WordPredictor

gsk
April 2015

What is it?

WordPredictor is a prototype of a statistical learning model for word prediction.

Given a certain text, WordPredictor suggests one or more words for the following word.

The current prototype version of WordPredictor has been trained inspecting more than 400,000 documents in a blogs data set.

WordPredictor has an average accuracy rate of of 15%, 3% and 0.1% for the first three words suggested correspondingly.

Highlights

WordPredictor has been completely developed from scratch in R. It includes the following capabilities.

  • Transformation, Training & Evaluation functions
  • Prediction functions
  • Data encoding functions
  • Data structures management functions

WordPredictor uses an algorithm for encoding n-grams as floating-point numbers that allows for fast searches and efficient memory usage. The prototype application, code and data structures included, requires no more than 25 MB of storage.

Technical Specifications

  • Prediction Model. Trigram Back-off with Witten-Belt smoothing probabilities.
  • Word encoding. N-grams stored as floating point numbers.
    Max Vocabulary size is 131,071 types.
  • Stepped Training functions. Designed to be run as a sequence of processes, allow systems with small processing power to train any amount of data.
  • Lean data structures. 5 R objects (3 data frames, 1 hash table and 1 Environment) required. Low memory requirements thanks to encoding scheme.
  • Evaluation functions. Randomized stepped functions. Supported metrics: Accuracy and Perplexity.
  • R Packages requirements. TM and Hash.

WordPredictor App

WordPredictor Shinny App

  1. Enter the number of words you want to get as predictions.
  2. Choose if you want probabilities for each suggested words.
  3. Enter text and press the PREDICT button.

Open WordPredictor Shiny App cliking here

References

You can find more information about WordPredictor here.