SwiftKey Word Predictor

Vlad Pascal
October 2016

Shiny web application that predicts next word.
A user enters the phrase into the search box and presses the submit button.
Input: The application takes a string as an input.
Output: Predicts next word given the relative frequency and probability given some history

Using corpus created by Swiftkey, which includes news, blog and twitter posts.
Draw a sample from each category and combine them. Clean the data by removing invalid characters.
Using N-gram model, split the data into trigrams, bigrams and unigrams. Compute relative frequency at each step.
Break user input into ngrams and look it up in the N-grams.
If N-gram has zero counts, “back off” to a lower-level N-gram i.e. use trigram if there's a match, otherwise proceed to a bigram, otherwise proceed to the unigram.

The sample is drawn to minimize the size of the application (approx. 3.1 million of trigrams and bigrams).
Lower frequency ngrams were removed from the sample.
All the data were pre-processed outside of the app.
The algorithm doesn't use interpolation (mixing of probabilities), smoothing and/or discounting.
A variation of stupid backoff based on relative frequency.

The app is easy to use and navigate.
Powerful yet computationally manageable method without utilizing “expensive” algorithm.
The model can be easily deployed and scaled up on other platforms.

Location:

or copy and paste the following link: