phrase alignment using posterior regularization

EM algorithm for phrase alignment training usually leads to degeneracy. How to design a posterior constraints to solve this problem ? If I can design efficient constraints , I can use the PR framework to solve this problem.