Why this product?

Typing on mobile devices is slow and error-prone.
Predictive text systems improve typing speed by suggesting the next likely word.

Goal:
Build a lightweight, fast, and accurate next-word prediction model using real-world text data.

Training Data

• Blogs
• News articles
• Twitter posts
  (SwiftKey corpus)

Model

• N-gram language model (1–4 grams)
• Text cleaning and normalization
• Frequency-based probability estimation
• Pruning to reduce model size

Backoff Algorithm

The model predicts the next word using: 1. 4-gram match (highest priority) 2. 3-gram backoff 3. 2-gram backoff 4. Unigram fallback

Each level is weighted to balance accuracy and coverage.

This allows predictions even when word sequences are unseen.

Accuracy (held-out test set)

• Top-1 accuracy: ~15–20%
• Top-3 accuracy: ~30–40%

Efficiency

• Pruned n-grams reduce memory usage
• Predictions run in milliseconds
• Suitable for real-time Shiny deployment

How the app works

• User enters text
• Model predicts the top 3 next words
• Results update instantly

🔗 Live App:

Summary

This project demonstrates how statistical language models can be used to build fast, interpretable predictive text systems suitable for real-world applications.