Sequence tagging with spaCy and crfsuite.
A port of 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.language import Language
from spacy_crfsuite import CRFEntityExtractor, CRFExtractor
@Language.factory("ner_crf")
def create_component(nlp, name):
crf_extractor = CRFExtractor().from_disk("spacy_crfsuite_conll03_sm.bz2")
return CRFEntityExtractor(nlp, crf_extractor=crf_extractor)
nlp = spacy.load("en_core_web_sm", disable=["ner"])
nlp.add_pipe("ner_crf")
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 - LOCVisualization (via Gradio)
Run the command below to launch a Gradio playground
$ pip install gradio
$ python spacy_crfsuite/visualize.py
You can download a pre-trained model.
| Dataset | F1 | 📥 Download |
|---|---|---|
| CoNLL03 | 82% | spacy_crfsuite_conll03_sm.bz2 |
Below is a command line to train a simple model for restaurants search bot with markdown annotations and save it to disk. If you prefer working on jupyter, follow this notebook.
$ 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.pklBelow is a command line to test the CRF model and print the classification report (In the example we use the training set, however normally we would use a held out set).
$ 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
ℹ Classification Report:
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 for named entity recognition in a spaCy pipeline!
import spacy
from spacy.language import Language
from spacy_crfsuite import CRFEntityExtractor, CRFExtractor
@Language.factory("ner_crf")
def create_component(nlp, name):
crf_extractor = CRFExtractor().from_disk("model/model.pkl")
return CRFEntityExtractor(nlp, crf_extractor=crf_extractor)
nlp = spacy.load("en_core_web_sm", disable=["ner"])
nlp.add_pipe("ner_crf")
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.
Start a web service
$ pip install uvicorn
$ uvicorn spacy_crfsuite.serve:app --host 127.0.0.1 --port 5000Notice: Set
$SPACY_MODELand$CRF_MODELin your environment to control the server configurations
cURL example
$ curl -X POST http://127.0.0.1:5000/parse -H 'Content-Type: application/json' -d '{"text": "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."}'
{
"data": [
{
"text": "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.",
"entities": [
{
"start": 0,
"end": 18,
"value": "George Walker Bush",
"entity": "PER"
},
{
"start": 45,
"end": 53,
"value": "American",
"entity": "MISC"
},
{
"start": 121,
"end": 134,
"value": "United States",
"entity": "LOC"
}
]
}
]
}Set up env
$ poetry install
$ poetry run spacy download en_core_web_smRun unit test
$ poetry run pytestRun black (code formatting)
$ poetry run black spacy_crfsuite/ --config=pyproject.toml