trf French model and NER

[theognis0110]

Hi,

the fr_dep_news_trf model has a better accuracy for dependency parsing but on the other hand it doesn't include a NER. Is it mandatory to use a different model such as fr_core_news_lg to get a NER ?
i attempted to add a ner to the trf but it doesn't seem to be possible or maybe i miss something about it ?

another question about the trf : i am only interested in the tagger and the dependency parse results; in case i can combine the trf model with a ner, can i just remove the transformer from the pipeline ? (at this point i don't intend to train anything)

[polm]

The components in the pretrained pipelines rely on either a tok2vec or transformer component to tranform the raw input into vector representations. The components are conditioned on the upstream component they come with, so you can't just swap that out.

i attempted to add a ner to the trf but it doesn't seem to be possible or maybe i miss something about it ?

You can't copy an NER component from another pipeline and have it work as-is unless you also include the tok2vec it relies on. See the section on sharing embedding layers in the docs for some more info on this.

can i just remove the transformer from the pipeline ?

No, the transformer creates the input for the other components in the trf pipeline, so it's required.