The Problem & Solution

Problem: Typing on mobile is slow and error-prone

Solution: A next-word prediction app trained on real English text — just like SwiftKey

• Trained on 150,000 lines of blogs, news and tweets
• Uses N-gram language modelling
• Instant predictions as you type
• Click any prediction to append it to your phrase

Try it: https://jjvqgo-farhan-bin.shinyapps.io/next-word-predictor/

How The Algorithm Works

Stupid Backoff N-gram Model — 4 layers of prediction:

User types → check QUADGRAMS (3,707,806 sequences)
No match?  → check TRIGRAMS  (2,997,658 sequences)
No match?  → check BIGRAMS   (1,361,416 pairs)
No match?  → fallback to most common UNIGRAMS

Example:

• Input: “happy new”
• Trigram lookup: happy + new → year ✓

This ensures a prediction is always returned — never blank!

The Training Data

Total N-grams built:

• 100,501 unique words
• 1,361,416 bigrams
• 2,997,658 trigrams
• 3,707,806 quadgrams

Data was lowercased, punctuation removed and symbols stripped before training

Using The App

3 simple steps:

1. Go to the app link below
2. Type any word or phrase in the text box
3. Click any predicted word to append it and keep typing!

Live demo results:

Summary & Links

What was built:

• Full EDA on 4 million lines of real English text
• N-gram model with nearly 8 million sequences
• Live Shiny web app with instant predictions

Links:

• App: https://jjvqgo-farhan-bin.shinyapps.io/next-word-predictor/
• RPubs Report: https://rpubs.com/keo/word_prediction

Johns Hopkins Data Science Capstone — Farhan Bin Hossain