Spacy: stop words missing for en_core_web_md and en_core_web_lg for spaCy v2.0

[Gauravtolani]

_en_core_web_md_ and _en_core_web_lg_ models are giving 'False' for all words in the sentence using "is_stop" attribute.

PS : _en_core_web_sm_ is working fine.

System Information :

• Python version: 2.7.12
• Platform: Linux-4.10.0-38-generic-x86_64-with-Ubuntu-16.04-xenial
• spaCy version: 2.0.2
• Models: en

[ines]

Thanks for the report and sorry about that – this should be fixed in the next update to the models.

In the meantime, here's a workaround:

nlp = spacy.load('en_core_web_lg')

for word in nlp.Defaults.stop_words:
    lex = nlp.vocab[word]
    lex.is_stop = True

[kbulygin]

With en_core_web_sm (spaCy 2.0.4), is_stop depends on casing:

>>> import spacy
>>> nlp = spacy.load('en_core_web_sm')
>>> [nlp(s)[0].is_stop for s in 'this This THIS tHIS the The THE tHE'.split()]
[True, False, False, False, True, False, False, False]
# Expected [True, True, False/True, False/True, True, True, False/True, False/True].

Info about spaCy

• spaCy version: 2.0.4
• Platform: Linux-3.16.0-4-amd64-x86_64-with-debian-8.8
• Python version: 3.6.0
• Models: en_core_web_lg, en_vectors_web_lg, en_core_web_sm

en_core_web_sm for spaCy 2.0.0a10 correctly returned t.is_stop == True for both this and This.

[brickpattern]

Bump. I'm facing the same using en_core_web_sm.
Is this the expected output for the is_stop ? Or should we be using a different approach?

[honnibal]

I know this has been an issue a long time --- the delay comes back to some infrastructure problems, which have made getting all the models retrained a hassle.

The following fix to the spacy train CLI command should make sure the issue doesn't reoccur: https://github.com/explosion/spaCy/commit/262d0a3148c9840b6df58ee955181c1cd486f8b1

The models for v2.1.0 are currently training, so fingers crossed updated models should be deployed soon.

[adrianog]

Hi

Is this still the way to go:

nlp = spacy.load('en_core_web_lg')

for word in nlp.Defaults.stop_words:
    lex = nlp.vocab[word]
    lex.is_stop = True

[randomsven]

This still leaves the is_stop property sensitive to case (ie "What" vs "what") - sounds like the fix needs to be applied upstream - as it is this is simple enough to handle outside of the token attribute system.

[rajhans]

Any updates or ETA on this?

[nathanwdavis]

Slightly better stopwords workaround (but still not a good solution):

for word in nlp.Defaults.stop_words:
    for w in (word, word[0].upper() + word[1:], word.upper()):
        lex = nlp.vocab[w]
        lex.is_stop = True

This sets is_stop on the lowercase, uppercase, and first-char uppercase form for each.

[KingNoosh]

Any progress?

I'm using the above snippet atm.

[adrianog]

Why does a stop word need to be in the vocabulary? (general question)

[ines]

We're currently training new models for the upcoming nightly release of the develop branch (spaCy v2.1.0). You can lurk the spacy-models repo for updates and progress, but it's all currently pre-alpha. Sorry this was taking so long – it really did come down to getting the infrastructure right to be able to train our current model family reliably (and be able to add more languages in the future).

@adrianog The is_stop attribute is an attribute on the lexeme, i.e. the context-independent entry in the vocabulary.

[adrianog]

@ines I see. Where vocabulary here is to be intended as spacy vocaculary i.e. lexeme.is_oov() could still return "False"?

[lock[bot]]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.