Is it possible to train two different NERs sourced from en_core_web_md ?

[oliviercwa]

Something like:

pipeline = ["tok2vec", "ner1", "ner2]

[components.ner1]
source = "en_core_web_md"

[components.ner2]
source = "en_core_web_md"
...

This does not work as is and I could not find a way to indicate that the "source name" in "en_core_web_md" is "ner"

Ideally we could do something like

pipeline = ["tok2vec", "ner1", "ner2]

[components.ner1]
source_model = "en_core_web_md"
source_name = "ner"

[components.ner2]
source_model = "en_core_web_md"
source_name = "ner"
...

Is there a clean way to achieve this ?

[adrianeboyd]

It's:

[components.ner2]
source = "en_core_web_md"
component = "ner"

Edit to fix both config settings, not just the component.

[oliviercwa]

Unfortunately this causes an exception when training.

I have a single NER components defined

[components.ner2]
source = "en_core_web_md"
component = "ner"

Causes the following exception

  File "spacy\spacy\training\loop.py", line 122, in train
    raise e
  File "spacy\spacy\training\loop.py", line 112, in train
    log_step(info if is_best_checkpoint is not None else None)
  File "spacy\spacy\training\loggers.py", line 63, in log_step
    for pipe_name in logged_pipes
  File "spacy\spacy\training\loggers.py", line 63, in <listcomp>
    for pipe_name in logged_pipes
KeyError: 'ner2'

[adrianeboyd]

That does look like a bug, we will take a look.

However, it doesn't really make sense to train two ner components in the same pipeline at the same time, since you can only have one source of NER annotations in your training corpus.

There's a slight possibility that the second ner component might make a few additional annotations because it will be starting from a different Doc state when the pipeline is run (existing entity spans in the Doc), but this would be an odd/inefficient setup overall.