This is Batch LDA (balda) implementation of the algorithm as described in

http://www.cs.umass.edu/~mimno/papers/fast-topic-model.pdf

This algorithm is based on Gibbs sampling, and runs a lot faster (typically by the order of magnitude) 
than 'conventional' samplers.

In our implementation, we also evaluate in-sample perplexity and use it as stopping criteria
for simulation. By default, the execution stops once changes in perplexity are under 5e-4.

To run the sampler, first, a dictionary  of terms  must be defined. Two classes - Words and TokenExtractor - are 
responsible for building the dictionary from the input documents. We remove non-alpha numeric characters and some
stop words from the text and then select words with the highest TF/IDF scores. 

The documents (bags of words) are sent to  SparseGibbsSampler, and it runs the simulation for a specific 
number of iterations, or until the stopping criteria (based on perplexity estimate) is reached. 

There is also a demo sample, please take a look at instructions in 'demo'  directory.

Have fun!