Pipeline with 2 NER: One trained, one from en_core_web_md

[oliviercwa]

Hello,
I am creating a pipeline with 2 NER:

1. the first one is Transformer-based and trained with my own dataset,
2. the second one is coming from Spacy en_core_web_md (component name = 'en-core-web-md-ner')

I am only training the first one but I want my final pipeline to have both: The trained transformer AND the en_core_web_md NER.

I don't have any problem with the first one, however for the second one, I need the model to be fully embedded as the en_core_web_md can't rely on the tensor created from the the first model.

Here is an excerpt from the config with the en-core-web-md-ner definition

[components.en-core-web-md-ner]
source = "en_core_web_md"
component = "ner"

frozen = ['en-core-web-md-ner']

However when I train my transformer, I am getting a UserWarning W113:

Sourced component 'en-core-web-md-ner' may not work as expected: source vectors are not identical to current pipeline vectors.

How can I tell the config to use en_core_web_md as source AND copy all embeddings (tok2vec) AND models from the source AND ignore the ones from the Transformer ?

Note #1 that I have also tried the following with the same outcome.

[components.en-core-web-md-ner]
source = "en_core_web_md"
component = "ner"

[components.en-core-web-md-ner.model]
@architectures = "spacy.TransitionBasedParser.v1"

[components.en-core-web-md-ner.model.tok2vec]
@architectures = "spacy.Tok2Vec.v2"

[components.en-core-web-md-ner.model.tok2vec.embed]
@architectures = "spacy.MultiHashEmbed.v2"

[components.en-core-web-md-ner.model.tok2vec.encode]
@architectures = "spacy.MaxoutWindowEncoder.v2"

Note #2: I have also tried to train first my transformer, load the resulting pipeline (my_pipeline) and add the en-core-web-md via APIs

    source_nlp = spacy.load('en_core_web_md')
    component = my_pipeline.add_pipe('ner', source=source_nlp)

When I do this, I am getting the following warning

Sourced component 'ner' may not work as expected: source vectors are not identical to current pipeline vectors.

[adrianeboyd]

You want to source the component (with no additional settings beyond changing its name) and initialize the vectors from the same model:

[components.en-core-web-md-ner]
source = "en_core_web_md"
component = "ner"

[initialize]
vectors = "en_core_web_md"

A pipeline can only contain one set of vectors, but it's no problem to have both a transformer that doesn't use vectors and a tok2vec that does.

[oliviercwa]

Thanks Adriane. Regarding your comment 'A pipeline can only contain one set of vectors'. I understand the the doc has only one property vector. However are you saying that there is no way within the same pipeline to have:

• a transformer that uses its own embeddings (e.g. from HuggingFace transformer wrapper)
• AND a pretrained NER from "en_core_web_md" with its own embeddings ?

I was under the impression that by specifying in the configuration that the tok2vec model was not shared, the "en_core_web_md" NER component would "embed" them in its own layer, thus making it entirely independent from the rest of the pipeline. Essentially, ending up with 2 set of vectors:

• 1 set from the transformer that are also accessible using doc.vector
• 1 set "embedded" into the "en_core_web_md" NER model, which can't be accessed.

I guess I am missing something ?

Thanks

[adrianeboyd]

By "vectors" I mean the static word vectors that are stored in nlp.vocab.vectors. The Tok2Vec architecture uses these vectors with include_static_vectors = true while the transformer component doesn't. The vectors aren't saved within the ner component, which is why you have to add them separately with [initialize.vectors]. Since they're usually large and shared by multiple components, they're in one central place nlp.vocab.vectors rather than duplicated in each component.