Skip to content


Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

CLDF: Cross-linguistic Data Formats

CLDF is a specification of data formats suitable to encode cross-linguistic data in a way that maximizes interoperability and reusability, thus contributing to FAIR Cross-Linguistic Data.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

Table of Contents

Conformance Levels

CLDF is based on W3C's suite of specifications for CSV on the Web, or short CSVW. A CLDF dataset is:

While the JSON-LD dialect to be used for metadata according to the Metadata Vocabulary for Tabular Data can be edited by hand, this may already be beyond what can be expected by regular users. Thus, CLDF specifies two conformance levels for datasets: metadata-free or extended.

Metadata-free conformance

A dataset can be CLDF conformant without providing a separate metadata description file. To do so, the dataset MUST follow the default specification for the appropriate module regarding:

  • file names
  • column names (for specified columns)
  • CSV dialect

Thus, rather than not having any metadata, the dataset does not specify any; instead it falls back to using the defaults, i.e. "free" as in "beer" not as in "gluten-free". The CSV file MAY contain additional columns not specified in the default module descriptions.

The default file names and column names are described in components. The default CSV dialect is RFC4180 using the UTF-8 character encoding, i.e. the CSV dialect specified as:

  "encoding": "utf-8",
  "lineTerminators": ["\r\n", "\n"],
  "quoteChar": "\"",
  "doubleQuote": true,
  "skipRows": 0,
  "commentPrefix": "#",
  "header": true,
  "headerRowCount": 1,
  "delimiter": ",",
  "skipColumns": 0,
  "skipBlankRows": false,
  "skipInitialSpace": false,
  "trim": false

Some of the effects of this metadata-free conformance are:

  • The first line of each file must contain the comma-separated list of column names.
  • No comment lines are allowed.

Extended conformance

A dataset is CLDF conformant if it uses a custom metadata file, derived from the default profile for the appropriate module. The metadata MUST specify a dc:conformsTo property with one of the CLDF module URIs as value.

Providing a metadata file allows for considerable flexibility in describing the data files, because the following aspects can be customized (within the boundaries of the CSVW specification):

  • the CSV dialect description (possibly per table), e.g. to:
    • allow comment lines (if appropriately prefixed with commentPrefix)
    • omit a header line (if appropriately indicated by "header": false)
    • use tab-separated data files (if appropriately indicated by "delimiter": "\t")
  • the table property url
  • the column property titles
  • the inherited column properties
  • adding common properties,
  • adding foreign keys, to specify relations between tables of the dataset.

Thus, using extended conformance via metadata, a dataset may:

  • use tab-separated data files,
  • use non-default file names,
  • use non-default column names,
  • add metadata describing attribution and provenance of the data,
  • specify relations between multiple tables in a dataset,
  • supply default values for required columns like Language_ID, using virtual columns.

In particular, since the metadata description resides in a separate file, it is often possible to retrofit existing CSV files into the CLDF framework by adding a metadata description.

Thus, conformant CLDF processing software MUST implement support for the CSVW specification to the extent necessary.

CLDF Ontology

CLDF data uses terms from the CLDF Ontology, as specified in the file terms.rdf, to mark TableGroup or Table objects which have special meaning (specifying modules or components respectively) within the CLDF framework.

The CLDF Ontology also provides a set of properties to attach semantics to individual columns. While many of these properties are similar (or identical) to properties defined elsewhere - most notably in the General Ontology for Linguistic Description - GOLD - we opted for inclusion to avoid ambiguity, but made sure to reference the related properties in the ontology.

Note that the column names in the default table descriptions (e.g. formTable) are not always the same as the column properties. Each column has both a csvw:name and a separate propertyURL linking the column to the CLDF Ontology. Each property also has a rdf:label which might also be different.


  • In an ill-advised attempt to version the ontology, v1.0 has been baked into the term URIs. While this may be a good idea in case of incompatible changes (e.g. if the semantics of a term changed), it presents an obstacle for initeroperability in case of backwards-compatible changes. So starting with CLDF 1.1, we will keep as namespace for all versions of the 1.x series, and specify the particular version when a term was introduced using dc:hasVersion properties per term.
  • For better human readability the CLDF Ontology should be visited with a browser capable of rendering XSLT - such as Firefox.

CLDF Dataset

CLDF Metadata file

A CLDF dataset is described with metadata provided as JSON file following the Metadata Vocabulary for Tabular Data. To make tooling simpler, we restrict the metadata specification as follows:

  • Metadata files MUST specify a tables property on top-level, i.e. MUST describe a TableGroup. While this adds a bit of verbosity to the metadata description, it makes it possible to describe multiple tables in one metadata file.
  • The common property dc:conformsTo of the TableGroup is used to indicate the CLDF module, e.g. "dc:conformsTo": ""
  • The common property dc:conformsTo of a Table is used to associate tables with a particular role in a CLDF module using appropriate classes from the CLDF Ontology.
  • If each row in the data file corresponds to a resource on the web (i.e. a resource identified by a dereferenceable HTTP URI), the tableSchema property SHOULD provide an aboutUrl property.
  • If individual cells in a row correspond to resources on the web, the corresponding column specification SHOULD provide a valueUrl property.

Each dataset SHOULD provide a dataset distribution description using the DCAT vocabulary. This will make it easy to
catalog cross-linguistic datasets. In particular, each dataset description SHOULD include these properties:

Thus, an example for a CLDF dataset description could look as follows:

  "@context": "",
  "dc:conformsTo": "",
  "dc:title": "The Dataset",
  "dc:bibliographicCitation": "Cite me like this!",
  "dc:license": "",
  "null": "?",
  "tables": [
      "url": "ds1.csv",
      "dc:conformsTo": "",
      "tableSchema": {
        "columns": [
            "name": "ID",
            "datatype": "string",
            "propertyUrl": ""
            "name": "Language_ID",
            "datatype": "string",
            "propertyUrl": "",
            "valueUrl": "{Language_ID}"
            "name": "Parameter_ID",
            "datatype": "string",
            "propertyUrl": ""
            "name": "Value",
            "datatype": "string",
            "propertyUrl": ""
            "name": "Comment",
            "datatype": "string",
            "propertyUrl": ""
            "name": "Source",
            "datatype": "string",
            "propertyUrl": ""
            "name": "Glottocode",
            "virtual": true,
            "propertyUrl": "",
            "valueUrl": "{Language_ID}"
        "aboutUrl": "{ID}",
        "primaryKey": "ID"

CLDF Data files

While it is possible to add any kind of CSV files to a CLDF dataset (by virtue of being an extension of CSVW), the CLDF standard recognizes (and attaches specified semantics) to tables described with a common property dc:conformsTo with one of the component URIs of the CLDF Ontology as value.

Additionally, CLDF semantics can be assigned to individual columns by assigning one of the property URIs defined in the CLDF Ontology as propertyUrl. CLDF conformant software MUST detect CLDF-specific columns by matching the propertyUrl to CLDF Ontology terms and NOT by matching column name to default column names recommended in the ontology.

Column specifications

  • CLDF column properties are assumed to have a complete row (or rather the entity a row stores data about) as scope; e.g. a source column is assumed to provide source information for any piece of data in the row. Thus, each property can be used only once per table, which makes processing simpler.
  • More generally, CLDF assumes column names in a table to be unique.
  • Since CLDF is designed to enable data reuse, data creators should assume that schema information like table or column names ends up in all sorts of environments, e.g. as names in SQL databases or as parts of URLs of a web application. Thus, it is RECOMMENDED to stick to ASCII characters in such names and avoid usage of punctuation other than :._-.
  • Cardinality: CSVW allows specifying columns as "multivalued", i.e. as containing a list of values (of the same datatype), using the separator property. Thus, CLDF consumers MUST consult a column's separator property, to figure out whether the value must be interpreted as list or not. Note that this also applies to foreign keys. However, CLDF may restrict the cardinality as follows:
    • The specification of a property in the ontology MAY contain a dc:extent property with value singlevalued or multivalued, fixing cardinality of any instance of such a column in any table.
    • The specification of a column in the default metadata of a component MAY contain a dc:extent property with value singlevalued or multivalued, fixing cardinality.


Each CLDF data table SHOULD contain a column which uniquely identifies a row in the table. This column SHOULD be marked using:

  • a propertyUrl of
  • the column name ID in the case of metadata-free conformance.

To allow usage of identifiers as path components of URIs and ensure they are portable across systems, identifiers SHOULD be composed of alphanumeric characters, underscore _ and hyphen - only, i.e. match the regular expression [a-zA-Z0-9\-_]+ (see RFC 3986).

Following our design goal to reference rather than duplicate data, identifiers may be used to reference existing entities (e.g. Glottolog languages, WALS features, etc.). This can be done as follows:

  • If the identifier can be interpreted as links to other entities, e.g. using the WALS three-lettered language codes to identify languages, this should be indicated by assigning the column an appropriate valueUrl property, e.g.{ID}
  • If the identifier follows a specified identification scheme, e.g. ISO 639-3 for languages, this can be indicated by adding a virtual column with a suitable propertyUrl to the table's list of columns.

Missing data

Data creators often want to distinguish two kinds of missing data, in particular when the data is extracted from sources:

  1. data that is missing/unknown because it was never extracted from the source,
  2. data that is indicated in the source as unknown.

The CSVW data model can be used to express this difference as follows:

  • Case 1 can be modeled by not including the relevant data as row at all.
  • Case 2 can be modeled using the null property of the relevant column specification (defaulting to the empty string) as value in a data row.


Considering that any single step in collecting (cross-)linguistic data involves some amount of analysis and judgement calls, it is essential to make it easy to trace assertions back to their source.

Each CLDF data table may contain a column listing sources for the data asserted in the row. This column MUST be marked using:

  • a propertyUrl of
  • the column name Source in the case of metadata-free conformance.

Sources are specified as semicolon-separated (unless the metadata specifies a different separator) source specifications, of the form source_ID[source context], e.g. meier2015[3-12] where meier2015 is a citation key in the accompanying sources file.

Foreign keys

Often cross-linguistic data is relational, e.g. cognate judgements group forms into cognate sets, creating a many-to-many relationship between a FormTable and a CognatesetTable.

To make such relations explicit, the CLDF Ontology provides a set of reference properties.

Reference properties MUST be interpreted as foreign keys, e.g. a propertyUrl specified for column Col1 of a table with url table1.csv is equivalent to a CSVW foreign key constraint

  "foreignKeys": [
           "columnReference": "Col1",
           "reference": {
               "resource": "languages.csv",
               "columnReference": "ID"

assuming that the LanguageTable component has url languages.csv and a column ID with propertyUrl

While spelling out foreign key constraints may feel cumbersome, it is still RECOMMENDED that metadata creators do so, to make the data compatible with CSVW tools. The foreign key constraints MUST be specified explicitly if the referenced column does not have a propertyUrl

Note that columns for reference properties may still be "nullable", i.e. contain NULL values, to allow for rows where no reference can be specified.

Sources reference file

References to sources - if not referenced by Glottolog ID - can be supplied as part of a CLDF dataset as an UTF-8 encoded BibTeX file (with the citation keys serving as local Source IDs). The filename of this BibTeX file MUST be either:

  • sources.bib in case of metadata-free conformance
  • or specified as top-level common property dc:source in the dataset's metadata.

CLDF Modules

Much like Dublin Core Application Profiles, CLDF Modules group terms of the CLDF Ontology into tables. Thus, CLDF module specifications are recommendations for groups of tables modeling typical cross-linguistic datatypes. Currently, the CLDF specification recognizes the following modules:

In addition, a CLDF dataset can be specified as Generic, imposing no requirements on tables or columns. Thus, Generic datasets are a way to evolve new data types (to become recognized modules), while already providing (generic) tool support.

In the CLDF Ontology modules are modeled as subclasses of dcat:Distribution, thus additional metadata as recommended in the DCAT specification should be provided.

For each type of CLDF dataset there is a CLDF module, i.e. a default metadata profile describing the required tables, columns and datatypes. metadata-free conformance means data files will be read as if they were accompanied by the corresponding default metadata.

CLDF Components

Some types of cross-linguistic data may be part of different CLDF modules. These types are specified as components in a way that can be re-used across modules (typically as table descriptions, which can be appended to the tables property of a module's metadata). A component is a CSVW table description with a dc:conformsTo property having one of the component terms in the CLDF Ontology as value. Each component listed below is described in a README and specified by the default metadata file in the respective directory.

A component corresponds to a certain type of data. Thus, to make sure all instances of such a type have the same set of properties, we allow at most one component for each type in a CLDF dataset.


  • Using UTF-8 as character encoding means editing these files with MS Excel is not completely trivial, because Excel assumes cp1252 as default character encoding - Libre Office Calc on the other hand handles these files just fine.
  • The tool support for CSV files is getting better and better due to increasing interest in data science. Some particularly useful tools are:


Changes to the CLDF specification will be released as new versions, using a Semantic Versioning number scheme. While older versions can be accessed via releases of this repository or from ZENODO, where releases will be archived, the latest released version is also reflected in the master branch of this repository, i.e. whatever you see navigating the directory tree at reflects the latest released version of the specification.


Work on this proposal for a cross-linguistic data format was triggered by the LANCLID 2 workshop held in April 2015 in Leipzig - in particular by Harald Hammarström's presentation A Proposal for Data Interface Formats for Cross-Linguistic Data.