A Pitch for Next Word Predictor Shiny App

N-Gram Next Word Predictor

This is a Shiny app that accepts a user input of a single word or a combination of words and predicts the most probable word(s)
The project was developed using the knowledge on Natural Language Processing (NLP)
The data from a Corpus called HC Corpora was used for buiding the app
In developing this app a few of R packages like tm-text mining, nlp-natural language processing and quanteda- Quantitative Analysis of Textual Data in R were used

A very simplified approach known as “Stupid Backoff” was used for implementation
It uses a four-gram data frame
The four-gram data frame was sorted from highest to lowest frequency.
The input words provided to the app are matched with all the four-grams and the ones matching all the first 3 words, or 2 words are shown on the basis of frequency
If no four-gram is found, the search backs off to tri-grams data frame
Progressively If no tri-gram is found, we back off to bi-gram
If no bi-gram is found, app backs off to uni-gram
If the user provides more than three words input, only the last three are considered as an initial parameter for predicting model function
If the user inputs profanity words, they are cleaned and changed to “*"
By sampling the corpus the file size of data shrunk to 6 Mb
Thus allowing the Shiny app to process and produce a prediction in less than a second

Stanford - Natural Language Processing Natural Language Processing: A Model to Predict a Sequence of Words Text Mining Infrastructure in R (pdf file)
Wikipedia - n-gram Bigrams and Trigrams
Boise State University Wikipedia
Katz’s back-off model Large Language Models in Machine Translation (Stupid Backoff) Smoothing, Interpolation and Backoff
Find my shiny app here