Text Prediction App

A capstone project for Coursera Data Science Specialization in collaboration with John Hopkins University and SwiftKey

Why This App

Typing on mobile = Considerable time = Frustration.
—> Need smart keyboard to predict next word.

How This App Was Built

1. Train model - using n-gram tokenization
   
2. Build predictive algorithm - Katz BackOff + Good Turing
   
3. Build Shiny App - clean text + predict next word
   

1. Build Corpus
   
   • Use Random Sampling
   • Split into train(60%), devtest(20%) & test(20%) sets.
     
2. Tokenization
   
   • Train set is cleansed (lower case, remove punctuations etc.)
     
   • 1,2,3,4-gram tokens built.
     
3. n-Gram Look-up Table - This is an (n x 6) matrix with
   
   • 4-gram tokens (w1-w2-w3-w4) as anchor + Frequency
     
   • + corresponding tri (w1-w2-w3) + bi (w2-w3) + uni (w3)
   • + corresponding next-word(w4).
     

Katz BackOff

1. Look for seen n-gram in look-up table.
   
2. If found, top 3 next-words are displayed.
   
3. If not, recursively back-off to (n-1)-gram & repeat 1 to 3.
   

Good Turing Smoothing

1. Probability discounted for seen n-grams with freq of freq <= 5.
   
2. Excess probabilities re-distributed to unseen n-grams.

What Shiny App Does

1. Clean a body of text (lower case, remove punctuations)
   
2. Deploy predictive algorithm in Step 2 above.
   

What the User Needs To Do

Go to https://honeydukes.shinyapps.io/Project/

1. Key in any phrase and click Predict button.
   
2. See output for top 3 most probable next words.