Word Predictor App

• People are spending more time on mobile devices for many types of activities.
• Typing on mobile devices can be inconvenient.
• The experience can be improved by making typing easier through text prediction.
• A prototype app is shown based on a predictive text model.
• Three corpora of texts were used from 3 sources: (blogs, twitter, news).
• Each set has 30+ million words. Only 1% of the total corpora were randomly sampled and used as reference to improve speed.

• The top most frequently appearing quadgrams, trigrams and bigrams were gathered. The word sets are arranged alphabetically first then sorted based on frequency of occurence
• The top most frequently occuring unigrams (excluding stop words) were are also gathered

Sample word set for quadgram

• The Katz's backoff model is used to predict the next word
• Inputs are truncated into the last three words.
• Given the last three words the algorithim first selects a quadgram which matches the first 3 words and supplies the fourth one
• If no quadgram exists, the app takes the last 2 words and finds a pattern in the trigram and so on.
• The highest occuring combination will provide the top ranked word.
• If additional combinations occur, the 2nd, 3rd and 4th most frequently occuring word will be provided
• A balance of speed and accuracy was used to build the model

• In case the app below doesn't work it could be accessed in this link : https://jodlgayatin.shinyapps.io/WordPredictorApp/.
• For queries you may contact jodl.gayatin@gmail.com.