Tweet Helper App

Jonathan Dorsey
June 19, 2018

The Tweet Helper App is designed to help users compose tweets more conveniently.
Just type into the box and the app will generate five suggested words.
Click on any of the five suggestions if you wish, and it will be added to the text box.
The app is located here.
The code is located on github here.

app screenshot

Here we see the interface.
Beneath the title and simple instructions, there is a text box to type into, and five buttons for adding suggested words
In this example, the top suggestion is “Wow thats pretty good”

The app generates its suggestions using a simple model
100,000 tweets were randomly selected and various n-grams were formed
An n-gram is just a list of n consecutive words. i.e. a 2-gram or bigram is just a pair of words
Using a list of the most common bigrams, one can predict the next word in a sentence by looking at the last word and then finding all bigrams that start with that word and predicting the most common second words in those bigrams.
For example, if “thank you” appears often in the text, then we predict “you” to follow “thank”
Similarly, one can use trigrams to predict using a two word context, and monograms to predict based on no words.

The model stores only the top 5 responses for common contexts. For a fuller explanation see the github repo linked to above.
When predicting text, we use the largest context stored in the model. That may be two words, one word, or no words.
A second set of 100,000 tweets was used to assess the model's accuracy
The best prediction (left most button in the app) guessed the correct word 11.6% of the time
The top 5 predictions contained the correct word 25.2% of the time.