SwiftKey Text Predictor: A Data Science Capstone

In the modern digital landscape, mobile users demand speed and efficiency.

• The Friction Point: Typing on touchscreens is error-prone and time-consuming. “Fat-finger” errors reduce communication velocity.
• The Opportunity: A smart, predictive algorithm can significantly reduce keystrokes.
• The Solution: We have developed a lightweight Natural Language Processing (NLP) engine that predicts user intent in real-time.

Our predictive engine is built on a robust statistical analysis of over 4 million lines of text (Blogs, News, Twitter).

1. N-Gram Hierarchies: We utilize a “Stupid Back-off” strategy for maximum efficiency:

• Quadgrams (4-words): Checks for deep context first.
• Trigrams (3-words): If no match, backs off to a 3-word sequence.
• Bigrams (2-words): Finally, defaults to the most common word pairs.

2. Optimization: The model is pruned to remove rare sequences, ensuring the final data structure is < 2MB, allowing for sub-millisecond query times.

The Shiny Application is designed with a “Mobile-First” philosophy.

• Zero-Latency Interface: The app computes predictions instantly as the user types—no “Submit” button required.
• Clean UI: A minimalist design focuses entirely on the user's text and the predicted result.
• Error Handling: The model gracefully handles unknown words by reverting to the most statistically probable English stop-words.

Key Metric: Average prediction time is < 0.1 seconds.

The application is fully deployed and ready for user testing.

Future Scalability:

• Personalization: Future versions will learn from specific user idiolects.
• Context Awareness: Integration of Recurrent Neural Networks (RNNs) for paragraph-level context.

Experience the App: [https://altamashali.shinyapps.io/swiftkey-predictor/]

Prepared for the Data Science Capstone • Johns Hopkins University