add prediction label type for span classifier

[helpmefindaname]

No description provided.

[alanakbik]

This currently does not work. To reproduce:

from flair.data import Sentence
from flair.models import MultitaskModel

# For comparison: This works since the label type is "ner" for both models in the multitask model
classifier: MultitaskModel = MultitaskModel.load("zelda")

sentence = Sentence("Kirk and Spock met on the Enterprise")

classifier.predict(sentence)

print(sentence)

# Giving them sensible label names, now made possible with this PR
classifier.tasks["Task_1"]._label_type = "nel"
classifier.tasks["Task_1"]._span_label_type = "ner"

# However, this no longer makes predictions
sentence = Sentence("Kirk and Spock met on the Enterprise")

classifier.predict(sentence)

print(sentence)

The problem lies in the _filter_data_point method of the DefaultClassifier.

[alanakbik]

@helpmefindaname I've pushed a small fix for the issue. I'll approve now, but can you double beck before you merge?

[helpmefindaname]

@alanakbik good catch, I did a small change so we allow both, setting the span_label_type immediately (but training on label_type if available) and having it set later instead.

[alanakbik]

Thanks @helpmefindaname!