realDonaldTweets

Inspiration

This man's tweets are insane, and since tweets by nature are pretty simple pieces of text, we wanted to see if we could generate our own using some sort of machine learning model. Once we saw that it actually worked well, it was only natural to make a game out of it.

What it does

Our project has two parts, the natural language generation of tweets and the front end web UI. Nikhil worked on the natural language generation and Sophia worked on the front end UI, but we both collaborated on each part a lot. Essentially we built a Markov model from a database of tweets and we both curated specifically for this project. We then used generated tweets from that model to put onto a web game where people can decide which tweet is real and which tweet is fake.

How I built it

We built it using dart, javascript, html, and css. The markov model was made in dart, and the front end UI was written in javascript, html, and css.

Challenges I ran into

We ran into several challenges on both the front end and with the model generation. The markov model is a fairly simple machine learning data structure, but we ran into some issues with entropy, since every single step of the markov model generation process uses a lot of system entropy. This meant that we had to restart our computer fairly often to get the algorithm to properly generate new tweets. On the front end, we had several issues with styling and formatting, but in the end we designed a website with a really nice twitter like feel. We hope to expand on this in the future and make the website more mobile friendly as well.

Accomplishments that I'm proud of

We built a super super cool website that we are both really really excited to share with others! And the markov model is also pretty nifty too.

What I learned

GPU acceleration with CUDA, promises in javascript, how to code properly

What's next for realDonaldTweets

Creating a mobile friendly view Creating a share button Adding ethical analytics Tracking to see which tweets are most and least realistic Starting a sharing campaign