Shiny App for word prediction - Background

People are spending an increasing amount of time on their mobile devices - but typing can be a serious pain. SwiftKey builds a smart keyboard that makes it easier for people to type on their mobile devices. Their cornerstone: using predictive text models!

Content

1. Background
2. Methodology
3. Predictive performance
4. That’s how it works

Methodology

In this presentation, we will present a shiny app that may be used to apply predictive text models like those used by SwiftKey.

Basis: Language examples collected from common internet sources are analysed

Principle:

• Language is split into n-grams (semantic entities)

• Irrelevant information is deleted

• Word frequencies are calculated

Based on this: The most probable next word is calculated by a mathematical prediction algorithm

Predictive performance

To calculate the probability of certain word combinations, the prediction algorithm uses trigrams, bigrams and unigrams. These are 3- and 2-word combinations derived from the language samples our study is based on. If you enter a phrase, the last two words are considered by the algorithm. They are matched with word combinations in the data set and the 5 most probable next words are searched for.

Predictive performance of the model evaluated by perplexity (PP=2−N1​∑log2​P(wi​∣context))
is:

3.91

–> The model has approximately 4 eqiprobable options per sentence, which is a very good performance - although not perfect!

That’s how it works

Using the Prediction Model

To use the prediction model do the following:

1. Start the app … [link below]

2. Enter a phrase you want the app to complete …

3. Find the most proper next word in the results table!

You find the application here:

Word Prediction App

Hopefully, you enjoy using it!