Various refactorings of the ensemble to enable it to be treated as a regular constituency parser model in the Pipeline and elsewhere, plus a learnable weight matrix which does some minor reweighting of the ensemble models.
More work needed to figure out the best way to merge the ensemble models' predictions - a current limitation is that the trained models tend to be very accurate on the training data, so there needs to be some way to actually give the model errors to learn from
Various refactorings of the ensemble to enable it to be treated as a regular constituency parser model in the Pipeline and elsewhere, plus a learnable weight matrix which does some minor reweighting of the ensemble models.
More work needed to figure out the best way to merge the ensemble models' predictions - a current limitation is that the trained models tend to be very accurate on the training data, so there needs to be some way to actually give the model errors to learn from