TextSmart

TextSmart is designed to predict the next word and/or the rest of the current word that's being typed.

• The user interface has the following input/output components
  . the text area for inputting the message
  . three buttons whose labels get updated with the predicted words
  . the output message area below the buttons
  
  
• The input from the text field is fed through the prediction routine after each character is typed
  
• The prediction routine returns three most probable words out of which one is expected to
  . autocomplete the rest of the word when a non-space character is typed or
  . match the next word when a sapce is typed
  
• Three buttons get updated with the predicted words with the word on the first button has the highest probability
  
• The user can hit the button that has the predicted word to add it to the input text and continue typing

• Built the corpora with about 4M lines of blogs, news, and twitter feed provided
  
• Split the corpora into three parts using tm package with random sampling
  . training (60%)
  . validation (20%)
  . testing (20%)
  
• Further split training set into 6 parts to process them in parallel on Linux running on 16 CPU x 64G hvm/AWS
  
• Used perl regular expressions to remove profanity words from the input datasets
  
• Used quanteda package to remove
  . non-ascii characters
  . punctuation
  . digits and white space
  . symbols and hyphens
  . URLs and separators

• Generated 1-4 ngrams from each part of the training data using quanteda
  
• Merged all ngrams into one final data table
  
• Calculated probabilities for the last word on every ngram via a copy of ngram-frequency hash table
  
• Merged the probabilities of ngrams from all 6 parts for training into one final table
  
• Validation and test parts were put through similar processing steps
  
• Tried interpolation on 1% sampled set and found no major improvement in accuracy
  
• Pruned final set of ngrams to limit the size of the object in memory to 86MB
  
• Used the back off technique
  . it tries to match on 4-gram first if the input has at least three prior words
  . it returns the last words of the top three matching 4-grams that start with the input string passed
  . if there is no matching 4-gram, it tries 3-gram, then 2-gram, and finally 1-gram
  
• 46.39% success rate with the test data when ngrams with the frequency==1 were discarded
  
• Predicted 50% words from Quiz 2 and 30% from Quiz 3

The UI has been built using Shiny package and is hosted at TextSmart

https://sontivr.shinyapps.io/WordPrediction