Data Science Capstone - Slide Deck

This presentation is about the App created for Coursera Data Science final course.

The app Word Oracle predicts the next word based on previous ones.

The app displays only the best macth for word prediction, as requested by assignement specification. If the word predicted is correct, it can be added to typed text with a simple double click of the mouse.

The app is minimal and is designed to be used by smartphones and to simplify the user interface and user experience as much as possible. You can see a screenshot in page 5

To create the final algorithm of the app I compared two very different approaches:

• Recurrent Neural Network with Long Short-Term Memory
• Katz's back-off model with Stemming and Good-Turing frequency estimation algorithms for enhance precision

On a sample of \( 1000 \) sentences taken from the dataset, the approach with the RNN allows a greater precision (\( 27.3\% \)) than the Backoff approach (\( 18.1\% \)) but requires a much greater computing power and in this specific application the user experience is more important than the precision, so the second approach was chosen.

When some text is inputted, the algorithm does the following steps:

• clean text. Remove stopwords, expand contractions, remove spaces.
• Calculate probability for 2-gram, and if applicable of 3-gram and 4-gram.
• Weight probability with Good-Turing frequency of frequencies table
• Return the best match that will be displayed

The Words Oracle app interface is very simple: