds creativity? For Dan Rockmore, a Dartmouth College professor who is taking a stab at this question, a little physical pain doesn’t hurt.

Rockmore, director of the Neukom Institute at Dartmouth College, enjoys spinning classes, workouts that take place on stationary bikes with pumping music and the shouted encouragement of a spin instructor.

One day, during a recent business trip to New York City, Rockmore sweated away at a popular spin studio, and his mind began to wander. He had always believed that the instructor made the class, and that a good instructor, besides having the right kind of personality, chose a good playlist.

For Rockmore, a professor of math and computational science, picking the right music involved more of both of those subjects than one might imagine. The energy of the exercise depended on its rhythm, on the shouting and the pounding music, and the speed of the cyclists (measured in revolutions per minute) needed to match the beat.

He wondered: Was there spin software? Do instructors analyze beats? Engineer them to optimal pump-up efficiency?

He asked an instructor he particularly liked. No, the man said, no software, but he did count the rhythms out by hand.

But Rockmore suspected he could engineer the same effect. He asked Michael Casey, a Dartmouth professor whose research spans music and computer science, about devising a Turing test: Could human dancers tell whether the music that moved them came from a person or from a machine?

Casey expanded the idea to other creative pursuits, and a contest was born.

Through early 2016, Rockmore and his colleagues are accepting submissions from programmers who believe their code can generate a piece of dance music, a sonnet or a short work of fiction that can pass as human before a human audience.

The Turing test, as conceived by British computer scientist Alan Turing, is a way to identify artificial intelligence. If a computer, chatting by text only, can convince a human that he or she is speaking to another person, the machine passes.

Rockmore’s three contests — DigiLit, for short stories; PoetiX, for poems; and AlgoRythms, for music — will extend the concept to each of the three art forms.

In a recent interview, Rockmore said the judging method was yet to be determined, though he said the music competition would likely take the form of a live dance party at Dartmouth.

“It just seems like so much fun, honestly,” the math professor said one recent morning in his sun-filled office in Kemeny Hall. Rather than a genteel conference, “our vision at the end is that there’s a party.”

As to whether Rockmore and his researchers had yet received any hate mail from literary snobs, he said he had received a thoughtful email from a student who wondered whether automating writing could put writers out of jobs.

To that question, the professor posed another: Does a writer really have what’s considered a traditional job?

There is precedent for a computational approach to writing, though so far it has fallen mostly on the side of analysis, rather than composition.

In April, for instance, two psychologists from the University of Texas at Austin claimed to have resolved the long-disputed authorship of Double Falsehood , an early 18th-century play possibly based on a lost work by William Shakespeare and John Fletcher, by analyzing period pronunciations and a category of so-called “social words” scattered throughout the manuscript. The researchers found that the patterns acted as psychological signatures, of sorts, that identify the two men as the original authors.

As it happens, Rockmore did similar research in 2012, when he tracked what he called “content-free” words in thousands of literary sources: prepositions, articles, conjunctions, “to be” verbs and some common nouns and pronouns. In the resulting paper, he and several colleagues dubbed their work “the first large-scale temporal stylometric study of literature.”

And in journalism, as Rockmore and Allen Riddell, a postdoctoral fellow at the Neukom Institute, pointed out last week, certain kinds of stories do literally write themselves: business articles, mainly, on such numbers-based subjects as profit reports and stock performance.

So will software soon replace the work of creative writers?

“As a writer,” Diana Whitney, a poet who lives in Brattleboro, Vt., joked, “my hope is that those poems are going to be disappointing, because otherwise we’re going to be obsolete.”

Though she said she would read computer-generated verse with great interest, Whitney expressed doubt that it could convey the subjectivity — “the awakening or evoking of an interior world,” as she put it — contained in a traditional poem.

“My first reaction was dismay,” she said. “Is that all poetry is — is the order and the meter and the sound of words on the page? Is there nothing else animating it, which hopefully comes from the writer?”

The poet and the scientists were largely in agreement, however.

“We’re not after the next great novelist here,” Riddell said. “We’re not trying to create a program that wins...”

“The next Nobel Prize,” Rockmore interjected.

“Right,” Riddell said.

The goal in all of this, Rockmore said, was simply to start a conversation.

“I’m sure we’ll do it at least twice, and then take stock,” he said of the contest.

People often assume, Rockmore said, that there’s one number for machine intelligence and another for the human variety and that one day the former will surpass the latter. But in fact, the possibility of similar tests always remains.

“There are so many dimensions of what it is to be human, you could pose a Turing test as to any of them,” he said.

Perhaps, he mused, the contest really was about whether a thoughtless, bloodless algorithm could spark a real human emotion.

But couldn’t an entrant also write a program that spat out worthless poetry — garbage, but seemingly human garbage?

Riddell laughed. It was a viable strategy, he said, though it would still require great skill to carry off.

“That’s what’s going to win,” he said, “but we’ll be happy.”

Rob Wolfe can be reached at rwolfe@vnews.com or 603-727-3242.

Correction

Allen Riddell is a postdoctoral fellow at Dartmouth’s Neukom Institute. His title was incorrect in an earlier version of this story.