Prior to publishing our first data poems, it seemed only fair to let the data have its say. In this post, we’ll look at some poetry that’s been generated by data.
A quick Google search for algorithmic poetry returns a strong crop of projects that have generated poetry from data using algorithms. However, quality is the key, so here are three of the better algorithms that generate poetry.
New York Times Haiku
The New York Times Haiku Generator is an absolute gem. It uses an algorithm to search within The NYT’s website and pull out poems that conform to the broad rules of haiku poetry, namely the five-seven-five syllable structure for which haikus are most famous.
A few factors combine to make this a successful project. The scale of copy generated by NYT is an obvious bonus — the more copy you produce, the likely it is that suitable haikus can be sourced from your content. The US journalistic style’s tendency toward punctuation heavy, slightly traditional sentence structures and the general use of simple, short, direct words also helps.
Finally, the superficially simple five-seven-five structure is relatively simple from a programming point of view; the starting point is essentially searching through data for a single pattern.
Swift-speare
This was an experiment by J Nathan Matias, a genius at MIT, who created an algorithm that learns through experience. Matias fed his algorithm poetry by Shakespeare, Milton and others. He then asked it to produce a sonnet and achieved some pretty impressive results.
This project is an excellent example of the potential of machine learning and the work it produced was quite complex. However, in working to produce a Shakespearean sonnet, the parameters guiding what needed to be produced were pretty specific.
Poetweet
A much more interactive project, if not as pleasing in output, is Poetweet, launched by cultural centre b_arco, based in Sao Paolo, Brazil. Poetweet takes a Twitter account’s entire tweet archive and turns it into poetry.
The project is incredibly ambitious, allowing users to choose to create a sonnet or two rarer forms, a rondel or indriso. It works for tweets in both English and Portuguese.
The results are mixed but a great achievement when you factor in the complexity involved and that fact that users can generate poetry from any twitter account.
These three projects show the amazing possibilities provided by data in terms of creating poetry. They also show the limitations and the derivative nature of what’s currently possible. The next step in our journey is to create poetry from data.