Creative computers

Can a computer be creative? Could one create art, even become an artist? We set out to explore these and other questions in a Creative Computations class lead by Ben Saunders. During this class, my lab partner and I created Morcom, an instagram poet with more social clout than me. In this blog, I will lead you through Morcoms story and share my thoughts on creative computers.

Yelpku's

We took inspiration from reddit haiku bots, which find comments that match a haiku syllable pattern and reformat them. A dull sentence can gain drama and emotion, sometimes even new meaning just by breaking it up. Though these poems are only haiku in the most shallow interpretation of the genre, we found this fitting for our purposes; generative art often emerges from following simple structural rules.

We decided to apply this technique to the Yelp review dataset, which contains XXX reviews of restaurants, bars and other (anonymized) establishments. We first split the reviews in sentences, counted the syllables, and filtered out the ones that could be formatted as a haiku.

This left us with XXX yelpku's, most of which were terrible. From an initial scan through these results, we found that many were incomprehensible, due to incorrect syllable counting or sentence splitting. Others are just uninspiring;

They offer Ladies Night Shooting on 2nd Thursday of the month and CCW classes .

But then we found a poem about pasta;

I was picking chunks of pasta noodles and meat off my shirt and hair .

This poem paints a very vivid scene. The haiku format also works out well, as each line builds up the image further. But most importantly, it made us laugh. We found several more of these gems;

I used the back end of the knife and fell apart so beautifully .

By removing context, our process creates room for new interpretations. In this example, this is aided further by a grammar mistake, forming a surreal picture, demanding to find deeper meaning -- instead of the banal scene of cracking open shelled food.

In a few hours, we found a handful of poems that we really enjoyed, but it was a very manual and labor intensive process. The harder we looked, the more it felt like we were generating the poems, instead of the computer.

Building a poet

We decided that our creative computer should not just create, but also have an opinion on what is good or bad. We used machine learning to model such an opinion. This was a fairly straight-forward process; we took a pretrained BERT model for classification, and finetuned it on a set of real, good haikus and a subset of our yelpku's. We collected the real haiku ourselves from across the web.

The resulting model was extremely good at deciding if a haiku was real, or taken from yelp. We could now use this model on the rest of the yelpku's, and filter out the false positives. We reasoned that the few yelpku's that fooled our model should look more like real haikus.

It's about the story

The value and appreciation of art does not happen in a vacuum. On the contrary, our experience of an art piece is strongly influenced by its context. By who made it, and how and why, by how it is presented and by the stories that surround it.