spaCy/website/docs/usage/adding-languages.md

title	next	menu
Adding Languages	/usage/training	Language Data language-data Testing testing Training training

Adding full support for a language touches many different parts of the spaCy library. This guide explains how to fit everything together, and points you to the specific workflows for each component.

Working on spaCy's source

To add a new language to spaCy, you'll need to modify the library's code. The easiest way to do this is to clone the repository and build spaCy from source. For more information on this, see the installation guide. Unlike spaCy's core, which is mostly written in Cython, all language data is stored in regular Python files. This means that you won't have to rebuild anything in between – you can simply make edits and reload spaCy to test them.

Obviously, there are lots of ways you can organize your code when you implement your own language data. This guide will focus on how it's done within spaCy. For full language support, you'll need to create a Language subclass, define custom language data, like a stop list and tokenizer exceptions and test the new tokenizer. Once the language is set up, you can build the vocabulary, including word frequencies, Brown clusters and word vectors. Finally, you can train the tagger and parser, and save the model to a directory.

For some languages, you may also want to develop a solution for lemmatization and morphological analysis.

• Language data 101
• The Language subclass
• Stop words
• Tokenizer exceptions
• Norm exceptions
• Lexical attributes
• Syntax iterators
• Lemmatizer
• Tag map
• Morph rules
• Testing the language
• Training

Language data

import LanguageData101 from 'usage/101/_language-data.md'

The individual components expose variables that can be imported within a language module, and added to the language's Defaults. Some components, like the punctuation rules, usually don't need much customization and can be imported from the global rules. Others, like the tokenizer and norm exceptions, are very specific and will make a big difference to spaCy's performance on the particular language and training a language model.

Variable	Type	Description
STOP_WORDS	set	Individual words.
TOKENIZER_EXCEPTIONS	dict	Keyed by strings mapped to list of one dict per token with token attributes.
TOKEN_MATCH	regex	Regexes to match complex tokens, e.g. URLs.
NORM_EXCEPTIONS	dict	Keyed by strings, mapped to their norms.
TOKENIZER_PREFIXES	list	Strings or regexes, usually not customized.
TOKENIZER_SUFFIXES	list	Strings or regexes, usually not customized.
TOKENIZER_INFIXES	list	Strings or regexes, usually not customized.
LEX_ATTRS	dict	Attribute ID mapped to function.
SYNTAX_ITERATORS	dict	Iterator ID mapped to function. Currently only supports 'noun_chunks'.
LOOKUP	dict	Keyed by strings mapping to their lemma.
LEMMA_RULES, LEMMA_INDEX, LEMMA_EXC	dict	Lemmatization rules, keyed by part of speech.
TAG_MAP	dict	Keyed by strings mapped to Universal Dependencies tags.
MORPH_RULES	dict	Keyed by strings mapped to a dict of their morphological features.

Should I ever update the global data?

Reusable language data is collected as atomic pieces in the root of the spacy.lang module. Often, when a new language is added, you'll find a pattern or symbol that's missing. Even if it isn't common in other languages, it might be best to add it to the shared language data, unless it has some conflicting interpretation. For instance, we don't expect to see guillemot quotation symbols (» and «) in English text. But if we do see them, we'd probably prefer the tokenizer to split them off.

In order for the tokenizer to split suffixes, prefixes and infixes, spaCy needs to know the language's character set. If the language you're adding uses non-latin characters, you might need to define the required character classes in the global char_classes.py. For efficiency, spaCy uses hard-coded unicode ranges to define character classes, the definitions of which can be found on Wikipedia. If the language requires very specific punctuation rules, you should consider overwriting the default regular expressions with your own in the language's Defaults.

Creating a Language subclass

Language-specific code and resources should be organized into a sub-package of spaCy, named according to the language's ISO code. For instance, code and resources specific to Spanish are placed into a directory spacy/lang/es, which can be imported as spacy.lang.es.

To get started, you can use our templates for the most important files. Here's what the class template looks like:

### __init__.py (excerpt)
# import language-specific data
from .stop_words import STOP_WORDS
from .tokenizer_exceptions import TOKENIZER_EXCEPTIONS
from .lex_attrs import LEX_ATTRS

from ..tokenizer_exceptions import BASE_EXCEPTIONS
from ...language import Language
from ...attrs import LANG
from ...util import update_exc

# Create Defaults class in the module scope (necessary for pickling!)
class XxxxxDefaults(Language.Defaults):
    lex_attr_getters = dict(Language.Defaults.lex_attr_getters)
    lex_attr_getters[LANG] = lambda text: "xx" # language ISO code

    # Optional: replace flags with custom functions, e.g. like_num()
    lex_attr_getters.update(LEX_ATTRS)

    # Merge base exceptions and custom tokenizer exceptions
    tokenizer_exceptions = update_exc(BASE_EXCEPTIONS, TOKENIZER_EXCEPTIONS)
    stop_words = STOP_WORDS

# Create actual Language class
class Xxxxx(Language):
    lang = "xx" # Language ISO code
    Defaults = XxxxxDefaults # Override defaults

# Set default export – this allows the language class to be lazy-loaded
__all__ = ["Xxxxx"]

Some languages contain large volumes of custom data, like lemmatizer lookup tables, or complex regular expression that are expensive to compute. As of spaCy v2.0, Language classes are not imported on initialization and are only loaded when you import them directly, or load a model that requires a language to be loaded. To lazy-load languages in your application, you can use the util.get_lang_class helper function with the two-letter language code as its argument.

Stop words

A "stop list" is a classic trick from the early days of information retrieval when search was largely about keyword presence and absence. It is still sometimes useful today to filter out common words from a bag-of-words model. To improve readability, STOP_WORDS are separated by spaces and newlines, and added as a multiline string.

What does spaCy consider a stop word?

There's no particularly principled logic behind what words should be added to the stop list. Make a list that you think might be useful to people and is likely to be unsurprising. As a rule of thumb, words that are very rare are unlikely to be useful stop words.

### Example
STOP_WORDS = set("""
a about above across after afterwards again against all almost alone along
already also although always am among amongst amount an and another any anyhow
anyone anything anyway anywhere are around as at

back be became because become becomes becoming been before beforehand behind
being below beside besides between beyond both bottom but by
""".split())

When adding stop words from an online source, always include the link in a comment. Make sure to proofread and double-check the words carefully. A lot of the lists available online have been passed around for years and often contain mistakes, like unicode errors or random words that have once been added for a specific use case, but don't actually qualify.

Tokenizer exceptions

spaCy's tokenization algorithm lets you deal with whitespace-delimited chunks separately. This makes it easy to define special-case rules, without worrying about how they interact with the rest of the tokenizer. Whenever the key string is matched, the special-case rule is applied, giving the defined sequence of tokens. You can also attach attributes to the subtokens, covered by your special case, such as the subtokens LEMMA or TAG.

Tokenizer exceptions can be added in the following format:

### tokenizer_exceptions.py (excerpt)
TOKENIZER_EXCEPTIONS = {
    "don't": [
        {ORTH: "do", LEMMA: "do"},
        {ORTH: "n't", LEMMA: "not", NORM: "not", TAG: "RB"}]
}

If an exception consists of more than one token, the ORTH values combined always need to match the original string. The way the original string is split up can be pretty arbitrary sometimes – for example "gonna" is split into "gon" (lemma "go") and "na" (lemma "to"). Because of how the tokenizer works, it's currently not possible to split single-letter strings into multiple tokens.

Unambiguous abbreviations, like month names or locations in English, should be added to exceptions with a lemma assigned, for example {ORTH: "Jan.", LEMMA: "January"}. Since the exceptions are added in Python, you can use custom logic to generate them more efficiently and make your data less verbose. How you do this ultimately depends on the language. Here's an example of how exceptions for time formats like "1a.m." and "1am" are generated in the English tokenizer_exceptions.py:

### tokenizer_exceptions.py (excerpt)
# use short, internal variable for readability
_exc = {}

for h in range(1, 12 + 1):
    for period in ["a.m.", "am"]:
        # always keep an eye on string interpolation!
        _exc["%d%s" % (h, period)] = [
            {ORTH: "%d" % h},
            {ORTH: period, LEMMA: "a.m."}]
    for period in ["p.m.", "pm"]:
        _exc["%d%s" % (h, period)] = [
            {ORTH: "%d" % h},
            {ORTH: period, LEMMA: "p.m."}]

# only declare this at the bottom
TOKENIZER_EXCEPTIONS = _exc

Generating tokenizer exceptions

Keep in mind that generating exceptions only makes sense if there's a clearly defined and finite number of them, like common contractions in English. This is not always the case – in Spanish for instance, infinitive or imperative reflexive verbs and pronouns are one token (e.g. "vestirme"). In cases like this, spaCy shouldn't be generating exceptions for all verbs. Instead, this will be handled at a later stage during lemmatization.

When adding the tokenizer exceptions to the Defaults, you can use the update_exc helper function to merge them with the global base exceptions (including one-letter abbreviations and emoticons). The function performs a basic check to make sure exceptions are provided in the correct format. It can take any number of exceptions dicts as its arguments, and will update and overwrite the exception in this order. For example, if your language's tokenizer exceptions include a custom tokenization pattern for "a.", it will overwrite the base exceptions with the language's custom one.

### Example
from ...util import update_exc

BASE_EXCEPTIONS =  {"a.": [{ORTH: "a."}], ":)": [{ORTH: ":)"}]}
TOKENIZER_EXCEPTIONS = {"a.": [{ORTH: "a.", LEMMA: "all"}]}

tokenizer_exceptions = update_exc(BASE_EXCEPTIONS, TOKENIZER_EXCEPTIONS)
# {"a.": [{ORTH: "a.", LEMMA: "all"}], ":)": [{ORTH: ":)"}]}

Unlike verbs and common nouns, there's no clear base form of a personal pronoun. Should the lemma of "me" be "I", or should we normalize person as well, giving "it" — or maybe "he"? spaCy's solution is to introduce a novel symbol, -PRON-, which is used as the lemma for all personal pronouns.

Norm exceptions

In addition to ORTH or LEMMA, tokenizer exceptions can also set a NORM attribute. This is useful to specify a normalized version of the token – for example, the norm of "n't" is "not". By default, a token's norm equals its lowercase text. If the lowercase spelling of a word exists, norms should always be in lowercase.

Norms vs. lemmas

doc = nlp(u"I'm gonna realise")
norms = [token.norm_ for token in doc]
lemmas = [token.lemma_ for token in doc]
assert norms == ["i", "am", "going", "to", "realize"]
assert lemmas == ["i", "be", "go", "to", "realise"]

spaCy usually tries to normalize words with different spellings to a single, common spelling. This has no effect on any other token attributes, or tokenization in general, but it ensures that equivalent tokens receive similar representations. This can improve the model's predictions on words that weren't common in the training data, but are equivalent to other words – for example, "realize" and "realize", or "thx" and "thanks".

Similarly, spaCy also includes global base norms for normalizing different styles of quotation marks and currency symbols. Even though $ and € are very different, spaCy normalizes them both to $. This way, they'll always be seen as similar, no matter how common they were in the training data.

Norm exceptions can be provided as a simple dictionary. For more examples, see the English norm_exceptions.py.

### Example
NORM_EXCEPTIONS = {
    "cos": "because",
    "fav": "favorite",
    "accessorise": "accessorize",
    "accessorised": "accessorized"
}

To add the custom norm exceptions lookup table, you can use the add_lookups() helper functions. It takes the default attribute getter function as its first argument, plus a variable list of dictionaries. If a string's norm is found in one of the dictionaries, that value is used – otherwise, the default function is called and the token is assigned its default norm.

lex_attr_getters[NORM] = add_lookups(Language.Defaults.lex_attr_getters[NORM],
                                     NORM_EXCEPTIONS, BASE_NORMS)

The order of the dictionaries is also the lookup order – so if your language's norm exceptions overwrite any of the global exceptions, they should be added first. Also note that the tokenizer exceptions will always have priority over the attribute getters.

Lexical attributes

spaCy provides a range of Token attributes that return useful information on that token – for example, whether it's uppercase or lowercase, a left or right punctuation mark, or whether it resembles a number or email address. Most of these functions, like is_lower or like_url should be language-independent. Others, like like_num (which includes both digits and number words), requires some customization.

Best practices

Keep in mind that those functions are only intended to be an approximation. It's always better to prioritize simplicity and performance over covering very specific edge cases.

English number words are pretty simple, because even large numbers consist of individual tokens, and we can get away with splitting and matching strings against a list. In other languages, like German, "two hundred and thirty-four" is one word, and thus one token. Here, it's best to match a string against a list of number word fragments (instead of a technically almost infinite list of possible number words).

Here's an example from the English lex_attrs.py:

### lex_attrs.py
_num_words = ["zero", "one", "two", "three", "four", "five", "six", "seven",
              "eight", "nine", "ten", "eleven", "twelve", "thirteen", "fourteen",
              "fifteen", "sixteen", "seventeen", "eighteen", "nineteen", "twenty",
              "thirty", "forty", "fifty", "sixty", "seventy", "eighty", "ninety",
              "hundred", "thousand", "million", "billion", "trillion", "quadrillion",
              "gajillion", "bazillion"]

def like_num(text):
    text = text.replace(",", "").replace(".", "")
    if text.isdigit():
        return True
    if text.count("/") == 1:
        num, denom = text.split("/")
        if num.isdigit() and denom.isdigit():
            return True
    if text.lower() in _num_words:
        return True
    return False

LEX_ATTRS = {
    LIKE_NUM: like_num
}

By updating the default lexical attributes with a custom LEX_ATTRS dictionary in the language's defaults via lex_attr_getters.update(LEX_ATTRS), only the new custom functions are overwritten.

Syntax iterators

Syntax iterators are functions that compute views of a Doc object based on its syntax. At the moment, this data is only used for extracting noun chunks, which are available as the Doc.noun_chunks property. Because base noun phrases work differently across languages, the rules to compute them are part of the individual language's data. If a language does not include a noun chunks iterator, the property won't be available. For examples, see the existing syntax iterators:

Noun chunks example

doc = nlp(u"A phrase with another phrase occurs.")
chunks = list(doc.noun_chunks)
assert chunks[0].text == u"A phrase"
assert chunks[1].text == u"another phrase"

Language	Code	Source
English	en	lang/en/syntax_iterators.py
German	de	lang/de/syntax_iterators.py
French	fr	lang/fr/syntax_iterators.py
Spanish	es	lang/es/syntax_iterators.py

Lemmatizer

As of v2.0, spaCy supports simple lookup-based lemmatization. This is usually the quickest and easiest way to get started. The data is stored in a dictionary mapping a string to its lemma. To determine a token's lemma, spaCy simply looks it up in the table. Here's an example from the Spanish language data:

### lang/es/lemmatizer.py (excerpt)
LOOKUP = {
    "aba": "abar",
    "ababa": "abar",
    "ababais": "abar",
    "ababan": "abar",
    "ababanes": "ababán",
    "ababas": "abar",
    "ababoles": "ababol",
    "ababábites": "ababábite"
}

To provide a lookup lemmatizer for your language, import the lookup table and add it to the Language class as lemma_lookup:

lemma_lookup = LOOKUP

Tag map

Most treebanks define a custom part-of-speech tag scheme, striking a balance between level of detail and ease of prediction. While it's useful to have custom tagging schemes, it's also useful to have a common scheme, to which the more specific tags can be related. The tagger can learn a tag scheme with any arbitrary symbols. However, you need to define how those symbols map down to the Universal Dependencies tag set. This is done by providing a tag map.

The keys of the tag map should be strings in your tag set. The values should be a dictionary. The dictionary must have an entry POS whose value is one of the Universal Dependencies tags. Optionally, you can also include morphological features or other token attributes in the tag map as well. This allows you to do simple rule-based morphological analysis.

### Example
from ..symbols import POS, NOUN, VERB, DET

TAG_MAP = {
    "NNS":  {POS: NOUN, "Number": "plur"},
    "VBG":  {POS: VERB, "VerbForm": "part", "Tense": "pres", "Aspect": "prog"},
    "DT":   {POS: DET}
}

Morph rules

The morphology rules let you set token attributes such as lemmas, keyed by the extended part-of-speech tag and token text. The morphological features and their possible values are language-specific and based on the Universal Dependencies scheme.

### Example
from ..symbols import LEMMA

MORPH_RULES = {
    "VBZ": {
        "am": {LEMMA: "be", "VerbForm": "Fin", "Person": "One", "Tense": "Pres", "Mood": "Ind"},
        "are": {LEMMA: "be", "VerbForm": "Fin", "Person": "Two", "Tense": "Pres", "Mood": "Ind"},
        "is": {LEMMA: "be", "VerbForm": "Fin", "Person": "Three", "Tense": "Pres", "Mood": "Ind"},
        "'re": {LEMMA: "be", "VerbForm": "Fin", "Person": "Two", "Tense": "Pres", "Mood": "Ind"},
        "'s": {LEMMA: "be", "VerbForm": "Fin", "Person": "Three", "Tense": "Pres", "Mood": "Ind"}
    }
}

In the example of "am", the attributes look like this:

Attribute	Description
LEMMA: "be"	Base form, e.g. "to be".
"VerbForm": "Fin"	Finite verb. Finite verbs have a subject and can be the root of an independent clause – "I am." is a valid, complete sentence.
"Person": "One"	First person, i.e. "I am".
"Tense": "Pres"	Present tense, i.e. actions that are happening right now or actions that usually happen.
"Mood": "Ind"	Indicative, i.e. something happens, has happened or will happen (as opposed to imperative or conditional).

The morphological attributes are currently not all used by spaCy. Full integration is still being developed. In the meantime, it can still be useful to add them, especially if the language you're adding includes important distinctions and special cases. This ensures that as soon as full support is introduced, your language will be able to assign all possible attributes.

Testing the new language

Before using the new language or submitting a pull request to spaCy, you should make sure it works as expected. This is especially important if you've added custom regular expressions for token matching or punctuation – you don't want to be causing regressions.

spaCy uses the pytest framework for testing. For more details on how the tests are structured and best practices for writing your own tests, see our tests documentation.

Writing language-specific tests

It's recommended to always add at least some tests with examples specific to the language. Language tests should be located in tests/lang in a directory named after the language ID. You'll also need to create a fixture for your tokenizer in the conftest.py. Always use the get_lang_class helper function within the fixture, instead of importing the class at the top of the file. This will load the language data only when it's needed. (Otherwise, all data would be loaded every time you run a test.)

@pytest.fixture
def en_tokenizer():
    return util.get_lang_class("en").Defaults.create_tokenizer()

When adding test cases, always parametrize them – this will make it easier for others to add more test cases without having to modify the test itself. You can also add parameter tuples, for example, a test sentence and its expected length, or a list of expected tokens. Here's an example of an English tokenizer test for combinations of punctuation and abbreviations:

### Example test
@pytest.mark.parametrize('text,length', [
    ("The U.S. Army likes Shock and Awe.", 8),
    ("U.N. regulations are not a part of their concern.", 10),
    ("“Isn't it?”", 6)])
def test_en_tokenizer_handles_punct_abbrev(en_tokenizer, text, length):
    tokens = en_tokenizer(text)
    assert len(tokens) == length

Training a language model

Much of spaCy's functionality requires models to be trained from labeled data. For instance, in order to use the named entity recognizer, you need to first train a model on text annotated with examples of the entities you want to recognize. The parser, part-of-speech tagger and text categorizer all also require models to be trained from labeled examples. The word vectors, word probabilities and word clusters also require training, although these can be trained from unlabeled text, which tends to be much easier to collect.

Creating a vocabulary file

spaCy expects that common words will be cached in a Vocab instance. The vocabulary caches lexical features. spaCy loads the vocabulary from binary data, in order to keep loading efficient. The easiest way to save out a new binary vocabulary file is to use the spacy init-model command, which expects a JSONL file with words and their lexical attributes. See the docs on the vocab JSONL format for details.

Training the word vectors

Word2vec and related algorithms let you train useful word similarity models from unlabeled text. This is a key part of using deep learning for NLP with limited labeled data. The vectors are also useful by themselves – they power the .similarity methods in spaCy. For best results, you should pre-process the text with spaCy before training the Word2vec model. This ensures your tokenization will match. You can use our word vectors training script, which pre-processes the text with your language-specific tokenizer and trains the model using Gensim. The vectors.bin file should consist of one word and vector per line.

https://github.com/explosion/spacy-dev-resources/tree/master/training/word_vectors.py

If you don't have a large sample of text available, you can also convert word vectors produced by a variety of other tools into spaCy's format. See the docs on converting word vectors for details.

Creating or converting a training corpus

The easiest way to train spaCy's tagger, parser, entity recognizer or text categorizer is to use the spacy train command-line utility. In order to use this, you'll need training and evaluation data in the JSON format spaCy expects for training.

You can now train the model using a corpus for your language annotated with If your data is in one of the supported formats, the easiest solution might be to use the spacy convert command-line utility. This supports several popular formats, including the IOB format for named entity recognition, the JSONL format produced by our annotation tool Prodigy, and the CoNLL-U format used by the Universal Dependencies corpus.

One thing to keep in mind is that spaCy expects to train its models from whole documents, not just single sentences. If your corpus only contains single sentences, spaCy's models will never learn to expect multi-sentence documents, leading to low performance on real text. To mitigate this problem, you can use the -N argument to the spacy convert command, to merge some of the sentences into longer pseudo-documents.

Training the tagger and parser

Once you have your training and evaluation data in the format spaCy expects, you can train your model use the using spaCy's train command. Note that training statistical models still involves a degree of trial-and-error. You may need to tune one or more settings, also called "hyper-parameters", to achieve optimal performance. See the usage guide on training for more details.