Word Prediction App with Stupid Backoff

Idea SwiftKey, the corporate partner in the capstone of Johns Hopkins University's Data Science Specialization, aims at making it easier for people to type on their mobile devices. One cornerstone of their technology is predictive text models. The idea of the capstone project is to build a shiny App that emulates this technique. Starting point is a large corpus of text documents from blogs, news and twitter sources provided by Swiftkey.

Goals

• Build a predictive text App that optimizes preformance and accuracy
• Explore a new data type and implement a useful model in a reasonable period of time
• Apply the entire data analysis pipline in a text mining context

Data base: Uni-, bi- and trigram tables

Developments: The App could be developed in different directions. In particular, we could think of feeding back the choices made by users into the n-gram tables, which would enable the App to learn from these choices and enhance predictive quality.

References (Selection)

• Theoretical background of stupid backoff modell: Large language models in machine translation by T. Brants et al, in EMNLP/CoNLL 2007
  
• Practical example of stupid backoff: Word Prediction Using Stupid Backoff With a 5-gram Language Model by Phil Ferriere on Rpubs

Thanks

Hope you'll like the App! (Try building the longest meaningful word sequence just by clicking on the word buttons, maybe choosing different starting words.)