There are hundreds of poetry generation programs on the Internet – that is, systems capable of producing poetry automatically – but what can they be used for? Do these programs have an interest, beyond that of satisfying their designer?

We have heard a lot about GPT2 or GPT3, these huge computer programs capable of producing very realistic texts, and even poetry. GPT2 and GPT3 are in fact “models”, species of knowledge bases, fed by billions of sentences and texts gleaned from the Internet, and “digested” in order to be able to produce new texts, inspired by old texts, but in very different times.

Add constraints

To produce poetry, it is “enough” to add constraints: monitor the rhymes and the length of the lines, respect the overall structure, the absence of repetition in rhymed position, etc. Automatic generation systems are having some success (we find a phenomenal number of them on the Internet) because the task is fun, playful, but also complex if we want to produce texts with meaning (and even more if we want to control what is said).

We therefore have here an ideal setting for experimenting (often without funding constraints): the generation of poetry is often a hobby and a hobby, for the researcher as well as for the enlightened amateur.

A question arises, however. Are these systems worthy of attention?

On the literary level, most systems are still, it must be admitted, quite rudimentary and have difficulty in competing with Baudelaire or Rimbaud. The most advanced are however impressive, and it is above all the training base that plays a crucial role (that is to say all the texts which have enabled the system to learn. We can indeed provide a system dataset reduced but specialized (works by poets of the XIX ^th century for example) to “specialize” a system, adapt cost. We can then get systems who write paragraphs on the way de Balzac, or poetry in the style of Baudelaire.

However, it should be noted that the results appearing in the press (whether it be the generation of prose or poetry) are often the fruit of multiple trials, or even the fruit of post-editing work on the part of the journalist.

The Oupoco project

The Oupoco project (Ouvroir de poésie combinatoire) that we developed with a team from the LATTICE laboratory , had a more modest goal. Like Queneau’s experience in One Hundred Thousand Billion Poems , our primary ambition was to produce billions of poems simply by recombining verses from a representative French poetic corpus.

One hundred thousand billion poems , Raymond Queneau, 1961. Médard Museum

To this end, we have assembled a database of more than 4000 sonnets by authors ranging from the beginning of the 19 ^th to the beginning of the 20 ^th century. While all of Queneau’s lines rhyme together, we had to automatically determine the rhyme of each line in order to be able to produce rhyming poetry. From the start, the project was therefore more of an analysis project than a generation project (as shown in this video .

This project may seem damaging, in that it would pass poetry for “rubbish”. But the goal is obviously quite different. Concrete experiences and the meeting with the public have shown us that this fear is largely unjustified. The public (young and old, women and men) are amused, intrigued, want to know more. An audience usually not attracted to poetry is interested in what is produced. The public is not naive, even when it comes to children: they can see the fabricated, strange and playful nature of the affair. He knows that behind what is produced hide other texts and the incongruity of an extraordinary verse often pushes to go and see the original context, that is to say the original poem.

The poetry generator, with the distribution devices that go with it (such as the Poetry Box , a work of art designed by Atelier Raffard-Roussel and allowing to obtain a portable object integrating the Oupoco poetry generator) , allow a wide audience to reconnect with poetry, while it is a form often neglected even by regular readers.

As for the experiences in pure generation (where the poetry produced is not composed from pre-existing verses, but is actually designed by computer), they lead to think about other aspects. On the text itself: how rich is the text produced? What makes the value of a poetic text? If we are within the framework of a generation “in the manner of” (in a way similar to the production of music “in the manner of”), one can wonder about the value of the result, about the characteristics of a author, on what makes the style of an author, ultimately.

Different levels of creativity

These questions finally lead to questioning the notion of creativity itself. Margaret Boden , an English computer scientist who developed a theory on the question, distinguishes three levels of creativity, in humans as in computers: “exploratory creativity”, which consists just of extending a little what already exists (write a poem to Hugo’s way); “combinatorial creativity”, which consists in combining in an original way elements existing around us, but of a distant nature (the work of the Oulipo, mixing literature and mathematical constraints are probably of this order). The third form of creativity, qualified as “transformational”, is of a different nature, it radically changes the way of seeing reality and generally produces a whole new line of works. Margaret Boden talks about Picasso’s invention of cubism; one can think of the abandonment of the codes of the novel in the 1950s, around the new novel, but the notion of rupture in literature would be a concept to be discussed in itself.

The Oupoco system, just recombining existing worms, is undeniably exploratory in nature, even if this exploration is based on combinatorics. The holy grail of computer creativity would be to achieve transformational creativity, in Boden’s sense of the word. Would a computer be able to reach this level? We can doubt it, because this level implies a certain awareness of oneself, a step back from reality, to imagine completely new mechanisms. Artificial learning, the source of most recent and media developments in AI (artificial intelligence), is very good at generalizing and recombining the billions of data received as input, but is incapable of “taking a step forward. side ”, to really transform reality.

Finally, note that humans also learn from stimuli and by imitation. The nature and reality of transformational creativity is not fully proven. Perhaps from the billions of perceptions received during his life man is able to recombine in a sufficiently free way to give the impression of transformational creativity. We are then at the heart of cognition!

Author Bio: Thierry Poibeau is DR CNRS at École normale supérieure (ENS) – PSL