refactor document

[ArneBinder]

This PR cleans up and simplifies the Document.

It contains the following changes:

• annotations and predictions as attributes of Document with type AnnotationCollection to reuse its functionality
• add class AnnotationCollection as mapping from layer names to AnnotationLayers with some helper methods.
• add class AnnotationLayer that derives from List, but with some sanity checks and typed getters (optional)
• add construct_document that allows to easily create a Document from a text, spans and binary_relations (both as mapping from names to iterables of the respective annotations). This was previously sitting in tests.helpers.document_utils, but seems to be useful also in other places than tests.
• fix pipeline test for TransformerRETextClassificationTaskModule: use pytest.approx to compare floats

[ArneBinder]

This should be ready now. @ChristophAlt could you have a look?

[ChristophAlt]

I find this to complicated and it will certainly confuse the user. Just automatically create the layer if it's not present and then there's also no need to make an annotation optional.

[ArneBinder]

Ok, I'm good with removing the create_layer parameter. However, there is the use case where we want to add just an empty layer. Do you think it makes more sense to provide an add_layer method instead of having annotation optional?

[ChristophAlt]

This is very interesting! I like it.

Two remarks:

• the add method is too complicated and should be simplified. I don't see a reason why a layer shouldn't be created automatically if it doesn't exist and adding an annotation to a layer should obviously not make the annotation optional.
• I honestly don't like the .as_{X} -- I know it's optional but let's be honest nobody is going to use it because it's quite unintuitive. I have to think a little bit about this one. Maybe we have to change our approach for how predictions and annotations are managed.

[ArneBinder]

Thanks for the feedback! I addressed your first point above above.

Regarding the .as_{X} getters in AnnotationLayer: I personally use the .as_{X} a lot. It is much better readable then adding cast(List[X], doc.annotations[xxx]) or #type: ignore which is even more annoying (often, the assigned variable has to be typed again in theses cases). Example:

entities = doc.annotations[self.entity_annotation].as_spans

vs.

entities = cast(List[SpanAnnotation], doc.annotations[self.entity_annotation])
# or
entities: List[SpanAnnotation] = doc.annotations[self.entity_annotation]  #type: ignore

If AnnotationLayer is to complicated in your opinion, I would be fine with stripping the internal type checking (every usage of _check_type) away, however, I would really like to have this (optional) functionality.

[ChristophAlt]

It's not about the usage of as_{X} -- it's certainly convenient given the necessity for the casts. The problem is that the casts are necessary in the first place, which might suggest our current approach of document annotations and predictions is flawed.

The same goes for the necessity of adding empty layers to control certain taskmodule behavior -- this doesn't seem the proper way to do it.

This is not a critique of your implementation but highlights some missing pieces in the current implementation.

[ArneBinder]

@ChristophAlt I just implemented another version of this which uses explicit types: #95. I like the other one more, what do you think?

[ArneBinder]

This is discarded in favor of #95.