Implementing auto stopping

[ljj7975]

Currently, the training process only uses positive dev/test sets for intermediate evaluation.
We will first need to update the code to use some of negative sets before we can implement proper auto-stopping