What is it?

• Goal: Predict the next word based on user input
• Method: N-gram language model with Stupid Backoff
• Technology: R, Shiny, NLP

Why it matters

• Text messaging: Predictive text
• Email composition: Auto-completion
• Accessibility: Help for users with disabilities

Data source

• SwiftKey en_US dataset (blogs, news, Twitter)
• ~4.2 million lines, ~100 million words

Prediction Strategy (Stupid Backoff)

Step 1: User types a phrase (e.g., “I want to”)

Step 2: Preprocess text

• Convert to lowercase
• Remove punctuation
• Remove extra spaces

Step 3: Predict using N-gram model

• Try 3-gram (last 2 words) → Return if found
• Backoff to 2-gram (last 1 word) → Return if found
• Final fallback → Most common word

Example

• Input: “I want to”
• Predict: “go”, “eat”, “see”, “be”, etc.
• Model selects highest probability word

Features

• Simple Interface: Clean, intuitive design
• Input: Text box for typing phrases
• Prediction: Click “Predict” button
• Output: Displays the most likely next word
• Examples: Click to auto-fill test phrases

How to Use

1. Type a phrase (e.g., “I love to”)
2. Click “Predict Next Word”
3. The app shows the predicted word

Try It!

https://tuenxu666.shinyapps.io/app904/

Dataset Summary

Model Training Process

1. Clean text: Lowercase, remove punctuation
2. Build N-grams: 1-gram, 2-gram, 3-gram models
3. Store frequencies: Data frames for fast lookup
4. Optimize: Lightweight model for Shiny deployment

Key Finding

• “love” appears ~4x more than “hate” in Twitter data