📘 Problem Statement

• Typing on mobile devices is slow and error-prone.
• Predictive text suggestions can speed up typing.
• Goal: Build a Next Word Prediction app trained on real-world text (blogs, news, Twitter).
• Deploy it as a Shiny App with a simple UI for public use.

⚙️ Data & Methodology

• Data Sources: en_US corpora (blogs, news, twitter)
• Preprocessing:
  • Lowercasing, punctuation & number removal
  • Stopword & profanity filtering
• N-Gram Models:
  • Built Unigrams, Bigrams, Trigrams
  • Used frequency-based filtering
• Prediction Logic:
  • Simple string matching on last 1–2 words
  • Top 3 next-word suggestions shown

💻 Shiny App Demo

• UI:
  • Text input box
  • ‘Predict’ button
  • Top 3 predictions shown in a table
• Model logic:
  • Look up trigram match on last 2 words
  • If none, fallback to bigram match
• Hosted at: https://khemanta.shinyapps.io/jhu-dsml-capstone

🔗 App Demo
🔗 GitHub Repo

📊 Results, Observations & Next Steps

• ✅ Observations:
  • Profanity filter ensured safe predictions
  • Cleaning reduced noise; realistic phrases emerged
  • Demo predicts well: “thanks for the”, “i love my”
• 🔁 Improvements:
  • Add Stupid Backoff or Kneser-Ney
  • Support for unseen n-grams
  • More UI features (e.g., autocomplete)
• 💡 Why It Matters:
  • Fast, real-world NLP app
  • Foundation for mobile/enterprise NLP tools

🔗 App Demo
🔗 GitHub Repo