View public_index.html for a sample.
These chains are generated with two independent dictionaries of 2-grams and 3-grams. Sentences are split into words and these words are used to index the following words. For instance, the following sentences:
to be or not to be that is the question
...would be recategorized as:
to: ['be', 'be']
be: ['or']
or: ['not']
not: ['to']
that: ['is']
...etc
Note that the duplicated be's are intentional; a random index in the to array
can then be used to pick a next word, weighted for probability.
3-grams are indexed as:
'to be': ['or']
'be or': ['not']
...etc
and so on.
It's horribly inefficient, but... So sue me.
More information here.