spaCy/website/docs/usage/processing-pipelines.md

title	next	menu
Language Processing Pipelines	/usage/embeddings-transformers	Processing Text processing Pipelines & Components pipelines Custom Components custom-components Component Data component-data Type Hints & Validation type-hints Trainable Components trainable-components Extension Attributes custom-components-attributes Plugins & Wrappers plugins

import Pipelines101 from 'usage/101/_pipelines.md'

Processing text

When you call nlp on a text, spaCy will tokenize it and then call each component on the Doc, in order. It then returns the processed Doc that you can work with.

doc = nlp("This is a text")

When processing large volumes of text, the statistical models are usually more efficient if you let them work on batches of texts. spaCy's nlp.pipe method takes an iterable of texts and yields processed Doc objects. The batching is done internally.

texts = ["This is a text", "These are lots of texts", "..."]
- docs = [nlp(text) for text in texts]
+ docs = list(nlp.pipe(texts))

• Process the texts as a stream using nlp.pipe and buffer them in batches, instead of one-by-one. This is usually much more efficient.
• Only apply the pipeline components you need. Getting predictions from the model that you don't actually need adds up and becomes very inefficient at scale. To prevent this, use the disable keyword argument to disable components you don't need – either when loading a pipeline, or during processing with nlp.pipe. See the section on disabling pipeline components for more details and examples.

In this example, we're using nlp.pipe to process a (potentially very large) iterable of texts as a stream. Because we're only accessing the named entities in doc.ents (set by the ner component), we'll disable all other components during processing. nlp.pipe yields Doc objects, so we can iterate over them and access the named entity predictions:

✏️ Things to try

1. Also disable the "ner" component. You'll see that the doc.ents are now empty, because the entity recognizer didn't run.

### {executable="true"}
import spacy

texts = [
    "Net income was $9.4 million compared to the prior year of $2.7 million.",
    "Revenue exceeded twelve billion dollars, with a loss of $1b.",
]

nlp = spacy.load("en_core_web_sm")
for doc in nlp.pipe(texts, disable=["tok2vec", "tagger", "parser", "attribute_ruler", "lemmatizer"]):
    # Do something with the doc here
    print([(ent.text, ent.label_) for ent in doc.ents])

When using nlp.pipe, keep in mind that it returns a generator that yields Doc objects – not a list. So if you want to use it like a list, you'll have to call list() on it first:

- docs = nlp.pipe(texts)[0]         # will raise an error
+ docs = list(nlp.pipe(texts))[0]   # works as expected

You can use the as_tuples option to pass additional context along with each doc when using nlp.pipe. If as_tuples is True, then the input should be a sequence of (text, context) tuples and the output will be a sequence of (doc, context) tuples. For example, you can pass metadata in the context and save it in a custom attribute:

### {executable="true"}
import spacy
from spacy.tokens import Doc

if not Doc.has_extension("text_id"):
    Doc.set_extension("text_id", default=None)

text_tuples = [
    ("This is the first text.", {"text_id": "text1"}),
    ("This is the second text.", {"text_id": "text2"})
]

nlp = spacy.load("en_core_web_sm")
doc_tuples = nlp.pipe(text_tuples, as_tuples=True)

docs = []
for doc, context in doc_tuples:
    doc._.text_id = context["text_id"]
    docs.append(doc)

for doc in docs:
    print(f"{doc._.text_id}: {doc.text}")

Multiprocessing

spaCy includes built-in support for multiprocessing with nlp.pipe using the n_process option:

# Multiprocessing with 4 processes
docs = nlp.pipe(texts, n_process=4)

# With as many processes as CPUs (use with caution!)
docs = nlp.pipe(texts, n_process=-1)

Depending on your platform, starting many processes with multiprocessing can add a lot of overhead. In particular, the default start method spawn used in macOS/OS X (as of Python 3.8) and in Windows can be slow for larger models because the model data is copied in memory for each new process. See the Python docs on multiprocessing for further details.

For shorter tasks and in particular with spawn, it can be faster to use a smaller number of processes with a larger batch size. The optimal batch_size setting will depend on the pipeline components, the length of your documents, the number of processes and how much memory is available.

# Default batch size is `nlp.batch_size` (typically 1000)
docs = nlp.pipe(texts, n_process=2, batch_size=2000)

Multiprocessing is not generally recommended on GPU because RAM is too limited. If you want to try it out, be aware that it is only possible using spawn due to limitations in CUDA.

In Linux, transformer models may hang or deadlock with multiprocessing due to an issue in PyTorch. One suggested workaround is to use spawn instead of fork and another is to limit the number of threads before loading any models using torch.set_num_threads(1).

Pipelines and built-in components

spaCy makes it very easy to create your own pipelines consisting of reusable components – this includes spaCy's default tagger, parser and entity recognizer, but also your own custom processing functions. A pipeline component can be added to an already existing nlp object, specified when initializing a Language class, or defined within a pipeline package.

config.cfg (excerpt)

 [nlp]
 lang = "en"
 pipeline = ["tok2vec", "parser"]

[components]

[components.tok2vec]
factory = "tok2vec"
# Settings for the tok2vec component

[components.parser]
factory = "parser"
# Settings for the parser component

When you load a pipeline, spaCy first consults the meta.json and config.cfg. The config tells spaCy what language class to use, which components are in the pipeline, and how those components should be created. spaCy will then do the following:

1. Load the language class and data for the given ID via get_lang_class and initialize it. The Language class contains the shared vocabulary, tokenization rules and the language-specific settings.
2. Iterate over the pipeline names and look up each component name in the [components] block. The factory tells spaCy which component factory to use for adding the component with add_pipe. The settings are passed into the factory.
3. Make the model data available to the Language class by calling from_disk with the path to the data directory.

So when you call this...

nlp = spacy.load("en_core_web_sm")

... the pipeline's config.cfg tells spaCy to use the language "en" and the pipeline ["tok2vec", "tagger", "parser", "ner", "attribute_ruler", "lemmatizer"]. spaCy will then initialize spacy.lang.en.English, and create each pipeline component and add it to the processing pipeline. It'll then load in the model data from the data directory and return the modified Language class for you to use as the nlp object.

spaCy v3.0 introduces a config.cfg, which includes more detailed settings for the pipeline, its components and the training process. You can export the config of your current nlp object by calling nlp.config.to_disk.

Fundamentally, a spaCy pipeline package consists of three components: the weights, i.e. binary data loaded in from a directory, a pipeline of functions called in order, and language data like the tokenization rules and language-specific settings. For example, a Spanish NER pipeline requires different weights, language data and components than an English parsing and tagging pipeline. This is also why the pipeline state is always held by the Language class. spacy.load puts this all together and returns an instance of Language with a pipeline set and access to the binary data:

### spacy.load under the hood
lang = "en"
pipeline = ["tok2vec", "tagger", "parser", "ner", "attribute_ruler", "lemmatizer"]
data_path = "path/to/en_core_web_sm/en_core_web_sm-3.0.0"

cls = spacy.util.get_lang_class(lang)  # 1. Get Language class, e.g. English
nlp = cls()                            # 2. Initialize it
for name in pipeline:
    nlp.add_pipe(name)                 # 3. Add the component to the pipeline
nlp.from_disk(data_path)               # 4. Load in the binary data

When you call nlp on a text, spaCy will tokenize it and then call each component on the Doc, in order. Since the model data is loaded, the components can access it to assign annotations to the Doc object, and subsequently to the Token and Span which are only views of the Doc, and don't own any data themselves. All components return the modified document, which is then processed by the next component in the pipeline.

### The pipeline under the hood
doc = nlp.make_doc("This is a sentence")  # Create a Doc from raw text
for name, proc in nlp.pipeline:           # Iterate over components in order
    doc = proc(doc)                       # Apply each component

The current processing pipeline is available as nlp.pipeline, which returns a list of (name, component) tuples, or nlp.pipe_names, which only returns a list of human-readable component names.

print(nlp.pipeline)
# [('tok2vec', <spacy.pipeline.Tok2Vec>), ('tagger', <spacy.pipeline.Tagger>), ('parser', <spacy.pipeline.DependencyParser>), ('ner', <spacy.pipeline.EntityRecognizer>), ('attribute_ruler', <spacy.pipeline.AttributeRuler>), ('lemmatizer', <spacy.lang.en.lemmatizer.EnglishLemmatizer>)]
print(nlp.pipe_names)
# ['tok2vec', 'tagger', 'parser', 'ner', 'attribute_ruler', 'lemmatizer']

Built-in pipeline components

spaCy ships with several built-in pipeline components that are registered with string names. This means that you can initialize them by calling nlp.add_pipe with their names and spaCy will know how to create them. See the API documentation for a full list of available pipeline components and component functions.

Usage

nlp = spacy.blank("en")
nlp.add_pipe("sentencizer")
# add_pipe returns the added component
ruler = nlp.add_pipe("entity_ruler")

String name	Component	Description
tagger	Tagger	Assign part-of-speech-tags.
parser	DependencyParser	Assign dependency labels.
ner	EntityRecognizer	Assign named entities.
entity_linker	EntityLinker	Assign knowledge base IDs to named entities. Should be added after the entity recognizer.
entity_ruler	EntityRuler	Assign named entities based on pattern rules and dictionaries.
textcat	TextCategorizer	Assign text categories: exactly one category is predicted per document.
textcat_multilabel	MultiLabel_TextCategorizer	Assign text categories in a multi-label setting: zero, one or more labels per document.
lemmatizer	Lemmatizer	Assign base forms to words.
morphologizer	Morphologizer	Assign morphological features and coarse-grained POS tags.
attribute_ruler	AttributeRuler	Assign token attribute mappings and rule-based exceptions.
senter	SentenceRecognizer	Assign sentence boundaries.
sentencizer	Sentencizer	Add rule-based sentence segmentation without the dependency parse.
tok2vec	Tok2Vec	Assign token-to-vector embeddings.
transformer	Transformer	Assign the tokens and outputs of a transformer model.

Disabling, excluding and modifying components

If you don't need a particular component of the pipeline – for example, the tagger or the parser, you can disable or exclude it. This can sometimes make a big difference and improve loading and inference speed. There are two different mechanisms you can use:

1. Disable: The component and its data will be loaded with the pipeline, but it will be disabled by default and not run as part of the processing pipeline. To run it, you can explicitly enable it by calling nlp.enable_pipe. When you save out the nlp object, the disabled component will be included but disabled by default.
2. Exclude: Don't load the component and its data with the pipeline. Once the pipeline is loaded, there will be no reference to the excluded component.

Disabled and excluded component names can be provided to spacy.load as a list.

💡 Optional pipeline components

The disable mechanism makes it easy to distribute pipeline packages with optional components that you can enable or disable at runtime. For instance, your pipeline may include a statistical and a rule-based component for sentence segmentation, and you can choose which one to run depending on your use case.

For example, spaCy's trained pipelines like en_core_web_sm contain both a parser and senter that perform sentence segmentation, but the senter is disabled by default.

# Load the pipeline without the entity recognizer
nlp = spacy.load("en_core_web_sm", exclude=["ner"])

# Load the tagger and parser but don't enable them
nlp = spacy.load("en_core_web_sm", disable=["tagger", "parser"])
# Explicitly enable the tagger later on
nlp.enable_pipe("tagger")

As of v3.0, the disable keyword argument specifies components to load but disable, instead of components to not load at all. Those components can now be specified separately using the new exclude keyword argument.

As a shortcut, you can use the nlp.select_pipes context manager to temporarily disable certain components for a given block. At the end of the with block, the disabled pipeline components will be restored automatically. Alternatively, select_pipes returns an object that lets you call its restore() method to restore the disabled components when needed. This can be useful if you want to prevent unnecessary code indentation of large blocks.

### Disable for block
# 1. Use as a context manager
with nlp.select_pipes(disable=["tagger", "parser", "lemmatizer"]):
    doc = nlp("I won't be tagged and parsed")
doc = nlp("I will be tagged and parsed")

# 2. Restore manually
disabled = nlp.select_pipes(disable="ner")
doc = nlp("I won't have named entities")
disabled.restore()

If you want to disable all pipes except for one or a few, you can use the enable keyword. Just like the disable keyword, it takes a list of pipe names, or a string defining just one pipe.

# Enable only the parser
with nlp.select_pipes(enable="parser"):
    doc = nlp("I will only be parsed")

The nlp.pipe method also supports a disable keyword argument if you only want to disable components during processing:

for doc in nlp.pipe(texts, disable=["tagger", "parser", "lemmatizer"]):
    # Do something with the doc here

Finally, you can also use the remove_pipe method to remove pipeline components from an existing pipeline, the rename_pipe method to rename them, or the replace_pipe method to replace them with a custom component entirely (more details on this in the section on custom components).

nlp.remove_pipe("parser")
nlp.rename_pipe("ner", "entityrecognizer")
nlp.replace_pipe("tagger", "my_custom_tagger")

The Language object exposes different attributes that let you inspect all available components and the components that currently run as part of the pipeline.

Example

nlp = spacy.blank("en")
nlp.add_pipe("ner")
nlp.add_pipe("textcat")
assert nlp.pipe_names == ["ner", "textcat"]
nlp.disable_pipe("ner")
assert nlp.pipe_names == ["textcat"]
assert nlp.component_names == ["ner", "textcat"]
assert nlp.disabled == ["ner"]

Name	Description
nlp.pipeline	(name, component) tuples of the processing pipeline, in order.
nlp.pipe_names	Pipeline component names, in order.
nlp.components	All (name, component) tuples, including disabled components.
nlp.component_names	All component names, including disabled components.
nlp.disabled	Names of components that are currently disabled.

Sourcing components from existing pipelines

Pipeline components that are independent can also be reused across pipelines. Instead of adding a new blank component, you can also copy an existing component from a trained pipeline by setting the source argument on nlp.add_pipe. The first argument will then be interpreted as the name of the component in the source pipeline – for instance, "ner". This is especially useful for training a pipeline because it lets you mix and match components and create fully custom pipeline packages with updated trained components and new components trained on your data.

When reusing components across pipelines, keep in mind that the vocabulary, vectors and model settings must match. If a trained pipeline includes word vectors and the component uses them as features, the pipeline you copy it to needs to have the same vectors available – otherwise, it won't be able to make the same predictions.

In training config

Instead of providing a factory, component blocks in the training config can also define a source. The string needs to be a loadable spaCy pipeline package or path.

[components.ner]
source = "en_core_web_sm"
component = "ner"

By default, sourced components will be updated with your data during training. If you want to preserve the component as-is, you can "freeze" it if the pipeline is not using a shared Tok2Vec layer:

[training]
frozen_components = ["ner"]

### {executable="true"}
import spacy

# The source pipeline with different components
source_nlp = spacy.load("en_core_web_sm")
print(source_nlp.pipe_names)

# Add only the entity recognizer to the new blank pipeline
nlp = spacy.blank("en")
nlp.add_pipe("ner", source=source_nlp)
print(nlp.pipe_names)

Analyzing pipeline components

The nlp.analyze_pipes method analyzes the components in the current pipeline and outputs information about them like the attributes they set on the Doc and Token, whether they retokenize the Doc and which scores they produce during training. It will also show warnings if components require values that aren't set by previous component – for instance, if the entity linker is used but no component that runs before it sets named entities. Setting pretty=True will pretty-print a table instead of only returning the structured data.

✏️ Things to try

1. Add the components "ner" and "sentencizer" before the "entity_linker". The analysis should now show no problems, because requirements are met.

### {executable="true"}
import spacy

nlp = spacy.blank("en")
nlp.add_pipe("tagger")
# This is a problem because it needs entities and sentence boundaries
nlp.add_pipe("entity_linker")
analysis = nlp.analyze_pipes(pretty=True)

### Structured
{
  "summary": {
    "tagger": {
      "assigns": ["token.tag"],
      "requires": [],
      "scores": ["tag_acc", "pos_acc", "lemma_acc"],
      "retokenizes": false
    },
    "entity_linker": {
      "assigns": ["token.ent_kb_id"],
      "requires": ["doc.ents", "doc.sents", "token.ent_iob", "token.ent_type"],
      "scores": [],
      "retokenizes": false
    }
  },
  "problems": {
    "tagger": [],
    "entity_linker": ["doc.ents", "doc.sents", "token.ent_iob", "token.ent_type"]
  },
  "attrs": {
    "token.ent_iob": { "assigns": [], "requires": ["entity_linker"] },
    "doc.ents": { "assigns": [], "requires": ["entity_linker"] },
    "token.ent_kb_id": { "assigns": ["entity_linker"], "requires": [] },
    "doc.sents": { "assigns": [], "requires": ["entity_linker"] },
    "token.tag": { "assigns": ["tagger"], "requires": [] },
    "token.ent_type": { "assigns": [], "requires": ["entity_linker"] }
  }
}

### Pretty
============================= Pipeline Overview =============================

#   Component       Assigns           Requires         Scores        Retokenizes
-   -------------   ---------------   --------------   -----------   -----------
0   tagger          token.tag                          tag_acc       False

1   entity_linker   token.ent_kb_id   doc.ents         nel_micro_f   False
                                      doc.sents        nel_micro_r
                                      token.ent_iob    nel_micro_p
                                      token.ent_type


================================ Problems (4) ================================
⚠ 'entity_linker' requirements not met: doc.ents, doc.sents,
token.ent_iob, token.ent_type

The pipeline analysis is static and does not actually run the components. This means that it relies on the information provided by the components themselves. If a custom component declares that it assigns an attribute but it doesn't, the pipeline analysis won't catch that.

Creating custom pipeline components

A pipeline component is a function that receives a Doc object, modifies it and returns it – for example, by using the current weights to make a prediction and set some annotation on the document. By adding a component to the pipeline, you'll get access to the Doc at any point during processing – instead of only being able to modify it afterwards.

Example

from spacy.language import Language

@Language.component("my_component")
def my_component(doc):
   # Do something to the doc here
   return doc

Argument	Type	Description
doc	Doc	The Doc object processed by the previous component.
RETURNS	Doc	The Doc object processed by this pipeline component.

The @Language.component decorator lets you turn a simple function into a pipeline component. It takes at least one argument, the name of the component factory. You can use this name to add an instance of your component to the pipeline. It can also be listed in your pipeline config, so you can save, load and train pipelines using your component.

Custom components can be added to the pipeline using the add_pipe method. Optionally, you can either specify a component to add it before or after, tell spaCy to add it first or last in the pipeline, or define a custom name. If no name is set and no name attribute is present on your component, the function name is used.

Example

nlp.add_pipe("my_component")
nlp.add_pipe("my_component", first=True)
nlp.add_pipe("my_component", before="parser")

Argument	Description
last	If set to True, component is added last in the pipeline (default). bool
first	If set to True, component is added first in the pipeline. bool
before	String name or index to add the new component before. Union[str, int]
after	String name or index to add the new component after. Union[str, int]

As of v3.0, components need to be registered using the @Language.component or @Language.factory decorator so spaCy knows that a function is a component. nlp.add_pipe now takes the string name of the component factory instead of the component function. This doesn't only save you lines of code, it also allows spaCy to validate and track your custom components, and make sure they can be saved and loaded.

- ruler = nlp.create_pipe("entity_ruler")
- nlp.add_pipe(ruler)
+ ruler = nlp.add_pipe("entity_ruler")

Examples: Simple stateless pipeline components

The following component receives the Doc in the pipeline and prints some information about it: the number of tokens, the part-of-speech tags of the tokens and a conditional message based on the document length. The @Language.component decorator lets you register the component under the name "info_component".

✏️ Things to try

1. Add the component first in the pipeline by setting first=True. You'll see that the part-of-speech tags are empty, because the component now runs before the tagger and the tags aren't available yet.
2. Change the component name or remove the name argument. You should see this change reflected in nlp.pipe_names.
3. Print nlp.pipeline. You'll see a list of tuples describing the component name and the function that's called on the Doc object in the pipeline.
4. Change the first argument to @Language.component, the name, to something else. spaCy should now complain that it doesn't know a component of the name "info_component".

### {executable="true"}
import spacy
from spacy.language import Language

@Language.component("info_component")
def my_component(doc):
    print(f"After tokenization, this doc has {len(doc)} tokens.")
    print("The part-of-speech tags are:", [token.pos_ for token in doc])
    if len(doc) < 10:
        print("This is a pretty short document.")
    return doc

nlp = spacy.load("en_core_web_sm")
nlp.add_pipe("info_component", name="print_info", last=True)
print(nlp.pipe_names)  # ['tagger', 'parser', 'ner', 'print_info']
doc = nlp("This is a sentence.")

Here's another example of a pipeline component that implements custom logic to improve the sentence boundaries set by the dependency parser. The custom logic should therefore be applied after tokenization, but before the dependency parsing – this way, the parser can also take advantage of the sentence boundaries.

✏️ Things to try

1. Print [token.dep_ for token in doc] with and without the custom pipeline component. You'll see that the predicted dependency parse changes to match the sentence boundaries.
2. Remove the else block. All other tokens will now have is_sent_start set to None (missing value), the parser will assign sentence boundaries in between.

### {executable="true"}
import spacy
from spacy.language import Language

@Language.component("custom_sentencizer")
def custom_sentencizer(doc):
    for i, token in enumerate(doc[:-2]):
        # Define sentence start if pipe + titlecase token
        if token.text == "|" and doc[i + 1].is_title:
            doc[i + 1].is_sent_start = True
        else:
            # Explicitly set sentence start to False otherwise, to tell
            # the parser to leave those tokens alone
            doc[i + 1].is_sent_start = False
    return doc

nlp = spacy.load("en_core_web_sm")
nlp.add_pipe("custom_sentencizer", before="parser")  # Insert before the parser
doc = nlp("This is. A sentence. | This is. Another sentence.")
for sent in doc.sents:
    print(sent.text)

Component factories and stateful components

Component factories are callables that take settings and return a pipeline component function. This is useful if your component is stateful and if you need to customize their creation, or if you need access to the current nlp object or the shared vocab. Component factories can be registered using the @Language.factory decorator and they need at least two named arguments that are filled in automatically when the component is added to the pipeline:

Example

from spacy.language import Language

@Language.factory("my_component")
def my_component(nlp, name):
    return MyComponent()

Argument	Description
nlp	The current nlp object. Can be used to access the shared vocab. Language
name	The instance name of the component in the pipeline. This lets you identify different instances of the same component. str

All other settings can be passed in by the user via the config argument on nlp.add_pipe. The @Language.factory decorator also lets you define a default_config that's used as a fallback.

### With config {highlight="4,9"}
import spacy
from spacy.language import Language

@Language.factory("my_component", default_config={"some_setting": True})
def my_component(nlp, name, some_setting: bool):
    return MyComponent(some_setting=some_setting)

nlp = spacy.blank("en")
nlp.add_pipe("my_component", config={"some_setting": False})

The @Language.component decorator is essentially a shortcut for stateless pipeline components that don't need any settings. This means you don't have to always write a function that returns your function if there's no state to be passed through – spaCy can just take care of this for you. The following two code examples are equivalent:

# Statless component with @Language.factory
@Language.factory("my_component")
def create_my_component():
    def my_component(doc):
        # Do something to the doc
        return doc

    return my_component

# Stateless component with @Language.component
@Language.component("my_component")
def my_component(doc):
    # Do something to the doc
    return doc

Yes, the @Language.factory decorator can be added to a function or a class. If it's added to a class, it expects the __init__ method to take the arguments nlp and name, and will populate all other arguments from the config. That said, it's often cleaner and more intuitive to make your factory a separate function. That's also how spaCy does it internally.

Language-specific factories

There are many use cases where you might want your pipeline components to be language-specific. Sometimes this requires entirely different implementation per language, sometimes the only difference is in the settings or data. spaCy allows you to register factories of the same name on both the Language base class, as well as its subclasses like English or German. Factories are resolved starting with the specific subclass. If the subclass doesn't define a component of that name, spaCy will check the Language base class.

Here's an example of a pipeline component that overwrites the normalized form of a token, the Token.norm_ with an entry from a language-specific lookup table. It's registered twice under the name "token_normalizer" – once using @English.factory and once using @German.factory:

### {executable="true"}
from spacy.lang.en import English
from spacy.lang.de import German

class TokenNormalizer:
    def __init__(self, norm_table):
        self.norm_table = norm_table

    def __call__(self, doc):
        for token in doc:
            # Overwrite the token.norm_ if there's an entry in the data
            token.norm_ = self.norm_table.get(token.text, token.norm_)
        return doc

@English.factory("token_normalizer")
def create_en_normalizer(nlp, name):
    return TokenNormalizer({"realise": "realize", "colour": "color"})

@German.factory("token_normalizer")
def create_de_normalizer(nlp, name):
    return TokenNormalizer({"daß": "dass", "wußte": "wusste"})

nlp_en = English()
nlp_en.add_pipe("token_normalizer")  # uses the English factory
print([token.norm_ for token in nlp_en("realise colour daß wußte")])

nlp_de = German()
nlp_de.add_pipe("token_normalizer")  # uses the German factory
print([token.norm_ for token in nlp_de("realise colour daß wußte")])

Under the hood, language-specific factories are added to the factories registry prefixed with the language code, e.g. "en.token_normalizer". When resolving the factory in nlp.add_pipe, spaCy first checks for a language-specific version of the factory using nlp.lang and if none is available, falls back to looking up the regular factory name.

Example: Stateful component with settings

This example shows a stateful pipeline component for handling acronyms: based on a dictionary, it will detect acronyms and their expanded forms in both directions and add them to a list as the custom doc._.acronyms extension attribute. Under the hood, it uses the PhraseMatcher to find instances of the phrases.

The factory function takes three arguments: the shared nlp object and component instance name, which are passed in automatically by spaCy, and a case_sensitive config setting that makes the matching and acronym detection case-sensitive.

✏️ Things to try

1. Change the config passed to nlp.add_pipe and set "case_sensitive" to True. You should see that the expanded acronym for "LOL" isn't detected anymore.
2. Add some more terms to the DICTIONARY and update the processed text so they're detected.
3. Add a name argument to nlp.add_pipe to change the component name. Print nlp.pipe_names to see the change reflected in the pipeline.
4. Print the config of the current nlp object with print(nlp.config.to_str()) and inspect the [components] block. You should see an entry for the acronyms component, referencing the factory acronyms and the config settings.

### {executable="true"}
from spacy.language import Language
from spacy.tokens import Doc
from spacy.matcher import PhraseMatcher
import spacy

DICTIONARY = {"lol": "laughing out loud", "brb": "be right back"}
DICTIONARY.update({value: key for key, value in DICTIONARY.items()})

@Language.factory("acronyms", default_config={"case_sensitive": False})
def create_acronym_component(nlp: Language, name: str, case_sensitive: bool):
    return AcronymComponent(nlp, case_sensitive)

class AcronymComponent:
    def __init__(self, nlp: Language, case_sensitive: bool):
        # Create the matcher and match on Token.lower if case-insensitive
        matcher_attr = "TEXT" if case_sensitive else "LOWER"
        self.matcher = PhraseMatcher(nlp.vocab, attr=matcher_attr)
        self.matcher.add("ACRONYMS", [nlp.make_doc(term) for term in DICTIONARY])
        self.case_sensitive = case_sensitive
        # Register custom extension on the Doc
        if not Doc.has_extension("acronyms"):
            Doc.set_extension("acronyms", default=[])

    def __call__(self, doc: Doc) -> Doc:
        # Add the matched spans when doc is processed
        for _, start, end in self.matcher(doc):
            span = doc[start:end]
            acronym = DICTIONARY.get(span.text if self.case_sensitive else span.text.lower())
            doc._.acronyms.append((span, acronym))
        return doc

# Add the component to the pipeline and configure it
nlp = spacy.blank("en")
nlp.add_pipe("acronyms", config={"case_sensitive": False})

# Process a doc and see the results
doc = nlp("LOL, be right back")
print(doc._.acronyms)

Initializing and serializing component data

Many stateful components depend on data resources like dictionaries and lookup tables that should ideally be configurable. For example, it makes sense to make the DICTIONARY in the above example an argument of the registered function, so the AcronymComponent can be re-used with different data. One logical solution would be to make it an argument of the component factory, and allow it to be initialized with different dictionaries.

config.cfg

[components.acronyms.data]
# 🚨 Problem: you don't want the data in the config
lol = "laugh out loud"
brb = "be right back"

@Language.factory("acronyms", default_config={"data": {}, "case_sensitive": False})
def create_acronym_component(nlp: Language, name: str, data: Dict[str, str], case_sensitive: bool):
    # 🚨 Problem: data ends up in the config file
    return AcronymComponent(nlp, data, case_sensitive)

However, passing in the dictionary directly is problematic, because it means that if a component saves out its config and settings, the config.cfg will include a dump of the entire data, since that's the config the component was created with. It will also fail if the data is not JSON-serializable.

Option 1: Using a registered function

• ✅ Pros: can load anything in Python, easy to add to and configure via config
• ❌ Cons: requires the function and its dependencies to be available at runtime

If what you're passing in isn't JSON-serializable – e.g. a custom object like a model – saving out the component config becomes impossible because there's no way for spaCy to know how that object was created, and what to do to create it again. This makes it much harder to save, load and train custom pipelines with custom components. A simple solution is to register a function that returns your resources. The registry lets you map string names to functions that create objects, so given a name and optional arguments, spaCy will know how to recreate the object. To register a function that returns your custom dictionary, you can use the @spacy.registry.misc decorator with a single argument, the name:

What's the misc registry?

The registry provides different categories for different types of functions – for example, model architectures, tokenizers or batchers. misc is intended for miscellaneous functions that don't fit anywhere else.

### Registered function for assets {highlight="1"}
@spacy.registry.misc("acronyms.slang_dict.v1")
def create_acronyms_slang_dict():
    dictionary = {"lol": "laughing out loud", "brb": "be right back"}
    dictionary.update({value: key for key, value in dictionary.items()})
    return dictionary

In your default_config (and later in your training config), you can now refer to the function registered under the name "acronyms.slang_dict.v1" using the @misc key. This tells spaCy how to create the value, and when your component is created, the result of the registered function is passed in as the key "dictionary".

config.cfg

[components.acronyms]
factory = "acronyms"

[components.acronyms.data]
@misc = "acronyms.slang_dict.v1"

- default_config = {"dictionary:" DICTIONARY}
+ default_config = {"dictionary": {"@misc": "acronyms.slang_dict.v1"}}

Using a registered function also means that you can easily include your custom components in pipelines that you train. To make sure spaCy knows where to find your custom @misc function, you can pass in a Python file via the argument --code. If someone else is using your component, all they have to do to customize the data is to register their own function and swap out the name. Registered functions can also take arguments, by the way, that can be defined in the config as well – you can read more about this in the docs on training with custom code.

Option 2: Save data with the pipeline and load it in once on initialization

• ✅ Pros: lets components save and load their own data and reflect user changes, load in data assets before training without depending on them at runtime
• ❌ Cons: requires more component methods, more complex config and data flow

Just like models save out their binary weights when you call nlp.to_disk, components can also serialize any other data assets – for instance, an acronym dictionary. If a pipeline component implements its own to_disk and from_disk methods, those will be called automatically by nlp.to_disk and will receive the path to the directory to save to or load from. The component can then perform any custom saving or loading. If a user makes changes to the component data, they will be reflected when the nlp object is saved. For more examples of this, see the usage guide on serialization methods.

About the data path

The path argument spaCy passes to the serialization methods consists of the path provided by the user, plus a directory of the component name. This means that when you call nlp.to_disk("/path"), the acronyms component will receive the directory path /path/acronyms and can then create files in this directory.

### Custom serialization methods {highlight="6-7,9-11"}
import srsly

class AcronymComponent:
    # other methods here...

    def to_disk(self, path, exclude=tuple()):
        srsly.write_json(path / "data.json", self.data)

    def from_disk(self, path, exclude=tuple()):
        self.data = srsly.read_json(path / "data.json")
        return self

Now the component can save to and load from a directory. The only remaining question: How do you load in the initial data? In Python, you could just call the pipe's from_disk method yourself. But if you're adding the component to your training config, spaCy will need to know how to set it up, from start to finish, including the data to initialize it with.

While you could use a registered function or a file loader like srsly.read_json.v1 as an argument of the component factory, this approach is problematic: the component factory runs every time the component is created. This means it will run when creating the nlp object before training, but also every a user loads your pipeline. So your runtime pipeline would either depend on a local path on your file system, or it's loaded twice: once when the component is created, and then again when the data is by from_disk.

### config.cfg
[components.acronyms.data]
# 🚨 Problem: Runtime pipeline depends on local path
@readers = "srsly.read_json.v1"
path = "/path/to/slang_dict.json"

### config.cfg
[components.acronyms.data]
# 🚨 Problem: this always runs
@misc = "acronyms.slang_dict.v1"

@Language.factory("acronyms", default_config={"data": {}, "case_sensitive": False})
def create_acronym_component(nlp: Language, name: str, data: Dict[str, str], case_sensitive: bool):
    # 🚨 Problem: data will be loaded every time component is created
    return AcronymComponent(nlp, data, case_sensitive)

To solve this, your component can implement a separate method, initialize, which will be called by nlp.initialize if available. This typically happens before training, but not at runtime when the pipeline is loaded. For more background on this, see the usage guides on the config lifecycle and custom initialization.

A component's initialize method needs to take at least two named arguments: a get_examples callback that gives it access to the training examples, and the current nlp object. This is mostly used by trainable components so they can initialize their models and label schemes from the data, so we can ignore those arguments here. All other arguments on the method can be defined via the config – in this case a dictionary data.

config.cfg

[initialize.components.my_component]

[initialize.components.my_component.data]
# ✅ This only runs on initialization
@readers = "srsly.read_json.v1"
path = "/path/to/slang_dict.json"

### Custom initialize method {highlight="5-6"}
class AcronymComponent:
    def __init__(self):
        self.data = {}

    def initialize(self, get_examples=None, nlp=None, data={}):
        self.data = data

When nlp.initialize runs before training (or when you call it in your own code), the [initialize] block of the config is loaded and used to construct the nlp object. The custom acronym component will then be passed the data loaded from the JSON file. After training, the nlp object is saved to disk, which will run the component's to_disk method. When the pipeline is loaded back into spaCy later to use it, the from_disk method will load the data back in.

Python type hints and validation

spaCy's configs are powered by our machine learning library Thinc's configuration system, which supports type hints and even advanced type annotations using pydantic. If your component factory provides type hints, the values that are passed in will be checked against the expected types. If the value can't be cast to an integer, spaCy will raise an error. pydantic also provides strict types like StrictFloat, which will force the value to be an integer and raise an error if it's not – for instance, if your config defines a float.

If you're not using strict types, values that can be cast to the given type will still be accepted. For example, 1 can be cast to a float or a bool type, but not to a List[str]. However, if the type is StrictFloat, only a float will be accepted.

The following example shows a custom pipeline component for debugging. It can be added anywhere in the pipeline and logs information about the nlp object and the Doc that passes through. The log_level config setting lets the user customize what log statements are shown – for instance, "INFO" will show info logs and more critical logging statements, whereas "DEBUG" will show everything. The value is annotated as a StrictStr, so it will only accept a string value.

✏️ Things to try

1. Change the config passed to nlp.add_pipe to use the log level "INFO". You should see that only the statement logged with logger.info is shown.
2. Change the config passed to nlp.add_pipe so that it contains unexpected values – for example, a boolean instead of a string: "log_level": False. You should see a validation error.
3. Check out the docs on pydantic's constrained types and write a type hint for log_level that only accepts the exact string values "DEBUG", "INFO" or "CRITICAL".

### {executable="true"}
import spacy
from spacy.language import Language
from spacy.tokens import Doc
from pydantic import StrictStr
import logging

@Language.factory("debug", default_config={"log_level": "DEBUG"})
class DebugComponent:
    def __init__(self, nlp: Language, name: str, log_level: StrictStr):
        self.logger = logging.getLogger(f"spacy.{name}")
        self.logger.setLevel(log_level)
        self.logger.info(f"Pipeline: {nlp.pipe_names}")

    def __call__(self, doc: Doc) -> Doc:
        is_tagged = doc.has_annotation("TAG")
        self.logger.debug(f"Doc: {len(doc)} tokens, is tagged: {is_tagged}")
        return doc

nlp = spacy.load("en_core_web_sm")
nlp.add_pipe("debug", config={"log_level": "DEBUG"})
doc = nlp("This is a text...")

Trainable components

spaCy's TrainablePipe class helps you implement your own trainable components that have their own model instance, make predictions over Doc objects and can be updated using spacy train. This lets you plug fully custom machine learning components into your pipeline.

You'll need the following:

1. Model: A Thinc Model instance. This can be a model implemented in Thinc, or a wrapped model implemented in PyTorch, TensorFlow, MXNet or a fully custom solution. The model must take a list of Doc objects as input and can have any type of output.
2. TrainablePipe subclass: A subclass of TrainablePipe that implements at least two methods: TrainablePipe.predict and TrainablePipe.set_annotations.
3. Component factory: A component factory registered with @Language.factory that takes the nlp object and component name and optional settings provided by the config and returns an instance of your trainable component.

Example

from spacy.pipeline import TrainablePipe
from spacy.language import Language

class TrainableComponent(TrainablePipe):
    def predict(self, docs):
        ...

    def set_annotations(self, docs, scores):
        ...

@Language.factory("my_trainable_component")
def make_component(nlp, name, model):
    return TrainableComponent(nlp.vocab, model, name=name)

Name	Description
predict	Apply the component's model to a batch of Doc objects (without modifying them) and return the scores.
set_annotations	Modify a batch of Doc objects, using pre-computed scores generated by predict.

By default, TrainablePipe.__init__ takes the shared vocab, the Model and the name of the component instance in the pipeline, which you can use as a key in the losses. All other keyword arguments will become available as TrainablePipe.cfg and will also be serialized with the component.

spaCy's config system resolves the config describing the pipeline components and models bottom-up. This means that it will first create a Model from a registered architecture, validate its arguments and then pass the object forward to the component. This means that the config can express very complex, nested trees of objects – but the objects don't have to pass the model settings all the way down to the components. It also makes the components more modular and lets you swap different architectures in your config, and re-use model definitions.

### config.cfg (excerpt)
[components]

[components.textcat]
factory = "textcat"
labels = []

# This function is created and then passed to the "textcat" component as
# the argument "model"
[components.textcat.model]
@architectures = "spacy.TextCatBOW.v2"
exclusive_classes = true
ngram_size = 1
no_output_layer = false

[components.other_textcat]
factory = "textcat"
# This references the [components.textcat.model] block above
model = ${components.textcat.model}
labels = []

Your trainable pipeline component factories should therefore always take a model argument instead of instantiating the Model inside the component. To register custom architectures, you can use the @spacy.registry.architectures decorator. Also see the training guide for details.

For some use cases, it makes sense to also overwrite additional methods to customize how the model is updated from examples, how it's initialized, how the loss is calculated and to add evaluation scores to the training output.

Name	Description
update	Learn from a batch of Example objects containing the predictions and gold-standard annotations, and update the component's model.
initialize	Initialize the model. Typically calls into Model.initialize and can be passed custom arguments via the [initialize] config block that are only loaded during training or when you call nlp.initialize, not at runtime.
get_loss	Return a tuple of the loss and the gradient for a batch of Example objects.
score	Score a batch of Example objects and return a dictionary of scores. The @Language.factory decorator can define the default_score_weights of the component to decide which keys of the scores to display during training and how they count towards the final score.

For more details on how to implement your own trainable components and model architectures, and plug existing models implemented in PyTorch or TensorFlow into your spaCy pipeline, see the usage guide on layers and model architectures.

Extension attributes

spaCy allows you to set any custom attributes and methods on the Doc, Span and Token, which become available as Doc._, Span._ and Token._ – for example, Token._.my_attr. This lets you store additional information relevant to your application, add new features and functionality to spaCy, and implement your own models trained with other machine learning libraries. It also lets you take advantage of spaCy's data structures and the Doc object as the "single source of truth".

Writing to a ._ attribute instead of to the Doc directly keeps a clearer separation and makes it easier to ensure backwards compatibility. For example, if you've implemented your own .coref property and spaCy claims it one day, it'll break your code. Similarly, just by looking at the code, you'll immediately know what's built-in and what's custom – for example, doc.sentiment is spaCy, while doc._.sent_score isn't.

Extension definitions – the defaults, methods, getters and setters you pass in to set_extension – are stored in class attributes on the Underscore class. If you write to an extension attribute, e.g. doc._.hello = True, the data is stored within the Doc.user_data dictionary. To keep the underscore data separate from your other dictionary entries, the string "._." is placed before the name, in a tuple.

---

There are three main types of extensions, which can be defined using the Doc.set_extension, Span.set_extension and Token.set_extension methods.

Description

1. Attribute extensions. Set a default value for an attribute, which can be overwritten manually at any time. Attribute extensions work like "normal" variables and are the quickest way to store arbitrary information on a Doc, Span or Token.

    Doc.set_extension("hello", default=True)
    assert doc._.hello
    doc._.hello = False
   

2. Property extensions. Define a getter and an optional setter function. If no setter is provided, the extension is immutable. Since the getter and setter functions are only called when you retrieve the attribute, you can also access values of previously added attribute extensions. For example, a Doc getter can average over Token attributes. For Span extensions, you'll almost always want to use a property – otherwise, you'd have to write to every possible Span in the Doc to set up the values correctly.

   Doc.set_extension("hello", getter=get_hello_value, setter=set_hello_value)
   assert doc._.hello
   doc._.hello = "Hi!"
   

3. Method extensions. Assign a function that becomes available as an object method. Method extensions are always immutable. For more details and implementation ideas, see these examples.

   Doc.set_extension("hello", method=lambda doc, name: f"Hi {name}!")
   assert doc._.hello("Bob") == "Hi Bob!"
   

Before you can access a custom extension, you need to register it using the set_extension method on the object you want to add it to, e.g. the Doc. Keep in mind that extensions are always added globally and not just on a particular instance. If an attribute of the same name already exists, or if you're trying to access an attribute that hasn't been registered, spaCy will raise an AttributeError.

### Example
from spacy.tokens import Doc, Span, Token

fruits = ["apple", "pear", "banana", "orange", "strawberry"]
is_fruit_getter = lambda token: token.text in fruits
has_fruit_getter = lambda obj: any([t.text in fruits for t in obj])

Token.set_extension("is_fruit", getter=is_fruit_getter)
Doc.set_extension("has_fruit", getter=has_fruit_getter)
Span.set_extension("has_fruit", getter=has_fruit_getter)

Usage example

doc = nlp("I have an apple and a melon")
assert doc[3]._.is_fruit      # get Token attributes
assert not doc[0]._.is_fruit
assert doc._.has_fruit        # get Doc attributes
assert doc[1:4]._.has_fruit   # get Span attributes

Once you've registered your custom attribute, you can also use the built-in set, get and has methods to modify and retrieve the attributes. This is especially useful it you want to pass in a string instead of calling doc._.my_attr.

Example: Pipeline component for GPE entities and country meta data via a REST API

This example shows the implementation of a pipeline component that fetches country meta data via the REST Countries API, sets entity annotations for countries and sets custom attributes on the Doc and Span – for example, the capital, latitude/longitude coordinates and even the country flag.

### {executable="true"}
import requests
from spacy.lang.en import English
from spacy.language import Language
from spacy.matcher import PhraseMatcher
from spacy.tokens import Doc, Span, Token

@Language.factory("rest_countries")
class RESTCountriesComponent:
    def __init__(self, nlp, name, label="GPE"):
        r = requests.get("https://restcountries.eu/rest/v2/all")
        r.raise_for_status()  # make sure requests raises an error if it fails
        countries = r.json()
        # Convert API response to dict keyed by country name for easy lookup
        self.countries = {c["name"]: c for c in countries}
        self.label = label
        # Set up the PhraseMatcher with Doc patterns for each country name
        self.matcher = PhraseMatcher(nlp.vocab)
        self.matcher.add("COUNTRIES", [nlp.make_doc(c) for c in self.countries.keys()])
        # Register attributes on the Span. We'll be overwriting this based on
        # the matches, so we're only setting a default value, not a getter.
        Span.set_extension("is_country", default=None)
        Span.set_extension("country_capital", default=None)
        Span.set_extension("country_latlng", default=None)
        Span.set_extension("country_flag", default=None)
        # Register attribute on Doc via a getter that checks if the Doc
        # contains a country entity
        Doc.set_extension("has_country", getter=self.has_country)

    def __call__(self, doc):
        spans = []  # keep the spans for later so we can merge them afterwards
        for _, start, end in self.matcher(doc):
            # Generate Span representing the entity & set label
            entity = Span(doc, start, end, label=self.label)
            # Set custom attributes on entity. Can be extended with other data
            # returned by the API, like currencies, country code, calling code etc.
            entity._.set("is_country", True)
            entity._.set("country_capital", self.countries[entity.text]["capital"])
            entity._.set("country_latlng", self.countries[entity.text]["latlng"])
            entity._.set("country_flag", self.countries[entity.text]["flag"])
            spans.append(entity)
        # Overwrite doc.ents and add entity – be careful not to replace!
        doc.ents = list(doc.ents) + spans
        return doc  # don't forget to return the Doc!

    def has_country(self, doc):
        """Getter for Doc attributes. Since the getter is only called
        when we access the attribute, we can refer to the Span's 'is_country'
        attribute here, which is already set in the processing step."""
        return any([entity._.get("is_country") for entity in doc.ents])

nlp = English()
nlp.add_pipe("rest_countries", config={"label": "GPE"})
doc = nlp("Some text about Colombia and the Czech Republic")
print("Pipeline", nlp.pipe_names)  # pipeline contains component name
print("Doc has countries", doc._.has_country)  # Doc contains countries
for ent in doc.ents:
    if ent._.is_country:
        print(ent.text, ent.label_, ent._.country_capital, ent._.country_latlng, ent._.country_flag)

In this case, all data can be fetched on initialization in one request. However, if you're working with text that contains incomplete country names, spelling mistakes or foreign-language versions, you could also implement a like_country-style getter function that makes a request to the search API endpoint and returns the best-matching result.

User hooks

While it's generally recommended to use the Doc._, Span._ and Token._ proxies to add your own custom attributes, spaCy offers a few exceptions to allow customizing the built-in methods like Doc.similarity or Doc.vector with your own hooks, which can rely on components you train yourself. For instance, you can provide your own on-the-fly sentence segmentation algorithm or document similarity method.

Hooks let you customize some of the behaviors of the Doc, Span or Token objects by adding a component to the pipeline. For instance, to customize the Doc.similarity method, you can add a component that sets a custom function to doc.user_hooks["similarity"]. The built-in Doc.similarity method will check the user_hooks dict, and delegate to your function if you've set one. Similar results can be achieved by setting functions to Doc.user_span_hooks and Doc.user_token_hooks.

Implementation note

The hooks live on the Doc object because the Span and Token objects are created lazily, and don't own any data. They just proxy to their parent Doc. This turns out to be convenient here – we only have to worry about installing hooks in one place.

Name	Customizes
user_hooks	Doc.similarity, Doc.vector, Doc.has_vector, Doc.vector_norm, Doc.sents
user_token_hooks	Token.similarity, Token.vector, Token.has_vector, Token.vector_norm, Token.conjuncts
user_span_hooks	Span.similarity, Span.vector, Span.has_vector, Span.vector_norm, Span.root

### Add custom similarity hooks
from spacy.language import Language


class SimilarityModel:
    def __init__(self, name: str, index: int):
        self.name = name
        self.index = index

    def __call__(self, doc):
        doc.user_hooks["similarity"] = self.similarity
        doc.user_span_hooks["similarity"] = self.similarity
        doc.user_token_hooks["similarity"] = self.similarity
        return doc

    def similarity(self, obj1, obj2):
        return obj1.vector[self.index] + obj2.vector[self.index]


@Language.factory("similarity_component", default_config={"index": 0})
def create_similarity_component(nlp, name, index: int):
    return SimilarityModel(name, index)

Developing plugins and wrappers

We're very excited about all the new possibilities for community extensions and plugins in spaCy, and we can't wait to see what you build with it! To get you started, here are a few tips, tricks and best practices. See here for examples of other spaCy extensions.

Usage ideas

• Adding new features and hooking in models. For example, a sentiment analysis model, or your preferred solution for lemmatization or sentiment analysis. spaCy's built-in tagger, parser and entity recognizer respect annotations that were already set on the Doc in a previous step of the pipeline.
• Integrating other libraries and APIs. For example, your pipeline component can write additional information and data directly to the Doc or Token as custom attributes, while making sure no information is lost in the process. This can be output generated by other libraries and models, or an external service with a REST API.
• Debugging and logging. For example, a component which stores and/or exports relevant information about the current state of the processed document, and insert it at any point of your pipeline.

Best practices

Extensions can claim their own ._ namespace and exist as standalone packages. If you're developing a tool or library and want to make it easy for others to use it with spaCy and add it to their pipeline, all you have to do is expose a function that takes a Doc, modifies it and returns it.

• Make sure to choose a descriptive and specific name for your pipeline component class, and set it as its name attribute. Avoid names that are too common or likely to clash with built-in or a user's other custom components. While it's fine to call your package "spacy_my_extension", avoid component names including "spacy", since this can easily lead to confusion.

  + name = "myapp_lemmatizer"
  - name = "lemmatizer"
  

• When writing to Doc, Token or Span objects, use getter functions wherever possible, and avoid setting values explicitly. Tokens and spans don't own any data themselves, and they're implemented as C extension classes – so you can't usually add new attributes to them like you could with most pure Python objects.

  + is_fruit = lambda token: token.text in ("apple", "orange")
  + Token.set_extension("is_fruit", getter=is_fruit)
  
  - token._.set_extension("is_fruit", default=False)
  - if token.text in ('"apple", "orange"):
  -     token._.set("is_fruit", True)
  

• Always add your custom attributes to the global Doc, Token or Span objects, not a particular instance of them. Add the attributes as early as possible, e.g. in your extension's __init__ method or in the global scope of your module. This means that in the case of namespace collisions, the user will see an error immediately, not just when they run their pipeline.

  + from spacy.tokens import Doc
  + def __init__(attr="my_attr"):
  +     Doc.set_extension(attr, getter=self.get_doc_attr)
  
  - def __call__(doc):
  -     doc.set_extension("my_attr", getter=self.get_doc_attr)
  

• If your extension is setting properties on the Doc, Token or Span, include an option to let the user to change those attribute names. This makes it easier to avoid namespace collisions and accommodate users with different naming preferences. We recommend adding an attrs argument to the __init__ method of your class so you can write the names to class attributes and reuse them across your component.

  + Doc.set_extension(self.doc_attr, default="some value")
  - Doc.set_extension("my_doc_attr", default="some value")
  

• Ideally, extensions should be standalone packages with spaCy and optionally, other packages specified as a dependency. They can freely assign to their own ._ namespace, but should stick to that. If your extension's only job is to provide a better .similarity implementation, and your docs state this explicitly, there's no problem with writing to the user_hooks and overwriting spaCy's built-in method. However, a third-party extension should never silently overwrite built-ins, or attributes set by other extensions.

• If you're looking to publish a pipeline package that depends on a custom pipeline component, you can either require it in the package's dependencies, or – if the component is specific and lightweight – choose to ship it with your pipeline package. Just make sure the @Language.component or @Language.factory decorator that registers the custom component runs in your package's __init__.py or is exposed via an entry point.

• Once you're ready to share your extension with others, make sure to add docs and installation instructions (you can always link to this page for more info). Make it easy for others to install and use your extension, for example by uploading it to PyPi. If you're sharing your code on GitHub, don't forget to tag it with spacy and spacy-extension to help people find it. If you post it on Twitter, feel free to tag @spacy_io so we can check it out.

Wrapping other models and libraries

Let's say you have a custom entity recognizer that takes a list of strings and returns their BILUO tags. Given an input like ["A", "text", "about", "Facebook"], it will predict and return ["O", "O", "O", "U-ORG"]. To integrate it into your spaCy pipeline and make it add those entities to the doc.ents, you can wrap it in a custom pipeline component function and pass it the token texts from the Doc object received by the component.

The training.biluo_tags_to_spans is very helpful here, because it takes a Doc object and token-based BILUO tags and returns a sequence of Span objects in the Doc with added labels. So all your wrapper has to do is compute the entity spans and overwrite the doc.ents.

How the doc.ents work

When you add spans to the doc.ents, spaCy will automatically resolve them back to the underlying tokens and set the Token.ent_type and Token.ent_iob attributes. By definition, each token can only be part of one entity, so overlapping entity spans are not allowed.

### {highlight="1,8-9"}
import your_custom_entity_recognizer
from spacy.training import biluo_tags_to_spans
from spacy.language import Language

@Language.component("custom_ner_wrapper")
def custom_ner_wrapper(doc):
    words = [token.text for token in doc]
    custom_entities = your_custom_entity_recognizer(words)
    doc.ents = biluo_tags_to_spans(doc, custom_entities)
    return doc

The custom_ner_wrapper can then be added to a blank pipeline using nlp.add_pipe. You can also replace the existing entity recognizer of a trained pipeline with nlp.replace_pipe.

Here's another example of a custom model, your_custom_model, that takes a list of tokens and returns lists of fine-grained part-of-speech tags, coarse-grained part-of-speech tags, dependency labels and head token indices. Here, we can use the Doc.from_array to create a new Doc object using those values. To create a numpy array we need integers, so we can look up the string labels in the StringStore. The doc.vocab.strings.add method comes in handy here, because it returns the integer ID of the string and makes sure it's added to the vocab. This is especially important if the custom model uses a different label scheme than spaCy's default models.

Example: spacy-stanza

For an example of an end-to-end wrapper for statistical tokenization, tagging and parsing, check out spacy-stanza. It uses a very similar approach to the example in this section – the only difference is that it fully replaces the nlp object instead of providing a pipeline component, since it also needs to handle tokenization.

### {highlight="1,11,17-19"}
import your_custom_model
from spacy.language import Language
from spacy.symbols import POS, TAG, DEP, HEAD
from spacy.tokens import Doc
import numpy

@Language.component("custom_model_wrapper")
def custom_model_wrapper(doc):
    words = [token.text for token in doc]
    spaces = [token.whitespace for token in doc]
    pos, tags, deps, heads = your_custom_model(words)
    # Convert the strings to integers and add them to the string store
    pos = [doc.vocab.strings.add(label) for label in pos]
    tags = [doc.vocab.strings.add(label) for label in tags]
    deps = [doc.vocab.strings.add(label) for label in deps]
    # Create a new Doc from a numpy array
    attrs = [POS, TAG, DEP, HEAD]
    arr = numpy.array(list(zip(pos, tags, deps, heads)), dtype="uint64")
    new_doc = Doc(doc.vocab, words=words, spaces=spaces).from_array(attrs, arr)
    return new_doc

If you create a Doc object with dependencies and heads, spaCy is able to resolve the sentence boundaries automatically. However, note that the HEAD value used to construct a Doc is the token index relative to the current token – e.g. -1 for the previous token. The CoNLL format typically annotates heads as 1-indexed absolute indices with 0 indicating the root. If that's the case in your annotations, you need to convert them first:

heads = [2, 0, 4, 2, 2]
new_heads = [head - i - 1 if head != 0 else 0 for i, head in enumerate(heads)]

For more details on how to write and package custom components, make them available to spaCy via entry points and implement your own serialization methods, check out the usage guide on saving and loading.