Next Word Prediction App

author: Lukhanyiso Bavuma date: 2025-10-15 autosize: true transition: rotate

Data Science Capstone Project
Coursera | Johns Hopkins University

Smart Text Prediction for Mobile Typing

======================================================== type: section

The Problem

Mobile typing is slow and error-prone

• 📱 Small screens make typing difficult
• ⌨️ Users want faster text input
• 🎯 Predictive keyboards help but need better algorithms

Our Solution:

Build an intelligent next-word prediction model using natural language processing and n-gram statistical modeling.

Similar technology powers: SwiftKey, Google Keyboard, iPhone QuickType

The Data & Approach

Dataset

• Source: HC Corpora - English text from blogs, news, and Twitter
• Size: ~4 million lines, ~100 million words
• Sample: Used 5% random sample for training

Methodology

1. Text Preprocessing
   • Lowercase conversion, remove punctuation/numbers
   • Tokenization into words
2. N-gram Generation
   • Unigrams (single words): ~50K unique
   • Bigrams (2-word phrases): ~100K unique
   • Trigrams (3-word phrases): ~100K unique

The Algorithm

Stupid Backoff with N-gram Model

Input: "I want to"
   ↓
Step 1: Search Trigrams ("want to" → ?)
   → Found: "go", "be", "see"
   ↓
Step 2: If not found, backoff to Bigrams ("to" → ?)
   ↓
Step 3: If still not found, use Unigrams
   → Most frequent words: "the", "a", "and"

Why This Works

• Fast: O(1) lookup with hash tables
• Accurate: Covers 90%+ of predictions with trigrams
• Robust: Always provides predictions via backoff
• Efficient: ~50MB model size

The App

left: 40%

Features

✨ Real-time Prediction - Type any phrase - Get instant suggestions - See top 3 predictions

📊 Transparent - Shows prediction source - Displays confidence levels

🚀 Fast & Responsive - < 0.5 second response time - Works on any device

---

App Screenshot

Try it: shinyapps.io/yourapp

Results & Future Work

Performance Metrics

Metric	Value
Accuracy (Top-1)	35%
Accuracy (Top-3)	65%
Avg Response Time	0.3 seconds
Model Size	48 MB

Key Achievements

✓ Successfully predicts next word in real-time
✓ Handles diverse inputs (news, social media, blogs)
✓ Efficient enough for mobile deployment

Future Improvements

• 🔮 Add 4-grams and 5-grams for better context
• 🧠 Implement neural language models (LSTM/GPT)
• 🌍 Multi-language support
• 👤 Personalization based on user history

---

Thank you! Questions?

View the code: github.com/yourname/capstone
Try the app: shinyapps.io/yourapp