Standard

Info

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.

This document defines a common standard for sharing lexicons of morphological paradigms. While the specifications tend to be rigid with respect to their formal aspects (how to format data and metadata), we understand that as linguists, we are most interested in the parts of language that are complex to analyze, and thus complex to code. Thus, the standard is more flexible regarding the exact content of the data, allowing linguists to make project-specific choices about content, while reaping other benefits of standardisation.

Documentation

Datasets must be accompanied by human-readable documentation, justifying important choices in the constitution of the dataset and guiding users in its interpretation. Minimally, a README.md MUST be present. It is RECOMMENDED to also add a data_sheet.md file, following the template. More documentation MAY be present if needed.

Metadata

Metadata MUST be provided in json, following the frictionless standard, and in conformity with the specs (also readable directly in json: paralex.package.json). See the tutorial on how to generate metadata automatically.

At the highest level, we add a languages_iso639 key, the value of which is a list of iso language codes.

The metadata is thus given as a json file. Here is what such a file might look like:

example of .json metadata file

.package.json file

    {
      "name": "paralex-min-chanter",
      "title": "Minimal example - Indicative present of The French Verb \"Chanter\"",
      "profile": "data-package",
      "licenses": [
        {
          "name": "CC-BY-SA-4.0",
          "path": "https://creativecommons.org/licenses/by-sa/4.0/",
          "title": "Attribution-ShareAlike 4.0 International"
        }
      ],
      "contributors": [
        {
          "title": "Sacha Beniamine"
        }
      ],
      "keywords": [
        "example",
        "toy",
        "French",
        "verbs",
        "paradigms",
        "paralex"
      ],
      "version": "1.0",
      "resources": [
        {
          "name": "readme",
          "type": "text",
          "title": "Read me",
          "description": "Basic documentation",
          "path": "readme.md",
          "scheme": "file",
          "format": "md",
          "mediatype": "text/markdown",
          "encoding": "utf-8"
        },
        {
          "name": "forms",
          "type": "table",
          "title": "Inflected forms",
          "path": "forms.csv",
          "scheme": "file",
          "format": "csv",
          "mediatype": "text/csv",
          "encoding": "utf-8",
          "schema": {
            "name": "forms-schema",
            "fields": [
              {
                "name": "form_id",
                "type": "string",
                "title": "Form table row identifiers",
                "description": "These identifiers are specific to form, lexeme, cell triples.",
                "constraints": {
                  "required": true,
                  "unique": true
                }
              },
              {
                "name": "lexeme",
                "type": "string",
                "title": "Reference to a lexeme identifier",
                "description": "Lexeme identifiers must be unique to paradigms.",
                "constraints": {
                  "required": true
                },
                "rdfProperty": "https://www.paralex-standard.org/paralex_ontology.xml#lexeme"
              },
              {
                "name": "cell",
                "type": "string",
                "title": "Reference to a cell identifier",
                "description": "The set of feature values as would appear in a gloss, separated by dots, eg. prs.ind.1sg or f.pl",
                "constraints": {
                  "required": true
                },
                "rdfProperty": "https://www.paralex-standard.org/paralex_ontology.xml#cell"
              },
              {
                "name": "phon_form",
                "type": "string",
                "title": "Inflected form (phonemic or phonetic)",
                "description": "The form, given in phonemic or phonetic notation, with sounds separated by spaces",
                "missingValues": [
                  "#DEF#"
                ],
                "rdfProperty": "https://www.paralex-standard.org/paralex_ontology.xml#phon_form"
              }
            ],
            "primaryKey": [
              "form_id"
            ]
          },
          "rdfType": "https://www.paralex-standard.org/paralex_ontology.xml#Form"
        }
      ],
      "languages_iso639": [
        "fr"
      ],
      "paralex-version": "2.1.0"
    }

The metadata specify information about the dataset (such as its contributors, title, relevant keywords, license, etc), and about what each table contains (its name, title, path, relation to other tables, detailed list of columns, etc.). It serves both as documentation, and as a tool to validate and manipulate the data.

For more information about what can be expressed by this metadata file, see the paralex specs and the frictionless specs.

Relation to other work

Frictionless does not provide a way to specify relations to other work. Instead, we adopt the format defined for zenodo metadata, which is also provided in json. A first level attribute related_identifiers MAY be present in the metadata json file. Its value is an array of objects. Each object has two attributes:

• identifier: the persistent identifier
• relation: One of: isCitedBy, cites, isSupplementTo, isSupplementedBy, isContinuedBy, continues, isDescribedBy, describes, hasMetadata, isMetadataFor, isNewVersionOf, isPreviousVersionOf, isPartOf, hasPart, isReferencedBy, references, isDocumentedBy, documents, isCompiledBy, compiles, isVariantFormOf, isOriginalFormof, isIdenticalTo, isAlternateIdentifier, isReviewedBy, reviews, isDerivedFrom, isSourceOf, requires, isRequiredBy, isObsoletedBy, obsoletes

For example, part of the json metadata file could look like this:

{
  ...
  'related_identifiers': [
    {
      'relation': 'isSupplementTo',
      'identifier': '10.1234/foo'
    },
    {
      'relation': 'cites',
      'identifier': 'https://doi.org/10.1234/bar',
    }
  ],
  ...
}

For citing academic sources in the data itself, we recommend to prepare a .bib file and to refer to the bibtex key.

Data format

Paradigmatic lexicons are tabular datasets. Paralex lexicons MUST be provided in long form, written as csv (comma separated value) tables using utf-8 encoding.

Example

Visual tableCSV table

form_id	lexeme	cell	phon_form
form-chanter-prs-1pl	chanter	ind.prs.1.pl	ʃ ɑ̃ t ɔ̃
form-chanter-prs-1sg	chanter	ind.prs.1.sg	ʃ ɑ̃ t
form-chanter-prs-2pl	chanter	ind.prs.2.pl	ʃ ɑ̃ t e
form-chanter-prs-2sg	chanter	ind.prs.2.sg	ʃ ɑ̃ t
form-chanter-prs-3pl	chanter	ind.prs.3.pl	ʃ ɑ̃ t
form-chanter-prs-3sg	chanter	ind.prs.3.sg	ʃ ɑ̃ t

forms.csv

 form_id,lexeme,cell,phon_form
 form-chanter-prs-1pl,chanter,ind.prs.1.pl,ʃ ɑ̃ t ɔ̃
 form-chanter-prs-1sg,chanter,ind.prs.1.sg,ʃ ɑ̃ t
 form-chanter-prs-2pl,chanter,ind.prs.2.pl,ʃ ɑ̃ t e
 form-chanter-prs-2sg,chanter,ind.prs.2.sg,ʃ ɑ̃ t
 form-chanter-prs-3pl,chanter,ind.prs.3.pl,ʃ ɑ̃ t
 form-chanter-prs-3sg,chanter,ind.prs.3.sg,ʃ ɑ̃ t

Files

Most of the data files expected are tables. A dataset MUST contain at least one forms table documenting inflected forms. Usually, a single table is not sufficient to provide all relevant information. The following tables SHOULD also be included:

• a sounds table documents the inventory of sounds used in the transcription
• a cells table documents the inventory of feature-value combinations (paradigm cells) for which lexemes inflect.
• a features-values table documents the inventory of grammatical features which compose the cells

Depending on the amount of information and granularity provided, the following files MAY also be included:

• a lexemes table documents the inventory of lexemes for which inflected forms are given
• a tags table provides a way of grouping rows of a table together, in paricular sets of inflected forms. Tags can indicate epistemic status (generated form vs attested), series of overabundant or defectiveness forms, specific dialectal or sociolinguistic variants, etc.
• a frequencies table provides complex frequency information where it would be insufficient to use frequency columns on lexemes and forms tables.
• a graphemes table documents the inventory of graphemes used in the orthographic form
• a sources file in bibtex or biblatex format (.bib) documents sources used in any table in the source column.

Further tables MAY be added to your resource as needed, for example to document languages, inflection classes, etc. The tables MAY be divided in several files to reduce their size.

The exact set of standard columns for each table is described in the specification. The dataset MUST NOT use aliases or alternate column names for columns described in the specification. It MAY use additional columns, which SHOULD strive to follow existing conventions and use known vocabularies. The columns MAY be given in any order, though as a convention, the identifier column (form_id, sound_id, cell_id, etc) SHOULD be the first column of their table.

The tables within the dataset have a range of specific, formal relationships to one another. That is to say, they are all linked together via a relational model. That model is summarized below. View the simple/advanced diagrams to see less or more detail.

Relations between tables: simpleRelations between tables: advanced

graph TD
    A[sounds.csv] --- B(forms.csv)
    F[graphemes.csv] --- B
    C[cells.csv] --- B
    D[tags.csv] --- B
    E[features-values.csv] --- C
    G[lexemes.csv] --- B
    H[frequencies.csv] --- C
    H --- B
    H --- G
    D --- G
    D --- H

• The phon_form entries of the forms table are composed of identifiers in the sounds table
• The orth_form entries of the forms table are composed of identifiers in the graphemes table
• The cell entries of the forms table match identifiers in the cells table
• IDs in the cells table are composed of identifiers in the features-values table
• Entries in the lexeme column of the forms table match identifiers in the lexemes table
• Entries in the _tags columns of the forms and lexemes table, separated by |, match identifiers in the tags table
• Entries in the frequencies tables may link to various other tables

erDiagram
    sounds }|..|{ forms : "compose the phon_form"
    graphemes }|..|{ forms : "compose the orth_form"
    features-values }|..|{ cells : "compose the cell_id"
    cells ||--|{ forms : "is referenced by (foreign key)"
    lexemes ||--|{ forms : "is referenced by (foreign key)"
    tags }|..|{ forms : "reference (multiple values possible)"
    tags }|..|{ lexemes : "reference (multiple values possible)"
    tags }|..|{ frequencies : "reference (multiple values possible)"
    frequencies }|--|{ lexemes : "is referenced by (foreign key)"
    frequencies }|--|{ forms : "is referenced by (foreign key)"
    frequencies }|--|{ cells : "is referenced by (foreign key)"
    sounds {
        string sound_id
        string label
        string comment
        string CLTS_id
        string PHOIBLE_id
    }
    forms {
        string form_id
        string lexeme
        string cell
        string phon_form
        string orth_form
    }
    features-values {
        string value_id
        string label
        string feature
        integer canonical_order
    }
    lexemes {
        string lexeme_id
        string inflection_class
        string label
        string gloss
    }
    cells {
        string cell_id
        string unimorph
        string ud
        string comment
    }
    tags {
        string tag_id
        string type
        string comment
    }
    graphemes {
        string grapheme_id
        string comment
        integer canonical_order
    }

    frequencies {
        string freq_id
        number frequency
    }

In this diagram:

• Solid lines indicate the "foreign key" relationship: the cell column in forms must refer exactly to rows of the cells table, identified by their cell_id. The lexeme column in forms must refer exactly to rows of the lexemes table, identified by their lexeme_id.
• Dotted lines indicate a vocabulary relationship: the phon_form column in forms must be composed of sound_ids from the sounds table, and the cell_id column in cells must be composed of value_ids from the features-values table.
• Only four columns are shown for each table, see the specs for the full column lists.

Forms

Each row of the forms table documents a single inflected form.

The forms table MUST have at least the columns: form_id, cell, lexeme and either a phon_form (phonological form), or orth_form (orthographic form). It SHOULD have a phon_form. We do not impose any particular phonological or phonetic analysis or sets of symbols, though datasets MUST follow some established convention.

The value of the phon_form MUST be a sequence of space-separated segments, e.g. not "dominoːrum" but "d o m i n oː r u m". If there is a sounds table, then all segments used in the phon_form MUST be documented by a row in sounds.

Values in the cell and lexeme columns MUST correspond to identifiers in the cells and lexemes tables respectively if these tables exist.

forms table

Here are a few rows representing the latin nominal paradigm of the word DOMINUS (master):

form_id	lexeme	cell	phon_form
f266	dominus	abl.pl	d o m i n iː s
f1299	dominus	abl.sg	d o m i n oː
f2325	dominus	acc.pl	d o m i n oː s
f3359	dominus	acc.sg	d o m i n u m
f4385	dominus	dat.pl	d o m i n iː s
f5418	dominus	dat.sg	d o m i n oː
f6444	dominus	gen.pl	d o m i n oː r u m
f7478	dominus	gen.sg	d o m i n iː
f8504	dominus	nom.pl	d o m i n iː
f9537	dominus	nom.sg	d o m i n u s
f11596	dominus	voc.sg	d o m i n e
f10563	dominus	voc.pl	d o m i n iː

If a frequency column is present, the source of the frequencies MUST be documented in the README.md or accompanying documentation. If providing multiple frequency measurements, it is best to use a separate frequencies table.

The table MAY have more columns, including those described in the specification.

Supra-segmental information in form transcriptions

Phonemic representations SHOULD include supra-segmentals such as stress, length, tones. They MAY also provide syllable structure.

There are many ways to annotate supra-segmental information. The RECOMMENDED way to do so is to mark supra-segmental information within the sequence of segments, as separate characters or diacritics attached to the segment which bears them (or the nucleus of the affected syllable).

suprasegmentals in phon_form

Here is an example for Nuer nominal plural, which involves breathy voice, length alternations, and tone.

form_id	lexeme	cell	phon_form
f1	‘ant’	nom.sg	c w ɔ̤́ x
f2	‘arrow’	nom.sg	b ʌ̤̀ːː r
f3	‘bead’	nom.sg	t ɪ̂ː k
f4	‘belt'	nom.sg	l à̤ːː ɣ
f5	‘ant’	nom.pl	c ṳ̌ːː ɣ
f6	‘arrow’	nom.pl	b ʌ̤̀ːː r í̤
f7	‘bead’	nom.pl	t j ɛ̂ x
f8	‘belt'	nom.pl	l ʌ̤̌ː k

Note

The choice of particular conventions is not constrained: you can choose to write a high tone as ́, ˦, 1, etc. However, a single dataset MUST adhere to a single, coherent notational choice which SHOULD be documented through the sounds table and follow a formally specified standard, such as CLTS BIPA.

Usage of tags

It is not uncommon for the cell of a paradigm to be associated with multiple variant forms. There are several kinds of variation that we would not wish to conflate. Accordingly, to mark forms which are related to each other in certain ways, we use tags. Any tag defined in the tags table can be used in the appropriate *_tags columns. Several tags can be used, in which case they are separated with a | (and no spaces). Their order is not meaningful.

Importantly, form variants MUST NOT be aggregated into a single entry such as 'learn{ed/t}' or 'learned~learnt'. Rather, each variant is represented in its own row in the forms table and indexed with a tag identifier (see why the long form ?).

overabundance can lead to two rows for the same lexeme and paradigm cell. Sometimes, overabundance is specific to a single lexeme. Sometimes, however, overabundant forms across cells of a single lexeme form coherent sets which share for example the same stem, the same register, or more generally the same type of variation. These forms can be linked together by a tag in the overabundance_tag column.

Overabundance

In the forms tableIn the tags table

form_id	lexeme	cell	phon_form	overabundance_tag
f1	dream	pst	d r ɛ m t	irreg;t-form
f2	dream	pst	d r iː m d	d-form
f3	learn	pst	l ɜː n d	d-form
f4	learn	pst	l ɜː n t	t-form
f5	leap	pst	l ɛ p t	irreg;t-form
f6	leap	pst	l iː p t	t-form
f7	sweat	pst	s w ɛ t	irreg
f8	sweat	pst	s w ɛ t ɪ d	d-form

tag_id	tag_column_name	comment
irreg	overabundance_tag	irregular form
d-form	overabundance_tag	past in /d/
t-form	overabundance_tag	past in /t/

defectiveness: Defectiveness happens when some forms do not exist in a paradigm: for example, what is the singular of the English word "scissors" ? There is none, because this word exists only in the plural, it is a plurale tantum.

When this happens, simply leaving out rows is insufficient, as it is ambiguous between defectiveness and missing data. Thus, defective forms MUST have their own rows, and SHOULD be identified by identifiers in the defectiveness_tag column. Defective rows use this column to group together either all defective forms using a general label such as defective, or sets of defective forms, according to some analysis or motivation, for example
pluralia_tantum. Any relevant labels and groupings can be used, and MUST be defined in the tags table. In most cases, defective forms will not have any defined or well-formed phonological or orthographic representation.

The phon_form of a defective entry MUST contain "#DEF#" (meaning defective). It is possible for a defective form to still provide a well-formed phonological or orthographic form (speakers could produce the form, it is just not used), using the defectiveness_tag to label the row.

Defectiveness: Latin pluralia_tantum

In the forms tableIn the tags table

In this example, the form of the latin noun "pauci" have no well formed singular, as the word is a "pluralia tantum" (it is only used in the plural). The table still provides rows for each singular case, but the form is replaced by "#DEF#", and the defectiveness_tag column is marked for these forms with the type of defectivity, pluralia_tantum.

form_id	lexeme	cell	phon_form	defectiveness_tag
f1122	pauci	abl.pl	p a w k iː s
f1299	pauci	abl.sg	#DEF#	pluralia_tantum
f2325	pauci	acc.pl	p aw k oː s
f3359	pauci	acc.sg	#DEF#	pluralia_tantum
f4385	pauci	dat.pl	p aw k iː s
f5418	pauci	dat.sg	#DEF#	pluralia_tantum
f6444	pauci	gen.pl	p aw k oː r u m
f7478	pauci	gen.sg	#DEF#	pluralia_tantum
f8504	pauci	nom.pl	p aw k iː
f9537	pauci	nom.sg	#DEF#	pluralia_tantum
f10563	pauci	voc.pl	p a w k iː
f11596	pauci	voc.sg	#DEF#	pluralia_tantum

tag_id	tag_column_name	comment
pluralia_tantum	defectiveness_tag	defective forms because the lexeme exist only in the plural

variants: Dataset MAY account for observed variation across speakers by providing multiple rows for the same lexeme and paradigm cell, each containing a distinct phonemic forms, as well as an annotation in the form of a tag (variants_tag).

epistemic status: The column epistemic_tag MAY be used to indicate the epistemic status of a row's form. Suggested levels are: manually_checked, controversial, uncertain, generated, attested. Custom levels may be chosen.

analytic choices: If providing various analyses of the phonological form, the forms file MUST have a separate row for each analysis, and the rows MUST be annotated using the column analysis_tag to identify which are part of the same analytic set.

If providing two levels of abstraction for the phon_form (eg. one more phonetic and one more phonemic), two rows should be present with distinct phon_form, one annotated as phonemic, and one as phonetic in the analysis_tag column.

Segmented forms

Lexicons MAY provide segmentation information, using the analysed_phon_form and analysed_orth_form columns. The default segmentation marker should be + within word, and # for word boundary. If a more complex set of markers is necessary, it MUST be documented explicitly. If multiple segmentations are provided, each alternative MUST constitute a separate row, and they MUST be annotated with tags in order to distinguish each series of forms.

segmented forms

Here is an example for the latin forms of dominus:

form_id	lexeme	cell	phon_form	analysed_phon_form
f266	dominus	abl.pl	d o m i n iː s	d o m i n + iː s
f1299	dominus	abl.sg	d o m i n oː	d o m i n + oː
f2325	dominus	acc.pl	d o m i n oː s	d o m i n + oː s
f3359	dominus	acc.sg	d o m i n u m	d o m i n + u m
f4385	dominus	dat.pl	d o m i n iː s	d o m i n + iː s
f5418	dominus	dat.sg	d o m i n oː	d o m i n + oː
f6444	dominus	gen.pl	d o m i n oː r u m	d o m i n + oː r u m
f7478	dominus	gen.sg	d o m i n iː	d o m i n + iː
f8504	dominus	nom.pl	d o m i n iː	d o m i n + iː
f9537	dominus	nom.sg	d o m i n u s	d o m i n + u s
f10563	dominus	voc.pl	d o m i n iː	d o m i n + iː
f11596	dominus	voc.sg	d o m i n e	d o m i n + e

Sounds

The sounds table describes the full inventory of sounds used in the transcriptions. In the phon_form, these sounds are separated by spaces.

The sounds table MUST have at least an identifier column sound_id.

The sounds table is the result of analytic choices. In order to detail the precise meaning of each sound in this particular dataset, the table SHOULD have a column label indicating what the sound means in natural language, and SHOULD provide ways to identify the sounds by linking to other resources (e.g. using the columns CLTS_id, PHOIBLE_id) or through other columns (for example, by defining a language-specific set of distinctive features). It is best to employ a multiplicities of these strategies to maximize compatibility of datasets. Moreover, this table MAY have a comment column to further annotate complex or unusual choices.

Graphemes

The Graphemes table describes the full inventory of graphemes used in the orthographic form column orth_form in the forms table. It MUST have at least an identifier column grapheme_id. It may have any other columns, in particular a label, comment and a canonical_order. It MAY have more columns, including those described in the specification.

Having this table ensures that orthographic forms will be checked automatically during validation.

Letters in French

grapheme_id	comment	canonical_order
a		1
à		2
b		3
c		4
d		5
e		6
é		7
è		8
ê		9
f		10
g		11
h		12
i		13
ï		14
j		15
k		16
l		17
m		18
n		19
o		20
ô		21
p		22
q		23
r		24
s		25
t		26
u		27
v		28
w		29
x		30
y		31
z		32

Cells

The cells table describes the full inventory of feature-value combinations for which lexemes inflect. These are usually called paradigm cells in morphology. Example of paradigm cells are: "indicative present first person singular" (abbreviated ind.prs.1sg) or "nominative plural" (abbreviated nom.pl).

The cells table MUST have at least an identifier column cell_id. It MAY have more columns, including those described in the specification.

If there is no features-values table, then the cells table MUST have at least one column mapping the cells to a widely used vocabulary, such as unimorph, universal dependencies, GOLD, etc. If there is a features-values table (properly linked to other vocabularies itself), then linking cells to other vocabularies is only RECOMMENDED. If there are several competing naming conventions for some cells, the authors may choose freely, but SHOULD provide alternate names in another column to maximize understandability and make translation from one convention to another as easy as possible.

We recognize that some vocabularies cannot account for specific phenomena (for example, when order is distinctive due to affix stacking, but a particular vocabulary does not distinguish between gen.du and du.gen). Dataset authors are free to pick any existing, formal vocabulary which they deem expressive enough. If none exist, it is strongly RECOMMENDED to raise the issue with the people maintaining these vocabularies, and they MAY provide imperfect mappings to the best vocabulary available. They MUST comment the issue in the documentation.

Identifiers for the cells, cell_id, MUST be feature values in lowercase, separated by dots, as in: "nom.sg" for nominative singular (this follows the Leipzig Glossing Rule convention). If there is a features-values table, all of the feature values composing these identifiers MUST be documented as separate rows in the features-values table.

This allows considerable flexibility in using ad-hoc cells, while ensuring compositional labels (made of features-values), and encouraging transparent semantics (linking to other vocabularies).

Choosing and labelling paradigm cells is often a tricky matter, with considerable disagreement in the literature. Whatever choices are made in this regard, they SHOULD be explicitly documented.

Example

cell_id	ud	POS	unimorph
nom.pl	NOUN:Nom+Plur	noun	N;NOM;PL
nom.sg	NOUN:Nom+Sing	noun	N;NOM;SG
voc.pl	NOUN:Voc+Plur	noun	N;VOC;PL
voc.sg	NOUN:Voc+Sing	noun	N;VOC;SG
acc.pl	NOUN:Acc+Plur	noun	N;ACC;PL
acc.sg	NOUN:Acc+Sing	noun	N;ACC;SG
gen.pl	NOUN:Gen+Plur	noun	N;GEN;PL
gen.sg	NOUN:Gen+Sing	noun	N;GEN;SG
dat.pl	NOUN:Dat+Plur	noun	N;DAT;SG
dat.sg	NOUN:Dat+Sing	noun	N;DAT;PL
abl.pl	NOUN:Abl+Plur	noun	N;ABL;PL
abl.sg	NOUN:Abl+Sing	noun	N;ABL;SG

Features-values

The features-values table MUST have at least a feature identifier value_id, a feature label (the full feature value in lowercase, e.g. nominative or past), and a feature dimension (e.g. case or tense). It MAY have more columns, including those described in the specification. Linking to other resources from this table is RECOMMENDED.

Feature identifiers are feature-values. They MUST be lowercase.

Feature values table for latin nouns

value_id	label	feature	POS	canonical_order
nom	nominative	case	noun	1
voc	vocative	case	noun	2
acc	accusative	case	noun	3
gen	genitive	case	noun	4
dat	dative	case	noun	5
abl	ablative	case	noun	6
sg	singular	case	noun	1
pl	plural	case	noun	2

While using case to distinguish feature values (eg. S for subject but s for singular) is not allowed, we recommend to provide mappings in the features table and cells table which provide mappings to other conventions, in particular those used in specific sources:

Feature values table with subject and singular

value_id	label	feature	POS	value-Author2020
sbj	subject	function	verb	S
sg	singular	number	noun	s

The meaning of these extra columns can be documented in the metadata.

Lexemes

The lexemes table MUST have at least a lexeme identifier lexeme_id. It documents any information which is valid for entire lexemes. It MAY have more columns, including those described in the specification. Additional columns MAY be added.

If a frequency column is present, the source of the frequencies MUST be documented in the README.md or accompanying documentation. If providing multiple frequency measurements, it is best to use a separate frequencies table.

In cases where a lexeme may take either of two inflection classes or stems, and has a full paradigm for each, it is preferable to divide it into two separate lexemes (or flexemes). This choice MUST be described in the documentation.

Two rows of a lexeme table for latin nouns

lexeme_id	label	inflection_class	POS	meaning	frequency
dominus	dominus	2	noun	master	10000
rosa	rosa	1	noun	rose	6000

Tags

The tags table MUST have at least a tag identifier tag_id, a reference to the column in which they occur (tag_column_name) and a comment. Tags are mainly used in the forms table but can also be used in the lexeme and frequency tables as needed. The standard pre-defines the tag column names analysis_tag, defectiveness_tag, epistemic_tag, variants_tag, overabundance_tag. More custom tag names can be added, and they MUST end in _tag.

It MAY have more columns, including those described in the specification.

A few rows of a possible tags table for English verbs

tag_id	tag_column_name	comment
uk_dialect	variants_tag	variant majoritarily used in UK English
us_dialect	variants_tag	variant majoritarily used in US English
irreg	overabundance_tag	irregular form
d-form	overabundance_tag	past in -ed pronounced /d/
t-form	overabundance_tag	past in -ed pronounced /t/

Note that one can not deduce information for rows which are not marked by tags. For example, if some rows of the forms table are marked as uk_dialect, some as us_dialect, and some are not marked, it is not possible to know where these unmarked forms are used. Thus, it is RECOMMENDED to be more explicit and always provide (and use) non-marked or default tags:

Tags including non-marked situations

tag_id	tag_column_name	comment
uk_specific	variants_tag	variant majoritarily used in UK English
us_specific	variants_tag	variant majoritarily used in US English
all_dialects	variants_tag	forms which are not specific to a particular English dialect

Frequencies

Sometimes, having frequencies in columns in the forms, lexeme, or even cells tables is not satisfactory. Possible reasons for this are:

• There are multiple sources of frequencies (either multiple corpora in which they were measured, or multiple sources providing frequencies directly) or methods through which they were measured.
• Frequencies were measured for units distinct from those that constitute rows in other tables: eg. combinations of which ignore orthographic or phonological variants documented in the forms table.

In that case, a dataset MAY provide a separate frequency table.

Frequencies tables MAY have a variety of shapes, as frequency can be measured or aggregated at various levels. They MUST link to relevant table(s) using at least one column with identifiers from another table. For example, this could be:

• Frequency of a lexeme (lexeme column linking to lexeme_id in the lexemes table)
• Frequency of a cell (cell column linking to cell_id in the cells table)
• Frequency of a specific form row (form column linking to form_id in the forms table)
• Frequency for a cell/lexeme combination, regardless of the phonological variant (two identifier columns),
• etc.

Each row MUST represent a single frequency value from a single source.

In this first example, frequencies are associated to specific rows from the forms table. Note that multiple sources for a same data point lead to separate rows:

Imaginary frequencies for possible variant forms, associated to forms

freq_id	form	value	source
9929	eat_prs_us	3888	source1
9930	eat_prs_uk	4900	source2
9931	eat_prs_uk	3000	source1

In the second example, frequencies are provided for lexeme/cell combinations (which might lead to multiple rows in the forms table):

Imaginary frequencies associated to lexemes and cells

freq_id	lexeme	cell	value	source
9929	eat	prs	3888	source1
9930	eat	prs	4900	source2

Note that for frequency, 0 means non-attested in the dataset used, whereas an empty cell means that the frequency for this form was never evaluated.

Sources

A sources file containing bibtex references MAY be added. The source columns of the tables MUST contain bibtex keys, which MUST be referenced in a sources.bib file.

The source columns are mostly used when specific data points (forms, lexemes, frequencies, etc) come from different sources, such that the dataset authors need to specify the source of each data point. They SHOULD NOT be used to provide the same single source for every data point (in that case the source is rather the source of the entire dataset and MUST be referenced in the JSON metadata).