A language of the resource.
Recommended practice is to use either a non-literal value representing a language from a controlled vocabulary such as ISO 639-2 or ISO 639-3, or a literal value consisting of an IETF Best Current Practice 47 language tag.
dc:language has the shortest definition in the element set, “a language of the resource”, and one of the most precise recommendations attached to it. The usage comment asks for either a code from a controlled vocabulary such as ISO 639-2 or ISO 639-3, or a language tag that follows IETF BCP 47. Both routes lead to machine-readable values; free text such as “English” or “Englisch” leads to facet lists with the same language in five spellings. The element describes the language of the intellectual content, so a French novel in an English-language catalog is coded French, whatever language the cataloging was done in.
From RFC 1766 to RFC 4646 and RFC 5646
BCP 47 is the stable label for whichever RFC currently defines language tags. The sequence matters because older Dublin Core records and schemas cite specific numbers:
| RFC | Year | What it brought |
|---|---|---|
| RFC 1766 | 1995 | two-letter ISO 639 codes plus an optional country code |
| RFC 3066 | 2001 | three-letter ISO 639-2 codes |
| RFC 4646 | 2006 | script and region subtags and the IANA subtag registry |
| RFC 5646 | 2009 | extended language subtags and ISO 639-3 and 639-5 coverage |
RFC 4646 was the turning point. Before it, a tag was a language code with an optional country. RFC 4646 introduced a structured grammar (language, script, region, variant) and a central registry maintained by IANA, which is why a tag like sr-Latn-RS (Serbian in Latin script, as used in Serbia) became possible. RFC 5646 kept that structure and widened the language inventory. DCMI Metadata Terms has an encoding scheme for each of these, so a qualified record can state exactly which rule set its tags follow.
Writing tags that validate
A few habits keep tags clean:
- Use the shortest valid code.
en, noteng, in a BCP 47 tag; the three-letter form is only correct where no two-letter code exists. - Add a region only when it changes something.
pt-BRandpt-PTdiffer in spelling conventions;en-USon every English record usually adds noise. - Add a script subtag for languages written in more than one script:
zh-Hansandzh-Hant,sr-Cyrlandsr-Latn,az-Latn. - Case does not affect meaning, but the conventional form (
zh-Hant-TW) is easier to read and to diff.
The site’s language code reference lists the code types that turn up in format and preservation records.
Codes as URIs
For linked data, the comment’s other option is a URI from a code list. The Library of Congress publishes ISO 639-2 at id.loc.gov, so Old English can be referenced as http://id.loc.gov/vocabulary/iso639-2/ang. That form works best with dcterms:language, whose range is dcterms:LinguisticSystem; in simple Dublin Core a literal tag is still the more portable choice. ISO 639-3 is the list to consult for languages without a 639-2 code, since it covers several thousand individual languages.
Edge cases archives meet
Multilingual resources take one statement per language. Content with no linguistic material, such as a wordless photograph or a binary sensor log, can be coded zxx. Ranges qaa to qtz are reserved for local use and are only meaningful inside the institution that assigned them.
Sign languages have their own ISO 639-3 codes, ase for American Sign Language and bfi for British Sign Language among them, so a video recorded in sign language can be described as precisely as a spoken interview. Leaving the element empty for such material, or coding it with the language of the subtitles alone, hides it from users searching for exactly that content.
Programming languages are a separate matter and do not belong here. The natural-language value for a Python script is the language of its comments and documentation, if any. The programming language is a format characteristic, better recorded in technical metadata or alongside dc:type, and confusing the two is a recurring problem when software collections are cataloged with the same templates as text.
Example in RDF
@prefix dc: <http://purl.org/dc/elements/1.1/> .
@prefix dcterms: <http://purl.org/dc/terms/> .
<https://example.org/texts/letters-volume-2>
dc:title "Collected letters, volume 2" ;
dc:language "en-GB", "sr-Latn" .
<https://example.org/texts/sermon-1680>
dcterms:language <http://id.loc.gov/vocabulary/iso639-2/ang> .
Questions about dc:language
What value should dc:language contain?
DCMI recommends either a BCP 47 language tag as a literal, such as en, de-AT or zh-Hant, or a URI for a language from a controlled list such as ISO 639-2 or ISO 639-3. Plain names like English are allowed but harder to process.
What is RFC 4646?
RFC 4646, published by the IETF in 2006, defined the structure of language tags and the IANA language subtag registry, adding script and region subtags. It replaced RFC 3066 and was itself replaced by RFC 5646 in 2009. All three have formed BCP 47.
Should new records use RFC 4646 or RFC 5646?
New records should follow BCP 47, which currently means RFC 5646 together with RFC 4647. Tags valid under RFC 4646 remain valid, so older records do not need to be rewritten when the newer rules are adopted.
How is a resource with several languages described?
Repeat the element once per language, for example one dc:language for en and one for fr in a bilingual report. A single value such as en/fr breaks validation and hides both languages from any filter.
Which code is used for content with no language?
ISO 639-2 provides zxx for no linguistic content, which suits instrumental music, wordless images or binary data files. It is better than leaving the element empty when a language field is mandatory in the schema.