Predict the Next Word from a Given String

• Collaboration between Johns Hopkins University and SwiftKey
• Objective: Build a functioning predictive text model
• Data: HC Corpora (English only)

• Sampled 1,000,000 lines from Twitter, Blogs, and News datasets
• Cleaned data by:
  • Removing non-ASCII characters (emojis)
  • Converting to lowercase
  • Removing contractions, punctuation, numbers, profanities, extra whitespaces
• Tokenized data to create MLE n-grams (up to 6-grams)

• Built Maximum Likelihood Estimation (MLE) matrices
• Used Back-off model for prediction
• Output: Top 3 predicted words for user input
• Accuracy enhanced by showing multiple predictions rather than 1

• Hosted at: Capstone Prediction App
• Features:
  • Clickable predicted words to append to input
  • Instant predictions
• UI preview: