Support asymmetric Dirichlet prior

[hiyijian]

According to Wallach’s paper, asymmetric, hierarchical Dirichlet prior over the document–topic distributions and a symmetric Dirichlet prior over the topic–word distributions results in significantly better model
This PR supports asymmetric alpha in following steps:

1. Add two extra tables to Multiverso. One is topic frequency table, a matrix to count each topics’ frequency. The other one is doc length table, a row to count how many document is with length k.
2. Initialize the two extra tables with random initialized documents
3. Learn alpha distribution with the two extra table every 5 iterations
4. Build alias table for leanred alpha distribution
5. Sample topics with learned alpha distribution and alias table. Meanwhile, update countings of topic frequency table if necessary

To use this new feature, please just run with an extra option "-num_alpha_iterations".

Please notice that there are two TODOs. One is Evaluation in asymmetric prior mode, the other is Inference with asymmetric prior.

[feiga]

Only need to request the table if iter % num_alpha_iteration == 0?

[feiga]

Thanks for the great work!@hiyijian I'm really sorry for my late response. The implementation should be OK. I review the code and add some notes. I think it's OK to merge to the master.

[lisendong]

@hiyijian I'm trying to use asymmetric lda. count you tell me your QQ or Wechat? I have some questions.

[hiyijian]

@lisendong My implemention is on the top of source codes provided by @feiga , which I think is used in Microsoft only. However, my implemention is incorrect, I think. Fell free to concact me via @hiyijian@qq.com

[tangzhenyu]

@hiyijian I'm tring to use your asymmetric prior version of lightlda that based on microsoft's code,I am seeing that you said "my implemention is incorrect",is that meaning your code still had some bug?

[koustuvsinha]

hi @hiyijian @feiga, why is this not merged yet? Are there are issues with the implementation?

[hiyijian]

hi @tangzhenyu and @koustuvsinha. I said "this implemention is incorrect" since I have not seen any improvement compared by feiga's original symmetric LDA, Sometimes seems even worse :( .