[Mt-list] Log-linear models, global or local optimum?

[david li]

Hi, all

I'm new here, and I've one question on log-linear models for SMT.

Since log-linear model is based on MaxEnt principle, it has the nice
property with global optimum.

However, I remember that during MERT training, one way to avoid local
maximum with log-linear training is to select several starting point
randomly. Then the max value will be chosen.

Is there anything wrong for me and thanks very much!


David

_______________________________________________
Mt-list mailing list