• Typing full sentences is slow
• Mobile keyboards already suggest the next word
• Goal: Build the same thing from scratch using real data

Given a phrase, predict the most likely next word.

Step 1 — Training data

• 3 sources: Blogs, News, Twitter (~555 MB, ~70M words)
• 5% random sample used for speed (~14M words after cleaning)

Step 2 — N-gram tables

• Counts how often word sequences appear together
• Builds lookup tables: bigram (2-word), trigram (3-word), quadgram (4-word)

Step 3 — Backoff prediction

Input: "thanks for the"
  → Look in quadgram table first
  → No match? Try trigram
  → No match? Try bigram
  → Return top 3 most frequent matches

Backoff ensures a prediction is always returned, even for rare phrases.

How to use it:

1. Type any word or phrase in the text box
2. Use the slider to choose 1, 2, or 3 suggestions
3. Predicted next words appear instantly below

Features:

• Handles contractions, mixed case, punctuation
• Filters profanity automatically
• Falls back gracefully for unknown words

Live app: shinyapps.io/next-word-prediction-shiny-app-swadhwa

Simple, fast, and always returns a prediction.

Created by Sumit Wadhwa