Predictive Text - Exploratory Analysis
Kris
2026-06-20
Overview
This report explores a large corpus of English text from blogs, news articles, and Twitter to build the foundation of a predictive text application — similar to the autocomplete feature on a smartphone keyboard.
The Data
The dataset comes from SwiftKey and contains three sources of English text. A 5% random sample was used for this analysis.
| Source | Lines | Words | Size_MB |
|---|---|---|---|
| Blogs | 899288 | 37334131 | 267.8 |
| News | 1010206 | 34371031 | 269.8 |
| 2360148 | 30373543 | 334.5 |
Most Frequent Words
Stopwords (the, a, is) are kept intentionally — they are critical for predicting natural language sequences.
Most Frequent Bigrams
Most Frequent Trigrams
Prediction Algorithm Plan
The prediction model uses a Stupid Backoff approach with n-grams:
- Given the last 3 words, search for a matching 4-gram
- If not found, back off to the last 2 words and search trigrams
- If not found, back off to the last word and search bigrams
- If nothing matches, return the most common words overall
This handles unseen word combinations gracefully without assigning zero probability to any input.
Shiny App Plan
The final app will:
- Accept user text input in real time
- Predict the next 3 most likely words instantly
- Run efficiently within Shiny’s memory limits by keeping only n-grams that appear 2 or more times in the corpus