ISO 639-2 language codes are the three-letter codes that library catalogs, archives and metadata records use to state the language of a resource. The standard is maintained by the Library of Congress as its registration authority, and the full list on this page reproduces it: 487 entries, from aar (Afar) to zza (Zaza). Search by code or by name, or filter to the special codes that cause most of the questions.
The special language codes: zxx, mul, und, mis and qaa-qtz
Five ISO 639-2 language codes do not name a language at all. They answer the awkward cases every collection eventually meets, and choosing the wrong one makes records less useful than leaving the field empty.
| Code | Meaning | Use it when |
|---|---|---|
| zxx | No linguistic content; Not applicable | The resource contains no language: instrumental music, photographs, raw data. |
| und | Undetermined | There is language, but nobody has identified it yet. |
| mul | Multiple languages | Many languages are present and listing them all is impractical. |
| mis | Uncoded languages | The language has no code in ISO 639-2 and no fitting collective code. |
| qaa-qtz | Reserved for local use | An institution needs its own documented code for a local purpose. |
The zxx language code is the most looked-up of the five and the most misused, so the full list below links it to its own page. It belongs on a record when there is nothing to read or hear in any language; a silent film with intertitles is not zxx, because the intertitles are text. A record that uses the zxx language code makes a claim about content, not about cataloging effort. Its opposite in practice is und: a field recording of speech in an unknown language is und, not zxx, because the language exists and only the identification is missing.
Bibliographic and terminology codes
For most languages ISO 639-2 assigns a single code. For 20 languages it keeps two: a bibliographic (B) code derived from the English name and a terminology (T) code derived from the language's own name. German is ger in the bibliographic form and deu in the terminology form; French is fre and fra; Chinese is chi and zho. Library systems built on MARC 21 use the B codes, while ISO 639-3, BCP 47 language tags and most software use the T codes. Mixing them inside one collection splits searches, so an application profile should state which form it expects.
The pairs are: alb/sqi, arm/hye, baq/eus, bur/mya, chi/zho, cze/ces, dut/nld, fre/fra, geo/kat, ger/deu, gre/ell, ice/isl, mac/mkd, mao/mri, may/msa, per/fas, rum/ron, slo/slk, tib/bod, wel/cym.
ISO 639-2 language codes in Dublin Core and other metadata
In Dublin Core the language of a resource goes in dc:language, and DCMI recommends a controlled vocabulary for the value. Three-letter ISO 639-2 language codes work in any record; on the web, RFC 5646 tags (BCP 47) such as en, pt-BR or sr-Latn are the norm, and the DCMI encoding schemes RFC5646, RFC4646 and ISO639-3 state which list a value comes from. In linked data the language can be a URI from the Library of Congress vocabulary service, for example http://id.loc.gov/vocabulary/iso639-2/zxx, which removes any doubt about B versus T forms.
BCP 47 uses the shortest available code: the two-letter ISO 639-1 code when one exists (183 of the 487 entries have one), otherwise the three-letter code. That is why a web page in German is tagged de, while the same language in a MARC record is ger.
Collective codes and when to use them
Many ISO 639-2 language codes stand for a group of languages, not a single one: art for artificial languages, cad for Caddo, tup for the Tupi languages, and dozens of names ending in “languages”. A collective code is the right choice when a resource is in a member language that has no code of its own in ISO 639-2. When ISO 639-3 offers an individual code, a record that needs precision is better served by that code with the ISO639-3 encoding scheme.
The ISO 639 family and where ISO 639-2 fits
ISO 639 is a family of code sets, each built for a different job, and records that mix them without saying so are the usual source of mismatched language codes.
- ISO 639-1 has two-letter codes (en, de, fr, ja) for a few hundred widely used languages. Its registration authority is Infoterm.
- ISO 639-2 has three-letter codes for a larger set of languages plus collective codes for language groups. The Library of Congress is its registration authority, and it is the only part with separate bibliographic and terminology forms.
- ISO 639-3 aims to cover every known individual language, living, extinct, ancient and constructed. SIL International maintains it. Where a language also has an ISO 639-2 code, ISO 639-3 uses the same three letters (the terminology form where B and T differ), so deu, fra and zho are valid in both.
- ISO 639-5 gives three-letter codes to language families and groups, such as gem for Germanic languages or sla for Slavic languages. Many of its codes are the collective codes of ISO 639-2, extended with further groups. The Library of Congress maintains this part too.
- ISO 639-4 sets out general principles for language coding and does not contain a code list.
- ISO 639-6 proposed four-letter codes for language variants. It was withdrawn and has no role in current records.
A consolidated edition of ISO 639 now gathers the parts into a single document and refers to the code lists as sets (Set 1, Set 2 and so on). The lists themselves, and the registration authorities that maintain them, carry on as before, so a record that says “ISO 639-2” remains perfectly clear.
In practice, ISO 639-2 language codes are the common ground. Library and archive systems were built on them, ISO 639-3 is a superset of their individual-language entries, and BCP 47 falls back to them whenever no two-letter code exists. Dublin Core states which set a value comes from through encoding schemes such as ISO639-3 and RFC5646.
Macrolanguages and individual languages in practice
ISO 639-3 introduced the macrolanguage: a code that stands for several closely related individual languages which, in some contexts, are treated as one. Three examples show why the distinction matters in a catalog.
| Macrolanguage | Some member individual languages |
|---|---|
| ara (Arabic) | arb (Standard Arabic), arz (Egyptian Arabic), apc (Levantine Arabic) |
| zho (Chinese) | cmn (Mandarin Chinese), yue (Cantonese), wuu (Wu Chinese), nan (Min Nan), hak (Hakka) |
| msa (Malay) | zsm (Standard Malay), ind (Indonesian) |
Norwegian works the same way: nor is a macrolanguage whose members are nob (Norwegian Bokmål) and nno (Norwegian Nynorsk).
A general library catalog is well served by chi (B form) or zho, the level ISO 639-2 language codes offer and the level most users search at. A speech corpus, a subtitle archive or a linguistic dataset needs the member code: a Cantonese film soundtrack tagged only zho tells a researcher almost nothing. The two levels can coexist in one system, provided the profile says which field holds which and the macrolanguage code is not used as a quiet substitute for “I did not check”.
On the web, the IANA Language Subtag Registry records that yue is preferred over the older form zh-yue, so software should canonicalize tags before comparing them.
Language codes in common metadata standards
Each metadata standard has its own place and its own expected form for language codes. The examples below use German, which conveniently has a different B code (ger) and T code (deu), so the choice of form is visible.
MARC 21
The main language sits in positions 35–37 of field 008, and additional languages go in field 041, with subfields that say what each code describes: $a for the text, $b for summaries, $d for sung or spoken text, $g for accompanying material, $h for the original language of a translation and $j for subtitles and captions.
008/35-37 ger
041 1# $a ger $h eng
The second line describes a German translation of an English original. MARC uses the B forms throughout.
Dublin Core
dc:language takes any controlled value. The cleanest records state the scheme, as the guide on writing Dublin Core metadata records shows in more detail:
<dcterms:language xsi:type="dcterms:ISO639-3">deu</dcterms:language>
<dcterms:language xsi:type="dcterms:RFC5646">de-CH</dcterms:language>
In DCMI terms, the value is an instance of LinguisticSystem, which is also why a URI from the Library of Congress vocabulary service is a valid value in linked data.
MODS
MODS names the authority on the element, so the B/T question is answered inside the record:
<language>
<languageTerm type="code" authority="iso639-2b">ger</languageTerm>
</language>
The authority values iso639-2t, iso639-3 and rfc5646 are also defined. A MODS record can repeat language for each language present, and the objectPart attribute can say which part of the resource (for example a summary or a table of contents) is in that language.
EAD
Finding aids in EAD 2002 put the language of the materials inside langmaterial, with the code in an attribute and the name as readable text:
<langmaterial>Materials in <language langcode="ger">German</language>
and <language langcode="fre">French</language>.</langmaterial>
The eadheader can declare the code list with the langencoding attribute, typically iso639-2b, so a consumer knows that fre and not fra is meant.
PREMIS
PREMIS deliberately leaves descriptive metadata out of scope, and it has no element for the language of an object’s content. In a METS package the language lives in the descriptive section (MODS or Dublin Core), next to the PREMIS preservation section.
HTML, XML and EPUB
Web and publishing formats use BCP 47 tags, never bibliographic codes:
<html lang="de">
<p lang="sr-Latn">Dobar dan</p>
<note xml:lang="de-AT">…</note>
An EPUB 3 package document must contain at least one dc:language element, also with a BCP 47 value:
<dc:language>de</dc:language>
Copying ger from a MARC record into an HTML lang attribute produces a tag no browser recognizes. Conversion means mapping the B code to the T code and then to the shortest form, which for German is de.
Choosing between zxx, und, mul and several language codes
The special codes on this page and the habit of listing codes one by one solve different problems. Four common cases show how the decision runs in a real record.
A silent film with intertitles
The film has no spoken language, but the intertitles are text, and text is language. If the intertitles are English, the language of the resource is eng. MARC catalogers often go further and record the intertitle language in 041 with a note explaining that the film is silent. The zxx language code would be correct only for a silent film with no intertitles, no credits in words and no visible text that carries meaning.
A bilingual dictionary
A German–Spanish dictionary contains two languages in substance, so the record lists both: ger and spa in MARC, two dc:language values in Dublin Core. mul is the wrong tool here. It loses information that is easy to record, and a user filtering for Spanish material would never find the book. mul earns its place on a 30-language phrasebook or a proceedings volume where enumerating every language would swamp the record.
An instrumental CD with a printed booklet
The music itself has no linguistic content, so the zxx language code fits the recording. The booklet, however, is text: if it is in English and French, MARC records zxx in 008/35-37 and the booklet languages in 041 $g. In Dublin Core, which has no subfield for accompanying material, a record can carry zxx for the audio and add the booklet languages only if the profile allows several values with a note.
A dataset of numbers with English column headers
A spreadsheet of measurements is mostly numbers, but its headers, units and codebook are English, and a user has to read them to use the data. The honest value is eng. zxx would claim that nothing in the file needs reading, which is false the moment someone has to interpret a column called “flow_rate_lps”. A dataset with no headers at all, such as raw sensor output documented elsewhere, is a fair candidate for zxx.
One rule covers all four: und means “language present, not yet identified”, mul means “too many languages to list”, zxx means “no language at all”, and in every other case the specific language codes beat all three.
Script and region: when language codes need more
A language code alone cannot always describe what a reader will see. BCP 47 adds two optional parts that matter for metadata: a script subtag from ISO 15924, four letters with an initial capital, maintained by the Unicode Consortium as registration authority; and a region subtag, either an ISO 3166-1 alpha-2 code or a UN M49 number (the region codes are covered on the country codes page).
- sr-Latn and sr-Cyrl. Serbian is written in both Latin and Cyrillic script. A text in one is not searchable in the other without transliteration, so the script subtag is part of the identity of the text, not a decoration.
- zh-Hant and zh-Hans. Traditional and Simplified Chinese characters differ for many words. For text, the script subtag usually says more than a region would: zh-Hant describes the writing system wherever it is used, while zh-TW says only where the text is from.
- pt-BR and pt-PT. Brazilian and European Portuguese differ in spelling, vocabulary and grammar enough that software localization and subtitle files are routinely separated by region.
- es-419. The numeric region 419 (Latin America and the Caribbean, from UN M49) marks Spanish written for that market as a whole, not for one country.
The registry also prevents redundant tags. Many languages carry a Suppress-Script field: English, for example, is almost always written in Latin script, so en-Latn is discouraged and plain en is the expected form. Adding a script or region subtag where it carries no distinction only makes matching harder.
Library standards differ again. MARC 21 records scripts separately (field 066 lists the character sets present, and subfield $6 identifies the script of linked 880 fields) and never folds script into the language code, which is another reason a straight copy between MARC and HTML fails.
How language codes are added, changed and deprecated
Each part of ISO 639 has its own route for change, and anyone maintaining a code list should follow the one that matches their data.
For ISO 639-2, the Library of Congress publishes change notices on its ISO 639-2 pages. Proposals for new codes are reviewed by a joint advisory committee drawn from the ISO 639 registration authorities, and a new code typically requires a body of literature in the language, since ISO 639-2 was designed around documents. Codes are deprecated, not deleted: mol (Moldavian), for instance, was deprecated and is now covered by rum/ron, and the old bibliographic codes scc and scr gave way to srp and hrv. A record holding a deprecated code is not wrong for the period it was made in, but a validator should flag it and propose the successor.
For ISO 639-3, SIL International runs an open change request process. Anyone can submit a request to add, split, merge, retire or rename a code, requests are posted for public review, and decisions are published at the end of each annual cycle. Retired codes stay in a separate table with the reason for retirement (merge, split, duplicate, non-existent) and, where possible, the code that replaces them.
For BCP 47, the IANA Language Subtag Registry is the authority. It picks up changes to ISO 639, ISO 15924 and ISO 3166 and records deprecations with a Preferred-Value field, which is exactly what tag-canonicalizing software reads. A registry entry never disappears; it only gains a Deprecated date and a pointer to its replacement.
A local list fed from these sources should store the source and the date it was last compared, keep deprecated entries with a status flag, and never reuse a code from the reserved local range qaa-qtz once a standard code exists for the same language. A record from any decade can then still be read against the full list it was made with.