Page 92 - AC/E Digital Culture Annual Report

92
This is based on the premise that the creation
of literature by human authors is the result of the application of a very extensive range of skills. Each of the experiments described below is designed to model a unique and specific skill. As in any act of engineering, the underlying idea is that once these specific skills are individually modeled, they can begin to be applied as com- binations in the development of more elaborate models. Until now, the number of skills that people have tried to model has been very limited and experiments in which skills have been combined have not yet been brought to any conclusion.
The WASP poetry generator combines the capacity of a poet to measure metric forms and iterate by applying successive modifications in search of the best fit.
The WASP (Wishful Automatic Spanish Poet) generator of poetry
With a generative mechanism based on the statistical model of language that is in turn based on n-grams, the WASP poetry generator [23, 25] was built using a developmental focus that combines the capacity of a poet to mea- sure metric forms and iterate on a draft by applying successive modifications in search of the best fit. WASP works like a group of fami- lies of automatic experts (generators, verse-makers, arbitrators and reviewers) who work together as a cooperative of readers, critics, editors and writers. Together they generate a number of drafts on which they all operate, modifying them and cleaning them up in a developmental manner until a final version is chosen, the effort being judged the best of the batch. The general style of the resulting poems is strongly determined by the statistical model of language that is used, worked out on the basis of different oeuvres: a collection of classic Spanish poems and a collection of journalistic articles taken from an online edition of a Spanish newspaper [23].
Fig. 1. Screenshot of the automatic poet WASP, showing a poem it created and its metric analysis.
The SPAR (Small Poem Automatic Rhymer) program
The SPAR program (Small Poem Automatic Rhymer) carries out its creative task in five indi- vidual stages. First, it builds a series of models from a work of reference that shows which words are usually followed by others and which words rhyme with others. These models are used to inform the next steps. Second, based on a word offered by the user, it builds a collection of related words that represents what the system considers could be mentioned in a poem with said word as a title. Third, it looks for possible connections between these words (and other words that could rhyme). Fourth, by exploring the space determined by these connections, it builds phrases that could be included in a poem and combines them in verses that end in words that rhyme. Finally, in a given stanza, it looks for combinations of the resulting verses that comply with the restrictions of rhyme and can be linked with the minimum of cohesion.
Each of these stages can take between one
and three hours to compute because the areas of probability involved are huge. With smaller areas, it could be done faster but the chances of finding valid combinations diminish proportion- ally. The density of correct verses that can be generated from a given work is very low and so there is an intrinsic difficulty in the generation poetry. For these reasons, this approximation to the automatic generation of poetry is still not at the point it can be used interactively. Putting
COMPUTER-DRIVEN CREATIVITY STANDS AT THE FOREFRONT OF ARTIFICIAL INTELLIGENCE... · PABLO GERVÁS
Digital Trends in Culture