libowen2121/vi-dependency-syntax

Dependency Grammar Induction

vi-dependency-syntax's Issues

Some questions about the paper and the code

Really appreciate it for your work.

There are three questions that make me confused:

1. In the paper, the eq.12 is
   
   and the here is derived by optimizing
   .
   But it seems both in the code and the following part of the paper the is treated with:
   
   i.e. without optimizing the to become first, and then use it to do the M-step (you seem to optimize , and jointly rather than first optimize and then optimize and ).

2. In the Supplementary Material, the revised EM algorithm seems to optimize
   
   And after applying the policy gradient formula, it becomes
   
   
   
   
   This is different from the eq.16
   

3. In the Supplementary Material, the Algorithm 1 and 2 seem to miss the factor in the score:
   

So is there anything that I understand in a wrong way? Thanks a lot and looking forward to your reply.