Model is predicting extra labels?

[siennathesane]

I noticed the pre-trained pipelines will predict labels which aren't explicitly mentioned in the models doc page, and aren't referenced in the models themselves.

Here's a reproducible example with JSON output:

import spacy
from json import dumps

models = ["sm", "md", "lg", "trf"]
text = "The third oldest French bridge."
all_known_labels = {}

for model in models:
    model = "en_core_web_{0}".format(model)
    nlp = spacy.load(model)
    all_known_labels[model] = {}
    all_known_labels[model]["labels"] = [{pipe[0]: list(nlp.get_pipe(pipe[0]).labels)} for pipe in nlp.components]
    all_known_labels[model]["parsed"] = nlp(text).to_json()

print(dumps(all_known_labels, indent=2))

The model's pipelines aren't aware/configured/trained on the ADJ label, but all 4 models are predicting old (oldest) and French (French) as ADJ parts of speech. The parser pipeline component is correctly marking them as amod dependencies, but none of any of the model's have any reference to the ADJ label.

I'm working on entity extraction, so adjectives describe the properties of a thing, and I noticed the model predicting things it didn't have labels for. I'm trying to get a list of all possible labels the models support. Is this a known issue? Am I doing something wrong?

Edit: I am using the CPU models on Apple Silicon with v3.5.1

[adrianeboyd]

It's true that the labels aren't all on the model docs page. It's a bit complicated because the pipelines include a mixture of statistical and rule-based components, and they don't all expose labels in the same way.

Are you looking for information about the English pipelines in particular or more generally?

For statistical components: the values in Pipe.labels are internal labels for the statistical models and may not be the exact labels that are saved on tokens or spans. You don't see this much with the components included in the current English pipelines, but more for components like the morphologizer or trainable lemmatizer.

For rule-based components like attribute_ruler and lemmatizer: these components don't store an internal set of labels (plus lemmatizers can generate new lemmas for novel words, so there's no way to store a list).

The English POS tags come from a tag+dep->pos mapping in the attribute_ruler component.

Unlike all the other attributes, token.pos has a hard-coded list of supported labels (SPACE and EOL are not standard UPOS tags. SPACE is used for whitespace tokens, EOL is unused):

spaCy/spacy/parts_of_speech.pxd

Lines 5 to 24 in 28de857

	ADJ = symbols.ADJ
	ADP
	ADV
	AUX
	CONJ
	CCONJ # U20
	DET
	INTJ
	NOUN
	NUM
	PART
	PRON
	PROPN
	PUNCT
	SCONJ
	SYM
	VERB
	X
	EOL
	SPACE

[siennathesane]

I'm looking for more information in general, I'm starting with English since that's my native language. All of this information is really helpful, thank you!