How to create NER pipeline with multiple models in Spacy 3.x version?

[Singhal-harsh]

I am loading these two models as follows:

nlp_en = spacy.load('en_core_web_lg')
nlp_xx = spacy.load('xx_ent_wiki_sm')

Now I want to use both of their ner components and have them both in a single pipeline.
Problem is that right now using the below approach:

text = "Jane lives in Brussels. Santos Carlos mora no Brasil"

@Language.component("ner_xx")
def ner_xx(doc):
   doc = nlp_xx(text)
   return doc

nlp_en.add_pipe("ner_xx", after="ner")
doc = nlp_en(text)
print([(ent.text, ent.label_) for ent in doc.ents])

gives me output with just the ner of nlp_xx model and not a combination of both nlp_en and nlp_xx model which is what I wanted.

Please help me correct this, and suggest the best and maybe the easiest way to do this?

Basically I need to use ner components of both the xx model and the en model and then package them as a single model to use in production for my enterprise.

Kindly help urgently.

[adrianeboyd]

You don't need a custom component to do this, you just need to add the ner component from one pipeline to another, which you can do with nlp.add_pipe(source=):

import spacy
nlp_en = spacy.load("en_core_web_sm")
nlp_xx = spacy.load("xx_ent_wiki_sm")
nlp_en.add_pipe("ner", name="ner_xx", source=nlp_xx)
doc = nlp_en(text)
print(doc.ents)  # (Jane, Brussels, Santos Carlos, Brasil)

You have to give the new component a custom name because there's already a pipeline component named ner in the pipeline.

Note that the order of the ner components matters. The second ner component will only add entities where the first component has not added any entities.

There's a demo project tutorials/ner_double that demonstrates this interaction: https://github.com/explosion/projects/tree/v3/tutorials/ner_double

[Singhal-harsh]

Hey I am getting the following erorr when adding the component as you said:

ValueError Traceback (most recent call last)

~\AppData\Local\Temp/ipykernel_2580/3338783104.py in

----> 1 nlp_en.add_pipe("ner", name="ner_xx", source=nlp_xx)

~\AppData\Roaming\Python\Python38\site-packages\spacy\language.py in add_pipe(self, factory_name, name, before, after, first, last, source, config, raw_config, validate)

754             # We're loading the component from a model. After loading the

755             # component, we know its real factory name

756             pipe_component, factory_name = self.create_pipe_from_source(

757                 factory_name, source, name=name

758             )

~\AppData\Roaming\Python\Python38\site-packages\spacy\language.py in create_pipe_from_source(self, source_name, source, name)

685         # TODO: handle errors and mismatches (vectors etc.)

686         if not isinstance(source, self.__class__):

687             raise ValueError(Errors.E945.format(name=source_name, source=type(source)))

688         if not source_name in source.component_names:

689             raise KeyError(

ValueError: [E945] Can't copy pipeline component 'ner' from source. Expected loaded nlp object, but got: <class 'spacy.lang.xx.MultiLanguage'>

Only thing different is I am using the en_core_web_lg model instead of small one, do you think that is causing the above error? if not what can be the reason and how can I resolve it?
Thanks for your help!

[Singhal-harsh]

Can the error be because the vocab is not the same for both of the models? and if yes then how can I make it same?

[adrianeboyd]

That particular error does not look it's related to the vocab or vectors. Can you show the full code you're using? (Please use code blocks with three backticks on a separate line before and after the code: https://docs.github.com/en/get-started/writing-on-github/getting-started-with-writing-and-formatting-on-github/basic-writing-and-formatting-syntax#quoting-code)

Loading with source handles the vocab correctly, but depending which pipeline you source from/into, you might need to add an extra step related to the vectors.

This example works fine for me (with slightly different results):

import spacy
nlp_en = spacy.load("en_core_web_md")
nlp_xx = spacy.load("xx_ent_wiki_sm")
nlp_en.add_pipe("ner", name="ner_xx", source=nlp_xx)
doc = nlp_en("Jane lives in Brussels. Santos Carlos mora no Brasil")
print(doc.ents)  # (Jane, Brussels, Santos Carlos mora, Brasil)

Can you also provide the output of spacy info --markdown?

[Singhal-harsh]

Hey,
So I am using almost the same code, with a change of using large model that is en_core_web_lg, rather than medium as you used.
Following is my code:

import spacy
nlp_en = spacy.load("en_core_web_lg")
nlp_xx = spacy.load("xx_ent_wiki_sm")
nlp_en.add_pipe("ner", name="ner_xx", source=nlp_xx)
doc = nlp_en("Jane lives in Brussels. Santos Carlos mora no Brasil")
print(doc.ents)

When I run the above, I get the error as shown above.

Following is the output of spacy info --markdown:

Info about spaCy

• spaCy version: 3.0.5

• Platform: Windows-10-*****

• Python version: 3.8.0

• Pipelines: en_core_web_lg (3.0.0), pt_core_news_md (3.0.0), xx_ent_wiki_sm (3.0.0)

Please let me know what can be the issue, and what can be done to solve it?

[adrianeboyd]

There have been a number of bug fixes and improvements since v3.0.5, so I would suggest upgrading. I tested this example with v3.2.3 on my end. It looks like the main bug fix related to sourcing from different languages is in v3.0.6, but I see that there have been a few other bug fixes since then, too.