Ever wanted to sound like your favorite celebrity, politician, or just nonsensically quote a movie? Well you can now generate contextually appropriate sentences from any corpus!

Markov Sentence Generator uses Markov chains and syntax parsing to choose contextually appropriate words and form them into sentences. Based on a seed that you put into the generator, the model can generate a sentence from a document or set of documents.

The module is composed of four main classes: SyntaxTree, SyntaxChain, MarkovDict and MarkovChain. SyntaxTree creates the syntax parsing, MarkovDict creates the contextual dictionaries from the syntactic elements that MarkovChain puts all that together to form sentences.

(SyntaxChain runs an even more naive model, it takes sentence grammar structures and fills in words that fit that grammar structure ignoring other context. It can get pretty silly.)

MarkovDict can be tuned using the chain_len parameter. Chain length is the number of words in your context. The longer the chain length, the more context you force. Longer chain lengths are more deterministic (more similar to the original text).

MarkovChain then offers you a choice, a naive version of the chain will not take into account syntax parsing and only give context based on word presence. A syntax version will take syntax into account.

MarkovChain has two tuning parameters that are put into the run method. Key Gram Size determines how much of the contextual phrase to pull from. This is a bit complex. If you set MarkovDict.gram_size to a small number, then context will be small and many possible values will follow each key. Key Gram Size determines how much of the context being searched for to put in the return text. Value Gram Size looks at the dictionary for that context and pulls a list of possible following phrases. It chooses one of those phrases and takes Value Gram Size words from it. The new key is now the Chain Length final words of the return text.

Run code.markov_chain to get a sense of how it works. Syntax model takes about 2 minutes to fit on the example text of 302KB.

You can also use the exploratory ipython notebook to get a feel for some of the properties.

I find the best parameters are gram_size=3, gtype=syntax_pos, key_gram_size=1, value_gram_size=2. It can get sillier from there so it depends on what you are going for.

If you want to play with it locally run:

python _code.markov_chain.py

I find its easier to play with the app.

python application.py

Finally the TrueCase class looks at a corpus and then truecases sentences.