Spacy lemmatizer to use blank model

[lizgzil]

Description

Fixes https://github.com/wellcometrust/WellcomeML/issues/271 by using blank spacy model.
Running on 1000 grant description: the previous method took 33 secs (when tagger is disabled) and 14 secs (when tagger is not disabled). This new method takes 2 secs.

I also noticed some unusual performance where other methods don't seem to actually lemmatize the text.

X = ['the cats sat on the sitting room floors enjoying the sun']

Original method (when the model is already downloaded)

nlp = spacy.load("en_core_web_sm", disable=["ner", "tagger", "parser", "textcat"])
output= [[token.lemma_.lower() for token in doc] for doc in nlp.pipe(X)]

gives loads of warning for every token of the form "spacy WARNING: [W108] The rule-based lemmatizer did not find POS annotation for the token 'sun'. Check that your pipeline includes components that assign token.pos, typically 'tagger'+'attribute_ruler' or 'morphologizer'" (allenai seem to have had a similar issue with it)
and the output is:

[['the', 'cats', 'sat', 'on', 'the', 'sitting', 'room', 'floors', 'enjoying', 'the', 'sun']]

I realise this had a bug in it where the tagger shouldn't have been disabled, i.e.

nlp = spacy.load("en_core_web_sm", disable=["ner", "parser", "textcat"])
output= [[token.lemma_.lower() for token in doc] for doc in nlp.pipe(X)]

does seem to lemmatize correctly.

I also tried the following method:

from spacy.lang.en import English
from spacy.pipeline.lemmatizer import Lemmatizer
nlp = English()
lemmatizer = Lemmatizer(nlp, model=None,  mode= "lookup")
output= [[lemmatizer.lookup_lemmatize(token)[0].lower() for token in doc] for doc in nlp.pipe(X)]

which also gave:

[['the', 'cats', 'sat', 'on', 'the', 'sitting', 'room', 'floors', 'enjoying', 'the', 'sun']]

The final method, which I've commited is:

nlp = spacy.blank("en")
nlp.add_pipe("lemmatizer", config={"mode": "lookup"})
nlp.initialize()
output= [[token.lemma_.lower() for token in doc] for doc in nlp.pipe(X)]

gives:

[['the', 'cat', 'sit', 'on', 'the', 'sit', 'room', 'floor', 'enjoy', 'the', 'sun']]

Checklist

• Added link to Github issue or Trello card
• Added tests

[aCampello]

Nice find. Is it much quicker then?

[aCampello]

Nice find. I think this needs spacy-lookup-data to be install (which is probably installed automatically with spacy[lookup].

[lizgzil]

just did a little experiment with a blank virtualenv and it seems like as long as spacy[lookups] is installed then in the script import spacy is enough for this to work. Or am I confusing something?

[nsorros]

can you try what happens if en_core_web_sm model is not installed? I think an IOError is thrown still so you should keep the exception handling that @aCampello had

[lizgzil]

seems to work without it being installed:

>>> import spacy
>>> nlp=spacy.load('en_core_web_sm')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/gallaghe/Code/nutrition-labels/build/virtualenv/lib/python3.7/site-packages/spacy/__init__.py", line 47, in load
    return util.load_model(name, disable=disable, exclude=exclude, config=config)
  File "/Users/gallaghe/Code/nutrition-labels/build/virtualenv/lib/python3.7/site-packages/spacy/util.py", line 329, in load_model
    raise IOError(Errors.E050.format(name=name))
OSError: [E050] Can't find model 'en_core_web_sm'. It doesn't seem to be a Python package or a valid path to a data directory.
>>> nlp = spacy.blank("en")
>>> nlp.add_pipe("lemmatizer", config={"mode": "lookup"})
<spacy.lang.en.lemmatizer.EnglishLemmatizer object at 0x15ad7ca00>
>>> nlp.initialize()
<thinc.optimizers.Optimizer object at 0x15aba8050>
>>> X = ['the cats sat on the sitting room floors enjoying the sun']
>>> [[token.lemma_.lower() for token in doc] for doc in nlp.pipe(X)]
[['the', 'cat', 'sit', 'on', 'the', 'sit', 'room', 'floor', 'enjoy', 'the', 'sun']]

[nsorros]

ok cool, just realised that you initialise a blank model, good 👍 thanks for checking

[aCampello]

Yeah, that's actually a big advantage. You just need to install the lookups, not the entire model! Much simpler code too.

[lizgzil]

Nice find. Is it much quicker then?

Running on 1000 grant descriptions the previous method took around 14 secs (when tagger is not disabled) and this new method takes 2 seconds.

[aCampello]

Nice find. Is it much quicker then?

Running on 1000 grant descriptions the previous method took around 14 secs (when tagger is not disabled) and this new method takes 2 seconds.

⚡

[aCampello]

Pipeline seems to fail with a weird error.

[lizgzil]

Pipeline seems to fail with a weird error.

@aCampello :( seems to be the same one as you got when you opened https://github.com/wellcometrust/WellcomeML/issues/270. Did you do anything to get your PR to pass in the end? As @nsorros said "Some non deterministic test fail here showing up again."

[aCampello]

Don't know, I re-ran and it passed! I assume some memory error again.

[aCampello]

Let me have a look. Does tox pass locally?

[lizgzil]

Let me have a look. Does tox pass locally?

@aCampello it's still running locally for 3.8, but has passed for 3.7

[aCampello]

I re-triggered your pipeline. Let's see.

[nsorros]

Blocking so we check whether the exception handling is still needed