Next Word Predictor App

Typing is one of the most common ways people interact with digital products. A useful next-word predictor can reduce the number of keystrokes required to complete a message and make text entry faster and more convenient.

This product predicts the next word after a user types a phrase.

It is designed for fast deployment in Shiny while still using more context than a simple n-gram model.

Context-aware hybrid n-gram model

The app combines five signals:

• 4-gram probability (4 words arrange)
• trigram probability (3 words arrange)
• bigram probability (2 words arrange)
• unigram frequency
• broader phrase-context score

Instead of relying only on the last one to three words, the model also considers recent earlier words in the phrase. This helps reduce weak fallback predictions when an exact n-gram is missing.

The final score is a weighted ranking score used to select the best next-word candidate.

Accuracy, speed, and model size

The final model was tuned to balance accuracy and deployment efficiency.

Key evaluation metrics used:

• Top-1, Top-2, and Top-3 accuracy on held-out text
• Median prediction time
• Model size in memory
• Number of retained n-gram parameters

The model uses pruning to remove rare n-grams and reduce memory usage. This sacrifices a small amount of accuracy but makes the application faster and easier to deploy on shinyapps.io.

Simple user experience

The Shiny app provides a simple multilingual workflow:

1. Select English, German, Finnish, or Russian from the permanent language panel.
2. Build or load the selected language model and wait for the ready message.
3. Type or paste a phrase and press Predict next word.
4. The app displays the best next-word prediction and a word cloud of other likely candidates.

The app also includes tabs for language-specific corpus summaries and model performance, including Top-1, Top-2, and Top-3 accuracy.

Small, fast, and explainable

This product is useful because it is:

• Fast: predictions are generated with compact lookup tables.
• Explainable: the model is based on interpretable n-gram probabilities and phrase-context scoring.
• Deployable: the trained model is saved as an RDS file and loaded directly by Shiny.
• Practical: the app demonstrates a real end-to-end NLP product from raw text to deployment.

The next improvement would be tuning the model weights and sample size further to improve context-sensitive predictions without making the app too large.