Next Word Predictor

A capstone project done for the Coursera Johns Hopkins Data Science Specialisation

Overview

Goal of this project

• To build a Natural Language Processing (NLP) model that takes in an input from the user, which can be a word, a phrase or an incomplete sentence.

What does it do?

• The algorithm will use the inputs given and predict what the next top matching word should be based on the model that its being trained on.
• The algorithm will also predict the next 3 best alternative words that should come after the input from the user.
• Users can start entering an input in the message box and wait for ~3-5 seconds to see its effect!

Examples

Input a word

Input a phrase

Input a sentence

Training Data

• Corpora from SwiftKey.
• Consist of text crawled from twitter, blogs and news in 4 languages (US, DE, FI and RU).
• As a starter, we will only build our model with the English dataset.

Algorithm

• Katz’s back-off Model.
• An n-gram language model that estimates the conditional probabilty of a word given its history in the n-gram.
• In the app’s context, we adopted a quad-gram (4 words) language model and use it to predict the last 3 words that the user have input.
• If there are less than 3 words, then it will based enitrely on what the user have input.

Brief Project Pipeline

• Web Scraping to create the raw datasets (already done by SwiftKey).
• Process the raw datasets by cleaning the dataset.
• Do exploratory data analysis to get insights from the datasets.
• Do ngram analysis and determine which combination of words happen more frequently than the other combinations.
• Apply Katz’s back-off algorithm to build the model.
• Set up the web application using Shiny.
• Deploy the application on shinyapps.io.

Software Used

Code

• R: Version 3.6.1
• RStudio: Version 1.1.456
• Shiny

Deployment

• shinyapps.io

Files

SlideDeck.Rmd

• The slides to pitch the idea of Next Word Predictor.

./next-word-predictor/app.R

• Entry point of the app.

./next-word-predictor/predict.R

• Helper function for server.R.

./next-word-predictor/ui.R

• User interface of the webapp.

./next-word-predictor/server.R

• The logic that make the webapp work.

./next-word-predictor/components/EDA.Rmd

• Basic Exploratory Data Analysis before starting on the webapp.

./next-word-predictor/components/about.html

• Description of the "About the Developer" tab in the webapp.

./next-word-predictor/components/descriptions.html

• Description of the "Brief Description" tab in the webapp.

./next-word-predictor/components/instructions.html

• Description of the side tab in the webapp.

./next-word-predictor/components/references.html

• Description of the "References" tab in the webapp.

Web App and Pitching Slide

Next Word Predictor App
Slide Deck