Combine "spacy evaluate" with "bluesearch.mining.eval"

[FrancescoCasalegno]

Currently our evaluation step in the dvc pipeline dedicated to NER models relies entirely on the following script, which calls functions from bluesearch.mining.eval. https://github.com/BlueBrain/Search/blob/fa1331c98c8823ec85c5b3d92d58e99ab6010574/data_and_models/pipelines/ner/eval.py#L1

It could be convenient to use instead the interface provided by spaCy's CLI, i.e. spacy evaluate. In particular, we would like that to potentially call our own evaluation functions (e.g. token-based and entity-based F1, Prec, Rec, ...).

• [ ] does spacy evaluate produce the same results (entity-based F1, Prec, Rec) than our bluesearch.mining.eval functions? if not, why?
• [ ] can we configure spacy evaluate in such a way that the spacy.Scorer calls our own user defined scoring functions?
• [ ] if we then don't need the annotations in JSONL format anymore, then (i) get rid of the preprocessing step and script (ii) use .spacy binaries instead of .jsonl.
• [ ] Simplify params.yaml by removing the use of etype_name (see here for details).

[pafonta]

UPDATE Actions

The section for evaluation in params.yaml could be simplified after #351.

There is indeed no more the need for mapping scispaCy labels to our labels.