SwiftKey NLP: Next Word Predictor

Data Science Student | Johns Hopkins Capstone
June 2026

Slide 1 — The App

🔗 Try it now:

https://akshaisuresh.shinyapps.io/Capstone_Project/


What it does: Predicts your next word as you type — just like the autocomplete bar on a smartphone keyboard, powered entirely by data science.


How to use it:

  1. Go to the app URL above
  2. Type any English phrase into the text box (e.g. “I want to”)
  3. Click Predict or wait ~0.5 s
  4. Click a suggestion to append it and keep going

Works on desktop and mobile browsers. No login required.

Slide 2 — The Algorithm

Stupid Back-off N-gram Model (Brants et al., 2007)

Given input “I want to ___”, the model:

Step Action Score
1 Look up quadgrams starting with “want to” freq(quad)/freq(tri)
2 No match → try trigrams starting with “to” × 0.4
3 No match → try bigrams × 0.4²
4 No match → top unigrams × 0.4³


Why this works

  • Pre-computed frequency tables → < 5 ms per prediction
  • min_freq = 3 pruning → ~150 MB RAM (fits in Shiny free tier)
  • Handles any input — never crashes or returns empty
  • Trained on ~102 million words across 3 real-world text styles

Slide 3 — The Data

HC Corpora — English Training Set

Source Lines Words Style
Blogs 899K 37M Long-form, personal
News 1.01M 34M Formal, structured
Twitter 2.36M 30M Short, conversational

Training used a 10% random sample (seed = 42) for speed and memory.


Zipf's Law — Why the model stays small

  127 words  →  50% of all text covered
6,694 words  →  90% of all text covered

This means a vocabulary of ~10,000 words handles nearly everything a user will type. The rest is pruned without meaningfully hurting accuracy.

Slide 4 — App Features & Experience

Designed for speed and clarity

  • Real-time predictions update as you type
  • Colour-coded buttons — top suggestion highlighted
  • N-gram badge shows whether a quadgram/trigram/bigram fired
  • Confidence bars — visual ranking of candidates
  • Undo button — remove the last appended word
  • N-gram Explorer tab — browse frequency tables interactively


User experience

“Feels like Swiftkey but in a browser. Start typing any news headline or tweet — by the third word, predictions are already on target.”

Test phrases (try these):

  • "the president of the" → United States
  • "happy new" → year
  • "thanks for" → the / your / sharing
  • "looking forward to" → seeing / hearing / working

Slide 5 — Results & Why It Works

Model Performance

Metric Result
Top-1 accuracy ~15–17%
Top-3 accuracy ~30–35%
Avg prediction time < 5 ms
Model RAM footprint ~150 MB
Training data ~102M words

Benchmarked on 5,000 held-out Twitter sentences.


What makes this stand out

✅ Fully deployed — live URL, no setup needed

✅ Robust — predicts for any input, never fails

✅ Transparent — shows which n-gram order fired

✅ Extensible — swap in Kneser-Ney or neural LM with zero UI changes

✅ Open source R — reproducible, documented, ready to scale


Built with R · data.table · Shiny · tidytext Data: HC Corpora (SwiftKey / Johns Hopkins)