WIP: Fix evaluation of trained tagger

This fixes bugs which made the evaluate_model.py script unusable for evaluating trained taggers.

Maybe we should also add an integration test for this?