Typing on phones? It’s a pain. Let’s fix that.
The problem: Typing on mobile devices is slow and frustrating.
- People type 40% slower on phones than on keyboards
- Predictive text saves time and reduces typos
- Existing solutions are often heavy or opaque
Our solution: A fast, transparent N-gram predictor.
- Runs on a 4 MB model (loads in under 1 second)
- Returns predictions in under 50 milliseconds
- Shows the top 5 candidates with full transparency
Under the hood: a clever N-gram engine
We use an N-gram language model with Katz-style backoff:
- 4-gram → uses the last 3 words
e.g. “I love to” → predicts “eat”
- 3-gram → falls back to the last 2 words
- 2-gram → falls back to the last 1 word
- Unigram → final fallback (most frequent word)
Training data: 10% sample of English blogs, news, and tweets
(more than 4 million lines, 10 million words).
Key trick: Pruning n-grams with frequency < 3 keeps the model
small enough for a free Shiny server.
Try it live – it’s fast!
Play with the app: https://adlinehaha.shinyapps.io/NextWordApp/
![]()
Shiny App Screenshot
Numbers that matter:
- Response time: under 50 milliseconds per prediction
- Top-5 accuracy: about 60% on test phrases
- Model size: 4 MB, loads in under 1 second
- Coverage: 107 words cover 50% of all word occurrences
Why you’ll love it (and what’s next)
Why it’s great:
- Fast: Sub-second loading on a free server
- Transparent: Shows all 5 candidates, not just one
- Extensible: Can be upgraded to RNNs or transformers later
- Practical: Ready for keyboards, autocomplete, writing assistants
Next steps:
- Train on full corpus for higher accuracy
- Add smartphone keyboard integration
- Explore deep learning (LSTM, BERT) for context-aware prediction
Thank you! Questions?
