I developed the algorithm by creating n-grams (combinations of n words) found within the training text, e.g. 'I live in'. I initially planned to do this on the basis of 2, 3 and 4 grams but for reasons of performance within Shiny, limited it to 2 and 3 grams. I also used a 20% sample of each of the 3 files (Twitter, News and Blogs).
I used the Quanteda package to create the 2 and 3 grams. For performance reasons, I did this in advance, saving them to files which are included as part of the App. Profanities (not exhaustive), punctuation, hyphens, Twitter handles, symbols and n-grams that only appear once are removed. The latter removes a very high volume, allowing a greater focus on those that appear more consistently.
The App takes the last 1 or 2 words input and uses the Stupid Backoff Method to score entries in the n-gram files and select those with the highest probability. See final slide for an example.
