Probably because they trained an LSTM model to learn conditional probabilities between words and then sampled using markov methods to actually generate the sequences.

It's actually pretty standard and not particularly innovative, technically, in that you can google already existing tutorials on exactly how to do that (even in Keras - forget lower level stuff).

At any rate, this doesn't work well, typically. I'm sure they generated a bunch of sample sequences and cherry picked ones that made even a little sense for inclusion in their script.