[WIP] Integrate predictions

[ArneBinder]

This implements AnnotationList.integrate_predictions(overwrite: bool = False) which allows to move predictions to the main annotations.

[ChristophAlt]

Sorry, this is clearly not how this is supposed to be handled. Annotations are groundtruth and predictions are predictions. There should be no need to merge one into the other.

[ArneBinder]

OK, I can also implement that in my project in a specific serializer.
The background was that Brat does not now the concept of predictions, it has just annotations, and with that PR I could just use my previous serialization scripts without much modification. But it is not so much important, sorry for the noise.

[ChristophAlt]

The problem isn't so much to use the predictions as annotations --which we need for multi-step pipelines anyway-- but to deliberately overwrite annotations with predictions. This process should be nondestructive, because there may be actual annotations which then get deleted. Instead one could have a context manager that temporarily maps predictions to annotations without actually deleting or overwriting anything.

with document.predictions_as_annotations(["entities", ...]):
  # do something with document (where document.X now points to document.X.predictions)

or

with document.entities.predictions_as_annotations():
  # do something with document (where document.entities now points to document.entities.predictions)

[ArneBinder]

OK, this sounds really good! Unfortunately, I don't have experience in writing context managers. So, if you want, you can take this over.