Next Word App

Capstone Project

Data Science Specialization

How does the app calculates probability of next word?

• It uses language modeling with trigrams and Modified Kneser-Ney Smoothing.

Why the app uses Modified Kneser-Ney Smoothing?

• Because it is one of the best known backoff smoothing methods. Reference: http://www.cs.jhu.edu/~jason/465/PDFSlides/lect05-smoothing.pdf

How does the app deals with unknown trigrams, bigram or unigrams?

• It performs backoff for unknown trigrams to bigrams, bigrams to unigrams, and uses only known unigrams.

The only difference for the modified is that the discount is different for each n-gram.

How to start using?

• Start by typing any text in the input available on the left part of this page.
• Right above the input the app shows three sugestions for next work based on the input text provided.

What are these sugestions for next word?

• They are from LEFT-TO-RIGHT the most probable words you would type based on the input provided.
• Therefore, the FIRST WORD on the LEFT is the MOST PROBABLE WORD you would type based on the input provided.

How is the app so fast to show the results?

• It already loads pre-computed probabilities, so it does not re-calculates every time, just performs a fast lookup.

What is the novel approach here?

• Modified Kneser-Ney Smoothing with backoff combined and super fast results for the suggested next word predictions.

Link for the app: https://ricardosc.shinyapps.io/NextWord/