17 June 2026

The problem & the product

Typing on phones is slow. Next-Word Predictor is a web app that guesses the next word as you type — like the suggestion bar on a smartphone keyboard.

  • Input: a phrase typed into a text box (e.g. “the president of the united”).
  • Output: the most likely next word (“states”), plus a couple of alternatives.
  • Built in R + Shiny, hosted on shinyapps.io so anyone can use it from a browser.

The model was trained on a large corpus of English blogs, news articles and tweets (the SwiftKey dataset, ~4 million lines / ~100 million words).

The algorithm: n-gram “stupid back-off”

The model is built from n-grams — sequences of 1, 2 and 3 consecutive words counted in the training text. To predict, it backs off from longer to shorter context:

  1. Look at the last two words → return the most frequent word that followed them (trigram).
  2. If unseen, back off to the last word → most frequent follower (bigram).
  3. If still unseen, fall back to the most common words overall (unigram).
predict_next <- function(phrase, k = 3) {
  # trigram (last 2 words) -> bigram (last word) -> unigram fallback
  ...
}

Tables are pruned (rare n-grams dropped, top-3 kept per prefix), so the whole model is under 1 MB and predicts in milliseconds.

The app: how to use it

Steps

  1. Open the app (link on the last slide).
  2. Type a phrase in the text box.
  3. Press “Predict next word”.
  4. The top prediction appears in large blue text, with other likely words below.

Why it’s nice to use

  • Loads instantly (tiny model).
  • Clean, single-screen layout — no setup.
  • Works on any phrase, always returns an answer.


Example: i can't wait to → see  (also: get, hear)

Performance

  • Trained on: 5% sample of the SwiftKey corpus (~214k lines) — enough for good coverage while keeping the app fast and within the free hosting tier.
  • Model size: < 1 MB across trigram, bigram and unigram tables.
  • Speed: predictions return effectively instantly (keyed data.table lookups).
  • Coverage: every test phrase drawn from news/Twitter returned a sensible prediction.
Phrase (last word removed) Prediction
the president of the united states
happy mothers day
looking forward to the next

Try it & summary