SwiftKey Next-Word Predictor

author: Igor Grybek date: Prototype evaluation complete

Turn a short phrase into one suggested next word.

Audience: People writing short messages or other text.

Data: English blogs, news, and Twitter posts from the course SwiftKey corpus.

Prototype: A focused Shiny app takes a phrase and returns one continuation.

Why next-word prediction?

• A useful suggestion can help a writer continue a phrase with fewer keystrokes.
• Blogs, news, and Twitter cover different lengths and styles of English text.
• A small, explainable prototype is a practical way to demonstrate language modeling.

Design goal: Keep the interaction simple, respond quickly, and make fallback behavior clear.

How the model works

1. Normalize the phrase into lowercase word tokens.
2. Search for the longest matching recent context, up to three words.
3. If no continuation is available, back off to a shorter context.
4. Return the most frequent observed continuation, or explain that no suggestion is available.

The model uses a deterministic sample: every 200th line from each source, with every 10th selected line held out. It trained on 19,105 lines and evaluated on 2,127 held-out lines.

Prototype results and app

Held-out source	Token-level top-1 accuracy
Blogs	11.8%
News	11.9%
Twitter	11.2%

These are modest sampled prototype estimates, not production benchmarks. The deployed Shiny app was tested in a browser: “I went to the” → “house.” Five corpus examples and their outputs are in model/smoke_test.csv.

Limitations and next steps

• Rare contexts can produce weak suggestions; source balance can affect style.
• The measured accuracy is modest, so the model should not be presented as a production keyboard.
• Evaluate the model on a larger sample and compare it with stronger baselines before making production claims.
• Review the live links and project claims before submitting the course assessment.

The source, model, smoke tests, and sampled evaluation are included with this project archive.