SwiftKey Next-Word Predictor
Faster, smarter typing on mobile devices

• Data: Sampled ~40 K lines from en_US blogs, news, Twitter
  
• Preprocessing:
  1. Clean & tokenize (lowercase, strip URLs/punctuation)
     
  2. Build unigram, bigram, trigram counts
     
• Model:
  • Back-off trigram: use last 2 words if seen, else bigram on last word
    
  • Returns top-3 next-word suggestions

• Coverage: ~95 % of real-world contexts found in our n-grams
  
• Top-1 Accuracy: 42 % (exact match)
  
• Top-3 Accuracy: 75 % (correct next-word in suggestions)
  
• Perplexity: 120 on held-out sample

• Try it yourself online!
  https://odysseasaccount.shinyapps.io/next-word-predictor/

• Type a phrase into the text box, click Predict, and see the top-3 suggestions instantly.

1.Personalization: adapt to individual typing habits

2.Multi-Lingual: extend to German, Russian, Finnish

3.Integration: embed into SwiftKey SDK

4.Investment Ask: $200K to productionize and scale to 1M users

Thank you—questions welcome!