Data Science Capstone

The app is here

Overview

• Predicts English word based on the preceding words
• Each predicted word is assigned a probability score
• The app outputs a list of probable words and a wordcloud

How to use

1. Enter a sentence into the text field (four or more words for better accuracy)
2. Voila!

• The app employs the simple Back-off algorithm (Brants, Popat, Xu, Och, & Dean, 2007)
• The corpus is used to construct a 5-gram model
• The algorithm first look for matching 5-grams in the 5-gram database
• Then recursively backs off to lower gram databases to look for additional matches
• Finally, look for most frequent words in the unigram database.

• The application takes advantage of pre-computation

• i.e. all the probability and scores of all possible word combinations from the corpus are ready to be pulled out, no calculation needed

• Since we only care the predictions with the highest probabilities, we only need to keep a handful of unique n-grams from each database below 5-gram. i.e. these database only contains identical starting features k times (I set k to be 5 in this application, this is completely arbitrary)

• Performance benchmark can be found here