Proposed Update Unicode Technical Standard #35

Unicode Locale Data Markup Language (LDML)

Summary

This document describes an XML format (vocabulary) for the exchange of structured locale data. This format is used in the Unicode Common Locale Data Repository.

Note: Some links may lead to in-development or older versions of the data files. See https://cldr.unicode.org for up-to-date CLDR release data.

Status

This is a draft document which may be updated, replaced, or superseded by other documents at any time. Publication does not imply endorsement by the Unicode Consortium. This is not a stable document; it is inappropriate to cite this document as other than a work in progress.

A Unicode Technical Standard (UTS) is an independent specification. Conformance to the Unicode Standard does not imply conformance to any UTS.

Please submit corrigenda and other comments with the CLDR bug reporting form [Bugs]. Related information that is useful in understanding this document is found in the References. For the latest version of the Unicode Standard see [Unicode]. For a list of current Unicode Technical Reports see [Reports]. For more information about versions of the Unicode Standard, see [Versions].

NOTE: The source for the LDML specification has been converted to GitHub Markdown (GFM) instead of HTML. The formatting is now simpler, but some features — such as formatting for table captions — may not be complete by the release date. Improvements in the formatting for the v39 specification are planned for after the release, but no substantive changes would be made to the content.

Parts

The LDML specification is divided into the following parts:

• Part 1: Core (languages, locales, basic structure)
• Part 2: General (display names & transforms, etc.)
• Part 3: Numbers (number & currency formatting)
• Part 4: Dates (date, time, time zone formatting)
• Part 5: Collation (sorting, searching, grouping)
• Part 6: Supplemental (supplemental data)
• Part 7: Keyboards (keyboard mappings)
• Part 8: Person Names (person names)

Contents of Part 1, Core

• 1 Introduction
  • 1.1 Conformance
• 2 What is a Locale?
• 3 Unicode Language and Locale Identifiers
  • 3.1 Unicode Language Identifier
  • 3.2 Unicode Locale Identifier
    • 3.2.1 Canonical Unicode Locale Identifiers
  • 3.3 BCP 47 Conformance
    • 3.3.1 BCP 47 Language Tag Conversion
      • Table: BCP 47 Language Tag to Unicode BCP 47 Locale Identifier Examples
      • Unicode Locale Identifier: CLDR to BCP 47
      • Unicode Locale Identifier: BCP 47 to CLDR
      • Truncation
  • 3.4 Language Identifier Field Definitions
    • unicode_language_subtag (also known as a Unicode base language code)
    • unicode_script_subtag (also known as a Unicode script code)
    • unicode_region_subtag (also known as a Unicode region code, or a Unicode territory code)
    • unicode_variant_subtag (also known as a Unicode language variant code)
  • 3.5 Special Codes
    • 3.5.1 Unknown or Invalid Identifiers
    • 3.5.2 Numeric Codes
    • 3.5.3 Private Use Codes
      • Table: Private Use Codes in CLDR
  • 3.6 Unicode BCP 47 U Extension
    • 3.6.1 Key And Type Definitions
      • Table: Key/Type Definitions
    • 3.6.2 Numbering System Data
    • 3.6.3 Time Zone Identifiers
    • 3.6.4 U Extension Data Files
    • 3.6.5 Subdivision Codes
      • 3.6.5.1 Validity
  • 3.7 Unicode BCP 47 T Extension
    • 3.7.1 T Extension Data Files
  • 3.8 Compatibility with Older Identifiers
    • 3.8.1 Old Locale Extension Syntax
      • Table: Locale Extension Mappings
    • 3.8.2 Legacy Variants
      • Table: Legacy Variant Mappings
    • 3.8.3 Relation to OpenI18n
  • 3.9 Transmitting Locale Information
    • 3.9.1 Message Formatting and Exceptions
  • 3.10 Unicode Language and Locale IDs
    • 3.10.1 Written Language
    • 3.10.2 Hybrid Locale Identifiers
  • 3.11 Validity Data
• 4 Locale Inheritance and Matching
  • 4.1 Lookup
    • 4.1.1 Bundle vs Item Lookup
      • Table: Lookup Differences
    • 4.1.2 Lateral Inheritance
      • Table: Count Fallback: normal
      • Table: Count Fallback: currency
    • 4.1.3 Parent Locales
  • 4.2 Inheritance and Validity
    • 4.2.1 Definitions
    • 4.2.2 Resolved Data File
    • 4.2.3 Valid Data
    • 4.2.4 Checking for Draft Status
    • 4.2.5 Keyword and Default Resolution
    • 4.2.6 Inheritance vs Related Information
  • 4.3 Likely Subtags
  • 4.4 Language Matching
    • 4.4.1 Enhanced Language Matching
• 5 XML Format
  • 5.1 Common Elements
    • 5.1.1 Element special
      • 5.1.1.1 Sample Special Elements
    • 5.1.2 Element alias
      • Table: Inheritance with source="locale"
    • 5.1.3 Element displayName
    • 5.1.4 Escaping Characters
  • 5.2 Common Attributes
    • 5.2.1 Attribute type
    • 5.2.2 Attribute draft
    • 5.2.3 Attribute alt
    • 5.2.4 Attribute references
  • 5.3 Common Structures
    • 5.3.1 Date and Date Ranges
    • 5.3.2 Text Directionality
    • 5.3.3 Unicode Sets
      • 5.3.3.1 Lists of Code Points
      • 5.3.3.2 Unicode Properties
      • 5.3.3.3 Boolean Operations
      • 5.3.3.4 UnicodeSet Examples
    • 5.3.4 String Range
  • 5.4 Identity Elements
  • 5.5 Valid Attribute Values
  • 5.6 Canonical Form
    • 5.6.1 Content
    • 5.6.2 Ordering
    • 5.6.3 Comments
  • 5.7 DTD Annotations
    • 5.7.1 Attribute Value Constraints
• 6 Property Data
  • 6.1 Script Metadata
  • 6.2 Extended Pictographic
  • 6.3 Labels.txt
  • 6.4 Segmentation Tests
• 7 Issues in Formatting and Parsing
  • 7.1 Lenient Parsing
    • 7.1.1 Motivation
    • 7.1.2 Loose Matching
  • 7.2 Handling Invalid Patterns
• Annex A Deprecated Structure
  • A.1 Element fallback
  • A.2 BCP 47 Keyword Mapping
  • A.3 Choice Patterns
  • A.4 Element default
  • A.5 Deprecated Common Attributes
    • A.5.1 Attribute standard
    • A.5.2 Attribute draft in non-leaf elements
  • A.6 Element base
  • A.7 Element rules
  • A.8 Deprecated subelements of <dates>
  • A.9 Deprecated subelements of <calendars>
  • A.10 Deprecated subelements of <timeZoneNames>
  • A.11 Deprecated subelements of <zone> and <metazone>
  • A.12 Renamed attribute values for <contextTransformUsage> element
  • A.13 Deprecated subelements of <segmentations>
  • A.14 Element cp
  • A.15 Attribute validSubLocales
  • A.16 Elements postalCodeData, postCodeRegex
  • A.17 Element telephoneCodeData
• Annex B Links to Other Parts
  • Table: Part 2 Links: General (display names & transforms, etc.)
  • Table: Part 3 Links: Numbers (number & currency formatting)
  • Table: Part 4 Links: Dates (date, time, time zone formatting)
  • Table: Part 5 Links: Collation (sorting, searching, grouping)
  • Table: Part 6 Links: Supplemental (supplemental data)
  • Table: Part 7 Links: Keyboards (keyboard mappings)
• Annex C. LocaleId Canonicalization
  • LocaleId Definitions
    • 1. Multimap interpretation
    • 2. Alias elements
    • 3. Matches
    • 4. Replacement
      • Territory Exception
    • 5. Canonicalizing Syntax
  • Preprocessing
  • Processing LanguageIds
• Processing LocaleIds
• Optimizations
• References
• Acknowledgments
• Modifications

1 Introduction

Not long ago, computer systems were like separate worlds, isolated from one another. The internet and related events have changed all that. A single system can be built of many different components, hardware and software, all needing to work together. Many different technologies have been important in bridging the gaps; in the internationalization arena, Unicode has provided a lingua franca for communicating textual data. However, there remain differences in the locale data used by different systems.

The best practice for internationalization is to store and communicate language-neutral data, and format that data for the client. This formatting can take place on any of a number of the components in a system; a server might format data based on the user's locale, or it could be that a client machine does the formatting. The same goes for parsing data, and locale-sensitive analysis of data.

But there remain significant differences across systems and applications in the locale-sensitive data used for such formatting, parsing, and analysis. Many of those differences are simply gratuitous; all within acceptable limits for human beings, but yielding different results. In many other cases there are outright errors. Whatever the cause, the differences can cause discrepancies to creep into a heterogeneous system. This is especially serious in the case of collation (sort-order), where different collation caused not only ordering differences, but also different results of queries! That is, with a query of customers with names between "Abbot, Cosmo" and "Arnold, James", if different systems have different sort orders, different lists will be returned. (For comparisons across systems formatted as HTML tables, see [Comparisons].)

Note: There are many different equally valid ways in which data can be judged to be "correct" for a particular locale. The goal for the common locale data is to make it as consistent as possible with existing locale data, and acceptable to users in that locale.

This document specifies an XML format for the communication of locale data: the Unicode Locale Data Markup Language (LDML). This provides a common format for systems to interchange locale data so that they can get the same results in the services provided by internationalization libraries. It also provides a standard format that can allow users to customize the behavior of a system. With it, for example, collation (sorting) rules can be exchanged, allowing two implementations to exchange a specification of tailored collation rules. Using the same specification, the two implementations will achieve the same results in comparing strings. Unicode LDML can also be used to let a user encapsulate specialized sorting behavior for a specific domain, or create a customized locale for a minority language. Unicode LDML is also used in the Unicode Common Locale Data Repository (CLDR). CLDR uses an open process for reconciling differences between the locale data used on different systems and validating the data, to produce with a useful, common, consistent base of locale data.

For more information, see the Common Locale Data Repository project page [LocaleProject].

As LDML is an interchange format, it was designed for ease of maintenance and simplicity of transformation into other formats, above efficiency of run-time lookup and use. Implementations should consider converting LDML data into a more compact format prior to use.

1.1 Conformance

There are many ways to use the Unicode LDML format and the data in CLDR, and the Unicode Consortium does not restrict the ways in which the format or data are used. However, an implementation may also claim conformance to LDML or to CLDR, as follows:

UAX35-C1. An implementation that claims conformance to this specification shall:

1. Identify the sections of the specification that it conforms to.
   • For example, an implementation might claim conformance to all LDML features except for transforms and segments.
2. Interpret the relevant elements and attributes of LDML documents in accordance with the descriptions in those sections.
   • For example, an implementation that claims conformance to the date format patterns must interpret the characters in such patterns according to Date Field Symbol Table.
3. Declare which types of CLDR data it uses.
   • For example, an implementation might declare that it only uses language names, and those with a draft status of contributed or approved.

UAX35-C2. An implementation that claims conformance to Unicode locale or language identifiers shall:

1. Specify whether Unicode locale extensions are allowed
2. Specify the canonical form used for identifiers in terms of casing and field separator characters.

External specifications may also reference particular components of Unicode locale or language identifiers, such as:

Field X can contain any Unicode region subtag values as given in Unicode Technical Standard #35: Unicode Locale Data Markup Language (LDML), excluding grouping codes.

2 What is a Locale?

Before diving into the XML structure, it is helpful to describe the model behind the structure. People do not have to subscribe to this model to use data in LDML, but they do need to understand it so that the data can be correctly translated into whatever model their implementation uses.

The first issue is basic: what is a locale? In this model, a locale is an identifier (id) that refers to a set of user preferences that tend to be shared across significant swaths of the world. Traditionally, the data associated with this id provides support for formatting and parsing of dates, times, numbers, and currencies; for measurement units, for sort-order (collation), plus translated names for time zones, languages, countries, and scripts. The data can also include support for text boundaries (character, word, line, and sentence), text transformations (including transliterations), and other services.

Locale data is not cast in stone: the data used on someone's machine generally may reflect the US format, for example, but preferences can typically set to override particular items, such as setting the date format for 2002.03.15, or using metric or Imperial measurement units. In the abstract, locales are simply one of many sets of preferences that, say, a website may want to remember for a particular user. Depending on the application, it may want to also remember the user's time zone, preferred currency, preferred character set, smoker/non-smoker preference, meal preference (vegetarian, kosher, and so on), music preference, religion, party affiliation, favorite charity, and so on.

Locale data in a system may also change over time: country boundaries change; governments (and currencies) come and go: committees impose new standards; bugs are found and fixed in the source data; and so on. Thus the data needs to be versioned for stability over time.

In general terms, the locale id is a parameter that is supplied to a particular service (date formatting, sorting, spell-checking, and so on). The format in this document does not attempt to represent all the data that could conceivably be used by all possible services. Instead, it collects together data that is in common use in systems and internationalization libraries for basic services. The main difference among locales is in terms of language; there may also be some differences according to different countries or regions. However, the line between locales and languages, as commonly used in the industry, are rather fuzzy. Note also that the vast majority of the locale data in CLDR is in fact language data; all non-linguistic data is separated out into a separate tree. For more information, see Section 3.10 Language and Locale IDs.

We will speak of data as being "in locale X". That does not imply that a locale is a collection of data; it is simply shorthand for "the set of data associated with the locale id X". Each individual piece of data is called a resource or field, and a tag indicating the key of the resource is called a resource tag.

3 Unicode Language and Locale Identifiers

Unicode LDML uses stable identifiers based on [BCP47] for distinguishing among languages, locales, regions, currencies, time zones, transforms, and so on. There are many systems for identifiers for these entities. The Unicode LDML identifiers may not match the identifiers used on a particular target system. If so, some process of identifier translation may be required when using LDML data.

The BCP 47 extensions (-u- and -t-) are described in Section 3.6 Unicode BCP 47 U Extension and Section 3.7 Unicode BCP 47 T Extension.

3.1 Unicode Language Identifier

A Unicode language identifier has the following structure (provided in EBNF (Perl-based)). The following table defines syntactically well-formed identifiers: they are not necessarily valid identifiers. For additional validity criteria, see the links on the right.

= "root"
| (unicode_language_subtag
    (sep unicode_script_subtag)?
  | unicode_script_subtag)
  (sep unicode_region_subtag)?
  (sep unicode_variant_subtag)* ;

= alpha{2,3} | alpha{5,8};

= alpha{4} ;

= (alpha{2} | digit{3}) ;

= (alphanum{5,8}
| digit alphanum{3}) ;

= [-_] ;

= [0-9] ;

= [A-Z a-z] ;

= [0-9 A-Z a-z] ;

The semantics of the various subtags is explained in Section 3.4 Language Identifier Field Definitions ; there are also direct links from unicode_language_subtag , etc. While theoretically the unicode_language_subtag may have more than 3 letters through the IANA registration process, in practice that has not occurred. The unicode_language_subtag "und" may be omitted when there is a unicode_script_subtag ; for that reason unicode_language_subtag values with 4 letters are not permitted. However, such unicode_language_id values are not intended for general interchange, because they are not valid BCP 47 tags. Instead, they are intended for certain protocols such as the identification of transliterators or font ScriptLangTag values. For more information on language subtags with 4 letters, see BCP 47 Language Tag to Unicode BCP 47 Locale Identifier.

For example, "en-US" (American English), "en_GB" (British English), "es-419" (Latin American Spanish), and "uz-Cyrl" (Uzbek in Cyrillic) are all valid Unicode language identifiers.

3.2 Unicode Locale Identifier

A Unicode locale identifier is composed of a Unicode language identifier plus (optional) locale extensions. It has the following structure. The semantics of the U and T extensions are explained in Section 3.6 Unicode BCP 47 U Extension and Section 3.7 Unicode BCP 47 T Extension. Other extensions and private use extensions are supported for pass-through. The following table defines syntactically well-formed identifiers: they are not necessarily valid identifiers. For additional validity criteria, see the links on the right.

As is often the case, the complete syntactic constraints are not easily captured by ABNF, so there is a further condition: There cannot be more than one extension with the same singleton (-a-, …, -t-, -u-, …). Note that the private use extension (-x-) must come after all other extensions.

For historical reasons, this is called a Unicode locale identifier. However, it really functions (with few exceptions) as a language identifier, and accesses language-based data. Except where it would be unclear, this document uses the term "locale" data loosely to encompass both types of data: for more information, see Section 3.10 Language and Locale IDs.

As of the release of this specification, there were no other_extensions defined. The other_extensions are present in the syntax to allow implementations to preserve that information.

As for terminology, the term code may also be used instead of "subtag", and "territory" instead of "region". The primary language subtag is also called the base language code. For example, the base language code for "en-US" (American English) is "en" (English). The type may also be referred to as a value or key-value.

The identifiers can vary in case and in the separator characters. The "-" and "_" separators are treated as equivalent, although "-" is preferred.

All identifier field values are case-insensitive. Although case distinctions do not carry any special meaning, an implementation of LDML should use the casing recommendations in [BCP47], especially when a Unicode locale identifier is used for locale data exchange in software protocols.

3.2.1 Canonical Unicode Locale Identifiers

A unicode_locale_id has canonical syntax when:

• It starts with a language subtag (those beginning with a script subtag are only for specialized use)
• Casing
  • Any script subtag inside unicode_language_id is in title case (eg, Hant)
  • Any region subtag inside unicode_language_id is in uppercase (eg, DE)
  • All other subtags are in lowercase (eg, en, fonipa)
• Order
  • Any variants are in alphabetical order (eg, en-fonipa-scouse, not en-scouse-fonipa)
  • Any extensions are in alphabetical order by their singleton (eg, en-t-xxx-u-yyy, not en-u-yyy-t-xxx)
  • All attributes are sorted in alphabetical order.
  • All keywords and tfields are sorted by alphabetical order of their keys, within their respective extensions.
  • Any type or tfield value "true" is removed.

For example, the canonical form of "en-u-foo-bar-nu-thai-ca-buddhist-kk-true" is "en-u-bar-foo-ca-buddhist-kk-nu-thai". The attributes "foo" and "bar" in this example are provided only for illustration; no attribute subtags are defined by the current CLDR specification.

NOTE: Some people may wonder why CLDR uses alphabetical order for variants, rather than the ordering in Section 4.1 of BCP 47. Here are the considerations that lead to that decision:

• The ordering in Section 4.1 is recommended, but not required for conformance. In particular, use of and ordering by Prefix is recommended but not required.
• Moreover, Section 4.5 states that “If more than one variant appears within a tag, processors MAY reorder the variants to obtain better matching behavior or more consistent presentation.”
• The best practices for internationalization have moved well beyond some of the guidelines and recommendations in BCP 47, especially for language matching and language fallback.
• Robust implementations will accept the variants in any order, just as they accept extensions in any order.
• A canonical order allows for determination of identity of identifiers via string comparison.
• The ordering in Section 4.1 does not result in a determinant order for canonicalization, since the mechanism for determining “importance” is not specified: ca-valencia-fonipa and ca-fonipa-valencia could both be ‘canonical’ variants of one another.
• Pure alphabetical order is determinant and simple to implement while the ordering in Section 4.1 is indeterminant, more complex, and provides no significant benefit in modern applications.

Note: The current version of CLDR data uses some non-preferred syntax for backward compatibility. This might be changed in future CLDR releases.

• It uses uppercase letters for variant subtags, while the preferred forms are all lowercase.
• It uses "_" as the separator, while the preferred form of the separator is "-".
• It uses "root", while the preferred form is "und".

A unicode_locale_id is in canonical form when it has canonical syntax and contains no aliased subtags. A unicode_locale_id can be transformed into canonical form according to Annex C. LocaleId Canonicalization.

A unicode_locale_id is maximal when the unicode_language_id and tlang (if any) have been transformed by the Add Likely Subtags operation in Section 4.3 Likely Subtags, excluding "und".

Example: the maxmal form of ja-Kana-t-it is ja-Kana-JP-t-it-latn-it

Note that the latn and final it don't use any uppercase characters, since they are not inside unicode_language_id.

Two unicode_locale_ids are equivalent when their maximal canonical forms are identical.

Example: "IW-HEBR-u-ms-imperial" ~ "he-u-ms-uksystem"

The equivalence relationship may change over time, such as when subtags are deprecated or likely subtag mappings change. For example, if two countries were to merge, then various subtags would become deprecated. These kinds of changes are generally very infrequent.

3.3 BCP 47 Conformance

Unicode language and locale identifiers inherit the design and the repertoire of subtags from [BCP47] Language Tags. There are some extensions and restrictions made for the use of the Unicode locale identifier in CLDR:

• It does not allow for the full syntax of [BCP47]:
  • No extlang subtags are allowed (as in the BCP 47 canonical form, see BCP 47 Section 4.5 and Section 3.1.7)
  • No irregular BCP 47 legacy language tags (marked as “Type: grandfathered” in BCP 47) are allowed (these are all deprecated in BCP 47)
  • A tag must not start with the subtag "x": thus a privateuse (eg x-abc) can only be after a language subtag, like "und"
• It allows for certain semantic additions and constraints:
  • Certain codes that are private-use in BCP 47 and ISO are given semantics by LDML
  • Each macrolanguage has an identified primary encompassed language, which is treated as an alias for the macrolanguage, and thus is replaced when canonicalizing (as allowed by BCP 47, see Section 4.1.2)
• It allows certain syntax for backwards compatibility (not BCP 47-compatible):
  • The "_" character for field separator characters, as well as the "-" used in [BCP47] (however, the canonical form is with "-")
  • The subtag "root" to indicate the generic locale used as the parent of all languages in the CLDR data model ("und" can be used instead)
  • The language tag may begin with a script subtag rather than a language subtag. This is specialized use only, and not required for CLDR conformance.

There are thus two subtypes of Unicode locale identifiers:

• the term Unicode CLDR locale identifier applies where the backwards compatibility syntax is used.
• the term Unicode BCP 47 locale identifier applies otherwise. A Unicode BCP 47 locale identifier is also a valid BCP 47 language tag.

3.3.1 BCP 47 Language Tag Conversion

The different identifiers can be converted to one another as described in this section.

A valid [BCP47] language tag can be converted to a valid Unicode BCP 47 locale identifier according to Annex C. LocaleId Canonicalization.

The result is a Unicode BCP 47 locale identifier, in canonical form. It is both a BCP 47 language tag and a Unicode locale identifier. Because the process maps from all BCP 47 language tags into a subset of BCP 47 language tags, the format changes are not reversible, much as a lowercase transformation of the string “McGowan” is not reversible.

Unicode Locale Identifier: CLDR to BCP 47

A Unicode CLDR locale identifier can be converted to a valid [BCP47] language tag (which is also a Unicode BCP 47 locale identifier) by performing the following transformation.

1. Replace the "_" separators with "-"
2. Replace the special language identifier "root" with the BCP 47 primary language tag "und"
3. Add an initial "und" primary language subtag if the first subtag is a script.

Examples:

Unicode Locale Identifier: BCP 47 to CLDR

A Unicode BCP 47 locale identifier can be transformed into a Unicode CLDR locale identifier by performing the following transformation.

1. the separator is changed to "_"
2. the primary language subtag "und" is replaced with "root" if no script, region, or variant subtags are present.

Examples:

Truncation

BCP 47 requires that implementations allow for language tags of at least 35 characters, in Section 4.1.1. To allow for use of extensions, CLDR extends that minimum to 255 for Unicode locale identifiers. Theoretically, a language tag could be far longer, due to the possibility of a large number of variants and extensions. In practice, the typical size of a locale or language identifier will be much smaller, so implementations can optimize for smaller sizes, as long as there is an escape mechanism allowing for up to 255.

3.4 Language Identifier Field Definitions

Unicode language and locale identifier field values are provided in the following table. Note that some private-use BCP 47 field values are given specific meanings in CLDR. While field values are based on [BCP47] subtag values, their validity status in CLDR is specified by means of machine-readable files in the common/validity/ subdirectory, such as language.xml. For the format of those files and more information, see Section 3.11 Validity Data.

unicode_language_subtag (also known as a Unicode base language code)

Subtags in the language.xml file (see Section 3.11 Validity Data ). These are based on [BCP47] subtag values marked as Type: language

ISO 639-3 introduces the notion of "macrolanguages", where certain ISO 639-1 or ISO 639-2 codes are given broad semantics, and additional codes are given for the narrower semantics. For backwards compatibility, Unicode language identifiers retain use of the narrower semantics for these codes. For example:

If a language subtag matches the type attribute of a languageAlias element, then the replacement value is used instead. For example, because "swh" occurs in <languageAlias type="swh" replacement="sw" /> , "sw" must be used instead of "swh". Thus Unicode language identifiers use "ar-EG" for Standard Arabic (Egypt), not "arb-EG"; they use "zh-TW" for Mandarin Chinese (Taiwan), not "cmn-TW".

The private use codes listed as excluded in Section 3.5.3 Private Use Codes will never be given specific semantics in Unicode identifiers, and are thus safe for use for other purposes by other applications.

The CLDR provides data for normalizing language/locale codes, including mapping overlong codes like "eng-840" or "eng-USA" to the correct code "en-US"; see the Aliases Chart.

The following are special language subtags:

unicode_script_subtag (also known as a Unicode script code)

Subtags in the script.xml file (see Section 3.11 Validity Data). These are based on [BCP47] subtag values marked as Type: script

In most cases the script is not necessary, since the language is only customarily written in a single script. Examples of cases where it is used are:

Unicode identifiers give specific semantics to certain Unicode Script values. For more information, see also [UAX24]:

The private use subtags listed as excluded in Section 3.5.3 Private Use Codes will never be given specific semantics in Unicode identifiers, and are thus safe for use for other purposes by other applications.

unicode_region_subtag (also known as a Unicode region code, or a Unicode territory code)

Subtags in the region.xml file (see Section 3.11 Validity Data). These are based on [BCP47] subtag values marked as Type: region

Unicode identifiers give specific semantics to the following subtags. (The alpha2 codes are used as Unicode region subtags. The alpha3 and numeric codes are derived according to Section 3.5.2 Numeric Codes and listed here for additional documentation.)

The private use subtags listed as excluded in Section 3.5.3 Private Use Codes will normally never be given specific semantics in Unicode identifiers, and are thus safe for use for other purposes by other applications. However, LDML may follow widespread industry practice in the use of some of these codes, such as for XK.

The CLDR provides data for normalizing territory/region codes, including mapping overlong codes like "eng-840" or "eng-USA" to the correct code "en-US".

Special Codes:

• The territory code 'UK' has a special status in ISO, and is used for the domain name instead of GB. It is thus recognized by CLDR as being an alternate (unnormalized) form of 'GB'.
• The territory code '001' (the World) is used to indicate a standardized form, such as "ar-001" for Modern Standard Arabic.

unicode_variant_subtag (also known as a Unicode language variant code)

Subtags in the variant.xml file (see Section 3.11 Validity Data). These are based on [BCP47] subtag values marked as Type: variant. The sequence of variant tags must not have any duplicates: thus de-1996-fonipa-1996 is invalid, while de-1996-fonipa and de-fonipa-1996 are both valid.

CLDR provides data for normalizing variant codes. About handling of the "POSIX" variant see Section 3.8.2, Legacy Variants.

Examples:

en
fr_BE
zh-Hant-HK

Deprecated codes—such as QU above—are valid, but strongly discouraged.

A locale that only has a language subtag (and optionally a script subtag) is called a language locale; one with both language and territory subtag is called a territory locale (or country locale).

3.5 Special Codes

3.5.1 Unknown or Invalid Identifiers

The following identifiers are used to indicate an unknown or invalid code in Unicode language and locale identifiers. For Unicode identifiers, the region code uses a private use ISO 3166 code, and Time Zone code uses an additional code; the others are defined by the relevant standards. When these codes are used in APIs connected with Unicode identifiers, the meaning is that either there was no identifier available, or that at some point an input identifier value was determined to be invalid or ill-formed.

When only the script or region are known, then a locale ID will use "und" as the language subtag portion. Thus the locale tag "und_Grek" represents the Greek script; "und_US" represents the US territory.

3.5.2 Numeric Codes

For region codes, ISO and the UN establish a mapping to three-letter codes and numeric codes. However, this does not extend to the private use codes, which are the codes 900-999 (total: 100), and AAA, QMA-QZZ, XAA-XZZ, and ZZZ (total: 1092). Unicode identifiers supply a standard mapping to these: for the numeric codes, it uses the top of the numeric private use range; for the 3-letter codes it doubles the final letter. These are the resulting mappings for all of the private use region codes:

For script codes, ISO 15924 supplies a mapping (however, the numeric codes are not in common use):

3.5.3 Private Use Codes

Private use codes fall into three groups.

• defined: those that are given particular semantics currently in CLDR
• reserved: those that may be given particular semantics in future versions of CLDR
• excluded: those that will never be given particular CLDR semantics in the future, and thus can normally be used by applications without worrying about collisions. However, CLDR may follow widespread industry practice in the use of some of these codes, such as for XA, XB, and XK.

See also Section 3.5.1 Unknown or Invalid Identifiers.

3.6 Unicode BCP 47 U Extension

[BCP47] Language Tags provides a mechanism for extending language tags for use in various applications by extension subtags. Each extension subtag is identified by a single alphanumeric character subtag assigned by IANA.

The Unicode Consortium has registered and is the maintaining authority for two BCP 47 language tag extensions: the extension 'u' for Unicode locale extension [RFC6067] and extension 't' for transformed content [RFC6497]. The Unicode BCP 47 extension data defines the complete list of valid subtags.

These subtags are all in lowercase (that is the canonical casing for these subtags), however, subtags are case-insensitive and casing does not carry any specific meaning. All subtags within the Unicode extensions are alphanumeric characters in length of two to eight that meet the rule extension in the [BCP47].

The -u- Extension. The syntax of 'u' extension subtags is defined by the rule unicode_locale_extensions in Section 3.2 Unicode locale identifier, except the separator of subtags sep must be always hyphen '-' when the extension is used as a part of BCP 47 language tag.

A 'u' extension may contain multiple attribute s or keyword s as defined in Section 3.2 Unicode locale identifier. The canonical syntax is defined as in Canonical Unicode Locale Identifiers.

See also Unicode Extensions for BCP 47 on the CLDR site.

3.6.1 Key And Type Definitions

The following chart contains a set of U extension key values that are currently available, with a description or sampling of the U extension type values. Each category is associated with an XML file in the bcp47 directory.

For the complete list of valid keys and types defined for Unicode locale extensions, see Section 3.6.4 U Extension Data Files. For information on the process for adding new key/type, see [LocaleProject].

Most type values are represented by a single subtag in the current version of CLDR. There are exceptions, such as types used for key "ca" (calendar) and "kr" (collation reordering). If the type is not included, then the type value "true" is assumed. Note that the default for key with a possible "true" value is often "false", but may not always be. Note also that "true"/"True" is not a valid script code, since the ISO 15924 Registration Authority has exceptionally reserved it, which means that it will not be assigned for any purpose.

Note that canonicalization does not change invalid locales to valid locales. For example, und-u-ka canonicalizes to und-u-ka-true, but:

1. "und-u-ka-true" — is invalid, since ‘yes’ is not a valid value for ka
2. "und-u-ka" — is invalid, since the value “true” is assumed whenever there is no value, and ‘true’ is not a valid value for ka

The BCP 47 form for keys and types is the canonical form, and recommended. Other aliases are included for backwards compatibility.

For more information on the allowed keys and types, see the specific elements below, and Section 3.6.4 U Extension Data Files.

Additional keys or types might be added in future versions. Implementations of LDML should be robust to handle any syntactically valid key or type values.

3.6.2 Numbering System Data

LDML supports multiple numbering systems. The identifiers for those numbering systems are defined in the file bcp47/number.xml. For example, for the latest version of the data see bcp47/number.xml.

Details about those numbering systems are defined in supplemental/numberingSystems.xml. For example, for the latest version of the data see supplemental/numberingSystems.xml.

LDML makes certain stability guarantees on this data:

1. Like other BCP 47 identifiers, once a numeric identifier is added to bcp47/number.xml or numberingSystems.xml, it will never be removed from either of those files.
2. If an identifier has type="numeric" in numberingSystems.xml, then
   1. It is a decimal, positional numbering system with an attribute digits=X, where X is a string with the 10 digits in order used by the numbering system.
   2. The values of the type and digits will never change.

3.6.3 Time Zone Identifiers

LDML inherits time zone IDs from the tz database [Olson]. Because these IDs from the tz database do not satisfy the BCP 47 language subtag syntax requirements, CLDR defines short identifiers for the use in the Unicode locale extension. The short identifiers are defined in the file common/bcp47/timezone.xml.

The short identifiers use UN/LOCODE [LOCODE] (excluding a space character) codes where possible. For example, the short identifier for "America/Los_Angeles" is "uslax" (the LOCODE for Los Angeles, US is "US LAX"). Identifiers of length not equal to 5 are used where there is no corresponding UN/LOCODE, such as "usnavajo" for "America/Shiprock", or "utcw01" for "Etc/GMT+1", so that they do not overlap with future UN/LOCODE.

Although the first two letters of a short identifier may match an ISO 3166 two-letter country code, a user should not assume that the time zone belongs to the country. The first two letters in an identifier of length not equal to 5 have no meaning. Also, the identifiers are stabilized, meaning that they will not change no matter what changes happen in the base standard. So if Hawaii leaves the US and joins Canada as a new province, the short time zone identifier "ushnl" would not change in CLDR even if the UN/LOCODE changes to "cahnl" or something else.

There is a special code "unk" for an Unknown or Invalid time zone. This can be expressed in the tz database style ID "Etc/Unknown", although it is not defined in the tz database.

Stability of Time Zone Identifiers

Although the short time zone identifiers are guaranteed to be stable, the preferred IDs in the tz database (as those found in zone.tab file) might be changed time to time. For example, "Asia/Culcutta" was replaced with "Asia/Kolkata" and moved to backward file in the tz database. CLDR contains locale data using a time zone ID from the tz database as the key, stability of the IDs is critical.

To maintain the stability of "long" IDs (for those inherited from the tz database), a special rule applied to the alias attribute in the <type> element for "tz" - the first "long" ID is the CLDR canonical "long" time zone ID.

For example:

<type name="inccu" alias="Asia/Calcutta Asia/Kolkata" description="Kolkata, India"/>

Above <type> element defines the short time zone ID "inccu" (for the use in the Unicode locale extension), corresponding CLDR canonical "long" ID "Asia/Culcutta", and an alias "Asia/Kolkata".

Links in the tz database

Not all TZDB links are in CLDR aliases. CLDR purposefully does not exactly match the Link structure in the TZDB.

1. The links are maintained in the TZDB, and it would duplicate information that could fall out of sync (especially because the TZDB can be updated many times in a single month).
2. The TZDB went though a change a few years ago where it dropped the mappings to countries, whereas CLDR still maintains that distinction.
3. Because there are several different timezones that all link together, that would make for a single long alias being an alias for several different short aliases.

CLDR doesn't alias across country boundaries because countries are useful for timezone selection. Even if, for example, Serbia and Croatia share the same rules, CLDR maintains the difference so that the user can either pick "Serbia time" or "Croatia time". The Croat is not forced to pick "Serbia time" (Europe/Belgrade) nor the Serb forced to pick “Croatia time” (Europe/Zagreb).

3.6.4 U Extension Data Files

The 'u' extension data is stored in multiple XML files located under common/bcp47 directory in CLDR. Each file contains the locale extension key/type values and their backward compatibility mappings appropriate for a particular domain. common/bcp47/collation.xml contains key/type values for collation, including optional collation parameters and valid type values for each key.

The 't' extension data is stored in common/bcp47/transform.xml.

<!ELEMENT keyword ( key* )>

<!ELEMENT key ( type* )>
<!ATTLIST key extension NMTOKEN #IMPLIED>
<!ATTLIST key name NMTOKEN #REQUIRED>
<!ATTLIST key description CDATA #IMPLIED>
<!ATTLIST key deprecated ( true | false ) "false">
<!ATTLIST key preferred NMTOKEN #IMPLIED>
<!ATTLIST key alias NMTOKEN #IMPLIED>
<!ATTLIST key valueType (single | multiple | incremental | any) #IMPLIED >
<!ATTLIST key since CDATA #IMPLIED>

<!ELEMENT type EMPTY>
<!ATTLIST type name NMTOKEN #REQUIRED>
<!ATTLIST type description CDATA #IMPLIED>
<!ATTLIST type deprecated ( true | false ) "false">
<!ATTLIST type preferred NMTOKEN #IMPLIED>
<!ATTLIST type alias CDATA #IMPLIED>
<!ATTLIST type since CDATA #IMPLIED>

<!ELEMENT attribute EMPTY>
<!ATTLIST attribute name NMTOKEN #REQUIRED>
<!ATTLIST attribute description CDATA #IMPLIED>
<!ATTLIST attribute deprecated ( true | false ) "false">
<!ATTLIST attribute preferred NMTOKEN #IMPLIED>
<!ATTLIST attribute since CDATA #IMPLIED>

The extension attribute in <key> element specifies the BCP 47 language tag extension type. The default value of the extension attribute is "u" (Unicode locale extension). The <type> element is only applicable to the enclosing <key>.

In the Unicode locale extension 'u' and 't' data files, the common attributes for the <key>, <type> and <attribute> elements are as follows:

name

The key or type name used by Unicode locale extension with 'u' extension syntax or the 't' extensions syntax. When alias below is absent, this name can be also used with the old style "@key=type" syntax.

Most type names are literal type names, which match exactly the same value. All of these have at least one lowercase letter, such as "buddhist". There are a small number of indirect type names, such as "RG_KEY_VALUE". These have no lowercase letters. The interpretation of each one is listed below.

CODEPOINTS

The type name "CODEPOINTS" is reserved for a variable representing Unicode code point(s). The syntax is:

	EBNF
codepoints	= codepoint (sep codepoint)?
codepoint	= [0-9 A-F a-f]{4,6}

In addition, no codepoint may exceed 10FFFF. For example, "00A0", "300b", "10D40C" and "00C1-00E1" are valid, but "A0", "U060C" and "110000" are not.

In the current version of CLDR, the type "CODEPOINTS" is only used for the deprecated locale extension key "vt" (variableTop). The subtags forming the type for "vt" represent an arbitrary string of characters. There is no formal limit in the number of characters, although practically anything above 1 will be rare, and anything longer than 4 might be useless. Repetition is allowed, for example, 0061-0061 ("aa") is a Valid type value for "vt", since the sequence may be a collating element. Order is vital: 0061-0062 ("ab") is different than 0062-0061 ("ba"). Note that for variableTop any character sequence must be a contraction which yields exactly one primary weight.

For example,

en-u-vt-00A4 : this indicates English, with any characters sorting at or below " ¤" (at a primary level) considered Variable.

By default in UCA, variable characters are ignored in sorting at a primary, secondary, and tertiary level. But in CLDR, they are not ignorable by default. For more information, see Collation: Section 3.3 Setting Options .

REORDER_CODE

The type name "REORDER_CODE" is reserved for reordering block names (e.g. "latn", "digit" and "others") defined in the Root Collation. The type "REORDER_CODE" is used for locale extension key "kr" (colReorder). The value of type for "kr" is represented by one or more reordering block names such as "latn-digit". For more information, see Collation: Section 3.12 Collation Reordering .

RG_KEY_VALUE

The type name "RG_KEY_VALUE" is reserved for region codes in the format required by the "rg" key; this is a subdivision code with idStatus='unknown' or 'regular' from the idValidity data in common/validity/subdivision.xml.

SCRIPT_CODE

The type name "SCRIPT_CODE" is reserved for unicode_script_subtag values (e.g. "thai", "laoo"). The type "SCRIPT_CODE" is used for locale extension key "dx". The value of type for "dx" is represented by one or more SCRIPT_CODEs, such as "thai-laoo".

SUBDIVISION_CODE

The type name "SUBDIVISION_CODE" is reserved for subdivision codes in the format required by the "sd" key; this is a subdivision code from the idValidity data in common/validity/subdivision.xml, excluding those with idStatus='unknown'. Codes with idStatus='deprecated' should not be generated, and those with idStatus='private_use' are only to be used with prior agreement.

PRIVATE_USE

The type name "PRIVATE_USE" is reserved for private use types. A valid type value is composed of one or more subtags separated by hyphens and each subtag consists of three to eight ASCII alphanumeric characters. In the current version of CLDR, "PRIVATE_USE" is only used for transform extension "x0".

valueType

The valueType attribute indicates how many subtags are valid for a given key:

Value	Description
single	Either exactly one type value, or no type value (but only if the value of "true" would be valid). This is the default if no valueType attribute is present.
incremental	Multiple type values are allowed, but only if a prefix is also present, and the sequence is explicitly listed. Each successive type value indicates a refinement of its prefix. For example: <key name="ca" description="Calendar algorithm key" valueType="incremental"> <type name="islamic" description="Islamic calendar"/> <type name="islamic-umalqura" description="Islamic calendar, Umm al-Qura"/> Thus ca-islamic-umalqura is valid. However, ca-gregory-japanese is not valid, because "gregory-japanese" is not listed as a type.
multiple	Multiple type values are allowed, but each may only occur once. For example: <key name="kr" description="Collation reorder codes" valueType="multiple"> <type name="REORDER_CODE" …/>
any	Any number of type values are allowed, with none of the above restrictions. For example: <key extension="t" name="x0" description="Private use transform type key." valueType="any"> <type name="PRIVATE_USE" …/>

description

The description of the key, type or attribute element. There is also some informative text about certain keys and types in the Section 3.5 Key And Type Definitions.

deprecated

The deprecation status of the key, type or attribute element. The value "true" indicates the element is deprecated and no longer used in the version of CLDR. The default value is "false".

preferred

The preferred value of the deprecated key, type or attribute element. When a key, type or attribute element is deprecated, this attribute is used for specifying a new canonical form if available.

alias (Not applicable to <attribute>)

The BCP 47 form is the canonical form, and recommended. Other aliases are included only for backwards compatibility.

Example:

<type name="phonebk" alias="phonebook" description="Phonebook style ordering (such as in German)"/>

The preferred term, and the only one to be used in BCP 47, is the name: in this example, "phonebk".

The alias is a key or type name used by Unicode locale extensions with the old "@key=type" syntax. The attribute value for type may contain multiple names delimited by ASCII space characters. Of those aliases, the first name is the preferred value.

since

The version of CLDR in which this key or type was introduced. Absence of this attribute value implies the key or type was available in CLDR 1.7.2.

Note: There are no values defined for the locale extension attribute in the current CLDR release.

For example,

<key name="co" alias="collation" description="Collation type key">
  <type name="pinyin" description="Pinyin ordering for Latin and for CJK characters (used in Chinese)"/>
</key>

<key name="ka" alias="colAlternate" description="Collation parameter key for alternate handling">
  <type name="noignore" alias="non-ignorable" description="Variable collation elements are not reset to ignorable"/>
  <type name="shifted" description="Variable collation elements are reset to zero at levels one through three"/>
</key>

<key name="tz" alias="timezone">
  ...
  <type name="aumel" alias="Australia/Melbourne Australia/Victoria" description="Melbourne, Australia"/>
  <type name="aumqi" alias="Antarctica/Macquarie" description="Macquarie Island Station, Macquarie Island" since="1.8.1"/>
  ...
</key>

The data above indicates:

• type "pinyin" is valid for key "co", thus "u-co-pinyin" is a valid Unicode locale extension.
• type "pinyin" is not valid for key "ka", thus "u-ka-pinyin" is not a valid Unicode locale extension.
• type "pinyin" has no alias, so "zh@collation=pinyin" is a valid Unicode locale identifier according to the old syntax.
• type "noignore" has an alias attribute, so "en@colAlternate=noignore" is not a valid Unicode locale identifier according to the old syntax.
• type "aumel" is valid for key "tz", supported by CLDR 1.7.2 (default value) or later versions.
• type "aumqi" is valid for key "tz", supported by CLDR 1.8.1 or later versions.

It is strongly recommended that all API methods accept all possible aliases for keywords and types, but generate the canonical form. For example, "ar-u-ca-islamicc" would be equivalent to "ar-u-ca-islamic-civil" on input, but the latter should be output. The one exception is where an alias would only be well-formed with the old syntax, such as "gregorian" (for "gregory").

3.6.5 Subdivision Codes

The subdivision codes designate a subdivision of a country or region. They are called various names, such as a state in the United States, or a province in Canada. The codes in CLDR are based on ISO 3166-2 subdivision codes. The ISO codes have a region code followed by a hyphen, then a suffix consisting of 1..3 ASCII letters or digits.

The CLDR codes are designed to work in a unicode_locale_id (BCP 47), and are thus all lowercase, with no hyphen. For example, the following are valid, and mean “English as used in California, USA”.

• en-u-sd-usca
• en-US-u-sd-usca

CLDR has additional subdivision codes. These may start with a 3-digit region code or use a suffix of 4 ASCII letters or digits, so they will not collide with the ISO codes. Subdivision codes for unknown values are the region code plus "zzzz", such as "uszzzz" for an unknown subdivision of the US. Other codes may be added for stability.

Like BCP 47, CLDR requires stable codes, which are not guaranteed for ISO 3166-2 (nor have the ISO 3166-2 codes been stable in the past). If an ISO 3166-2 code is removed, it remains valid (though marked as deprecated) in CLDR. If an ICU 3166-2 code is reused (for the same region), then CLDR will define a new equivalent code using these as 4-character suffixes.

3.6.5.1 Validity

A unicode_subdivision_id is only valid when it is present in the subdivision.xml file as described in Section 3.11 Validity Data. The data is in a compressed form, and thus needs to be expanded before such a test is made.

Examples:

• usca is valid — there is an id element <id type="subdivision"…>… usca …</id>
• ussct is invalid — there is no id element <id type="subdivision"…>… ussct …</id>

If a unicode_locale_id contains both a unicode_region_subtag and a unicode_subdivision_id, it is only valid if the unicode_subdivision_id starts with the unicode_region_subtag (case-insensitively).

It is recommended that a unicode_locale_id contain a unicode_region_subtag if it contains a unicode_subdivision_id and the region would not be added by adding likely subtags. That produces better behavior if the unicode_subdivision_id is ignored by an implementation or if the language tag is truncated.

Examples:

• en-US-u-sd-usca is valid — the region "US" matches the first part of "usca"
• en-u-sd-usca is valid — it still works after adding likely subtags.
• en-CA-u-sd-gbsct is invalid — the region "CA" does not match the first part of "gbsct". An implementation should disregard the subdivision id (or return an error).
• en-u-sd-gbsct is valid but not recommended — an implementation that ignores the unicode_subdivision_id can get the wrong fallback behavior, or could add likely subtags and get the invalid en-Latn-US-u-sd-gbsct

In version 28.0, the subdivisions in the validity files used the ISO format, uppercase with a hyphen separating two components, instead of the BCP 47 format.

3.7 Unicode BCP 47 T Extension

The Unicode Consortium has registered and is the maintaining authority for two BCP 47 language tag extensions: the extension 'u' for Unicode locale extension [RFC6067] and extension 't' for transformed content [RFC6497]. The Unicode BCP 47 extension data defines the complete list of valid subtags. While the title of the RFC is “Transformed Content”, the abstract makes it clear that the scope is broader than the term "transformed" might indicate to a casual reader: “including content that has been transliterated, transcribed, or translated, or _in some other way influenced by the source. It also provides for additional information used for identification._”

The -t- Extension. The syntax of 't' extension subtags is defined by the rule unicode_locale_extensions in Section 3.2 Unicode locale identifier, except the separator of subtags sep must be always hyphen '-' when the extension is used as a part of BCP 47 language tag. For information about the registration process, meaning, and usage of the 't' extension, see [RFC6497].

These subtags are all in lowercase (that is the canonical casing for these subtags), however, subtags are case-insensitive and casing does not carry any specific meaning. All subtags within the Unicode extensions are alphanumeric characters in length of two to eight that meet the rule extension in the [BCP47].

The following keys are defined for the -t- extension:

3.7.1 T Extension Data Files

The overall structure of the data files is the similar to the U Extension, with the following exceptions.

In the transformed content 't' data file, the name attribute in a <key> element defines a valid field separator subtag. The name attribute in an enclosed <type> element defines a valid field subtag for the field separator subtag. For example:

<key extension="t" name="m0" description="Transform extension mechanism">
    <type name="ungegn" description="United Nations Group of Experts on Geographical Names" since="21"/>
</key>

The data above indicates:

• "m0" is a valid field separator for the transformed content extension 't'.
• field subtag "ungegn" is valid for field separator "m0".
• field subtag "ungegn" was introduced in CLDR 21.

The attributes are:

name

The name of the mechanism, limited to 3-8 characters (or sequences of them). Any indirect type names are listed in 3.6.4 U Extension Data Files.

description

A description of the name, with all and only that information necessary to distinguish one name from others with which it might be confused. Descriptions are not intended to provide general background information.

since

Indicates the first version of CLDR where the name appears. (Required for new items.)

alias

Alternative name, not limited in number of characters. Aliases are intended for compatibility, not to provide all possible alternate names or designations. (Optional)

For information about the registration process, meaning, and usage of the 't' extension, see [RFC6497].

3.8 Compatibility with Older Identifiers

LDML version before 1.7.2 used slightly different syntax for variant subtags and locale extensions. Implementations of LDML may provide backward compatible identifier support as described in following sections.

3.8.1 Old Locale Extension Syntax

LDML 1.7 or older specification used different syntax for representing Unicode locale extensions. The previous definition of Unicode locale extensions had the following structure:

The new specification mandates keys to be two alphanumeric characters and types to be three to eight alphanumeric characters. As the result, new codes were assigned to all existing keys and some types. For example, a new key "co" replaced the previous key "collation", a new type "phonebk" replaced the previous type "phonebook". However, the existing collation type "big5han" already satisfied the new requirement, so no new type code was assigned to the type. All new keys and types introduced after LDML 1.7 satisfy the new requirement, so they do not have aliases dedicated for the old syntax, except time zone types. The conversion between old types and new types can be done regardless of key, with one known exception (old type "traditional" is mapped to new type "trad" for collation and "traditio" for numbering system), and this relationship will be maintained in the future versions unless otherwise noted.

The new specification introduced a new field attribute in addition to key/type pairs in the Unicode locale extension. When it is necessary to map a new Unicode locale identifier with attribute field to a well-formed old locale identifier, a special key name attribute with the value of entire attribute subtags in the new identifier is used. For example, a new identifier ja-u-xxx-yyy-ca-japanese is mapped to an old identifier ja@attribute=xxx-yyy;calendar=japanese .

The chart below shows some example mappings between the new syntax and the old syntax.

Where the old API is supplied the bcp47 language code, or vice versa, the recommendation is to:

1. Have all methods that take the old syntax also take the new syntax, interpreted correctly. For example, "zh-TW-u-co-pinyin" and "zh_TW@collation=pinyin" would both be interpreted as meaning the same.
2. Have all methods (both for old and new syntax) accept all possible aliases for keywords and types. For example, "ar-u-ca-islamicc" would be equivalent to "ar-u-ca-islamic-civil".
   • The one exception is where an alias would only be well-formed with the old syntax, such as "gregorian" (for "gregory").
3. Where an API cannot successfully accept the alternate syntax, throw an exception (or otherwise indicate an error) so that people can detect that they are using the wrong method (or wrong input).
4. Provide a method that tests a purported locale ID string to determine its status:
   1. well-formed - syntactically correct
   2. valid - well-formed and only uses registered language subtags, extensions, keywords, types...
   3. canonical - valid and no deprecated codes or structure.

3.8.2 Legacy Variants

Old LDML specification allowed codes other than registered [BCP47] variant subtags used in Unicode language and locale identifiers for representing variations of locale data. Unicode locale identifiers including such variant codes can be converted to the new [BCP47] compatible identifiers by following the descriptions below:

When converting to old syntax, the Unicode locale extension "-u-va-posix" should be converted to the "POSIX" variant, not to old extension syntax like "@va=posix". This is an exception: The other mappings above should not be reversed.

Examples:

• en_US_POSIX ↔ en-US-u-va-posix
• en_US_POSIX@colNumeric=yes ↔ en-US-u-kn-va-posix
• en-US-POSIX-u-kn-true → en-US-u-kn-va-posix
• en-US-POSIX-u-kn-va-posix → en-US-u-kn-va-posix

👉 Note that the mapping between en_US_POSIX and en-US-u-va-posix is a conversion process, not a canonicalization process.

3.8.3 Relation to OpenI18n

The locale id format generally follows the description in the OpenI18N Locale Naming Guideline [NamingGuideline], with some enhancements. The main differences from those guidelines are that the locale id:

1. does not include a charset (since the data in LDML format always provides a representation of all Unicode characters. The repository is stored in UTF-8, although that can be transcoded to other encodings as well.)
2. adds the ability to have a variant, as in Java
3. adds the ability to discriminate the written language by script (or script variant).
4. is a superset of [BCP47] codes.

3.9 Transmitting Locale Information

In a world of on-demand software components, with arbitrary connections between those components, it is important to get a sense of where localization should be done, and how to transmit enough information so that it can be done at that appropriate place. End-users need to get messages localized to their languages, messages that not only contain a translation of text, but also contain variables such as date, time, number formats, and currencies formatted according to the users' conventions. The strategy for doing the so-called JIT localization is made up of two parts:

1. Store and transmit neutral-format data wherever possible.
   • Neutral-format data is data that is kept in a standard format, no matter what the local user's environment is. Neutral-format is also (loosely) called binary data, even though it actually could be represented in many different ways, including a textual representation such as in XML.
   • Such data should use accepted standards where possible, such as for currency codes.
   • Textual data should also be in a uniform character set (Unicode/10646) to avoid possible data corruption problems when converting between encodings.
2. Localize that data as "close" to the end-user as possible.

There are a number of advantages to this strategy. The longer the data is kept in a neutral format, the more flexible the entire system is. On a practical level, if transmitted data is neutral-format, then it is much easier to manipulate the data, debug the processing of the data, and maintain the software connections between components.

Once data has been localized into a given language, it can be quite difficult to programmatically convert that data into another format, if required. This is especially true if the data contains a mixture of translated text and formatted variables. Once information has been localized into, say, Romanian, it is much more difficult to localize that data into, say, French. Parsing is more difficult than formatting, and may run up against different ambiguities in interpreting text that has been localized, even if the original translated message text is available (which it may not be).

Moreover, the closer we are to end-user, the more we know about that user's preferred formats. If we format dates, for example, at the user's machine, then it can easily take into account any customizations that the user has specified. If the formatting is done elsewhere, either we have to transmit whatever user customizations are in play, or we only transmit the user's locale code, which may only approximate the desired format. Thus the closer the localization is to the end user, the less we need to ship all of the user's preferences around to all the places that localization could possibly need to be done.

Even though localization should be done as close to the end-user as possible, there will be cases where different components need to be aware of whatever settings are appropriate for doing the localization. Thus information such as a locale code or time zone needs to be communicated between different components.

3.9.1 Message Formatting and Exceptions

Windows (FormatMessage, String.Format), Java (MessageFormat) and ICU (MessageFormat, umsg) all provide methods of formatting variables (dates, times, etc) and inserting them at arbitrary positions in a string. This avoids the manual string concatenation that causes severe problems for localization. The question is, where to do this? It is especially important since the original code site that originates a particular message may be far down in the bowels of a component, and passed up to the top of the component with an exception. So we will take that case as representative of this class of issues.

There are circumstances where the message can be communicated with a language-neutral code, such as a numeric error code or mnemonic string key, that is understood outside of the component. If there are arguments that need to accompany that message, such as a number of files or a datetime, those need to accompany the numeric code so that when the localization is finally at some point, the full information can be presented to the end-user. This is the best case for localization.

More often, the exact messages that could originate from within the component are not known outside of the component itself; or at least they may not be known by the component that is finally displaying text to the user. In such a case, the information as to the user's locale needs to be communicated in some way to the component that is doing the localization. That locale information does not necessarily need to be communicated deep within the component; ideally, any exceptions should bundle up some language-neutral message ID, plus the arguments needed to format the message (for example, datetime), but not do the localization at the throw site. This approach has the advantages noted above for JIT localization.

In addition, exceptions are often caught at a higher level; they do not end up being displayed to any end-user at all. By avoiding the localization at the throw site, it the cost of doing formatting, when that formatting is not really necessary. In fact, in many running programs most of the exceptions that are thrown at a low level never end up being presented to an end-user, so this can have considerable performance benefits.

3.10 Unicode Language and Locale IDs

People have very slippery notions of what distinguishes a language code versus a locale code. The problem is that both are somewhat nebulous concepts.

In practice, many people use [BCP47] codes to mean locale codes instead of strictly language codes. It is easy to see why this came about; because [BCP47] includes an explicit region (territory) code, for most people it was sufficient for use as a locale code as well. For example, when typical web software receives a [BCP47] code, it will use it as a locale code. Other typical software will do the same: in practice, language codes and locale codes are treated interchangeably. Some people recommend distinguishing on the basis of "-" versus "_" (for example, zh-TW for language code, zh_TW for locale code), but in practice that does not work because of the free variation out in the world in the use of these separators. Notice that Windows, for example, uses "-" as a separator in its locale codes. So pragmatically one is forced to treat "-" and "_" as equivalent when interpreting either one on input.

Another reason for the conflation of these codes is that very little data in most systems is distinguished by region alone; currency codes and measurement systems being some of the few. Sometimes date or number formats are mentioned as regional, but that really does not make much sense. If people see the sentence "You will have to adjust the value to १,२३४.५६७ from ૭૧,૨૩૪.૫૬" (using Indic digits), they would say that sentence is simply not English. Number format is far more closely associated with language than it is with region. The same is true for date formats: people would never expect to see intermixed a date in the format "2003年4月1日" (using Kanji) in text purporting to be purely English. There are regional differences in date and number format — differences which can be important — but those are different in kind than other language differences between regions.

As far as we are concerned — as a completely practical matter — two languages are different if they require substantially different localized resources. Distinctions according to spoken form are important in some contexts, but the written form is by far and away the most important issue for data interchange. Unfortunately, this is not the principle used in [ISO639], which has the fairly unproductive notion (for data interchange) that only spoken language matters (it is also not completely consistent about this, however).

[BCP47] can express a difference if the use of written languages happens to correspond to region boundaries expressed as [ISO3166] region codes, and has recently added codes that allow it to express some important cases that are not distinguished by [ISO3166] codes. These written languages include simplified and traditional Chinese (both used in Hong Kong S.A.R.); Serbian in Latin script; Azerbaijani in Arab script, and so on.

Notice also that currency codes are different than currency localizations. The currency localizations should largely be in the language-based resource bundles, not in the territory-based resource bundles. Thus, the resource bundle en contains the localized mappings in English for a range of different currency codes: USD → US$, RUR → Rub, AUD → $A and so on. Of course, some currency symbols are used for more than one currency, and in such cases specializations appear in the territory-based bundles. Continuing the example, en_US would have USD → $, while en_AU would have AUD → $. (In protocols, the currency codes should always accompany any currency amounts; otherwise the data is ambiguous, and software is forced to use the user's territory to guess at the currency. For some informal discussion of this, see JIT Localization.)

3.10.1 Written Language

Criteria for what makes a written language should be purely pragmatic; what would copy-editors say? If one gave them text like the following, they would respond that is far from acceptable English for publication, and ask for it to be redone:

1. "Theatre Center News: The date of the last version of this document was 2003年3月20日. A copy can be obtained for $50,0 or 1.234,57 грн. We would like to acknowledge contributions by the following authors (in alphabetical order): Alaa Ghoneim, Behdad Esfahbod, Ahmed Talaat, Eric Mader, Asmus Freytag, Avery Bishop, and Doug Felt."

So one would change it to either B or C below, depending on which orthographic variant of English was the target for the publication:

2. "Theater Center News: The date of the last version of this document was 3/20/2003. A copy can be obtained for $50.00 or 1,234.57 Ukrainian hryvni. We would like to acknowledge contributions by the following authors (in alphabetical order): Alaa Ghoneim, Ahmed Talaat, Asmus Freytag, Avery Bishop, Behdad Esfahbod, Doug Felt, Eric Mader."
3. "Theatre Centre News: The date of the last version of this document was 20/3/2003. A copy can be obtained for $50.00 or 1,234.57 Ukrainian hryvni. We would like to acknowledge contributions by the following authors (in alphabetical order): Alaa Ghoneim, Ahmed Talaat, Asmus Freytag, Avery Bishop, Behdad Esfahbod, Doug Felt, Eric Mader."

Clearly there are many acceptable variations on this text. For example, copy editors might still quibble with the use of first versus last name sorting in the list, but clearly the first list was not acceptable English alphabetical order. And in quoting a name, like "Theatre Centre News", one may leave it in the source orthography even if it differs from the publication target orthography. And so on. However, just as clearly, there are limits on what is acceptable English, and "2003年3月20日", for example, is not.

Note that the language of locale data may differ from the language of localized software or web sites, when those latter are not localized into the user's preferred language. In such cases, the kind of incongruous juxtapositions described above may well appear, but this situation is usually preferable to forcing unfamiliar date or number formats on the user as well.

3.10.2 Hybrid Locale Identifiers

Hybrid locales have intermixed content from 2 (or more) languages, often with one language's grammatical structure applied to words in another. These are commonly referred to with portmanteau words such as Franglais, ​Spanglish or Denglish. Hybrid locales do not not reference text simply containing two languages: a book of parallel text containing English and French, such as the following, is not Franglais:

While text in a document can be tagged as partly in one language and partly in another, that is not the same having a hybrid locale. There is a difference between having a Spanglish document, and a Spanish document that has some passages quoted in English. Fine-grained tagging doesn't handle grammatical combinations like Tanglish “Enna matteru?” (What’s the matter?), which is neither standard Tamil nor standard English. More importantly, it doesn’t work for the very common use case for a unicode_locale_id: locale selection.

To communicate requests for localized content and internationalization services, locales are used. When people pick a language from a menu, internally they are picking a locale (en-GB, es-419, etc.). To allow an application to support Spanglish or Hinglish locale selection, unicode_locale_ids can represent hybrid locales using the T Extension key-value 'h0-hybrid'. (For more information on the T extension, see Section 3.7 Unicode BCP 47 T Extension.)

However, if users typically expect a their language in non-default script to contain a significant amount of text due to lexical borrowing, then the -t- and hybrid subtags may be omitted. An example of this is when Hindi is written in Latin script since Romanized Hindi typically contains a significant amount of English text, ‘hi-Latn’ can be used instead of ‘hi-Latn-t-en-h0-hybrid’. This tends to work better in implementations that don't yet handle the -t- extension.

Examples:

* When used as a request for international services (such as date formatting), the request is for everything to be in the base script if possible. When used to tag arbitrary content on a coarse level, the expectation is that it be the predominant script — that is, there may be certain passages or phrases that are in the other script but are not tagged on a fine-grained level.

Note: The unicode_language_id should be the language used as the ‘scaffold’: for the fallback locale for internationalization services, typically used for more of the core vocabulary/structure in the content. Thus where Hindi is the scaffold, Hinglish should be represented as hi-t-en-h0-hybrid (when written in Devanagari script) or hi-Latn-t-en-h0-hybrid (when written in Latin characters). Where English is the scaffold, Hinglish should be represented as en-t-hi-h0-hybrid (or possibly en-Deva-t-hi-h0-hybrid).

The value of -t- is a full unicode_language_id, and can contain a subtag for the region where it is important to include it, as in the following. The value can also include the script, although that is not normally included: the only instance where it should be is where the content of the source text varies by script. So because zh-Hant has different vocabulary and expressions, it could make sense to have en-t-zh-hant to make that distinction.

Note: The default script for the language is computed without reference to the hybrid subtags. Thus the default script for 'ru' is “Cyrl”, no matter what the source is in the -t- tag.

Should there ever be strong need for hybrids of more than two languages or for other purposes such as hybrid languages as the source of translated content, additional structure could be added.

3.11 Validity Data

<!ELEMENT idValidity (id*) >
<!ELEMENT id ( #PCDATA ) >
<!ATTLIST id type NMTOKEN #REQUIRED >
<!ATTLIST id idStatus NMTOKEN #REQUIRED >

The directory common/validity contains machine-readable data for validating the language, region, script, and variant subtags, as well as currency, subdivisions and measure units. Each file contains a number of subtags with the following idStatus values:

• regular — the standard codes used for the specific type of subtag
• special — certain exceptional language codes like 'mul' (languages only)
• unknown — the code used to indicate the "unknown", "undetermined" or "invalid" values. For more information, see Section 3.5.1 Unknown or Invalid Identifiers.
• macroregion — the standard codes that are macroregions (for regions only).
  • Note that some two-letter region codes are macroregions, and (in the future) some three-digit codes may be regular codes.
  • For details as to which regions are contained within which macroregions, see the <containment> element of the supplemental data.
• deprecated — codes that should not be used. The <alias> element in the supplementalMeta file contains more information about these codes, and which codes should be used instead.
• private_use — codes that, for CLDR, are considered private use. Note that some private-use codes in a source standard such as BCP 47 have defined CLDR semantics, and are considered regular codes. For more information, see Section 3.5.3 Private Use Codes.
• reserved — codes that are private use in a source standard, but are reserved for future use as regular codes by CLDR.

The list of subtags for each idStatus use a compact format as a space-delimited list of StringRanges, as defined in Section 5.3.4 String Range. The separator for each StringRange is a "~".

Each measure unit is a sequence of subtags, such as “angle-arc-minute”. The first subtag provides a general “category” of the unit.

In version 28.0, the subdivisions in the validity files used the ISO format, uppercase with a hyphen separating two components, instead of the BCP 47 format.

4 Locale Inheritance and Matching

The XML format relies on an inheritance model, whereby the resources are collected into bundles, and the bundles organized into a tree. Data for the many Spanish locales does not need to be duplicated across all of the countries having Spanish as a national language. Instead, common data is collected in the Spanish language locale, and territory locales only need to supply differences. The parent of all of the language locales is a generic locale known as root. Wherever possible, the resources in the root are language & territory neutral. For example, the collation (sorting) order in the root is based on the [DUCET] (see Root Collation). Since English language collation has the same ordering as the root locale, the 'en' locale data does not need to supply any collation data, nor do the 'en_US', 'en_GB' or the any of the various other locales that use English.

Given a particular locale id "en_US_someVariant", the default search chain for a particular resource is the following.

en_US_someVariant
en_US
en
root

The inheritance is often not simple truncation, as will be seen later in this section.

The default search chain is slighly different for multiple variants. In that case, the inheritance chain covers all combinations of variants, with longest number of variants first, and otherwise in alphabetical order. For example, where the requested locale ID is en_fonipa_scouse, the inheritance chain is as follows:

en_GB_fonipa_scouse
en_GB_scouse_fonipa // extra step, only needed if not canonical
en_GB_fonipa
en_GB_scouse // extra step
en_GB
en

If the data for the implementation performing the inheritance doesn't require canonical locale identifiers, then extra locale IDs need to be inserted in the chain. That is indicated in the example above, marked with "only needed if not canonical". These would would include all combinations of variants that are not in canonical order, inserted in alphabetical order. Note that the order of multiple variants in canonical locale identifiers is alphabetical, as per 5. Canonicalizing Syntax in Annex C. LocaleId Canonicalization.

If a type and key are supplied in the locale id, then logically the chain from that id to the root is searched for a resource tag with a given type, all the way up to root. If no resource is found with that tag and type, then the chain is searched again without the type.

Thus the data for any given locale will only contain resources that are different from the parent locale. For example, most territory locales will inherit the bulk of their data from the language locale: "en" will contain the bulk of the data: "en_IE" will only contain a few items like currency. All data that is inherited from a parent is presumed to be valid, just as valid as if it were physically present in the file. This provides for much smaller resource bundles, and much simpler (and less error-prone) maintenance. At the script or region level, the "primary" child locale will be empty, since its parent will contain all of the appropriate resources for it. For more information see CLDR Information: Section 9.3 Default Content.

Certain data items depend only on the region specified in a locale id (by a unicode_region_subtag or an “rg” Region Override key), and are obtained from supplemental data rather than through locale resources. For example:

• The currency for the specified region (see Supplemental Currency Data)
• The measurement system for the specified region (see Measurement System Data)
• The week conventions for the specified region (see Week Data)

(For more information on the specific items handled this way, see Territory-Based Preferences.) These items will be correct for the specified region regardless of whether a locale bundle actually exists with the same combination of language and region as in the locale id. For example, suppose data is requested for the locale id "fr_US" and there is no bundle for that combination. Data obtained via locale inheritance, such as currency patterns and currency symbols, will be obtained from the parent locale "fr". However, currency amounts would be formatted by default using US dollars, just displayed in the manner governed by the locale "fr". When a locale id does not specify a region, the region-specific items such as those above are obtained from the likely region for the locale (obtained via Likely Subtags).

For the relationship between Inheritance, DefaultContent, LikelySubtags, and LocaleMatching, see Section 4.2.6 Inheritance vs Related Information.

4.1 Lookup

If a language has more than one script in customary modern use, then the CLDR file structure in common/main follows the following model:

lang
lang_script
lang_script_region
lang_region (aliases to lang_script_region)

4.1.1 Bundle vs Item Lookup

There are actually two different kinds of inheritance fallback: resource bundle lookup and resource item lookup. For the former, a process is looking to find the first, best resource bundle it can; for the later, it is fallback within bundles on individual items, like the translated name for the region "CN" in Breton.

These are closely related, but distinct, processes. They are illustrated in the table Lookup Differences, where "key" stands for zero or more key/type pairs. Logically speaking, when looking up an item for a given locale, you first do a resource bundle lookup to find the best bundle for the locale, then you do an inherited item lookup starting with that resource bundle.

The table Lookup Differences uses the naïve resource bundle lookup for illustration. More sophisticated systems will get far better results for resource bundle lookup if they use the algorithm described in Section 4.4 Language Matching. That algorithm takes into account both the user’s desired locale(s) and the application’s supported locales, in order to get the best match.

If the naïve resource bundle lookup is used, the desired locale needs to be canonicalized using 4.3 Likely Subtags and the supplemental alias information, so that locales that CLDR considers identical are treated as such. Thus eng-Latn-GB should be mapped to en-GB, and cmn-TW mapped to zh-Hant-TW.

For the purposes of CLDR, everything with the <ldml> dtd is treated logically as if it is one resource bundle, even if the implementation separates data into separate physical resource bundles. For example, suppose that there is a main XML file for Nama (naq), but there are no <unit> elements for it because the units are all inherited from root. If the <unit> elements are separated into a separate data tree for modularity in the implementation, the Nama <unit> resource bundle would be empty. However, for purposes of resource-bundle lookup the resource bundle lookup still stops at naq.xml.

Both the resource bundle inheritance and the inherited item inheritance use the parentLocale data, where available, instead of simple truncation.

The fallback is a bit different for these two cases; internal aliases and keys are not involved in the bundle lookup, and the default locale is not involved in the item lookup. If the default-locale were used in the resource-item lookup, then strange results will occur. For example, suppose that the default locale is Swedish, and there is a Nama locale but no specific inherited item for collation. If the default-locale were used in resource-item lookup, it would produce odd and unexpected results for Nama sorting.

The default locale is not even always used in resource bundle inheritance. For the following services, the fallback is always directly to the root locale rather than through default locale.

• collation
• break iteration
• case mapping
• transliteration
  • The lookup for transliteration is yet more complicated because of the interplay of source and target locales: see Part 2 General, Section 10.1 Inheritance.

Thus if there is no Akan locale, for example, asking for a collation for Akan should produce the root collation, not the Swedish collation.

The inherited item lookup must remain stable, because the resources are built with a certain fallback in mind; changing the core fallback order can render the bundle structure incoherent.

Resource bundle lookup, on the other hand, is more flexible; changes in the view of the "best" match between the input request and the output bundle are more tolerant, when represent overall improvements for users. For more information, see A.1 Element fallback.

Where the LDML inheritance relationship does not match a target system, such as POSIX, the data logically should be fully resolved in converting to a format for use by that system, by adding all inherited data to each locale data set.

For a more complete description of how inheritance applies to data, and the use of keywords, see Section 4.2 Inheritance .

The locale data does not contain general character properties that are derived from the Unicode Character Database [UAX44]. That data being common across locales, it is not duplicated in the bundles. Constructing a POSIX locale from the CLDR data requires use of UCD data. In addition, POSIX locales may also specify the character encoding, which requires the data to be transformed into that target encoding.

Warning: If a locale has a different script than its parent (for example, sr_Latn), then special attention must be paid to make sure that all inheritance is covered. For example, auxiliary exemplar characters may need to be empty ("[]") to block inheritance.

Empty Override: There is one special value reserved in LDML to indicate that a child locale is to have no value for a path, even if the parent locale has a value for that path. That value is "∅∅∅". For example, if there is no phrase for "two days ago" in a language, that can be indicated with:

<field type="day">
  <relative type="-2">∅∅∅</relative>

4.1.2 Lateral Inheritance

Lateral Inheritance is where resources are inherited from within the same locale, before inheriting from the parent. This is used for the following element@attribute instances:

* The "unless otherwise specified" clause is for when an API or other context indicates a different choice, such as currencyFormat type="accounting".

For example, with /currency [@type="CVE"], the decimal symbol for almost all locales is the value from symbols/decimal, but for pt_CV it is explicitly <decimal>$</decimal>.

The following attributes use lateral inheritance for all elements with the DTD root = ldml, except where otherwise noted. The process is applied recursively.

The gender fallback is to neuter if the locale has a neuter gender, otherwise masculine. This may be extended in the future if necessary. See also Part 2, Section 15, Grammatical Features.

For example, if there is no value for a path, and that path has a [@count="x"] attribute and value, then:

1. If "x" is numeric, the path falls back to the path with [@count=«the plural rules category for x for that locale»], within that the same locale.
   1. For example, [@count="0"] for English falls back to [@count="other"], while for French falls back to [@count="one"].
2. If "x" is anything but "other", it falls back to a path [@count="other"], within that the same locale.
3. If "x" is "other", it falls back to the path that is completely missing the count item, within that the same locale.
4. If there is no value for that path the same locale, the same process is used for the original path in the parent locale.

A path may have multiple attributes with lateral inheritance. In such a case, all of the combinations are tried, and in the order supplied above. For example (this is an extreme case):

/compoundUnitPattern1[@count="few"][@gender="feminine"][@case="accusative">] →
/compoundUnitPattern1[@count="few"][@gender="feminine"][@case="nominative">] →
/compoundUnitPattern1[@count="few"][@gender="feminine"] →
/compoundUnitPattern1[@count="few"][@gender="neuter"][@case="accusative">] →
/compoundUnitPattern1[@count="few"][@gender="neuter"][@case="nominative">] →
/compoundUnitPattern1[@count="few"][@gender="neuter"] →
/compoundUnitPattern1[@count="few"][@case="accusative">] →
/compoundUnitPattern1[@count="few"][@case="nominative">] →
/compoundUnitPattern1[@count="few"] →

/compoundUnitPattern1[@count="other"][@gender="feminine"][@case="accusative">] →
/compoundUnitPattern1[@count="other"][@gender="feminine"][@case="nominative">] →
/compoundUnitPattern1[@count="other"][@gender="feminine"] →
/compoundUnitPattern1[@count="other"][@gender="neuter"][@case="accusative">] →
/compoundUnitPattern1[@count="other"][@gender="neuter"][@case="nominative">] →
/compoundUnitPattern1[@count="other"][@gender="neuter"] →
/compoundUnitPattern1[@count="other"][@case="accusative">] →
/compoundUnitPattern1[@count="other"][@case="nominative">] →
/compoundUnitPattern1[@count="other"] →

/compoundUnitPattern1[@gender="feminine"][@case="accusative">] →
/compoundUnitPattern1[@gender="feminine"][@case="nominative">] →
/compoundUnitPattern1[@gender="feminine"] →
/compoundUnitPattern1[@gender="neuter"][@case="accusative">] →
/compoundUnitPattern1[@gender="neuter"][@case="nominative">] →
/compoundUnitPattern1[@gender="neuter"] →
/compoundUnitPattern1[@case="accusative">] →
/compoundUnitPattern1[@case="nominative">] →
/compoundUnitPattern1

Examples:

Note that there may also be an alias in root that changes the path and starts again from the requested locale, such as:

<unitLength type="narrow">
   <alias source="locale" path="../unitLength[@type='short']"/>
</unitLength>

4.1.3 Parent Locales

<!ELEMENT parentLocales ( parentLocale* ) >
<!ELEMENT parentLocale EMPTY >
<!ATTLIST parentLocale parent NMTOKEN #REQUIRED >
<!ATTLIST parentLocale locales NMTOKENS #REQUIRED >

In some cases, the normal truncation inheritance does not function well. This happens when:

1. The child locale is of a different script. In this case, mixing elements from the parent into the child data results in a mishmash.
2. A large number of child locales behave similarly, and differently from the truncation parent.

The parentLocale element is used to override the normal inheritance when accessing CLDR data.

For case 1, the children are script locales, and the parent is "root". For example:

<parentLocale parent="root" locales="az_Cyrl ha_Arab … zh_Hant"/>

For case 2, the children and parent share the same primary language, but the region is changed. For example:

<parentLocale parent="es_419" locales="es_AR es_BO … es_UY es_VE"/>

Collation data, however, is an exception. Since collation rules do not truly inherit data from the parent, the parentLocale element is not necessary and not used for collation. Thus, for a locale like zh_Hant in the example above, the parentLocale element would dictate the parent as "root" when referring to main locale data, but for collation data, the parent locale would still be "zh", even though the parentLocale element is present for that locale.

Since parentLocale information is not localizable on a per locale basis, the parentLocale information is contained in CLDR’s supplemental data.

When a parentLocale element is used to override normal inheritance, the following guidelines apply in most cases:

1. If X is the parentLocale of Y, then either X is the root locale, or X has the same base language code as Y. For example, the parent of en cannot be fr, and the parent of en_YY cannot be fr or fr_XX.
2. If X is the parentLocale of Y, Y must not be a base language locale. For example, the parent of en cannot be en_XX.

There may be specific exceptions to these for certain closely-related languages or language-script combinations, for example:

• no may be the parent of nb and nn.
• en_IN may be the parent of hi_Latn (the parent is one of the languages for a child that is effectively a hybrid of two languages in Latn script)

There are certain invariants that must always be true:

3. The parent must either be the root locale or have the same script as the child.
4. There must never be cycles, such as: X parent of Y ... parent of X.
5. Following the inheritance path, using parentLocale where available and otherwise truncating the locale, must always lead eventually to the root locale.

4.2 Inheritance and Validity

The following describes in more detail how to determine the exact inheritance of elements, and the validity of a given element in LDML.

4.2.1 Definitions

Blocking elements are those whose subelements do not inherit from parent locales. For example, a <collation> element is a blocking element: everything in a <collation> element is treated as a single lump of data, as far as inheritance is concerned. For more information, see Section 5.5 Valid Attribute Values.

Attributes that serve to distinguish multiple elements at the same level are called distinguishing attributes. For example, the type attribute distinguishes different elements in lists of translations, such as:

<language type="aa">Afar</language>
<language type="ab">Abkhazian</language>

Distinguishing attributes affect inheritance; two elements with different distinguishing attributes are treated as different for purposes of inheritance. For more information, see Section 5.5 Valid Attribute Values. Other attributes are called value attributes. Value attributes do not affect inheritance, and elements with value attributes may not have child elements (see XML Format).

Non-distinguishing attributes are identified by DTD Annotations such as @VALUE.

For any element in an XML file, an element chain is a resolved [XPath] leading from the root to an element, with attributes on each element in alphabetical order. So in, say, https://github.com/unicode-org/cldr/blob/main/common/main/el.xml we may have:

<ldml>
    <identity>
        <version number="1.1" />
        <language type="el" />
    </identity>
    <localeDisplayNames>
        <languages>
            <language type="ar">Αραβικά</language>
...

Which gives the following element chains (among others):

• //ldml/identity/version[@number="1.1"]
• //ldml/localeDisplayNames/languages/language[@type="ar"]

An element chain A is an extension of an element chain B if B is equivalent to an initial portion of A. For example, #2 below is an extension of #1. (Equivalent, depending on the tree, may not be "identical to". See below for an example.)

1. //ldml/localeDisplayNames
2. //ldml/localeDisplayNames/languages/language[@type="ar"]

An LDML file can be thought of as an ordered list of element pairs: <element chain, data>, where the element chains are all the chains for the end-nodes. (This works because of restrictions on the structure of LDML, including that it does not allow mixed content.) The ordering is the ordering that the element chains are found in the file, and thus determined by the DTD.

For example, some of those pairs would be the following. Notice that the first has the null string as element contents.

• <//ldml/identity/version[@number="1.1"], "">
• <//ldml/localeDisplayNames/languages/language[@type="ar"], "Αραβικά">

Note: There are two exceptions to this:

1. Blocking nodes and their contents are treated as a single end node.
2. In terms of computing inheritance, the element pair consists of the element chain plus all distinguishing attributes; the value consists of the value (if any) plus any nondistinguishing attributes.

Thus instead of the element pair being (a) below, it is (b):

1. <//ldml/dates/calendars/calendar[@type='gregorian']/week/weekendStart[@day='sun'][@time='00:00'],"">
2. <//ldml/dates/calendars/calendar[@type='gregorian']/week/weekendStart,[@day='sun'][@time='00:00']>

Two LDML element chains are equivalent when they would be identical if all attributes and their values were removed — except for distinguishing attributes. Thus the following are equivalent:

• //ldml/localeDisplayNames/languages/language[@type="ar"]
• //ldml/localeDisplayNames/languages/language[@type="ar"][@draft="unconfirmed"]

For any locale ID, a locale chain is an ordered list starting with the root and leading down to the ID. For example:

<root, de, de_DE, de_DE_xxx>

4.2.2 Resolved Data File

To produce fully resolved locale data file from CLDR for a locale ID L, you start with L, and successively add unique items from the parent locales until you get up to root. More formally, this can be expressed as the following procedure.

1. Let Result be initially L.
2. For each Li in the locale chain for L, starting at L and going up to root:
   1. Let Temp be a copy of the pairs in the LDML file for Li
   2. Replace each alias in Temp by the resolved list of pairs it points to.
      1. The resolved list of pairs is obtained by recursively applying this procedure.
      2. That alias now blocks any inheritance from the parent. (See Section 5.1 Common Elements for an example.)
   3. For each element pair P in Temp:
      1. If P does not contain a blocking element, and Result does not have an element pair Q with an equivalent element chain, add P to Result.

Notes:

• When adding an element pair to a result, it has to go in the right order for it to be valid according to the DTD.
• The identity element and its children are unaffected by resolution.
• The LDML data must be constructed so as to avoid circularity in step 2.2.

4.2.3 Valid Data

The attribute draft="x" in LDML means that the data has not been approved by the subcommittee. (For more information, see Process). However, some data that is not explicitly marked as draft may be implicitly draft, either because it inherits it from a parent, or from an enclosing element.

Example 2. Suppose that new locale data is added for af (Afrikaans). To indicate that all of the data is unconfirmed, the attribute can be added to the top level.

<ldml version="1.1" draft="unconfirmed">
    <identity>
        <version number="1.1" />
        <language type="af" />
    </identity>
    <characters>...</characters>
    <localeDisplayNames>...</localeDisplayNames>
</ldml>

Any data can be added to that file, and the status will all be draft="unconfirmed". Once an item is vetted—_whether it is inherited or explicitly in the file_—then its status can be changed to approved. This can be done either by leaving draft="unconfirmed" on the enclosing element and marking the child with draft="approved", such as:

<ldml version="1.1" draft="unconfirmed">
    <identity>
        <version number="1.1" />
        <language type="af" />
    </identity>
    <characters draft="approved">...</characters>
    <localeDisplayNames>...</localeDisplayNames>
    <dates />
    <numbers />
    <collations />
</ldml>

However, normally the draft attributes should be canonicalized, which means they are pushed down to leaf nodes as described in Section 5.6 Canonical Form. If an LDML file does have draft attributes that are not on leaf nodes, the file should be interpreted as if it were the canonicalized version of that file.

More formally, here is how to determine whether data for an element chain E is implicitly or explicitly draft, given a locale L. Sections 1, 2, and 4 are simply formalizations of what is in LDML already. Item 3 adds the new element.

4.2.4 Checking for Draft Status

1. Parent Locale Inheritance
   1. Walk through the locale chain until you find a locale ID L' with a data file D. (L' may equal L).
   2. Produce the fully resolved data file D' for D.
   3. In D', find the first element pair whose element chain E' is either equivalent to or an extension of E.
   4. If there is no such E', return true
   5. If E' is not equivalent to E, truncate E' to the length of E.
2. Enclosing Element Inheritance
   1. Walk through the elements in E', from back to front.
      1. If you ever encounter draft=x, return x
   2. If L' = L, return false
3. Missing File Inheritance
   1. Otherwise, walk again through the elements in E', from back to front.
      1. If you encounter a validSubLocales attribute (deprecated):
         1. If L is in the attribute value, return false
         2. Otherwise return true
4. Otherwise
   1. Return true

The validSubLocales in the most specific (farthest from root file) locale file "wins" through the full resolution step (data from more specific files replacing data from less specific ones).

4.2.5 Keyword and Default Resolution

When accessing data based on keywords, the following process is used. Consider the following example:

• The locale 'de' has collation types A, B, C, and no <default> element
• The locale 'de_CH' has <default type='B'>

Here are the searches for various combinations.

Examples of "search" collator lookup; 'de' has a language-specific version, but 'en' does not:

Examples of lookup for Chinese collation types. Note:

• All of the Chinese-specific collation types are provided in the 'zh' locale
• For 'zh' the <default> element specifies "pinyin"; for 'zh_Hant' the <default> element specifies "stroke". However any of the available Chinese collation types can be explicitly requested for any Chinese locale.

Note: It is an invariant that the default in root for a given element must always be a value that exists in root. So you can not have the following in root:

<someElements>
    <default type='a'/>
    <someElement type='b'>...</someElement>
    <someElement type='c'>...</someElement>
    <!-- no 'a' -->
</someElements>

For identifiers, such as language codes, script codes, region codes, variant codes, types, keywords, currency symbols or currency display names, the default value is the identifier itself whenever no value is found in the root. Thus if there is no display name for the region code 'QA' in root, then the display name is simply 'QA'.

4.2.6 Inheritance vs Related Information

There are related types of data and processing that are easy to confuse:

4.3 Likely Subtags

<!ELEMENT likelySubtag EMPTY >
<!ATTLIST likelySubtag from NMTOKEN #REQUIRED>
<!ATTLIST likelySubtag to NMTOKEN #REQUIRED>

There are a number of situations where it is useful to be able to find the most likely language, script, or region. For example, given the language "zh" and the region "TW", what is the most likely script? Given the script "Thai" what is the most likely language or region? Given the region TW, what is the most likely language and script?

Conversely, given a locale, it is useful to find out which fields (language, script, or region) may be superfluous, in the sense that they contain the likely tags. For example, "en_Latn" can be simplified down to "en" since "Latn" is the likely script for "en"; "ja_Jpan_JP" can be simplified down to "ja".

The likelySubtag supplemental data provides default information for computing these values. This data is based on the default content data, the population data, and the suppress-script data in [BCP47]. It is heuristically derived, and may change over time.

For the relationship between Inheritance, DefaultContent, LikelySubtags, and LocaleMatching, see Section 4.2.6 Inheritance vs Related Information.

To look up data in the table, see if a locale matches one of the from attribute values. If so, fetch the corresponding to attribute value. For example, the Chinese data looks like the following:

<likelySubtag from="zh" to="zh_Hans_CN" />
<likelySubtag from="zh_HK" to="zh_Hant_HK" />
<likelySubtag from="zh_Hani" to="zh_Hani_CN" />
<likelySubtag from="zh_Hant" to="zh_Hant_TW" />
<likelySubtag from="zh_MO" to="zh_Hant_MO" />
<likelySubtag from="zh_TW" to="zh_Hant_TW" />

So looking up "zh_TW" returns "zh_Hant_TW", while looking up "zh" returns "zh_Hans_CN".

In more detail, the data is designed to be used in the following operations.

Note that as of CLDR v24, any field present in the 'from' field is also present in the 'to' field, so an input field will not change in "Add Likely Subtags" operation. The data and operations can also be used with language tags using [BCP47] syntax, with the appropriate changes. In addition, certain common 'denormalized' language subtags such as 'iw' (for 'he') may occur in both the 'from' and 'to' fields. This allows for implementations that use those denormalized subtags to use the data with only minor changes to the operations.

An implementation may choose to exclude language tags with the language subtag "und" from the following operation. In such a case, only the canonicalization is done. An implementation can declare that it is doing the exclusion, or can take a parameter that controls whether or not to do it.

Add Likely Subtags: Given a source locale X, to return a locale Y where the empty subtags have been filled in by the most likely subtags. This is written as X ⇒ Y ("X maximizes to Y").

A subtag is called empty if it is a missing script or region subtag, or it is a base language subtag with the value "und". In the description below, a subscript on a subtag x indicates which tag it is from: xs is in the source, xm is in a match, and xr is in the final result.

This operation is performed in the following way.

1. Canonicalize.
   1. Make sure the input locale is in canonical form: uses the right separator, and has the right casing.
   2. Replace any deprecated subtags with their canonical values using the <alias> data in supplemental metadata. Use the first value in the replacement list, if it exists. Language tag replacements may have multiple parts, such as "sh" ➞ "sr_Latn" or "mo" ➞ "ro_MD". In such a case, the original script and/or region are retained if there is one. Thus "sh_Arab_AQ" ➞ "sr_Arab_AQ", not "sr_Latn_AQ".
   3. If the tag is a legacy language tag (marked as “Type: grandfathered” in BCP 47; see <variable id="$grandfathered" type="choice"> in the supplemental data), then return it.
   4. Remove the script code 'Zzzz' and the region code 'ZZ' if they occur.
   5. Get the components of the cleaned-up source tag (languages, scripts, and regions), plus any variants and extensions.
2. Lookup. Look up each of the following in order, and stop on the first match:
   1. languages_scripts_regions
   2. languages_regions
   3. languages_scripts
   4. languages
   5. und_scripts
3. Return
   1. If there is no match, either return
      1. an error value, or
      2. the match for "und" (in APIs where a valid language tag is required).
   2. Otherwise there is a match = languagem_scriptm_regionm
   3. Let xr = xs if xs is not empty, and xm otherwise.
   4. Return the language tag composed of languager _ scriptr _ regionr + variants + extensions .

The lookup can be optimized. For example, if any of the tags in Step 2 are the same as previous ones in that list, they do not need to be tested.

Example1:

• Input is ZH-ZZZZ-SG.
• Normalize to zh_SG.
• Look up in table. No match.
• Look up zh, and get the match (zh_Hans_CN). Substitute SG, and return zh_Hans_SG.

To find the most likely language for a country, or language for a script, use "und" as the language subtag. For example, looking up "und_TW" returns zh_Hant_TW.

A goal of the algorithm is that if X ⇒ Y, and X' results from replacing an empty subtag in X by the corresponding subtag in Y, then X' ⇒ Y. For example, if und_AF ⇒ fa_Arab_AF, then:

• fa_Arab_AF ⇒ fa_Arab_AF
• und_Arab_AF ⇒ fa_Arab_AF
• fa_AF ⇒ fa_Arab_AF

There are a small number of exceptions to this goal in the current data, where X ∈ {und_Bopo, und_Brai, und_Cakm, und_Limb, und_Shaw}.

Remove Likely Subtags: Given a locale, remove any fields that Add Likely Subtags would add.

The reverse operation removes fields that would be added by the first operation.

1. First get max = AddLikelySubtags(inputLocale). If an error is signaled, return it.
2. Remove the variants from max.
3. Get the components of the max (languagemax, scriptmax, regionmax).
4. Then for trial in {languagemax, languagemax_regionmax, languagemax_scriptmax}
   • If AddLikelySubtags(trial) = max, then return trial + variants.
5. If you do not get a match, return max + variants.

Example:

• Input is zh_Hant. Maximize to get zh_Hant_TW.
• zh => zh_Hans_CN. No match, so continue.
• zh_TW => zh_Hant_TW. Matches, so return zh_TW.

A variant of this favors the script over the region, thus using {language, language_script, language_region} in the above. If that variant is used, then the result in this example would be zh_Hant instead of zh_TW.

4.4 Language Matching

<!ELEMENT languageMatching ( languageMatches* ) >
<!ELEMENT languageMatches ( paradigmLocales*, matchVariable*, languageMatch* ) >
<!ATTLIST languageMatches type NMTOKEN #REQUIRED >

<!ELEMENT languageMatch EMPTY >
<!ATTLIST languageMatch desired CDATA #REQUIRED >
<!ATTLIST languageMatch supported CDATA #REQUIRED >
<!ATTLIST languageMatch percent NMTOKEN #REQUIRED >
<!ATTLIST languageMatch distance NMTOKEN #IMPLIED >
<!ATTLIST languageMatch oneway ( true | false ) #IMPLIED >

<!ELEMENT languageMatches ( paradigmLocales*, matchVariable*, languageMatch* ) >
<!ATTLIST languageMatches type NMTOKEN #REQUIRED >

<!ELEMENT paradigmLocales EMPTY >
<!ATTLIST paradigmLocales locales NMTOKENS #REQUIRED >

Implementers are often faced with the issue of how to match the user's requested languages with their product's supported languages. For example, suppose that a product supports {ja-JP, de, zh-TW}. If the user understands written American English, German, French, Swiss German, and Italian, then de would be the best match; if s/he understands only Chinese (zh), then zh-TW would be the best match.

The standard truncation-fallback algorithm does not work well when faced with the complexities of natural language. The language matching data is designed to fill that gap. Stated in those terms, language matching can have the effect of a more complex fallback, such as:

sr-Cyrl-RS
sr-Cyrl
sr-Latn-RS
sr-Latn
sr
hr-Latn
hr

Language matching is used to find the best supported locale ID given a requested list of languages. The requested list could come from different sources, such as the user's list of preferred languages in the OS Settings, or from a browser Accept-Language list. For example, if my native tongue is English, I can understand Swiss German and German, my French is rusty but usable, and Italian basic, ideally an implementation would allow me to select {gsw, de, fr} as my preferred list of languages, skipping Italian because my comprehension is not good enough for arbitrary content.

Language Matching can also be used to get fallback data elements. In many cases, there may not be full data for a particular locale. For example, for a Breton speaker, the best fallback if data is unavailable might be French. That is, suppose we have found a Breton bundle, but it does not contain translation for the key "CN" (for the country China). It is best to return "chine", rather than falling back to the value default language such as Russian and getting "Китай". The language matching data can be used to get the closest fallback locales (of those supported) to a given language.

For the relationship between Inheritance, DefaultContent, LikelySubtags, and LocaleMatching, see Section 4.2.6 Inheritance vs Related Information.

When such fallback is used for inherited item lookup, the normal order of inheritance is used for inherited item lookup, except that before using any data from root, the data for the fallback locales would be used if available. Language matching does not interact with the fallback of resources within the locale-parent chain. For example, suppose that we are looking for the value for a particular path P in nb-NO. In the absence of aliases, normally the following lookup is used.

nb-NO → nb → root

That is, we first look in nb-NO. If there is no value for P there, then we look in nb. If there is no value for P there, we return the value for P in root (or a code value, if there is nothing there). Remember that if there is an alias element along this path, then the lookup may restart with a different path in nb-NO (or another locale).

However, suppose that nb-NO has the fallback values [nn da sv en], derived from language matching. In that case, an implementation may progressively look up each of the listed locales, with the appropriate substitutions, returning the first value that is not found in root. This follows roughly the following pseudocode:

value = lookup(P, nb-NO); if (locationFound != root) return value;
value = lookup(P, nn-NO); if (locationFound != root) return value;
value = lookup(P, da-NO); if (locationFound != root) return value;
value = lookup(P, sv-NO); if (locationFound != root) return value;
value = lookup(P, en-NO); return value;

The locales in the fallback list are not used recursively. For example, for the lookup of a path in nb-NO, if fr were a fallback value for da, it would not matter for the above process. Only the original language matters.

The language matching data is intended to be used according to the following algorithm. This is a logical description, and can be optimized for production in many ways. In this algorithm, the languageMatching data is interpreted as an ordered list.

Distances between given pair of subtags can be larger or smaller than the typical distances. For example, the distance between en and en-GB can be greater than those between en-GB and en-IE. In some cases, language and/or script differences can be as small as the typical region difference. (Example: sr-Latn vs. sr-Cyrl).

The distances resulting from the table are not linear, but are rather chosen to produce expected results. So a distance of 10 is not necessarily twice as "bad" as a distance of 5. Implementations may want to have a mode where script distances should swamp language distances. The tables are built such that this can be accomplished by multiplying the language distance by 0.25.

The language matching algorithm takes a list of a user’s desired languages, and a list of the application’s supported languages.

• Set the best weighted distance BWD to ∞
• Set the best desired language BD to null
• Set the best supported language BS to null
• For each desired language D
  • Compute a demotion value F, based on the position in the list.
    • This demotion value is up to the implementation, but is typically a positive value that increases according to how far D is from the start of the desired language list.
  • For each supported language S
    • Find the matching distance MD as described below.
    • Compute the weighted distance as F + MD
    • If WD < BD
      • BWD = WD
      • BD = D
      • BS = S
• If the BWD is less than a threshold, return <BD, BS>
  • The threshold is implementation-defined, typically set to greater than a default region difference, and less than a default script difference.
• Otherwise BD = the default supported language (like English); return <BD, null>

To find the matching distance MD between any two languages, perform the following steps.

1. Maximize each language using Section 4.3 Likely Subtags.
   • und is a special case: see below.
2. Set the match-distance MD to 0
3. For each subtag in {language, script, region}
   1. If respective subtags in each language tag are identical, remove the subtag from each (logically) and continue.
   2. Traverse the languageMatching data until a match is found.
      • * matches any field.
      • If the oneway flag is false, then the match is symmetric; otherwise only match one direction.
      • For region matching, use the mechanisms in Section 4.4.1 Enhanced Language Matching.
   3. Add the distance attribute value to MD.
      • This used to be a percent attribute value, which was 100 - the distance attribute value.
   4. Remove the subtag from each (logically)
4. Return MD

It is typically useful to set the discount factor between successive elements of the desired languages list to be slightly greater than the default region difference. That avoids the following problem:

Supported languages: "de, fr, ja"

User's desired languages: "de-AT, fr"

This user would expect to get "de", not "fr". In practice, when a user selects a list of preferred languages, they don't include all the regional variants ahead of their second base language. Yet while the user's desired languages really doesn't tell us the priority ranking among their languages, normally the fall-off between the user's languages is substantially greater than regional variants. But unless F is greater than the distance between de-AT and de-DE, then the user’s second-choice language would be returned.

The base language subtag "und" is a special case. Suppose we have the following situation:

• desired languages: {und, it}
• supported languages: {en, it}
• resulting language: en

Part of this is because 'und' has a special function in BCP 47; it stands in for 'no supplied base language'. To prevent this from happening, if the desired base language is und, the language matcher should not apply likely subtags to it.

Examples:

For example, suppose that nn-DE and nb-FR are being compared. They are first maximized to nn-Latn-DE and nb-Latn-FR, respectively. The list is searched. The first match is with "*-*-*", for a match of 96%. The languages are truncated to nn-Latn and nb-Latn, then to nn and nb. The first match is also for a value of 96%, so the result is 92%.

Note that language matching is orthogonal to the how closely two languages are related linguistically. For example, Breton is more closely related to Welsh than to French, but French is the better match (because it is more likely that a Breton reader will understand French than Welsh). This also illustrates that the matches are often asymmetric: it is not likely that a French reader will understand Breton.

The "*" acts as a wild card, as shown in the following example:

<languageMatch desired="es-*-ES" supported="es-*-ES" percent="100" />
<!-- Latin American Spanishes are closer to each other. Approximate by having es-ES be further from everything else. -->

<languageMatch desired="es-*-ES" supported="es-*-*" percent="93" />

<languageMatch desired="*" supported="*" percent="1" />
<!-- [Default value - must be at end!] Normally there is no comprehension of different languages. -->

<languageMatch desired="*-*" supported="*-*" percent="20" />
<!-- [Default value - must be at end!] Normally there is little comprehension of different scripts. -->

<languageMatch desired="*-*-*" supported="*-*-*" percent="96" />
<!-- [Default value - must be at end!] Normally there are small differences across regions. -->

When the language+region is not matched, and there is otherwise no reason to pick among the supported regions for that language, then some measure of geographic "closeness" can be used. The results may be more understandable by users. Looking for en-SK, for example, should fall back to something within Europe (eg en-GB) in preference to something far away and unrelated (eg en-SG). Such a closeness metric does not need to be exact; a small amount of data can be used to give an approximate distance between any two regions. However, any such data must be used carefully; although Hong Kong is closer to India than to the UK, it is unlikely that en-IN would be a better match to en-HK than en-GB would.

4.4.1 Enhanced Language Matching

The enhanced format for language matching adds structure to enable better matching of languages. It is distinguished by having a suffix "_new" on the type, as in the example below. The extended structure allows matching to take into account broad similarities that would give better results. For example, for English the regions that are or inherit from US (AS|GU|MH|MP|PR|UM|VI|US) form a “cluster”. Each region in that cluster should be closer to each other than to any other region. And a region outside the cluster should be closer to another region outside that cluster than to one inside. We get this issue with the “world languages” like English, Spanish, Portuguese, Arabic, etc.

Example:

<languageMatches type="written_new">
    <paradigmLocales locales="en en-GB es es-419 pt-BR pt-PT" />
    <matchVariable id="$enUS" value="AS+GU+MH+MP+PR+UM+US+VI" />
    <matchVariable id="$cnsar" value="HK+MO" />
    <matchVariable id="$americas" value="019" />
    <matchVariable id="$maghreb" value="MA+DZ+TN+LY+MR+EH" />
    <languageMatch desired="no" supported="nb" distance="1" /><!-- no ⇒ nb -->
    …
    <languageMatch desired="ar_*_$maghreb" supported="ar_*_$maghreb" distance="4" />
    <!-- ar; *; $maghreb ⇒ ar; *; $maghreb -->
    <languageMatch desired="ar_*_$!maghreb" supported="ar_*_$!maghreb" distance="4" />
    <!-- ar; *; $!maghreb ⇒ ar; *; $!maghreb -->
    …

The matchVariable allows for a rule to match to multiple regions, as illustrated by $maghreb. The syntax is simple: it allows for + for union and - for set difference, but no precedence. So A+B-A+D is interpreted as (((A+B)-A)+D), not as (A+B)-(A+D). The variable id has a value of the form [$][a-zA-Z0-9]+. If $X is defined, then $!X automatically means all those regions that are not in $X.

When the set is interpreted, then macrolanguages are (logically) transformed into a list of their contents, so “053+GB” → “AU+GB+NF+NZ”. This is done recursively, so 009 → “053+054+057+061+QO” → “AU+NF+NZ+FJ+NC+PG+SB +VU...”. Note that we use 019 for all of the Americas in the variables above, because en-US should be in the same cluster as es-419 and its contents.

In the rules, the percent value (100..0) is replaced by a distance value, which is the inverse (0..100).

These new variables and rules divide up the world into clusters, where items in the same clusters (for specific languages) get the normal regional difference, and items in different clusters get different weights.

Each cluster can have one or more associated paradigmLocales. These are locales that are preferred within a cluster. So when matching desired=[en-SA] against [en-GU en en-IN en-GB], the value en-GB is returned. Both of {en-GU en} are in a different cluster. While {en-IN en-GB} are in the same cluster, and the same distance from en-SA, the preference is given to en-GB because it is in the paradigm locales. It would be possible to express this in rules, but using this mechanism handles these very common cases without bulking up the tables.

The paradigmLocales also allow matching to macroregions. For example, desired=[es-419] should match to {es-MX} more closely than to {es}, and vice versa: {es-MX} should match more closely to {es-419} than to {es}. But es-MX should match more closely to es-419 than to any of the other es-419 sublocales. In general, in the absence of other distance data, there is a ‘paradigm’ in each cluster that the others should match more closely to: en(-US), en-GB, es(-ES), es-419, ru(-RU)...

5 XML Format

There are two kinds of data that can be expressed in LDML: language-dependent data and supplementary data. In either case, data can be split across multiple files, which can be in multiple directory trees.

For example, the language-dependent data for Japanese in CLDR is present in the following files:

• common/collation/ja.xml
• common/main/ja.xml
• common/rbnf/ja.xml
• common/segmentations/ja.xml

Data for cased languages such as French are in files like:

• common/casing/fr.xml

The status of the data is the same, whether or not data is split. That is, for the purpose of validation and lookup, all of the data for the above ja.xml files is treated as if it was in a single file. These files have the <ldml> root element and use ldml.dtd. The file name must match the identity element. For example, the <ldml> file pa_Arab_PK.xml must contain the following elements:

<ldml>
    <identity>
    …
        <language type="pa" />
        <script type="Arab" />
        <territory type="PK" />
    </identity>
    …

Supplemental data can have different root elements, currently: ldmlBCP47, supplementalData, keyboard, and platform. Keyboard and platform files are considered distinct. The ldmlBCP47 files and supplementalData files that have the same root are all logically part of the same file; they are simply split into separate files for convenience. Implementations may split the files in different ways, also for their convenience. The files in /properties are also supplemental data files, but are structured like UCD properties.

For example, supplemental data relating to Japan or the Japanese writing are in:

• common/supplemental/ (in many files, such as supplementalData.xml)
• common/transforms/Hiragana-Katakana.xml
• common/transforms/Hiragana-Latin.xml
• common/properties/scriptMetadata.txt
• common/bcp47/calendar.xml
• uca/allkeys_CLDR.txt (sorting)
• /keyboards/chromeos/ja-t-k0-chromeos.xml
• ...

Like the <ldml> files, the keyboard file names must match internal data: in particular, the locale attribute on the keyboard element must have a value that corresponds to the file name, such as <keyboard locale="af-t-k0-android"> for the file af-t-k0-android.xml.

The following sections describe the structure of the XML format for language-dependent data. The more precise syntax is in the ldml.dtd file; however, the DTD does not describe all the constraints on the structure.

To start with, the root element is <ldml>, with the following DTD entry:

<!ELEMENT ldml (identity,(alias|(fallback*,localeDisplayNames?,layout?,contextTransforms?,characters?,
delimiters?,measurement?,dates?,numbers?,units?,listPatterns?,collations?,posix?,
segmentations?,rbnf?,annotations?,metadata?,references?,special*)))>

The XML structure is stable over releases. Elements and attributes may be deprecated: they are retained in the DTD but their usage is strongly discouraged. In most cases, an alternate structure is provided for expressing the information. There is only one exception: newer DTDs cannot be used with version 1.1 files, without some modification.

In general, all translatable text in this format is in element contents, while attributes are reserved for types and non-translated information (such as numbers or dates). The reason that attributes are not used for translatable text is that spaces are not preserved, and we cannot predict where spaces may be significant in translated material.

There are two kinds of elements in LDML: rule elements and structure elements.

For structure elements, there are restrictions to allow for effective inheritance and processing:

1. There is no "mixed" content: if an element has textual content, then it cannot contain any elements.
2. The [XPath] leading to the content is unique; no two different pieces of textual content have the same [XPath].
3. An element that has value attributes MUST NOT also have have child elements.

To illustrate these restrictions, consider the below chunk of XML:

<!-- Not correct LDML -->
<unit type="duration-day"
      displayName="days"> <!-- #3: @VALUE attribute AND children -->
  {0} per day <!-- #1: Mixed content -->
  <unitPattern>{0} day</unitPattern>  <!-- #2 same XPath /unit[@type="duration-day"]/unitPattern -->
  <unitPattern>{0} days</unitPattern> <!-- #2 same XPath /unit[@type="duration-day"]/unitPattern -->
</unit>

LDML is actually structured as below (from en.xml):

<unit type="duration-day">  <!-- OK: "type" is distinguishing -->
  <displayName>days</displayName>
  <unitPattern count="one">{0} day</unitPattern> <!-- "count" is distinguishing -->
  <unitPattern count="other">{0} days</unitPattern>
  <perUnitPattern>{0} per day</perUnitPattern> <!-- mixed content in an element -->
</unit>

Rule elements do not have these restrictions, but also do not inherit, except as an entire block. Items which are ordered have the DTD Annotation @ORDERED. See DTD Annotations and Section 4.2 Inheritance and Validity. For more technical details, see Updating-DTDs.

Note that the data in examples given below is purely illustrative, and does not match any particular language. For a more detailed example of this format, see [Example]. There is also a DTD for this format, but remember that the DTD alone is not sufficient to understand the semantics, the constraints, nor the interrelationships between the different elements and attributes. You may wish to have copies of each of these to hand as you proceed through the rest of this document.

In particular, all elements allow for draft versions to coexist in the file at the same time. Thus most elements are marked in the DTD as allowing multiple instances. However, unless an element is annotated as @ORDERED, or has a distinguishing attribute, it can only occur once as a subelement of a given element. Thus, for example, the following is illegal even though allowed by the DTD:

<languages>
    <language type="aa">...</language>
    <language type="aa">..</language>

There must be only one instance of these per parent, unless there are other distinguishing attributes (such as an alt element).

In general, LDML data should be in NFC format. Normalization forms are defined by [UAX15]. However, certain elements may need to contain characters that are not in NFC, including exemplars, transforms, segmentations, and p/s/t/i/pc/sc/tc/ic rules in collation. These elements must not be normalized (either to NFC or NFD), or their meaning may be changed. Thus LDML documents must not be normalized as a whole. To prevent problems with normalization, no element value can start with a combining slash (U+0338 COMBINING LONG SOLIDUS OVERLAY).

Lists, such as singleCountries are space-delimited. That means that they are separated by one or more XML whitespace characters:

• singleCountries
• preferenceOrdering
• references

5.1 Common Elements

At any level in any element, two special elements are allowed.

5.1.1 Element special

This element is designed to allow for arbitrary additional annotation and data that is product-specific. It has one required attribute xmlns, which specifies the XML namespace of the special data. For example, the following used the version 1.0 POSIX special element.

<!DOCTYPE ldml SYSTEM "https://www.unicode.org/cldr/dtd/1.0/ldml.dtd" [
    <!ENTITY % posix SYSTEM "https://www.unicode.org/cldr/dtd/1.0/ldmlPOSIX.dtd">
%posix;
]>
<ldml>
...
    <special xmlns:posix="https://www.opengroup.org/regproducts/xu.htm">
        <!-- old abbreviations for pre-GUI days -->
        <posix:messages>
            <posix:yesstr>Yes</posix:yesstr>
            <posix:nostr>No</posix:nostr>
            <posix:yesexpr>^[Yy].*</posix:yesexpr>
            <posix:noexpr>^[Nn].*</posix:noexpr>
        </posix:messages>
    </special>
</ldml>

5.1.1.1 Sample Special Elements

The elements in this section are not part of the Locale Data Markup Language 1.0 specification. Instead, they are special elements used for application-specific data to be stored in the Common Locale Repository. They may change or be removed in future versions of this document, and are present here more as examples of how to extend the format. (Some of these items may move into a future version of the Locale Data Markup Language specification.)

• https://www.unicode.org/cldr/dtd/1.1/ldmlICU.dtd
• https://www.unicode.org/cldr/dtd/1.1/ldmlOpenOffice.dtd

The above examples are old versions: consult the documentation for the specific application to see which should be used.

These DTDs use namespaces and the special element. To include one or more, use the following pattern to import the special DTDs that are used in the file:

<?xml version="1.0" encoding="UTF-8" ?>
<!DOCTYPE ldml SYSTEM "https://www.unicode.org/cldr/dtd/1.1/ldml.dtd" [
    <!ENTITY % icu SYSTEM "https://www.unicode.org/cldr/dtd/1.1/ldmlICU.dtd">
    <!ENTITY % openOffice SYSTEM "https://www.unicode.org/cldr/dtd/1.1/ldmlOpenOffice.dtd">
%icu;
%openOffice; ]>

Thus to include just the ICU DTD, one uses:

<?xml version="1.0" encoding="UTF-8" ?>
<!DOCTYPE ldml SYSTEM "https://www.unicode.org/cldr/dtd/1.1/ldml.dtd" [
    <!ENTITY % icu SYSTEM "https://www.unicode.org/cldr/dtd/1.1/ldmlICU.dtd">
%icu; ]>

Note: A previous version of this document contained a special element for ISO TR 14652 compatibility data. That element has been withdrawn, pending further investigation, since 14652 is a Type 1 TR: "when the required support cannot be obtained for the publication of an International Standard, despite repeated effort". See the ballot comments on 14652 Comments for details on the 14652 defects. For example, most of these patterns make little provision for substantial changes in format when elements are empty, so are not particularly useful in practice. Compare, for example, the mail-merge capabilities of production software such as Microsoft Word or OpenOffice.

Note: While the CLDR specification guarantees backwards compatibility, the definition of specials is up to other organizations. Any assurance of backwards compatibility is up to those organizations.

A number of the elements above can have extra information for openoffice.org, such as the following example:

<special xmlns:openOffice="https://www.openoffice.org">
    <openOffice:search>
        <openOffice:searchOptions>
            <openOffice:transliterationModules>IGNORE_CASE</openOffice:transliterationModules>
        </openOffice:searchOptions>
    </openOffice:search>
</special>

5.1.2 Element alias

<!ELEMENT alias (special*) >
<!ATTLIST alias source NMTOKEN #REQUIRED >
<!ATTLIST alias path CDATA #IMPLIED>

The contents of any element in root can be replaced by an alias, which points to the path where the data can be found.

Aliases will only ever appear in root with the form //ldml/.../alias[@source="locale"][@path="..."].

Consider the following example in root:

<calendar type="gregorian">
    <months>
        <default choice="format" />
        <monthContext type="format">
            <default choice="wide" />
            <monthWidth type="abbreviated">
                <alias source="locale" path="../monthWidth[@type='wide']"/>
            </monthWidth>

If the locale "de_DE" is being accessed for a month name for format/abbreviated, then a resource bundle at "de_DE" will be searched for a resource element at that path. If not found there, then the resource bundle at "de" will be searched, and so on. When the alias is found in root, then the search is restarted, but searching for format/wide element instead of format/abbreviated.

If the path attribute is present, then its value is an [XPath] that points to a different node in the tree. For example:

<alias source="locale" path="../monthWidth[@type='wide']"/>

The default value if the path is not present is the same position in the tree. All of the attributes in the [XPath] must be distinguishing elements. For more details, see Section 4.2 Inheritance and Validity.

There is a special value for the source attribute, the constant source="locale". This special value is equivalent to the locale being resolved. For example, consider the following example, where locale data for 'de' is being resolved:

<x>
  <a>1</a>
  <b>2</b>
  <c>3</c>

</x>

<x>
 <a>11</a>
 <b>12</b>

 <d>14</d>
</x>

<x>
 <a>11</a>
 <b>12</b>
 <c>3</c>
 <d>14</d>
</x>

<y>
 <alias source="locale" path="../x">
</y>

<y>

 <b>22</b>


 <e>25</e>
</y>

<y>
 <a>11</a>
 <b>22</b>
 <c>3</c>
 <d>14</d>
 <e>25</e>
</y>

The first row shows the inheritance within the <x> element, whereby <c> is inherited from root. The second shows the inheritance within the <y> element, whereby <a>, <c>, and <d> are inherited also from root, but from an alias there. The alias in root is logically replaced not by the elements in root itself, but by elements in the 'target' locale.

For more details on data resolution, see Section 4.2 Inheritance and Validity.

Aliases must be resolved recursively. An alias may point to another path that results in another alias being found, and so on. For example, looking up Thai buddhist abbreviated months for the locale xx-YY may result in the following chain of aliases being followed:

../../calendar[@type="buddhist"]/months/monthContext[@type="format"]/monthWidth[@type="abbreviated"]

xx-YY → xx → root // finds alias that changes path to:

../../calendar[@type="gregorian"]/months/monthContext[@type="format"]/monthWidth[@type="abbreviated"]

xx-YY → xx → root // finds alias that changes path to:

../../calendar[@type="gregorian"]/months/monthContext[@type="format"]/monthWidth[@type="wide"]

xx-YY → xx // finds value here

It is an error to have a circular chain of aliases. That is, a collection of LDML XML documents must not have situations where a sequence of alias lookups (including inheritance and lateral inheritance) can be followed indefinitely without terminating.

5.1.3 Element displayName

Many elements can have a display name. This is a translated name that can be presented to users when discussing the particular service. For example, a number format, used to format numbers using the conventions of that locale, can have translated name for presentation in GUIs.

<numberFormat>
    <displayName>Prozentformat</displayName>
    ...
<numberFormat>

Where present, the display names must be unique; that is, two distinct codes would not get the same display name. (There is one exception to this: in time zones, where parsing results would give the same GMT offset, the standard and daylight display names can be the same across different time zone IDs.) Any translations should follow customary practice for the locale in question. For more information, see [Data Formats].

5.1.4 Escaping Characters

Unfortunately, XML does not have the capability to contain all Unicode code points. Due to this, in certain instances extra syntax is required to represent those code points that cannot be otherwise represented in element content. The escaping syntax is only defined on a few types of elements, such as in collation or exemplar sets, and uses the appropriate syntax for that type.

The element <cp>, which was formerly used for this purpose, has been deprecated.

5.2 Common Attributes

5.2.1 Attribute type

The attribute type is also used to indicate an alternate resource that can be selected with a matching type=option in the locale id modifiers, or be referenced by a default element. For example:

<ldml>
    ...
    <currencies>
        <currency>...</currency>
        <currency type="preEuro">...</currency>
    </currencies>
</ldml>

5.2.2 Attribute draft

If this attribute is present, it indicates the status of all the data in this element and any subelements (unless they have a contrary draft value), as per the following:

• approved: fully approved by the technical committee (equals the CLDR 1.3 value of false, or an absent draft attribute). This does not mean that the data is guaranteed to be error-free—this is the best judgment of the committee.
• contributed: partially approved by the technical committee.
• provisional: partially confirmed. Implementations may choose to accept the provisional data, especially if there is no translated alternative.
• unconfirmed: no confirmation available.

For more information on precisely how these values are computed for any given release, see Data Submission and Vetting Process on the CLDR website.

The draft attribute should only occur on "leaf" elements, and is deprecated elsewhere. For a more formal description of how elements are inherited, and what their draft status is, see Section 4.2 Inheritance and Validity.

5.2.3 Attribute alt

This attribute labels an alternative value for an element. The value is a descriptor that indicates what kind of alternative it is, and takes one of the following

• variantname means that the value is a variant of the normal value, and may be used in its place in certain circumstances. If a variant value is absent for a particular locale, the normal value is used. The variant mechanism should only be used when such a fallback is acceptable.
• proposed, optionally followed by a number, indicating that the value is a proposed replacement for an existing value.
• variantname-proposed, optionally followed by a number, indicating that the value is a proposed replacement variant value.

proposed should only be present if the draft status is not approved. It indicates that the data is proposed replacement data that has been added provisionally until the differences between it and the other data can be vetted. For example, suppose that the translation for September for some language is "Settembru", and a bug report is filed that that should be "Settembro". The new data can be entered in, but marked as alt="proposed" until it is vetted.

...
<month type="9">Settembru</month>
<month type="9" draft="unconfirmed" alt="proposed">Settembro</month>
<month type="10">...

Now assume another bug report comes in, saying that the correct form is actually "Settembre". Another alternative can be added:

...
<month type="9" draft="unconfirmed" alt="proposed2">Settembre</month>
...

The values for variantname at this time include "variant", "list", "email", "www", "short", and "secondary".

For a more complete description of how draft applies to data, see Section 4.2 Inheritance and Validity.

5.2.4 Attribute references

The value of this attribute is a token representing a reference for the information in the element, including standards that it may conform to. <references>. (In older versions of CLDR, the value of the attribute was freeform text. That format is deprecated.)

Example:

<territory type="UM" references="R222">USAs yttre öar</territory>

The reference element may be inherited. Thus, for example, R222 may be used in sv_SE.xml even though it is not defined there, if it is defined in sv.xml.

<... allow="verbatim" ...> (deprecated)

This attribute was originally intended for use in marking display names whose capitalization differed from what was indicated by the now-deprecated <inText> element (perhaps, for example, because the names included a proper noun). It was never supported in the dtd and is not needed for use with the new <contextTransforms> element.

5.3 Common Structures

5.3.1 Date and Date Ranges

When attribute specify date ranges, it is usually done with attributes from and to. The from attribute specifies the starting point, and the to attribute specifies the end point. The deprecated time attribute was formerly used to specify time with the deprecated weekEndStart and weekEndEnd elements, which were themselves inherently from or to.

The data format is a restricted ISO 8601 format, restricted to the fields year, month, day, hour, minute, and second in that order, with "-" used as a separator between date fields, a space used as the separator between the date and the time fields, and : used as a separator between the time fields. If the minute or minute and second are absent, they are interpreted as zero. If the hour is also missing, then it is interpreted based on whether the attribute is from or to.

• from defaults to "00:00:00" (midnight at the start of the day).
• to defaults to "24:00:00" (midnight at the end of the day).

That is, Friday at 24:00:00 is the same time as Saturday at 00:00:00. Thus when the hour is missing, the from and to are interpreted inclusively: the range includes all of the day mentioned.

For example, the following are equivalent:

<usesMetazone from="1991-10-27" to="2006-04-02" .../>
<usesMetazone from="1991-10-27 00:00:00" to="2006-04-02 24:00:00" .../>
<usesMetazone from="1991-10-26 24:00:00" to="2006-04-03 00:00:00" .../>

If the from element is missing, it is assumed to be as far backwards in time as there is data for; if the to element is missing, then it is from this point onwards, with no known end point.

The dates and times are specified in local time, unless otherwise noted. (In particular, the metazone values are in UTC (also known as GMT).

5.3.2 Text Directionality

The content of certain elements, such as date or number formats, may consist of several sub-elements with an inherent order (for example, the year, month, and day for dates). In some cases, the order of these sub-elements may be changed depending on the bidirectional context in which the element is embedded.

For example, short date formats in languages such as Arabic may contain neutral or weak characters at the beginning or end of the element content. In such a case, the overall order of the sub-elements may change depending on the surrounding text.

Element content whose display may be affected in this way should include an explicit direction mark, such as U+200E LEFT-TO-RIGHT MARK or U+200F RIGHT-TO-LEFT MARK, at the beginning or end of the element content, or both.

5.3.3 Unicode Sets

Some attribute values or element contents use UnicodeSet notation. A UnicodeSet represents a finite set of Unicode code points and strings, and is defined by lists of code points and strings, Unicode property sets, and set operators, all bounded by square brackets. In this context, a code point means a string consisting of exactly one code point.

A UnicodeSet implements the semantics in UTS #18: Unicode Regular Expressions [UTS18] Levels 1 & 2 that are relevant to determining sets of characters. Note however that it may deviate from the syntax provided in [UTS18], which is illustrative rather than a requirement. There is one exception to the supported semantics, Section RL2.6 Wildcards in Property Values. That feature can be supported in clients such as ICU by implementing a “hook” as is done in the online UnicodeSet utilities.

A UnicodeSet may be cited in specifications outside of the domain of LDML. In such a case, the specification may specify a subset of the syntax provided here.

The following provides EBNF syntax for a UnicodeSet:

= prop
| '[-]'
| '[' [\-\^]? s seq+ ']'

= root (s [\&\-] s root)* s
| range s

= char ('-' char)?
| '{' (s char)+ s '}'

= '\' [pP] '{' propName ([≠=] s value1+)? '}'
| '[:' '^'? propName ([≠=] s value2+)? ':]'

= s [A-Za-z0-9] [A-Za-z0-9_\x20]* s

= [^\}]
| '\' quoted

= [^:]
| '\' quoted

= [^\& \- \[ \[ \] \\ \} \{ [:Pat_WS:]]
| '\' quoted

= 'u' (hex{4} | bracketedHex)
| 'x' (hex{2} | bracketedHex)
| 'U00' ('0' hex{5} | '10' hex{4})
| 'N{' propName '}'
| [[\u0000-\U00010FFFF]-[uxUN]]

= s [A-Za-z0-9] [-A-Za-z0-9_\x20]* s

= '{' s hexCodePoint (s hexCodePoint)* s '}'

= hex{1,5} | '10' hex{4}

= [0-9A-Fa-f]

= [:Pattern_White_Space:]*

Some constraints on UnicodeSet syntax are not captured by this EBNF. Notably, property names and values are restricted to those supported by the implementation, and have additional constraints imposed by [UAX44]. In addition, quoted values that resolve to more than one code point are disallowed in ranges of the form char '-' char.

The syntax characters are listed in the table below:

5.3.3.1 Lists of Code Points

Lists are a sequence of strings that may include ranges, which are indicated by a '-' between two code points, as in "a-z". The sequence start-end specifies the range of all code points from the start to end, inclusive, in Unicode order. For example, [a c d-f m] is equivalent to [a c d e f m]. Whitespace can be freely used for clarity, as [a c d-f m] means the same as [acd-fm].

A string with multiple code points is represented in a list by being surrounded by curly braces, such as in [a-z {ch}]. It can be used with the range notation, as described in Section 5.3.4 String Range . There is an additional restriction on string ranges in a UnicodeSet: the number of codepoints in the first string of the range must be identical to the number in the second. Thus [{ab}-{c}] and [{ab}-c] are invalid.

In UnicodeSets, there are two ways to quote syntax code points:

Outside of single quotes, certain backslashed code point sequences can be used to quote code points:

Anything else following a backslash is mapped to itself, except the property syntax described below, or in an environment where it is defined to have some special meaning.

Any code point formed as the result of a backslash escape loses any special meaning and is treated as a literal. In particular, note that \x, \u and \U escapes create literal code points. (In contrast, Java treats Unicode escapes as just a way to represent arbitrary code points in an ASCII source file, and any resulting code points are not tagged as literals.)

Unicode property sets are defined as described in UTS #18: Unicode Regular Expressions [UTS18], Level 1 and RL2.5, including the syntax where given. For an example of a concrete implementation of this, see [ICUUnicodeSet].

5.3.3.2 Unicode Properties

Briefly, Unicode property sets are specified by any Unicode property and a value of that property, such as [:General_Category=Letter:] for Unicode letters or \p{uppercase} for the set of upper case letters in Unicode. The property names are defined by the PropertyAliases.txt file and the property values by the PropertyValueAliases.txt file. For more information, see [UAX44]. The syntax for specifying the property sets is an extension of either POSIX or Perl syntax, by the addition of "=<value>". For example, you can match letters by using the POSIX-style syntax:

[:General_Category=Letter:]

or by using the Perl-style syntax

\p{General_Category=Letter}.

Property names and values are case-insensitive, and whitespace, "-", and "_" are ignored. The property name can be omitted for the General_Category and Script properties, but is required for other properties. If the property value is omitted, it is assumed to represent a boolean property with the value "true". Thus [:Letter:] is equivalent to [:General_Category=Letter:], and [:Wh-ite-s pa_ce:] is equivalent to [:Whitespace=true:].

The table below shows the two kinds of syntax: POSIX and Perl style. Also, the table shows the "Negative" version, which is a property that excludes all code points of a given kind. For example, [:^Letter:] matches all code points that are not [:Letter:].

5.3.3.3 Boolean Operations

The low-level lists or properties then can be freely combined with the normal set operations (union, inverse, difference, and intersection):

• To union two sets, simply concatenate them. For example, [[:letter:] [:number:]]
• To intersect two sets, use the '&' operator. For example, [[:letter:] & [a-z]]
• To take the set-difference of two sets, use the '-' operator. For example, [[:letter:] - [a-z]]
• To invert a set, place a '^' immediately after the opening '['. For example, [^a-z]. In any other location, the '^' does not have a special meaning. The inversion [^X] is equivalent to [[\x{0}-\x{10FFFF}]-[X]]. Thus multi-code point strings are discarded.
• Symmetric difference (~) is not supported.

The binary operators '&', '-', and the implicit union have equal precedence and bind left-to-right. Thus [[:letter:]-[a-z]-[\u0100-\u01FF]] is equal to [[[:letter:]-[a-z]]-[\u0100-\u01FF]]. Another example is the set [[ace][bdf] - [abc][def]], which is not the empty set, but instead equal to [[[[ace] [bdf]] - [abc]] [def]], which equals [[[abcdef] - [abc]] [def]], which equals [[def] [def]], which equals [def].

One caution: the '&' and '-' operators operate between sets. That is, they must be immediately preceded and immediately followed by a set. For example, the pattern [[:Lu:]-A] is illegal, since it is interpreted as the set [:Lu:] followed by the incomplete range -A. To specify the set of upper case letters except for 'A', enclose the 'A' in brackets: [[:Lu:]-[A]].

5.3.3.4 UnicodeSet Examples

The following table summarizes the syntax that can be used.

5.3.4 String Range

A String Range is a compact format for specifying a list of strings.

Syntax:

X sep Y

The separator and the format of strings X, Y may vary depending on the domain. For example,

• for the validity files the separator is ~,
• for UnicodeSet the separator is -, and any multi-codepoint string is enclosed in {…}.

Validity:

A string range X sep Y is valid iff len(X) ≥ len(Y) > 0, where len(X) is the length of X in code points.

There may be additional, domain-specific requirements for validity of the expansion of the string range.

Interpretation:

1. Break X into P and S, where len(S) = len(Y)
   • Note that P will be an empty string if the lengths of X and Y are equal.
2. Form the combinations of all P+(s₀..y₀)+(s₁..y₁)+...(sₙ..yₙ)
   • s₀ is the first code point in S, etc.

Examples:

5.4 Identity Elements

<!ELEMENT identity (alias | (version, generation?, language, script?, territory?, variant?, special*) ) >

The identity element contains information identifying the target locale for this data, and general information about the version of this data.

<version number="$Revision: 1.227 $">

The version element provides, in an attribute, the version of this file. The contents of the element can contain textual notes about the changes between this version and the last. For example:

<version number="1.1">Various notes and changes in version 1.1</version>

This is not to be confused with the version attribute on the ldml element, which tracks the dtd version.

<generation date="$Date: 2007/07/17 23:41:16 $" />

The generation element is now deprecated. It was used to contain the last modified date for the data. This could be in two formats: ISO 8601 format, or CVS format (illustrated by the example above).

<language type="en" />

The language code is the primary part of the specification of the locale id, with values as described above.

<script type="Latn" />

The script code may be used in the identification of written languages, with values described above.

<territory type="US" />

The territory code is a common part of the specification of the locale id, with values as described above.

<variant type="NYNORSK" />

The variant code is the tertiary part of the specification of the locale id, with values as described above.

When combined according to the rules described in Section 3, Unicode Language and Locale Identifiers, the language element, along with any of the optional script, territory, and variant elements, must identify a known, stable locale identifier. Otherwise, it is an error.

5.5 Valid Attribute Values

The DTD Annotations in Section 5.7 are used to determine whether elements, attributes, or attribute values are valid (or deprecated).

5.6 Canonical Form

The following are restrictions on the format of LDML files to allow for easier parsing and comparison of files.

Peer elements have consistent order. That is, if the DTD or this specification requires the following order in an element foo:

<foo>
    <pattern>
    <somethingElse>
</foo>

It can never require the reverse order in a different element bar.

<bar>
    <somethingElse>
    <pattern>
</bar>

Note that there was one case that had to be corrected in order to make this true. For that reason, pattern occurs twice under currency:

<!ELEMENT currency (alias | (pattern*, displayName?, symbol?, pattern*, decimal?, group?, special*)) >

XML files can have a wide variation in textual form, while representing precisely the same data. By putting the LDML files in the repository into a canonical form, this allows us to use the simple diff tools used widely (and in CVS) to detect differences when vetting changes, without those tools being confused. This is not a requirement on other uses of LDML; just simply a way to manage repository data more easily.

5.6.1 Content

1. All start elements are on their own line, indented by depth tabs.

2. All end elements (except for leaf nodes) are on their own line, indented by depth tabs.

3. Any leaf node with empty content is in the form <foo/>.

4. There are no blank lines except within comments or content.

5. Spaces are used within a start element. There are no extra spaces within elements.

   • <version number="1.2"/>, not <version number = "1.2" />
   • </identity>, not </identity >

6. All attribute values use double quote ("), not single (').

7. There are no CDATA sections, and no escapes except those absolutely required.

   • no ' since it is not necessary
   • no 'a', it would be just 'a'

8. All attributes with defaulted values are suppressed.

9. The draft and alt="proposed.*" attributes are only on leaf elements.

10. The tzid are canonicalized in the following way:

    • All tzids as of CLDR 1.1 (2004.06.08) in zone.tab are canonical.
    • After that point, the first time a tzid is introduced, that is the canonical form.

    That is, new IDs are added, but existing ones keep the original form. The TZ timezone database keeps a set of equivalences in the "backward" file. These are used to map other tzids to the canonical form. For example, when America/Argentina/Catamarca was introduced as the new name for the previous America/Catamarca , a link was added in the backward file.

    Link America/Argentina/Catamarca America/Catamarca

Example:

<ldml draft="unconfirmed" >
    <identity>
        <version number="1.2" />
        <language type="en" />
        <territory type="AS" />
    </identity>
    <numbers>
        <currencyFormats>
            <currencyFormatLength>
                <currencyFormat>
                    <pattern>¤#,##0.00;(¤#,##0.00)</pattern>
                </currencyFormat>
            </currencyFormatLength>
        </currencyFormats>
    </numbers>
</ldml>

5.6.2 Ordering

An element is ordered first by the element name, and then if the element names are identical, by the sorted set of attribute-value pairs. For the latter, compare the first pair in each (in sorted order by attribute pair). If not identical, go to the second pair, and so on.

Elements and attributes are ordered according to their order in the respective DTDs. Attribute value comparison is a bit more complicated, and may depend on the attribute and type. This is currently done with specific ordering tables.

Any future additions to the DTD must be structured so as to allow compatibility with this ordering. See also Section 5.5 Valid Attribute Values.

5.6.3 Comments

1. Comments are of the form <!-- stuff -->.
2. They are logically attached to a node. There are 4 kinds:
   1. Inline always appear after a leaf node, on the same line at the end. These are a single line.
   2. Preblock comments always precede the attachment node, and are indented on the same level.
   3. Postblock comments always follow the attachment node, and are indented on the same level.
   4. Final comment, after </ldml>
3. Multiline comments (except the final comment) have each line after the first indented to one deeper level.

Examples:

<eraAbbr>
    <era type="0">BC</era> <!-- might add alternate BDE in the future -->
...
<timeZoneNames>
    <!-- Note: zones that do not use daylight time need further work -->
    <zone type="America/Los_Angeles">
    ...
    <!-- Note: the following is known to be sparse,
            and needs to be improved in the future -->
    <zone type="Asia/Jerusalem">

5.7 DTD Annotations

The information in a standard DTD is insufficient for use in CLDR. To make up for that, DTD annotations are added. These are of the form

<!--@...-->

and are included below the !ELEMENT or !ATTLIST line that they apply to. The current annotations are:

There is additional information in the attributeValueValidity.xml file that is used internally for testing. For example, the following line indicates that the 'currency' element in the ldml dtd must have values from the bcp47 'cu' type.

<attributeValues dtds='ldml' elements='currency' attributes='type'>$_bcp47_cu</attributeValues>

The element values may be literals, regular expressions, or variables (some of which are set programmatically according to other CLDR data, such as the above). However, the information at this point does not cover all attribute values, is used only for testing, and should not be used in implementations since the structure may change without notice.

5.7.1 Attribute Value Constraints

The following are constraints on the attribute values. Note: in future versions, the format may change, and/or the constraints may be tightened.

6 Property Data

Some data in CLDR does not use an XML format, but rather a semicolon-delimited format derived from that of the Unicode Character Database. That is because the data is more likely to be parsed by implementations that already parse UCD data. Those files are present in the common/properties directory.

Each file has a header that explains the format and usage of the data.

6.1 Script Metadata

scriptMetadata.txt

This file provides general information about scripts that may be useful to implementations processing text. The information is the best currently available, and may change between versions of CLDR. The format is similar to Unicode Character Database property file, and is documented in the header of the data file.

6.2 Extended Pictographic

ExtendedPictographic.txt

This file was used to define the ExtendedPictographic data used for “future-proofing” emoji behavior, especially in segmentation. As of Emoji version 11.0, the set of Extended_Pictographic is incorporated into the emoji data files found at unicode.org/Public/emoji/.

6.3 Labels.txt

labels.txt

This file provides general information about associations of labels to characters that may be useful to implementations of character-picking applications. The information is the best currently available, and may change between versions of CLDR. The format is similar to Unicode Character Database property file, and is documented in the header of the data file.

Initially, the contents are focused on emoji, but may be expanded in the future to other types of characters. Note that a character may have multiple labels.

6.4 Segmentation Tests

CLDR provides a tailoring to the Grapheme Cluster Break (gcb) algorithm to avoid splitting Indic aksaras. The corresponding test files for that are located in common/properties/segments/, along with a readme.txt that provides more details. There are also specific test files for the supported Indic scripts in the unittest directory.

7 Issues in Formatting and Parsing

7.1 Lenient Parsing

7.1.1 Motivation

User input is frequently messy. Attempting to parse it by matching it exactly against a pattern is likely to be unsuccessful, even when the meaning of the input is clear to a human being. For example, for a date pattern of "MM/dd/yy", the input "June 1, 2006" will fail.

The goal of lenient parsing is to accept user input whenever it is possible to decipher what the user intended. Doing so requires using patterns as data to guide the parsing process, rather than an exact template that must be matched. This informative section suggests some heuristics that may be useful for lenient parsing of dates, times, and numbers.

7.1.2 Loose Matching

Loose matching ignores attributes of the strings being compared that are not important to matching. It involves the following steps:

• Remove "." from currency symbols and other fields used for matching, and also from the input string unless:
  • "." is in the decimal set, and
  • its position in the input string is immediately before a decimal digit
• Ignore all format characters: in particular, ignore any RLM, LRM or ALM used to control BIDI formatting.
• Ignore all characters in [:Zs:] unless they occur between letters. (In the heuristics below, even those between letters are ignored except to delimit fields)
• Map all characters in [:Dash:] to U+002D HYPHEN-MINUS
• Use the data in the <character-fallback> element to map equivalent characters (for example, curly to straight apostrophes). Other apostrophe-like characters should also be treated as equivalent, especially if the character actually used in a format may be unavailable on some keyboards. For example:
  • U+02BB MODIFIER LETTER TURNED COMMA (ʻ) might be typed instead as U+2018 LEFT SINGLE QUOTATION MARK (‘).
  • U+02BC MODIFIER LETTER APOSTROPHE (ʼ) might be typed instead as U+2019 RIGHT SINGLE QUOTATION MARK (’), U+0027 APOSTROPHE, etc.
  • U+05F3 HEBREW PUNCTUATION GERESH (‎׳) might be typed instead as U+0027 APOSTROPHE.
• Apply mappings particular to the domain (i.e., for dates or for numbers, discussed in more detail below)
• Apply case folding (possibly including language-specific mappings such as Turkish i)
• Normalize to NFKC; thus no-break space will map to space; half-width katakana will map to full-width.

Loose matching involves (logically) applying the above transform to both the input text and to each of the field elements used in matching, before applying the specific heuristics below. For example, if the input number text is " - NA f. 1,000.00", then it is mapped to "-naf1,000.00" before processing. The currency signs are also transformed, so "NA f." is converted to "naf" for purposes of matching. As with other Unicode algorithms, this is a logical statement of the process; actual implementations can optimize, such as by applying the transform incrementally during matching.

7.2 Handling Invalid Patterns

Processes sometimes encounter invalid number or date patterns, such as a number pattern with “¤¤¤¤¤” (valid pattern character but invalid length in current CLDR), a date pattern with “nn” (invalid pattern character in current CLDR), or a date pattern with “MMMMMM” (invalid length in current CLDR). The recommended behavior for handling such an invalid pattern field is:

• For a field using a currently-invalid length for a valid pattern character:
  • In formatting, emit U+FFFD REPLACEMENT CHARACTER for the invalid field.
  • In parsing, the field may be parsed as if it had a valid length.
• For a pattern that contains a currently-invalid pattern character (applies only to date patterns, for which A-Za-z are reserved as pattern characters but not all defined as valid):
  • Produce an error (set an error code or throw an exception) when an attempt is made to create a formatter with such a pattern or to apply such a pattern to an existing formatter.

Annex A Deprecated Structure

The DTD Annotations in Section 5.7 are used to determine whether DTD items such as elements, attributes, or attribute values are deprecated.

Though such deprecated items are still valid LDML, they are strongly discouraged, and are no longer used in CLDR.

The CLDR DTD Deltas chart shows which DTD items have been deprecated in which version of CLDR.

The remainder of this section describes selected cases of deprecated structure, and what (if any) should be used instead.

A.1 Element fallback

Implementations should use instead the information in Section 4.4 Language Matching for doing language fallback.

A.2 BCP 47 Keyword Mapping

Instead use the mechanisms descibed in Section 3.6.4 U Extension Data Files.

A.3 Choice Patterns

Instead use count attributes.

A.4 Element default

Instead use replacement structure, for example:

• For <collations>, now use the <defaultCollation> element.
• For <calendars>, the default calendar type for a locale is now specified by Calendar Preference Data.

A.5 Deprecated Common Attributes

A.5.1 Attribute standard

Instead, use a reference element with the attribute standard="true".

A.5.2 Attribute draft in non-leaf elements

The draft attribute is deprecated except in leaf elements (elements that do not have any subelements)

A.6 Element base

Instead use the collation <import> element.

A.7 Element rules

Instead use the basic collation syntax with the <cr> element.

A.8 Deprecated subelements of <dates>

• <localizedPatternChars>
• <dateRangePattern>, replaced by <intervalFormats>.

A.9 Deprecated subelements of <calendars>

• The deprecated <monthNames> and <monthAbbr> are replaced by the months element with the context type="format" and the width type="wide" (for ...Names) and type="narrow" (for ...Abbr), respectively.
• The deprecated <dayNames> and <dayAbbr> are replaced by the days element with the context type="format" and the width type="wide" (for ...Names) and type="narrow" (for ...Abbr), respectively.
• <week> is deprecated in the main LDML files, because the data is more appropriately organized as connected to territories, not to linguistic data. Use the supplemental <weekData> element instead.
• The standalone <am> and <pm> are deprecated, and the data are instead included as part of the <dayPeriods> element
• <fields> is deprecated as a subelement of <calendars> instead, a <fields> element should be located just under a <dates> element. See Calendar Fields.

A.10 Deprecated subelements of <timeZoneNames>

• <preferenceOrdering>: use metazones instead.
• <singleCountries>:use Primary Zones
• <hoursFormat>, <fallbackRegionFormat>, <abbreviationFallback>

A.11 Deprecated subelements of <zone> and <metazone>

• <commonlyUsed>, formerly used to indicate whether a zone was commonly used in the locale.

A.12 Renamed attribute values for <contextTransformUsage> element

The <contextTransformUsage> element was introduced in CLDR 21. The values for its type attribute are documented in <contextTransformUsage> type attribute values. In CLDR 25, some of these values were renamed from their previous values for improved clarity:

• type was renamed to keyValue
• displayName was renamed to currencyName
• displayName-count was renamed to currencyName-count
• tense was renamed to relative

A.13 Deprecated subelements of <segmentations>

• <exceptions> and <exception>: Replaced with <suppressions> and <suppression>.

A.14 Element cp

The cp element was used in certain elements to escape characters that cannot be represented in XML, even with NCRs. This mechanism was replaced by specialized syntax:

A.15 Attribute validSubLocales

Instead of using validSubLocales, it is recommended to simply add empty files to specify which sublocales are valid. This convention is used throughout the CLDR.

A.16 Elements postalCodeData, postCodeRegex

Instead please see other services that are kept up to date, such as:

• https://i18napis.appspot.com/address/data/US
• https://i18napis.appspot.com/address/data/CH
• ...

A.17 Element telephoneCodeData

The element <telephoneCodeData> and its subelements have been deprecated and the data removed.

Annex B Links to Other Parts

The LDML specification is split into several parts by topic, with one HTML document per part. The following tables provide redirects for links to specific topics. Please update your links and bookmarks.

Part 1 Links: Core (this document): No redirects needed.

Annex C. LocaleId Canonicalization

The languageAlias, scriptAlias, territoryAlias, and variantAlias elements are used as rules to transform an input source localeId. The first step is to transform the languageId portion of the localeId.

Note: in the following discussion, the separator '-' is used. That is also used in examples of XML alias data, even though for compatibility reasons that alias data actually uses '_' as a separator. The processing can also be applied to syntax while maintaining the separator '_', mutatis mutandis. CLDR also uses “territory” and “region” interchangeably.

Also note that the discussion of canonicalization assumes BCP 47 input data. If input data is a CLDR or ICU locale ID such as en_US_POSIX, a conversion step must be done prior to canonicalization. See §3.8.2 Legacy Variants.

LocaleId Definitions

1. Multimap interpretation

Interpret each languageId as a multimap from a fieldId (language, script, region, variants) to a sorted set of field values.

Examples:

• This can be represented as an abbreviated format: {L={ja}, S={Latn}, R={YU}, V={hepburn, heploc}}, skipping empty sets.
• “und” is a special language code that is treated as an empty set.
• Of course, only the Variants can contain more than one item: the others are either empty or contain exactly 1 item.

2. Alias elements

For the languageAlias elements, the type and replacements are languageIds.

For the script-, territory- (aka region), and variant- Alias elements, the type and replacements are interpreted as a languageId, after prefixing with “und-”. Thus

<territoryAlias type="AN" replacement="CW SX BQ" reason="deprecated" />

is interpreted as:

<territoryAlias type="und-AN" replacement="und-CW und-SX und-BQ" reason="deprecated" />

Note that for the case of territoryAlias, there may be multiple replacement values separated by spaces in the text (such as replacement="und-CW und-SX und-BQ"); other rules only ever have a single replacement value.

3. Matches

A rule matches a source if and only for all fields, each source field ⊇ type field.

Examples:

source="ja-heploc-hepburn" and type="und-hepburn"

so the rule matches the source. (Note that order of variants is immaterial to matching)

source="ja-hepburn" and type="und-hepburn-heploc"

so the rule does not match the source.

4. Replacement

A matching rule can be used to transform the source fields as follows

• if type.field ≠ {}
  • source.field = (source.field - type.field) ∪ replacement.field
• else if source.field = {} and replacement.field ≠ {}
  • source.field = replacement.field

Example:

source=ja-Latn-fonipa-hepburn-heploc

rule ="<languageAlias type="und-hepburn-heploc"

replacement="und-alalc97">"

result="ja-Latn-alalc97-fonipa" // note that CLDR canonical order of variants is alphabetical

Territory Exception

If the field = territory, and the replacement.field has more than one value, then look up the most likely territory for the base language code (and script, if there is one). If that likely territory is in the list of replacements, use it. Otherwise, use the first territory in the list.

5. Canonicalizing Syntax

To canonicalize the syntax of source:

• Initial Script Subtag
  • If the first subtag has 4 letters, prepend the source with "und-"
  • Note: These are only for specialized use.
• Casing
  • Put any script subtag inside unicode_language_id into title case (eg, Hant)
  • Put any region subtag inside unicode_language_id into uppercase (eg, DE)
  • Put all other subtags into lowercase (eg, en, fonipa)
• Order
  • Put any variants into alphabetical order (eg, en-fonipa-scouse, not en-scouse-fonipa)
  • Put any extensions into alphabetical order by their singleton (eg, en-t-xxx-u-yyy, not en-u-yyy-t-xxx)
  • Put all attributes into alphabetical order.
  • Put all ufields (<ukey, uvalue>) and tfields (<tkey, tvalue>) into alphabetical order according to their keys (ukey or tkey), within their respective extensions.
  • Remove any uvalue (aka type) equal to "true". Note that "true" values cannot be removed from tvalues.
• Separator
  • Replace '_' by '-'

Preprocessing

The data from supplementalMetadata is (logically) preprocessed as follows.

1. Load the rules from supplementalMetadata.xml, replacing '_' by '-', and adding “und-” as described in Definition 2. Alias Elements.
2. Capture all languageAlias rules where the type is an invalid languageId into a set of BCP47 LegacyRules. Example:
   1. <languageAlias type="i-mingo" replacement="see-x-i-mingo" reason="legacy" />
3. Discard all rules where the type is an invalid languageId. Examples are
   1. <languageAlias type="i-mingo" replacement="see-x-i-mingo" reason="legacy" />
   2. <territoryAlias type="und-AAA" replacement="und-AA" reason="overlong" />
4. Change the type and replacement values in the remaining rules into multimap rules, as per Definition 1. Multimap Interpretation.
   1. Note that the “und” value disappears.
5. Order the set of rules by the following levels
   1. First order by the size of the union of all field value sets, with larger sizes before smaller sizes.
   • So V={hepburn, heploc}} is before {R={CA}}
   • V={hepburn, heploc}} and {L={en}, R={GB}} are not ordered at this level
   2. And then order by field, where L < S < R < V. Thus L is first and V is last.
   • So {L={fr}, R={CA}} is before {V={fonipa, heploc}}.
   • V={hepburn, heploc}} and {V={hepburn, heploc}} are not ordered at this level
   • After this point we are guaranteed to have the same set of fields, with possibly different field value sets.
   3. And then order by field value sets, traversing also in the order of their fields L < S < R < V.
   • To determine the ordering between a field value set A and B, traverse each in parallel
   • If the corresponding field value sets for A and B are identical, then the next pair of field value sets is processed
   • Otherwise at the first pair of differing field values, A is before B if its field value is alphabetically less, otherwise B is before.
6. The result is the set of Alias Rules

So using the examples above, we get the following order:

Processing LanguageIds

To canonicalize a given source:

1. Canonicalize the syntax of source as per Definition 5. Canonicalizing Syntax.
2. Where the source could be an arbitrary BCP 47 language tag, first process as follows:
   1. If the source is identical to one of the types in the BCP47 LegacyRules, replace the entire source by the replacement value.
   2. Else if there is an extlang subtag, then apply Step 3 of BCP 47 Section 4.5 to remove the extlang subtag (possibly adjusting the language subtag).
      1. Don’t apply any of the other canonicalization steps in that section, however.
   3. Else if the first subtag is "x", prefix by "und-".
   4. Note: there are currently no valid 4-letter primary language subtags. While it is extremely unlikely that BCP 47 would ever register them, if so then languageAlias mappings will be supplied for them, mapping to defined CLDR language subtags (from the idStatus="reserved" set).
3. Find the first matching rule in Alias Rules (from Preprocessing)
   1. If there are none, return source
4. Transform source according to that rule
5. loop (goto #3)

Processing LocaleIds

The canonicalization of localeIds is done by first canonicalizing the languageId portion, then handling extensions in the following way:

1. Replace any tlang languageId value by its canonicalization.
2. Use the bcp47 data to replace keys, types, tfields, and tvalues by their canonical forms. See Section 3.6.4 U Extension Data Files and Section 3.7.1 T Extension Data Files. The matches are in the alias attribute value, while the canonical replacement is in the name attribute value. For example:
   1. Because of the following bcp47 data: <key name="ms"…>…<type name="uksystem" … alias="imperial" … />…</key>
   2. We get the following transformation: en-u-ms-imperial ⇒ en-u-ms-uksystem
3. Replace any unicode_subdivision_id that is a subdivision alias by its replacement value in the same way, using subdivisionAlias data. This applies, for example, to the values for the 'sd' and 'rg' keys. However, where the replacement value is a two-letter region code, also append zzzz so that the result is syntactically correct. For example:
   1. Because of the following bcp47 data: <subdivisionAlias type="fi01" replacement="AX"…
   2. We get the following transformation: en-u-rg-fi01 ⇒ en-u-rg-axzzzz

Optimizations

The above algorithm is a logical statement of the process, but would obviously not be directly suited to production code. Production-level code can use many optimizations for efficiency while achieving the same result. For example, the Alias Rules can be further preprocessed to avoid indefinite looping, instead doing a rule lookup once per subtag. As another example, the small number of Territory Exceptions can be preprocessed to avoid the likely subtags processing.

References

Acknowledgments

Special thanks to the following people for their continuing overall contributions to the CLDR project, and for their specific contributions in the following areas. These descriptions only touch on the many contributions that they have made.

• Mark Davis for creating the initial version of LDML, and adding to and maintaining this specification, and for his work on the LDML code and tests, much of the supplemental data and overall structure, and transforms and keyboards.
• John Emmons for the POSIX conversion tool and metazones.
• Deborah Goldsmith for her contributions to LDML architecture and this specification.
• Chris Hansten for coordinating and managing data submissions and vetting.
• Erkki Kolehmainen and his team for their work on Finnish.
• Steven R. Loomis for development of the survey tool and database management.
• Peter Nugent for his contributions to the POSIX tool and from Open Office, and for coordinating and managing data submissions and vetting.
• George Rhoten for his work on currencies.
• Roozbeh Pournader (روزبه پورنادر) for his work on South Asian countries.
• Ram Viswanadha (రఘురామ్ విశ్వనాధ) for all of his work on LDML code and data integration, and for coordinating and managing data submissions and vetting.
• Vladimir Weinstein (Владимир Вајнштајн) for his work on collation.
• Yoshito Umaoka (馬岡 由人) for his work on the timezone architecture.
• Rick McGowan for his work gathering language, script and region data.
• Xiaomei Ji (吉晓梅) for her work on time intervals and plural formatting.
• David Bertoni for his contributions to the conversion tools.
• Mike Tardif for reviewing this specification and for coordinating and vetting data submissions.
• Peter Edberg for work on this specification, monthPatterns, cyclicNameSets, contextTransforms and other items.
• Raymond Wainman and Cibu Johny for their work on keyboards.
• Jennifer Chye for her contributions to the conversion tools.
• Markus Scherer for a major rewrite of Part 5, Collation.
• Shane Carr for his work on numbers and measurement units.
• Robin Leroy for his work on compact plurals: Part 3, Section 5, Language Plural Rules
• Rich Gillam for work on Person Names.
• Alex Kolisnychenko for work on Person Names.
• Mike McKenna for work on Person Names.

Other contributors to CLDR are listed on the CLDR Project Page.

Modifications

Revision 67

Proposed Update for CLDR Version 42.

• Parent Locales
  • Updated the description of guidelines and invariants for parentLocale data.
• Currency Formats and Currencies
  • Described the new alt="alphaNextToNumber" and alt="noCurrency" variants for patterns used with currencyFormat elements
  • Described the new currencyPatternAppendISO element under currencyFormats
  • Discouraged the use of the old currencySpacing element (and its subelements) in favor of the alt="alphaNextToNumber" variant
• Element dateTimeFormat
  • Described the new dateTimeFormat type="atTime" pattern and when to use it versus the standard dateTimeFormat pattern.
• Matching Skeletons
  • Provided more detailed recommendations on matching pattern field length to field length in the requested skeleton.
• Unit Preferences
  • Added a new subsection to specify the interaction of the unit Preferences data with the locale keys mu, ms, and rg, and the base locale.
• Plurals
  • In Plural rules syntax, allow sample values to have positive and negative signs. `* Units of measurement
  • Unit Preferences
    • Added a new subsection to specify the interaction of the unit Preferences data with the locale keys mu, ms, and rg, and the base locale.
  • Unit Elements, Unit_Conversion
    • For simpler and cleaner parsing, add a new element (unitIdComponent) and restructured the EBNF for parsing unit identifiers.
    • As part of this work, the identifier metric-ton was deprecated in favor of tonne. As usual, the older identifier remains for compatibility, and is aliased to the new one. `
• Person Names
  • Added a new Part 8, Person Names.

Note that small changes such as typos and link fixes are not listed above. Modifications in previous versions are listed in those respective versions. Click on Previous Version in the header until you get to the desired version.

Copyright © 2001–2022 Unicode, Inc. All Rights Reserved. The Unicode Consortium makes no expressed or implied warranty of any kind, and assumes no liability for errors or omissions. No liability is assumed for incidental and consequential damages in connection with or arising out of the use of the information or programs contained or accompanying this technical report. The Unicode Terms of Use apply.

Unicode and the Unicode logo are trademarks of Unicode, Inc., and are registered in some jurisdictions.