Contents:

Langcodes: a library for language codes

langcodes knows what languages are. It knows the standardized codes that refer to them, such as en for English, es for Spanish and hi for Hindi. Often, it knows what these languages are called in a language, and that language doesn’t have to be English.

It may sound to you like langcodes solves a pretty boring problem. At one level, that’s right. Sometimes you have a boring problem, and it’s great when a library solves it for you.

But there’s an interesting problem hiding in here. How do you work with language codes? How do you know when two different codes represent the same thing? How should your code represent relationships between codes, like the following?

  • eng is equivalent to en.

  • fra and fre are both equivalent to fr.

  • en-GB might be written as en-gb or en_GB. Or as ‘en-UK’, which is erroneous, but should be treated as the same.

  • en-CA is not exactly equivalent to en-US, but it’s really, really close.

  • en-Latn-US is equivalent to en-US, because written English must be written in the Latin alphabet to be understood.

  • The difference between ar and arb is the difference between “Arabic” and “Modern Standard Arabic”, a difference that may not be relevant to you.

  • You’ll find Mandarin Chinese tagged as cmn on Wiktionary, but many other resources would call the same language zh.

  • Chinese is written in different scripts in different territories. Some software distinguishes the script. Other software distinguishes the territory. The result is that zh-CN and zh-Hans are used interchangeably, as are zh-TW and zh-Hant, even though occasionally you’ll need something different such as zh-HK or zh-Latn-pinyin.

  • The Indonesian (id) and Malaysian (ms or zsm) languages are mutually intelligible.

  • jp is not a language code. (The language code for Japanese is ja, but people confuse it with the country code for Japan.)

One way to know is to read IETF standards and Unicode technical reports. Another way is to use a library that implements those standards and guidelines for you, which langcodes does.

langcodes is maintained by Robyn Speer, and is released as free software under the MIT license.

Standards implemented

Although this is not the only reason to use it, langcodes will make you more acronym-compliant.

langcodes implements BCP 47, the IETF Best Current Practices on Tags for Identifying Languages. BCP 47 is also known as RFC 5646. It subsumes standards such as ISO 639, and it also implements recommendations from the Unicode CLDR.

The package also comes with a database of language properties and names, built from Unicode CLDR and the IANA subtag registry.

In summary, langcodes takes language codes and does the Right Thing with them, and if you want to know exactly what the Right Thing is, there are some documents you can go read.

Documentation

Standardizing language tags

This function standardizes tags, as strings, in several ways.

It replaces overlong tags with their shortest version, and also formats them according to the conventions of BCP 47:

>>> from langcodes import *
>>> standardize_tag('eng_US')
'en-US'

It removes script subtags that are redundant with the language:

>>> standardize_tag('en-Latn')
'en'

It replaces deprecated values with their correct versions, if possible:

>>> standardize_tag('en-uk')
'en-GB'

Sometimes this involves complex substitutions, such as replacing Serbo-Croatian (sh) with Serbian in Latin script (sr-Latn), or the entire tag sgn-US with ase (American Sign Language).

>>> standardize_tag('sh-QU')
'sr-Latn-EU'

>>> standardize_tag('sgn-US')
'ase'

If macro is True, it uses macrolanguage codes as a replacement for the most common standardized language within that macrolanguage.

>>> standardize_tag('arb-Arab', macro=True)
'ar'

Even when macro is False, it shortens tags that contain both the macrolanguage and the language:

>>> standardize_tag('zh-cmn-hans-cn')
'zh-Hans-CN'

If the tag can’t be parsed according to BCP 47, this will raise a LanguageTagError (a subclass of ValueError):

>>> standardize_tag('spa-latn-mx')
'es-MX'

>>> standardize_tag('spa-mx-latn')
Traceback (most recent call last):
    ...
langcodes.tag_parser.LanguageTagError: This script subtag, 'latn', is out of place. Expected variant, extension, or end of string.

Language objects

This package defines one class, named Language, which contains the results of parsing a language tag. Language objects have the following fields, any of which may be unspecified:

  • language: the code for the language itself.

  • script: the 4-letter code for the writing system being used.

  • territory: the 2-letter or 3-digit code for the country or similar region whose usage of the language appears in this text.

  • extlangs: a list of more specific language codes that follow the language code. (This is allowed by the language code syntax, but deprecated.)

  • variants: codes for specific variations of language usage that aren’t covered by the script or territory codes.

  • extensions: information that’s attached to the language code for use in some specific system, such as Unicode collation orders.

  • private: a code starting with x- that has no defined meaning.

The Language.get method converts a string to a Language instance, and the Language.make method makes a Language instance from its fields. These values are cached so that calling Language.get or Language.make again with the same values returns the same object, for efficiency.

By default, it will replace non-standard and overlong tags as it interprets them. To disable this feature and get the codes that literally appear in the language tag, use the normalize=False option.

>>> Language.get('en-Latn-US')
Language.make(language='en', script='Latn', territory='US')

>>> Language.get('sgn-US', normalize=False)
Language.make(language='sgn', territory='US')

>>> Language.get('und')
Language.make()

Here are some examples of replacing non-standard tags:

>>> Language.get('sh-QU')
Language.make(language='sr', script='Latn', territory='EU')

>>> Language.get('sgn-US')
Language.make(language='ase')

>>> Language.get('zh-cmn-Hant')
Language.make(language='zh', script='Hant')

Use the str() function on a Language object to convert it back to its standard string form:

>>> str(Language.get('sh-QU'))
'sr-Latn-EU'

>>> str(Language.make(territory='IN'))
'und-IN'

Checking validity

A language code is valid when every part of it is assigned a meaning by IANA. That meaning could be “private use”.

In langcodes, we check the language subtag, script, territory, and variants for validity. We don’t check other parts such as extlangs or Unicode extensions.

For example, ja is a valid language code, and jp is not:

>>> Language.get('ja').is_valid()
True

>>> Language.get('jp').is_valid()
False

If one subtag is invalid, the entire code is invalid:

>>> Language.get('en-000').is_valid()
False

iw is valid, though it’s a deprecated alias for he:

>>> Language.get('iw', normalize=False).is_valid()
True

The empty language code (und) is valid:

>>> Language.make().is_valid()
True

Private use codes are valid:

>>> Language.get('qaa-Qaai-AA-x-what-even-is-this').is_valid()
True

Language codes that are very unlikely are still valid:

>>> Language.get('fr-Cyrl').is_valid()
True

Language codes that don’t parse should be invalid, but it’s moot, because this method only exists on Language objects:

>>> Language.get('C').is_valid()
Traceback (most recent call last):
...
langcodes.tag_parser.LanguageTagError: Expected a language code, got 'c'

Getting alpha3 codes

Before there was BCP 47, there was ISO 639-2. The ISO tried to make room for the variety of human languages by assigning every language a 3-letter code, including the ones that already had 2-letter codes.

Unfortunately, this just led to more confusion. Some languages ended up with two different 3-letter codes for legacy reasons, such as French, which is fra as a “terminology” code, and fre as a “biblographic” code. And meanwhile, fr was still a code that you’d be using if you followed ISO 639-1.

In BCP 47, you should use 2-letter codes whenever they’re available, and that’s what langcodes does. Fortunately, all the languages that have two different 3-letter codes also have a 2-letter code, so if you prefer the 2-letter code, you don’t have to worry about the distinction.

But some applications want the 3-letter code in particular, so langcodes provides a method for getting those, Language.to_alpha3(). It returns the ‘terminology’ code by default, and passing variant='B' returns the bibliographic code.

When this method returns, it always returns a 3-letter string.

>>> Language.get('fr').to_alpha3()
'fra'
>>> Language.get('fr-CA').to_alpha3()
'fra'
>>> Language.get('fr-CA').to_alpha3(variant='B')
'fre'
>>> Language.get('de').to_alpha3()
'deu'
>>> Language.get('no').to_alpha3()
'nor'
>>> Language.get('un').to_alpha3()
Traceback (most recent call last):
    ...
LookupError: 'un' is not a known language code, and has no alpha3 code.

For many languages, the terminology and bibliographic alpha3 codes are the same.

>>> Language.get('en').to_alpha3(variant='T')
'eng'
>>> Language.get('en').to_alpha3(variant='B')
'eng'

When you use any of these “overlong” alpha3 codes in langcodes, they normalize back to the alpha2 code:

>>> Language.get('zho')
Language.make(language='zh')

Working with language names

The methods in this section require an optional package called language_data. You can install it with pip install language_data, or request the optional “data” feature of langcodes with pip install langcodes[data].

The dependency that you put in setup.py should be langcodes[data].

Describing Language objects in natural language

It’s often helpful to be able to describe a language code in a way that a user (or you) can understand, instead of in inscrutable short codes. The display_name method lets you describe a Language object in a language.

The .display_name(language, min_score) method will look up the name of the language. The names come from the IANA language tag registry, which is only in English, plus CLDR, which names languages in many commonly-used languages.

The default language for naming things is English:

>>> Language.make(language='fr').display_name()
'French'

>>> Language.make().display_name()
'Unknown language'

>>> Language.get('zh-Hans').display_name()
'Chinese (Simplified)'

>>> Language.get('en-US').display_name()
'English (United States)'

But you can ask for language names in numerous other languages:

>>> Language.get('fr').display_name('fr')
'français'

>>> Language.get('fr').display_name('es')
'francés'

>>> Language.make().display_name('es')
'lengua desconocida'

>>> Language.get('zh-Hans').display_name('de')
'Chinesisch (Vereinfacht)'

>>> Language.get('en-US').display_name('zh-Hans')
'英语（美国）'

Why does everyone get Slovak and Slovenian confused? Let’s ask them.

>>> Language.get('sl').display_name('sl')
'slovenščina'
>>> Language.get('sk').display_name('sk')
'slovenčina'
>>> Language.get('sl').display_name('sk')
'slovinčina'
>>> Language.get('sk').display_name('sl')
'slovaščina'

If the language has a script or territory code attached to it, these will be described in parentheses:

>>> Language.get('en-US').display_name()
'English (United States)'

Sometimes these can be the result of tag normalization, such as in this case where the legacy tag ‘sh’ becomes ‘sr-Latn’:

>>> Language.get('sh').display_name()
'Serbian (Latin)'

>>> Language.get('sh', normalize=False).display_name()
'Serbo-Croatian'

Naming a language in itself is sometimes a useful thing to do, so the .autonym() method makes this easy, providing the display name of a language in the language itself:

>>> Language.get('fr').autonym()
'français'
>>> Language.get('es').autonym()
'español'
>>> Language.get('ja').autonym()
'日本語'
>>> Language.get('en-AU').autonym()
'English (Australia)'
>>> Language.get('sr-Latn').autonym()
'srpski (latinica)'
>>> Language.get('sr-Cyrl').autonym()
'српски (ћирилица)'

The names come from the Unicode CLDR data files, and in English they can also come from the IANA language subtag registry. Together, they can give you language names in the 196 languages that CLDR supports.

Describing components of language codes

You can get the parts of the name separately with the methods .language_name(), .script_name(), and .territory_name(), or get a dictionary of all the parts that are present using the .describe() method. These methods also accept a language code for what language they should be described in.

>>> shaw = Language.get('en-Shaw-GB')
>>> shaw.describe('en')
{'language': 'English', 'script': 'Shavian', 'territory': 'United Kingdom'}

>>> shaw.describe('es')
{'language': 'inglés', 'script': 'shaviano', 'territory': 'Reino Unido'}

Recognizing language names in natural language

As the reverse of the above operations, you may want to look up a language by its name, converting a natural language name such as “French” to a code such as ‘fr’.

The name can be in any language that CLDR supports (see “Ambiguity” below).

>>> import langcodes
>>> langcodes.find('french')
Language.make(language='fr')

>>> langcodes.find('francés')
Language.make(language='fr')

However, this method currently ignores the parenthetical expressions that come from .display_name():

>>> langcodes.find('English (Canada)')
Language.make(language='en')

There is still room to improve the way that language names are matched, because some languages are not consistently named the same way. The method currently works with hundreds of language names that are used on Wiktionary.

Ambiguity

For the sake of usability, langcodes.find() doesn’t require you to specify what language you’re looking up a language in by name. This could potentially lead to a conflict: what if name “X” is language A’s name for language B, and language C’s name for language D?

We can collect the language codes from CLDR and see how many times this happens. In the majority of cases like that, B and D are codes whose names are also overlapping in the same language and can be resolved by some general principle.

For example, no matter whether you decide “Tagalog” refers to the language code tl or the largely overlapping code fil, that distinction doesn’t depend on the language you’re saying “Tagalog” in. We can just return tl consistently.

>>> langcodes.find('tagalog')
Language.make(language='tl')

In the few cases of actual interlingual ambiguity, langcodes won’t match a result. You can pass in a language= parameter to say what language the name is in.

For example, there are two distinct languages called “Tonga” in various languages. They are to, the language of Tonga which is called “Tongan” in English; and tog, a language of Malawi that can be called “Nyasa Tonga” in English.

>>> langcodes.find('tongan')
Language.make(language='to')

>>> langcodes.find('nyasa tonga')
Language.make(language='tog')

>>> langcodes.find('tonga')
Traceback (most recent call last):
...
LookupError: Can't find any language named 'tonga'

>>> langcodes.find('tonga', language='id')
Language.make(language='to')

>>> langcodes.find('tonga', language='ca')
Language.make(language='tog')

Other ambiguous names written in Latin letters are “Kiga”, “Mbundu”, “Roman”, and “Ruanda”.

Demographic language data

The Language.speaking_population() and Language.writing_population() methods get Unicode’s estimates of how many people in the world use a language.

As with the language name data, this requires the optional language_data package to be installed.

.speaking_population() estimates how many people speak a language. It can be limited to a particular territory with a territory code (such as a country code).

>>> Language.get('es').speaking_population()
487664083

>>> Language.get('pt').speaking_population()
237135429

>>> Language.get('es-BR').speaking_population()
76218

>>> Language.get('pt-BR').speaking_population()
192661560

>>> Language.get('vo').speaking_population()
0

Script codes will be ignored, because the script is not involved in speaking:

>>> Language.get('es-Hant').speaking_population() ==\
... Language.get('es').speaking_population()
True

.writing_population() estimates how many people write a language.

>>> all = Language.get('zh').writing_population()
>>> all
1240326057

>>> traditional = Language.get('zh-Hant').writing_population()
>>> traditional
37019589

>>> simplified = Language.get('zh-Hans').writing_population()
>>> all == traditional + simplified
True

The estimates for “writing population” are often overestimates, as described in the CLDR documentation on territory data. In most cases, they are derived from published data about literacy rates in the places where those languages are spoken. This doesn’t take into account that many literate people around the world speak a language that isn’t typically written, and write in a different language.

Like .speaking_population(), this can be limited to a particular territory:

>>> Language.get('zh-Hant-HK').writing_population()
6439733
>>> Language.get('zh-Hans-HK').writing_population()
338933

Comparing and matching languages

The tag_distance function returns a number from 0 to 134 indicating the distance between the language the user desires and a supported language.

The distance data comes from CLDR v38.1 and involves a lot of judgment calls made by the Unicode consortium.

Distance values

This table summarizes the language distance values:

Value

Meaning

Example

0

These codes represent the same language, possibly after filling in values and normalizing.

Norwegian Bokmål → Norwegian

1-3

These codes indicate a minor regional difference.

Australian English → British English

4-9

These codes indicate a significant but unproblematic regional difference.

American English → British English

10-24

A gray area that depends on your use case. There may be problems with understanding or usability.

Afrikaans → Dutch, Wu Chinese → Mandarin Chinese

25-50

These languages aren’t similar, but there are demographic reasons to expect some intelligibility.

Tamil → English, Marathi → Hindi

51-79

There are large barriers to understanding.

Japanese → Japanese in Hepburn romanization

80-99

These are different languages written in the same script.

English → French, Arabic → Urdu

100+

These languages have nothing particularly in common.

English → Japanese, English → Tamil

See the docstring of tag_distance for more explanation and examples.

Finding the best matching language

Suppose you have software that supports any of the supported_languages. The user wants to use desired_language. The best_match(desired_language, supported_languages) function lets you choose the right language, even if there isn’t an exact match.

The min_score parameter sets the minimum score that will be allowed to match. If all the scores are less than min_score, the result will be ‘und’ with a strength of 0.

When there is a tie for the best matching language, the first one in the tie will be used.

Setting min_score lower will enable more things to match, at the cost of possibly mis-handling data or upsetting users.

Here are some examples. (If you want to know what these language tags mean, scroll down and learn about the language_name method!)

>>> closest_match('fr', ['de', 'en', 'fr'])
('fr', 0)

>>> closest_match('sh', ['hr', 'bs', 'sr-Latn', 'sr-Cyrl'])
('sr-Latn', 0)

>>> closest_match('zh-CN', ['cmn-Hant', 'cmn-Hans', 'gan', 'nan'])
('cmn-Hans', 0)

>>> closest_match('pt', ['pt-BR', 'pt-PT'])
('pt-BR', 0)

>>> closest_match('en-AU', ['en-GB', 'en-US'])
('en-GB', 3)

>>> closest_match('af', ['en', 'nl', 'zu'])
('nl', 24)

>>> closest_match('id', ['zsm', 'mhp'])
('zsm', 14)

>>> closest_match('ja', ['ja-Latn-hepburn', 'en'])
('und', 1000)

>>> closest_match('ja', ['ja-Latn-hepburn', 'en'], max_distance=60)
('ja-Latn-hepburn', 50)

Further API documentation

There are many more methods for manipulating and comparing language codes, and you will find them documented thoroughly in the code itself.

The interesting functions all live in this one file, with extensive docstrings and annotations. Making a separate Sphinx page out of the docstrings would be the traditional thing to do, but here it just seems redundant. You can go read the docstrings in context, in their native habitat, and they’ll always be up to date.

Code with documentation

Changelog

Version 3.1 (February 2021)

  • Added the Language.to_alpha3() method, for getting a three-letter code for a language according to ISO 639-2.

  • Updated the type annotations from obiwan-style to mypy-style.

Version 3.0 (February 2021)

  • Moved bulky data, particularly language names, into a separate language_data package. In situations where the data isn’t needed, langcodes becomes a smaller, pure-Python package with no dependencies.

  • Language codes where the language segment is more than 4 letters no longer parse: Language.get(‘nonsense’) now returns an error.

  • Added a method for checking the validity of a language code.

  • Added methods for estimating language population.

  • Updated to CLDR 38.1, which includes differences in language matching.

  • Tested on Python 3.6 through 3.9; no longer tested on Python 3.5.

Version 2.2 (February 2021)

  • Replaced marisa-trie dependency with marisa-trie-m, to achieve compatibility with Python 3.9.

Version 2.1 (June 2020)

  • Added the display_name method to be a more intuitive way to get a string describing a language code, and made the autonym method use it instead of language_name.

  • Updated to CLDR v37.

  • Previously, some attempts to get the name of a language would return its language code instead, perhaps because the name was being requested in a language for which CLDR doesn’t have name data. This is unfortunate because names and codes should not be interchangeable.

    Now we fall back on English names instead, which exists for all IANA codes. If the code is unknown, we return a string such as “Unknown language [xx]”.

Version 2.0 (April 2020)

Version 2.0 involves some significant changes that may break compatibility with 1.4, in addition to updating to version 36.1 of the Unicode CLDR data and the April 2020 version of the IANA subtag registry.

Match scores replaced with distances

Originally, the goodness of a match between two different language codes was defined in terms of a “match score” with a maximum of 100. Around 2016, Unicode started replacing this with a different measure, the “match distance”, which was defined much more clearly, but we had to keep using the “match score”.

As of langcodes version 2.0, the “score” functions (such as Language.match_score, tag_match_score, and best_match) are deprecated. They’ll keep using the deprecated language match tables from around CLDR 27.

For a better measure of the closeness of two language codes, use Language.distance, tag_distance, and closest_match.

‘region’ renamed to ‘territory’

We were always out of step with CLDR here. Following the example of the IANA database, we referred to things like the ‘US’ in ‘en-US’ as a “region code”, but the Unicode standards consistently call it a “territory code”.

In langcodes 2.0, parameters, dictionary keys, and attributes named region have been renamed to territory. We try to support a few common cases with deprecation warnings, such as looking up the region property of a Language object.

A nice benefit of this is that when a dictionary is displayed with ‘language’, ‘script’, and ‘territory’ keys in alphabetical order, they are in the same order as they are in a language code.

Python version support

The minimum required version of Python has been raised from 3.3 to 3.5, because of the end-of-life of Python 3.4 and older.

Indices and tables