add ensemble-style training to evals?

[Tim Miller]

A few months ago I was able to implement a bare bones ensemble training
using different views for the component classifiers. Then it was pointed
out to me that this is not best practice for ensemble training, because
the labels you're getting from the component classifiers (features for
the ensemble classifier) are not optimal because they are obtained by
running the component systems on the same data they were trained on.
What you really want is component labels from data the component
classifiers haven't seen before along with the gold labels to give the
ensemble method.

This can be done similarly to cross-validation, but I think it's
different enough that it would be hard to jam into Evaluation_ImplBase.
I'm thinking about it but thought I'd put it our there if anyone has
already put any thought into it.

Tim

[Steven Bethard]

Can you nest your Evaluation_ImplBases? That is, inside the train
method, have another Evaluation_ImplBase implementation that, say,
produces the component labels by running cross-validation, and then
uses the cross-validation labels as the input to the ensemble
classifier that is trained next?

Steve