Scientists of Australia together with colleagues from the University of Toronto, department of English, have designed an algorithm to write poetry that follows the rules of rhythm and measure. To test their results, the researchers asked people in line if they could distinguish between the verses written by the software and by human beings. The best version of the algorithm showed that the 50% of the people came to the conclusion that the verses written by the program were written by a human being. Could this be considered a new proof of Turing?
Somehow, the verses of the computer were better than those of Shakespeare. The rhythm and the metric on the poetry generated by the software was more precise than the one found in poems written by human beings. However, “Deep-speare” still has a long way before you write something that is good enough.
“It is very easy for me to know if it was written by the computer or not, it is ridiculously simple,” says Adam Hammond, a professor of English and co-author of the article that outlines the algorithm used. “We have resolved two of four problems,” says Hammond, referring to the metric and the rhythm. “The other two are much more difficult: to make something readable, and to evoke some emotion in the reader.”
Scientists from IBM Research Australia, the University of Melbourne and Thomson Reuters, trained a neural network of about 2,700 sonnets at Project Gutenberg, which is an open-access library. The computer uses three models: language, meter, and rhythm, and some probability to be the correct word for their poems. Produces “quatrains”, or four line verses.
The researchers measured the results by asking people in line about whether the poems were human, or created by the software, which -frankly – is a test of little value really, because there is no parameter to judge the level of literature they have these people.
One of the verses of the system was this:
“With joyous gambols gay and still array
No longer when I twas, while in his day
At first to pass in all delightful ways
Around him, charming and of all his days”
Deep-speare could not fool the expert. Hammond says that it was easy to realize the verses made by the computer because they are often inconsistent and contain grammatical errors like in this one: “I twas”. However, in the measurements of the metric and rhythm, the “quatrains”, the 4 lines that make up the verse, showed that are best seen in the poems human. The reason is that “it has a metric in a poem to create the expectation,” he explained. “Don’t break the pattern. It is about creating expectation and then violating it at some point”. The signal of the break is a change in the emotional state of the speaker, adds Hammond.
And if the poets could worry about these advances in the software, Hammond says that there is a lot of which a sonnet can’t be imitated on the computer, at least not yet. A sonnet contains three quatrains” and ends with a “couplet”, that is to say, two lines of verse. “It’s really a very complicated because usually the first bit proposes a problem and then there is a twist, then the “couplet” is the one who gives the solution”, he says. “Imagine teaching a computer to define a problem, to express it and offer a solution. Only that it is impossible to imagine”, stated Hammond.
The co-authors of the project are now looking to refine the algorithm so that it can be used in a single topic as for example, to write a fiction short. “Our results suggest that research in future should focus beyond the metrics, and improve what the program writes,” the researchers said.
The article can be read here: Deep-speare: To joint neural model of poetic language, meter and rhyme.
Coolest-hacks.com and Partners.