Overview

• Goal: Predict the next word from a user-typed phrase
• Built using Shiny for NLP interaction
• Real-time prediction using N-gram models
• Useful for mobile typing suggestions, chatbots, etc.

Data Source & Cleaning

• Dataset: HC Corpora (Blogs, News, Twitter)
• Steps:
  • Removed punctuation, numbers, stopwords
  • Converted to lowercase
  • Tokenized into N-grams (uni-, bi-, tri-)
• Sampled data (~5%) for efficiency

Prediction Algorithm

• Built N-gram language models
• Used Stupid Backoff strategy:
  • Try trigram → if not found, backoff to bigram → then unigram
• Fast, simple, works well with sparse data
• Implemented in R with dplyr, stringr, tidytext

The Shiny App

• Input: user types a phrase (≥1 word)
• Output: app predicts and shows next word
• Hosted at: https://shreyashk.shinyapps.io/project/
• Lightweight UI, real-time results

Reflection

• Works well for many common phrases
• Could improve with:
  • Smarter backoff (e.g. Kneser-Ney)
  • Contextual models (e.g. RNN, transformers)
• Great experience building full NLP pipeline!