Word Up! ... a simple yet powerful word prediction app

• Do you want to type faster?
• Do you want to avoid inaccuracies while typing when using a small keyboard on your phone or tablet?
• We have got an app for you!!

App Features

• A simple and intuitive interface - simply start typing in the text-box
• Fast response time
• Under 10 MB of final dictionary that loads quickly
• Displays top 3 most likely words almost instantaneously

Building the data dictionary

• 80% of the entire data is used to build the final dictionary
• To overcome the limitations of RAM, the data was divided into 6 chunks (500,000 entries each) for further processing
• Each dataset is first cleaned by converting to lowercase, removing extra whitespace, numbers, punctuation and non-english characters
• After cleaning, one through five grams are generated and stored in a frequency-sorted data table to allow for efficient processing
• To get a better trade-off between accuracy and model-size, only n-grams that occur 5 or more times are kept in the final dictionary. This allowed to keep the size of final dictionary under 10 MB but with a good accuracy.

• Last four words from input string are used for predicting the next word
• These four words are converted to unigram, bigram, trigram and four-gram to search for the next most likely word using the backoff approach
• Backoff approach first tries to find the most likely match in a four-gram, followed by trigram, bigram and finally unigram
• If no match is found, three of the top 10 frequently occurring unigrams from training set are presented.

Acknowledgements

• John Hopkins University's Data Science Specialization Team (Jeff Leek, Roger Peng and Brian Caffo) for teaching wonderful courses in the specialization
• SwiftKey for providing the data
• Fellow learners in the specialization and the vast and active online R community

Thanks for checking the app out!