CapStone - NLP App

“NLP” based on “Stupid Backoff Prediction” [1].

Publicly available blog, twitter, & news data used.

Project information available from Coursera web site.

“SwiftKey.zip” available from the link under References.

There are three raw data files [2]:

en_US.blogs.txt

en_US.news.txt

en_US.twitter.txt

Prediction Algorithm

Stupid Backoff Prediction Methodology (SBPM):

1. Input text word or phrase truncate to four words.

2. Possible matches determined using “SBPM”.

3. Attempt to match input in pentagram, backing off to tetra-, tri-, and bigrams as required.

4. If not found in bigram return the top 10 most frequently used unigram words.

App Screenshot:

Application available at: https://rforreal.shinyapps.io/shiny_app/

1. Model shows acceptable predictive capability, but not great.

2. Good one and two word phrase predictions (found in top 5).

3. Three, and larger word phrases, predicts with poor accuracy.

4. Positively, the model response is quick.

5. Model consistently “backs off” through ngrams.

6. “Stupid Backoff” approach, as implemented, needs attention.

References:

[1] Milestone report and source code can be found at: https://github.com/djduvenhage/CapStone

[2] Brants T. et al., Large Language Models in …, http://www.aclweb.org/anthology/D07-1090.pdf

[3] https://www.coursera.org/learn/data-science-project/supplement/Iimbd/task-0-understanding-the-problem