Use the latest StanfordNLP research models directly in spaCy

spaCy + StanfordNLP

This package wraps the StanfordNLP library, so you can use Stanford’s models as 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 labelled dependency parsing in 58 languages.

Using this wrapper, you’ll be able to use the following annotations, computed by your pretrained stanfordnlp 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_)
• Dependency parsing (token.dep, token.dep_, token.head)
• Sentence segmentation (doc.sents)

️️️⌛️ Installation

pip install spacy-stanfordnlp

Make sure to also install one of the pre-trained StanfordNLP models. It’s recommended to run StanfordNLP on Python 3.6.8+ or Python 3.7.2+.

📖 Usage & Examples

The StanfordNLPLanguage class can be initialized with a loaded StanfordNLP pipeline and returns a spaCy Language object, i.e. the nlp object you can use to process text and create a Doc object.

import stanfordnlp
from spacy_stanfordnlp import StanfordNLPLanguage

snlp = stanfordnlp.Pipeline(lang="en")
nlp = StanfordNLPLanguage(snlp)

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_)

If language data for the given language is available in spaCy, the respective language class will 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

# Efficient processing with nlp.pipe
for doc in nlp.pipe(["Lots of texts", "Even more texts", "..."]):
    print(doc.text)

# Combine with your own custom pipeline components
def custom_component(doc):
    # Do something to the doc here
    return doc

nlp.add_pipe(custom_component)

# Serialize it to a numpy array
np_array = doc.to_array(['ORTH', 'LEMMA', 'POS'])

Experimental: Mixing and matching pipeline components

By default, the nlp object’s pipeline will be empty, because all attributes are computed once and set in the custom Tokenizer. But since it’s a regular nlp object, you can add your own components to the pipeline.

For example, the entity recognizer from one of spaCy’s pre-trained models:

import spacy
import spacy_stanfordnlp
import stanfordnlp

snlp = stanfordnlp.Pipeline(lang="en", models_dir="./models")
nlp = StanfordNLPLanguage(snlp)

# Load spaCy's pre-trained en_core_web_sm model, get the entity recognizer and
# add it to the StanfordNLP model's pipeline
spacy_model = spacy.load("en_core_web_sm")
ner = spacy_model.get_pipe("ner")
nlp.add_pipe(ner)

doc = nlp("Barack Obama was born in Hawaii. He was elected president in 2008.")
print([(ent.text, ent.label_) for ent in doc.ents])
# [('Barack Obama', 'PERSON'), ('Hawaii', 'GPE'), ('2008', 'DATE')]

You could also add and train your own custom text classification component.

Advanced: serialization and entry points

The spaCy nlp object created by StanfordNLPLanguage exposes its language as stanfordnlp_xx.

from spacy.util import get_lang_class
lang_cls = get_lang_class("stanfordnlp_en")

Normally, the above would fail because spaCy doesn’t include a language class stanfordnlp_en. But because this package exposes a spacy_languages entry point in its setup.py that points to StanfordNLPLanguage, spaCy knows how to initialize it.

This means that saving to and loading from disk works:

snlp = stanfordnlp.Pipeline(lang="en")
nlp = StanfordNLPLanguage(snlp)
nlp.to_disk("./stanfordnlp-spacy-model")

Additional arguments on spacy.load are automatically passed down to the language class and pipeline components. So when loading the saved model, you can pass in the snlp argument:

snlp = stanfordnlp.Pipeline(lang="en")
nlp = spacy.load("./stanfordnlp-spacy-model", snlp=snlp)

Note that this will not save any model data by default. The StanfordNLP models are very large, so for now, this package expects that you load them separately.