Sequence tagging with spaCy and crfsuite.
Copied from Rasa NLU.
- Simple but tough to beat CRF entity tagger (via sklearn-crfsuite)
- spaCy NER component
- Command line interface for training & evaluation and example notebook
- CoNLL, JSON and Markdown annotations
- Pre-trained NER component
pip install spacy_crfsuiteimport spacy
from spacy_crfsuite import CRFEntityExtractor, CRFExtractor
nlp = spacy.load("en_core_web_sm", disable=["ner"])
crf_extractor = CRFExtractor().from_disk("spacy_crfsuite_conll03_sm.bz2")
pipe = CRFEntityExtractor(nlp, crf_extractor=crf_extractor)
nlp.add_pipe(pipe)
doc = nlp(
"George Walker Bush (born July 6, 1946) is an American politician and businessman "
"who served as the 43rd president of the United States from 2001 to 2009.")
for ent in doc.ents:
print(ent, "-", ent.label_)
# Output:
# George Walker Bush - PER
# American - MISC
# United States - LOCYou can download a pre-trained model.
| Dataset | F1 | 📥 Download |
|---|---|---|
| CoNLL03 | 82% | spacy_crfsuite_conll03_sm.bz2 |
Let's train a simple model for restaurent search bot with markdown annotations and the command line. You can also try this notebook.
So we start by training a model and saving it to disk.
$ python -m spacy_crfsuite.train examples/restaurent_search.md -c examples/default-config.json -o model/ -lm en_core_web_sm
ℹ Loading config from disk
✔ Successfully loaded config from file.
examples/default-config.json
ℹ Loading training examples.
✔ Successfully loaded 15 training examples from file.
examples/restaurent_search.md
ℹ Using spaCy model: en_core_web_sm
ℹ Training entity tagger with CRF.
ℹ Saving model to disk
✔ Successfully saved model to file.
model/model.pklWe can also evaluate on a dev set to get f1 & classification report. Below we use the training examples.
$ python -m spacy_crfsuite.eval examples/restaurent_search.md -m model/model.pkl -lm en_core_web_sm
ℹ Loading model from file
model/model.pkl
✔ Successfully loaded CRF tagger
<spacy_crfsuite.crf_extractor.CRFExtractor object at 0x126e5f438>
ℹ Loading dev dataset from file
examples/example.md
✔ Successfully loaded 15 dev examples.
ℹ Using spaCy model: en_core_web_sm
⚠ f1 score: 1.0
precision recall f1-score support
B-cuisine 1.000 1.000 1.000 2
I-cuisine 1.000 1.000 1.000 1
L-cuisine 1.000 1.000 1.000 2
U-cuisine 1.000 1.000 1.000 5
U-location 1.000 1.000 1.000 7
micro avg 1.000 1.000 1.000 17
macro avg 1.000 1.000 1.000 17
weighted avg 1.000 1.000 1.000 17Now we can use the tagger in a spaCy pipeline!
import spacy
from spacy_crfsuite import CRFEntityExtractor
nlp = spacy.load('en_core_web_sm')
pipe = CRFEntityExtractor(nlp).from_disk("model/model.pkl")
nlp.add_pipe(pipe)
doc = nlp("show mexican restaurents up north")
for ent in doc.ents:
print(ent.text, "--", ent.label_)
# Output:
# mexican -- cuisine
# north -- locationOr alternatively as a standalone component.
from spacy_crfsuite import CRFExtractor
from spacy_crfsuite.tokenizer import SpacyTokenizer
crf_extractor = CRFExtractor().from_disk("model/model.pkl")
tokenizer = SpacyTokenizer()
example = {"text": "show mexican restaurents up north"}
tokenizer.tokenize(example, attribute="text")
crf_extractor.process(example)
# Output:
# [{'start': 5,
# 'end': 12,
# 'value': 'mexican',
# 'entity': 'cuisine',
# 'confidence': 0.5823148506311286},
# {'start': 28,
# 'end': 33,
# 'value': 'north',
# 'entity': 'location',
# 'confidence': 0.8863076478494413}]We can also take a look at what model learned.
Use the .explain() method to understand model decision.
print(crf_extractor.explain())
# Output:
#
# Most likely transitions:
# O -> O 1.637338
# B-cuisine -> I-cuisine 1.373766
# U-cuisine -> O 1.306077
# I-cuisine -> L-cuisine 0.915989
# O -> U-location 0.751463
# B-cuisine -> L-cuisine 0.698893
# O -> U-cuisine 0.480360
# U-location -> U-cuisine 0.403487
# O -> B-cuisine 0.261450
# L-cuisine -> O 0.182695
#
# Positive features:
# 1.976502 O 0:bias:bias
# 1.957180 U-location -1:low:the
# 1.216547 B-cuisine -1:low:for
# 1.153924 U-location 0:prefix5:centr
# 1.153924 U-location 0:prefix2:ce
# 1.110536 U-location 0:digit
# 1.058294 U-cuisine 0:prefix5:chine
# 1.058294 U-cuisine 0:prefix2:ch
# 1.051457 U-cuisine 0:suffix2:an
# 0.999976 U-cuisine -1:low:meNotice: You can also access the
crf_extractordirectly withnlp.get_pipe("crf_ner").crf_extractor.
Set up virtualenv
$ pipenv sync -dRun unit test
$ pipenv run pytestRun black (code formatting)
$ pipenv run black spacy_crfsuite/ --config=pyproject.toml