spaCy + Stanza (formerly StanfordNLP)
This package wraps the Stanza (formerly StanfordNLP) library, so you can use Stanford's models in a spaCy pipeline. The Stanford models achieved top accuracy in the CoNLL 2017 and 2018 shared task, which involves tokenization, part-of-speech tagging, morphological analysis, lemmatization and labeled dependency parsing in 68 languages. As of v1.0, Stanza also supports named entity recognition for selected languages.
โ ๏ธ Previous version of this package were available asspacy-stanfordnlp
.
Using this wrapper, you'll be able to use the following annotations, computed by
your pretrained stanza
model:
- Statistical tokenization (reflected in the
Doc
and its tokens) - Lemmatization (
token.lemma
andtoken.lemma_
) - Part-of-speech tagging (
token.tag
,token.tag_
,token.pos
,token.pos_
) - Morphological analysis (
token.morph
) - Dependency parsing (
token.dep
,token.dep_
,token.head
) - Named entity recognition (
doc.ents
,token.ent_type
,token.ent_type_
,token.ent_iob
,token.ent_iob_
) - Sentence segmentation (
doc.sents
)
โ๏ธ Installation
๏ธ๏ธ๏ธAs of v1.0.0 spacy-stanza
is only compatible with spaCy v3.x. To install
the most recent version:
pip install spacy-stanza
For spaCy v2, install v0.2.x and refer to the v0.2.x usage documentation:
pip install "spacy-stanza<0.3.0"
Make sure to also download one of the pre-trained Stanza models.
๐ Usage & Examples
โ ๏ธ Important note: This package has been refactored to take advantage of spaCy v3.0. Previous versions that were built for spaCy v2.x worked considerably differently. Please see previous tagged versions of this README for documentation on prior versions.
Use spacy_stanza.load_pipeline()
to create an nlp
object that you can use to
process a text with a Stanza pipeline and create a spaCy
Doc
object. By default, both the spaCy pipeline
and the Stanza pipeline will be initialized with the same lang
, e.g. "en":
import stanza
import spacy_stanza
# Download the stanza model if necessary
stanza.download("en")
# Initialize the pipeline
nlp = spacy_stanza.load_pipeline("en")
doc = nlp("Barack Obama was born in Hawaii. He was elected president in 2008.")
for token in doc:
print(token.text, token.lemma_, token.pos_, token.dep_, token.ent_type_)
print(doc.ents)
If language data for the given language is available in spaCy, the respective
language class can be used as the base for the nlp
object โ for example,
English()
. This lets you use spaCy's lexical attributes like is_stop
or
like_num
. The nlp
object follows the same API as any other spaCy Language
class โ so you can visualize the Doc
objects with displaCy, add custom
components to the pipeline, use the rule-based matcher and do pretty much
anything else you'd normally do in spaCy.
# Access spaCy's lexical attributes
print([token.is_stop for token in doc])
print([token.like_num for token in doc])
# Visualize dependencies
from spacy import displacy
displacy.serve(doc) # or displacy.render if you're in a Jupyter notebook
# Process texts with nlp.pipe
for doc in nlp.pipe(["Lots of texts", "Even more texts", "..."]):
print(doc.text)
# Combine with your own custom pipeline components
from spacy import Language
@Language.component("custom_component")
def custom_component(doc):
# Do something to the doc here
print(f"Custom component called: {doc.text}")
return doc
nlp.add_pipe("custom_component")
doc = nlp("Some text")
# Serialize attributes to a numpy array
np_array = doc.to_array(['ORTH', 'LEMMA', 'POS'])
Stanza Pipeline options
Additional options for the Stanza
Pipeline
can be
provided as keyword arguments following the Pipeline
API:
-
Provide the Stanza language as
lang
. For Stanza languages without spaCy support, use "xx" for the spaCy language setting:# Initialize a pipeline for Coptic nlp = spacy_stanza.load_pipeline("xx", lang="cop")
-
Provide Stanza pipeline settings following the
Pipeline
API:# Initialize a German pipeline with the `hdt` package nlp = spacy_stanza.load_pipeline("de", package="hdt")
-
Tokenize with spaCy rather than the statistical tokenizer (only for English):
nlp = spacy_stanza.load_pipeline("en", processors= {"tokenize": "spacy"})
-
Provide any additional processor settings as additional keyword arguments:
# Provide pretokenized texts (whitespace tokenization) nlp = spacy_stanza.load_pipeline("de", tokenize_pretokenized=True)
The spaCy config specifies all Pipeline
options in the [nlp.tokenizer]
block. For example, the config for the last example above, a German pipeline
with pretokenized texts:
[nlp.tokenizer]
@tokenizers = "spacy_stanza.PipelineAsTokenizer.v1"
lang = "de"
dir = null
package = "default"
logging_level = null
verbose = null
use_gpu = true
[nlp.tokenizer.kwargs]
tokenize_pretokenized = true
[nlp.tokenizer.processors]
Serialization
The full Stanza pipeline configuration is stored in the spaCy pipeline
config, so you can save and load the
pipeline just like any other nlp
pipeline:
# Save to a local directory
nlp.to_disk("./stanza-spacy-model")
# Reload the pipeline
nlp = spacy.load("./stanza-spacy-model")
Note that this does not save any Stanza model data by default. The Stanza
models are very large, so for now, this package expects you to download the
models separately with stanza.download()
and have them available either in the
default model directory or in the path specified under [nlp.tokenizer.dir]
in
the config.
Adding additional spaCy pipeline components
By default, the spaCy pipeline in the nlp
object returned by
spacy_stanza.load_pipeline()
will be empty, because all stanza
attributes
are computed and set within the custom tokenizer,
StanzaTokenizer
. But since it's a regular nlp
object, you can add your own components to the pipeline. For example, you could
add
your own custom text classification component
with nlp.add_pipe("textcat", source=source_nlp)
, or augment the named entities
with your own rule-based patterns using the
EntityRuler
component.