diff --git a/readme.md b/readme.md index b9cb0ef47c2004f12897d9e7201840b28c4bf4fc..67bc4c65bdd5cd1596e39b6653de3ddadfa662c1 100644 --- a/readme.md +++ b/readme.md @@ -1,11 +1,41 @@ +# Python - Trump Tweet Predictions -# CS5300 a3 +## Description +Recurrent Neural Networks using the Keras library. -## Author -Brandon Rodriguez +This project reads in archives of trump tweets (specifically, the dataset was his 2017 tweets). +If I recall correctly, the network would handle character by character, starting with the first one of the sentence. +With each progressive character (hence the "recurrent" part), it would read all predicted characters of the tweet so +far, up until the current character. It would then use this data to predict the next character. -## Description -Neural Networks using Keras (High level Tensorflow library). +All tweets would have a "starting delimiter" character as the first character by default. And then either end at once +the network output an "ending delimiter" character, or if it hit the maximum possible length of tweets. + +### Notes on Future Iterations +While it's been years, so I would need to do research again to examine current trends, I remember specifically wanting +to try a new iteration of this network, back when I took this class. + +Essentially, I wanted to try teaching the network to output an entire word at a time, instead of a single character at a +time. + +Obviously, this would be more complicated and likely take longer to train, but my thought was that it would result in +more coherent sentence structure, and better understanding of grammar. + + +## Keras +Keras is a high-level library built ontop of TensorFlow. The intention is to make it even easier and more user-friendly +than ever to build a Neural Network. + + +## Why Twitter and Trump? +There were two main reasons for the class-wide choice: +1) Twitter was chosen due to generally limiting tweet size, which inherently limits complexity of the network. +2) Trump was specifically chosen due to being a very active account that happens to also be fairly well known and +archived. + +Thus, working with data from this account was simpler, and very little formatting or normalization was needed. Twitter + +known archives took care of most of that for us. +Of course, at the time, we had no idea that Trump would raise to the level of controversy and infamy that he now has.