1. Algorithm Description

• Model Used: The model used is N-gram language model (bigram + trigram).
• Training Data: I have cleaned corpus of English tweets, blogs, and news.
• Steps:
  • Input text is tokenized.
  • Algorithm approach is: When last 2 words are matched in trigram data (w1, w2 → w3) it fetches the word with the same.
  • If not found, it fallbacks to bigram (w1 → w2).
  • Else finally it fallback: most frequent word `“the”.
  • For speed, Optimized with pre-saved .RDS` models.
• Workflow:
  • User enters text input (e.g., “I am”).
  • Text gets tokenized and converted to lowercase.
  • Trigram lookup: the trigram lookups works and matches last two words (w1, w2) → w3.
  • If found no match in trigram looks into bigram lookups.
  • *after searching through both trigram and bigram if found no match it returns the most frequent word the.

2. App Description & Instructions

The app provides a simple, user-friendly experience:

• The Word Predict App is built in R Shiny App for web-based usage.

• How to use:

  • User Input: Type any English phrase to test.
  • Action: Click “Predict”.
  • Output: The app displays the most likely next word.

• Snapshot for the User Intrface:

  

3. User Experience

• Simple and Minimal UI: We have tried to keep the UI simple and minimal With one input textbox and one submit button. So, that it would be easier for any user to undersand and use it.
• Quick Predictions: It is optimised in a way with the help of .RDS models to give quick and acurate response.
• Responsiveness: It is compatible with both desktop and mobile browsers, can work effortlessly on both.
• Educational: It can be helpful in understanding how word prediction model works.
• Internally: it uses dplyr to filter and extract word suggestions based on frequency.
• Data models: “bigram_model.RDS” and “trigram_model.RDS” are used to support the prediction algorithm.
• Great entry point for users who are curious about NLP.

4. Predictive Performance

• Evaluation Metrics:
  • Top-1 Accuracy on holdout test set: ~23%
  • Top-3 Accuracy: ~41%
  • Baseline (always guessing “the”): ~6%
• Data Used:
• English corpus (News, Blogs, Twitter).
• Cleaned, tokenized, and frequency sorted.
• Models capped at top 20K rows for deployment efficiency.
• Light-weight, scalable for embedded use cases (mobile, survey tools).
• We can enhance the performance and accuracy by increasing the number of raws used to calculate.

5. Why It’s Awesome (and What’s Next)

• End-to-end product from model → app → deployment:
  • from the beginning of building the prediction model, to creating a working app, to putting it online so anyone can use it. It’s a complete, usable product from start to finish.
• Fully built in R with reproducible pipeline:
  • The entire product is built using R, which keeps costs low and development fast. Our pipeline is fully reproducible - meaning anyone on the team can rerun it and trust the results.
• Open-source, lightweight, and customizable:
  • The app is built with open-source tools, so there are no licensing costs. It’s lightweight, runs quickly, and can be easily customized for different needs or industries.
• Next Steps:
• Add Top-3 predictions: Show the top three most likely next words instead of just one, giving users more choices.
• Personalization by user typing history: Improve accuracy by learning from each user’s writing style and frequently used words.
• Enhancement with Deep Learning (e.g., LSTM or Transformers): Use advanced models to understand context better and make smarter predictions over longer text.