For my final project for computational approaches to narrative, I explored the limitations of GPT-2's ability to understand human connections. Missed connections is a personal advertisement of an encounter someone had with another person, but failed to exchange numbers. Usually it describes the person (their feature or what they were wearing), the place that it happened and the exchange that happened between those two people.
An example taken from Wikipedia:
"You were on your bike on the sidewalk (Sunday). I was walking the opposite direction. You had a plastic parrot on your handlebars. I smiled. You said "Hi" and smiled back. You won my undying admiration for not wearing a football jersey. That's it. Coffee or a drink?"
A missed connection from San Diego Magazine, December 2011, p.42
There is something deeply human about reading these missed connections. Some of them feel like a peek into someone's most intimate thoughts that are out on public display. Missed connections originated from Craigslist but has also since sprouted in other places such a Reddit. It existed as something largely of the past and since then it's started to populate with more inappropriate comments. I was able to find archived old missed connection posts from back when it was just about encounters between two people that failed to amount to anything more after, but always that wonder of what if.
I fine tuned a GPT-2 model with this collection of missed connections. I thought there was something interesting about juxtaposing a model that was highly praised for its ability to replicate human text, yet also fell short when it came to understanding human connections and stories. The outputs fail to make any real sense, but there are some that make you read twice. It is almost by accident that, every once in a while, the model is able to output text that touches upon something for the reader. Below are a selection of outputs I generated. Some are funny while others thoughtful.
The visuals are made using Runway through GANs and style transfer. The source photos were taken from artificially generated photos of people. Again, I wanted the photos to be distorted and look slightly disconnected from the real world in order to match the outputs. I still am playing around with the images (i'm not 100% satisfied with them currently). But this project is still a work in progress.