Is the multitask tagger still predicts supertags?

[oroszgy]

Hi!

I am wondering whether this description of the tagger model is still true. I am particularly interested in the supertag part, as training configuration facilities seemingly contradicts with this approach.

Could anyone give me some guidance here?

[adrianeboyd]

The tagger doesn't currently predict supertags. There is support for multi-task learning with a shared tok2vec or transformer component as described here: https://spacy.io/usage/embeddings-transformers#embedding-layers

The provided trained non-trf pipelines have a shared tok2vec for all the syntactic components (tagger, morphologizer, parser) and the ner component is separate. For the trf pipelines, the transformer component is shared for all trainable components. A few more details: https://spacy.io/models#design

[oroszgy]

I see, thanks for the quick answer!