Implement a interface for external datasets.

[rajarshd]

Currently, it is hard to run experiments with new datasets, as the interface to create multi_corpus object is missing. This makes starting up with the code very difficult.

Originally reported by @manzilzaheer.

[shehzaadzd]

Use protocol buffers. Suggested and told to comment by Manzil Zaheer (http://www.manzil.ml/)

[yeliu918]

How can you get the correct_paras.json? Do you deal with the whole Wikipedia dataset? Can you provide the preprocessing of data?