Does your business need to improve it’s technology game? …. the following text prediction app could help you do just that!

How

• This simple to use application hosted on ShinyApps.io uses a wide corpus of tweets, blogs and news articles to help predict the next word in a sentance.

• The app has been created in order to balance both speed for the user whilst also aiming to retain accuracy

The Text Prediction App

Data

• The course data was kindly supplied by SwiftKey in the form of numerous Tweets, Blogs and News articles
• In total there were over four million lines of data to process, which made the balance between speed and accuracy all the more important

Speed

• To make the app as quick as possible the processing was completed using the following R Packages:
  • Quanteda: A text processing package used to create Tokens and N-Gram data
  • Data.Table: A faster processing alternative of using dplyr and data.frames in R
  • SQLdf: In order to quickly match input text to the N-Gram data using SQL calls
  • DoParallel: Parallelising processing tasks to speed up run time

Complexity reduction

• All text provided by Swiftkey was processed initially and then reduced by only including N-Grams which occur greater than 5 or 10 times in the text

Prediction process

The App is based on pre-computed N-Gram tables, from bigrams upto 5-Grams, containing prediction words and scores

1. An input string is converted to tokens with punctuation, stopwords etc.. removed
2. Each N-Gram table contains strings of words in one column, a predictor word and a frequency / score - e.g. String: “once_upon_a”, Predictor: “time”, Score: n
3. A backoff model is used to find the most likely next word prediction
   • Match the last four words of the input string to the 5-Gram table and select the most frequent word
   • If no match exists backoff to a lower order N-Gram using a stupid backoff score and match
   • Failing this if no matches are found, the most popular unigrams are selected

Description

• The application is hosted on Shiny Apps and predicts the next word given a user input string

How to Use the App

• It is simple to use, just type in a text string on the left navigation bar, and the predicted word will appear on the right hand side
• The app only updates once the button on the left hand side is clicked to refresh
• Note that the app will only work at present with English as the language

How to Find the App

• The application can be run from the following address: https://ma3pab.shinyapps.io/TextPredictR/