In Mikuana/NextWordR: An R-package/Shiny-application for word prediction

knitr::opts_chunk$set(echo = FALSE)

Overview

• We developed an application that accepts user text input, and returns suggestions for the next word. This can be added to the text by clicking on the suggestion button.

• Suggestions are based upon probability models using large text training data sets

• The algorithm can accept any text input, and will fall back on more naive suggestions in the case of input text that was unseen in the training data

• Model data, functions, and Shiny application code are all available in a repository

• The Shiny application is hosted online via ShinyApps.io, and embedded in this deck

Data

• The source data for our prediction models is provided by the Swiftkey company in collaboration with:
  • Coursera
  • Johns Hopkins University
• These data includes millions of posts from :
  • Twitter
  • various blogs
  • news articles
• We randomly sampled 10% of these data due to resource constraints on the development machine

Algorithm

• Word prediction is achieved using probabilistic estimates of the next word, based on up to 5 preceding words
• Input words are processed into lower-case, without special characters, and accordingly suggestions are formatted in the same way
• Good-Turing estimation used to predict for unseen or rarely seen combinations of up to 3 preceding words
  • Good-Turing not used for higher level ngrams due to computational requirements
• Back-off modeling reduces the number of preceding words provided to the model until at least 3 suggestions are returned
  • in some cases the combinations are completely novel to the model, and the suggestion is the same as if there were no preceding words

App

To use the app, simply type in the text box. Suggestions will appear as buttons below, which can be clicked to insert them into your text.

Mikuana/NextWordR documentation built on May 8, 2019, 10:49 a.m.