I spent some time this weekend feeding a recurrent neural network (torch-rnn) a dump of Urban Dictionary definitions that I found posted online. Once you train the neural network on a text corpus, you can ask it to generate a block text based on the rules it learned. Janelle Shane over at aiweirdness.com leverages this or a similar nn to great humorous effect. Maybe I’ll write a more technical post later to explain how it works behind the scenes. I’d have to understand that myself first.
I chopped the corpus in a few different ways in an attempt to get better results, but I never quite got it past the gibberish phase of training. The base idea is just feeding in the words themselves, sans definitions, and seeing what the nn makes of it. Here are a few I hand selected from its generated output:
These would all be better if the nn generated definitions for each, but when I included the definitions in the training set, it just spit out effortless nonsense.
The kernel of what makes Janelle Shane’s work so fun and interesting lies in the anthropomorphizing of the rnn. We ask it to do something, and then when it does a terrible job at it, we can’t help but say “Oh computer, at least you tried to understand humans.” With this in mind, I decided to try one more thing before hanging this up for the day. What if I isolated just the definitions that were sex acts and fed them into the nn? Would the computer come up with something sexual that humans have never considered?
A sex act involves the felled-umbing this imerican inserted it