Next Word Prediction App

A lightweight Shiny product that predicts the next word after a user enters an English phrase.

The goal is practical: fast, simple, and understandable enough for a non-technical user.

The model is built from the Coursera SwiftKey English corpus:

• Blogs: longer informal writing
• News: formal edited writing
• Twitter: short conversational text

Preprocessing steps:

• Convert text to lowercase
• Remove URLs, numbers, and punctuation noise
• Normalize whitespace
• Tokenize text into word sequences
• Build n-gram frequency tables

The app uses an n-gram back-off model.

Example input:

I am going to

The algorithm checks contexts from longest to shortest:

1. am going to -> likely next word
2. going to -> likely next word
3. to -> likely next word

This gives a good balance between coverage and speed.

The app is intentionally simple:

• User types a phrase into a text input box
• User clicks Predict next word
• The app returns one predicted word

The app does not require statistical knowledge. It behaves like a small predictive keyboard prototype.

Typical examples:

• I love -> you
• new york -> city
• thank you for -> the

Why this approach works:

• Compact enough for shinyapps.io
• Fast enough for interactive use
• Easy to explain and maintain
• Can improve by training on larger samples

Next improvements:

• Tune n-gram pruning thresholds
• Add profanity filtering
• Evaluate accuracy on held-out text
• Display top three predictions instead of only one

This product demonstrates a working data product pipeline: corpus processing, model compression, prediction logic, and a public Shiny interface.