• People are spending more time using their mobile devices
• Typing with mobile devices is not always easy
• Tech companies create applications that help people type more easily
• Providing further text suggestions to the application user is achieved with predictive text modeling
• For example, if we have a sentence:

Mary had a little …

• Google search text prediction function tells us the next most likely word is lamb
• The purpose of this project is to create an application that would predict the following words in a similar way
• Three files with texts from Twitter messages, blogs and internet news in US English provided by SwiftKey were used

• The next word is predicted using n-grams - contiguous sequences of last n words from given phrases
• This assumption of the next word depending only on a smaller number of previous words comes from the concept of Markov chains
• For the given n-gram, most frequent next words are selected
• If there are no such words, we “back-off” to lower n values
• Katz back-off is used to calculate next word probabilities
• It is used with absolute discounting to save some probability mass for unseen words/n-grams
• Absolute discounting means subtracting a fixed discount from each count

• For each data source, accuracy rates were calculated for both train and test data
• It was checked if the actual next word is among top n (1, 3, 5, 10) words predicted by the algorithm
• For example, 11.3 for Twitter top 3 train predictions means that in 11.3% of tested cases for train data, the actual next word was among top 3 predictions for the given phrase
• The higher the n value, the higher the accuracy rates

• Enter the phrase for which you want to predict the next word
• Select the number of predicted words (default is 1)
• Choose if you want to see prediction probabilities
• Push “Show Results” button