Skip to main content

Use the latest Stanza (StanfordNLP) research models directly in spaCy

Project description

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.

⚠️ Previous version of this package were available as spacy-stanfordnlp.

Azure Pipelines PyPi GitHub Code style: black

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 and token.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)

️️️⌛️ Installation

As 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.

📖 Usage & Examples

⚠️ 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 – 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 – 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.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

spacy-stanza-1.0.2.tar.gz (11.9 kB view hashes)

Uploaded Source

Built Distribution

spacy_stanza-1.0.2-py3-none-any.whl (9.7 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page