CoNLL-C field reference

A deeper, field-by-field reference for the .conllc format introduced in The CoNLL-C format, sourced from the project’s wiki. The format has 13 fields: ID, FORM, LEMMA, UPOS, FEATS, HEAD, DEPREL, REQUIRED, EXCLUSION, SEM_FEATS, SEM_ROLES, ADJACENCY, IDENTITY.

LEMMA

For morphological constructions, the lemma at the word level is expressed as a regular expression, while the lemma(s) for sub-word tokens are the morphemes themselves. Regexes should use the constructs Grew requests accept. Allomorphy can sometimes be captured this way — e.g. the prefix in- ‘in, un’ has four allomorphs (in-, im-, ir-, il-), expressible as a single regex:

ID   UD.FORM        LEMMA                     UPOS     ...
A       _       in.+|im.+|irr.+|ill.+         ADJ      ...
A-1     _           in                       BMORPH    ...
A-2     _            _                        ADJ      ...

UPOS

For word-level tokens, values come from the Universal part-of-speech tags. For sub-word tokens, there are two possibilities: elements that also exist as free lexemes keep their normal part of speech; bound forms (affixes, combining forms, affixoids) are tagged BMORPH (“Bound MORPHeme”). In a morphological construction, the word-level token still carries its own UPOS, so the output category of the morphological process stays specified. Two more examples beyond the semi- + ADJ case already shown:

Noun+Noun compounding with capo:

ID   UD.FORM     LEMMA     UPOS   ...
A       _       capo.+     NOUN   ...
A-1     _        capo      NOUN   ...
A-2     _         _        NOUN   ...

Neoclassical compounding (two combining forms, one lexically fixed):

ID   UD.FORM     LEMMA     UPOS     ...
A       _       .+logia    NOUN     ...
A-1     _         _        BMORPH   ...
A-2     _        logia     BMORPH   ...

FEATS

A list of lexical types and morphosyntactic features from the universal feature inventory or the Italian-specific extension, used to express morphosyntactic/lexical constraints on open slots, or morphosyntactic constraints on lexically-filled slots left underspecified for form. This field does not annotate every applicable feature the way a UD treebank would — only what’s needed to constrain matching. Beyond the standard inventory, two extensions:

Animacy and Aspect are used here to express semantic constraints even though Italian corpora don’t generally annotate them — for now this only expresses the constraint formally, without helping the actual matching process.
Definite is extended beyond the Italian UD tagset’s definite/indefinite to also allow specific indefinite, when needed.

HEAD

The ID of the syntactic head, or 0 for the construction’s root. In morphological constructions, the full word is, by convention, the head of the root sub-word element (subword and above-word relations form two separate graph layers).

DEPREL

Values come from Universal Dependency Relations and their Italian-specific subtypes. ItCon generally follows the annotation found in UD-annotated Italian treebanks, even where a different theoretical analysis might be preferred (mostly for chunks/MWEs) — theoretical considerations belong in the entry itself, while CoNLL-C’s job is matching UD-annotated patterns. Below the word level, four additional relations apply:

root/m — the root within the morphological construction (the stem in derivation, the head of the compound in compounding)
der/m — links a derivational affix to the stem
case/m — links the complement to the head in subordinating compounds (e.g. capostazione ‘station master’)
mod/m — links the attribute to the head in attributive compounds (e.g. altopiano ‘upland’)
conj/m — links the second constituent to the first (the head) in coordinating compounds (e.g. cartongesso ‘drywall’)

REQUIRED

Encodes whether a token must be obligatorily expressed (1) or can be omitted (0) — used to capture formal variants like subject omission in the matching process. For example, in the Passive Construction with venire, both the subject (Patient) and the da N prepositional phrase (Agent) are optional:

ID    UD.FORM   LEMMA     UPOS                 ...    HEAD   DEPREL       REQUIRED
A        _       _        NOUN, PROPN, PRON    ...    C      nsubj:pass   0
B        _       venire   AUX                  ...    C      aux:pass     1
C        _       _        VERB                 ...    0      root         1
D        _       da       ADP                  ...    E      case         0
E        _       _        NOUN, PROPN, PRON    ...    C      obl          0

EXCLUSION

Beyond the CHILDREN:DEPREL=value pattern shown in the main guide, EXCLUSION supports excluding specific values from the token’s own fields directly, and combining more than one constraint with a pipe (|):

COLUMN_NAME=value1,value2|COLUMN_NAME2=value1

For example, to exclude zero and uno from the open numeral slot of Num N in croce ‘barely/only Num of Ns’:

ID    UD.FORM    LEMMA     UPOS    ...   REQUIRED    EXCLUSION
A       _         _        NUM           1           LEMMA=zero,uno
B       _         _        NOUN          0           _
C      in         in       ADP           1           _
D      croce      croce    NOUN          1           _

The CHILDREN: prefix constrains a token’s children rather than the token itself — e.g. an intransitive verb must have no obj child, or (using a regex as the value) a parenthetical verb like sembra must have no children of any kind:

ID    UD.FORM    LEMMA      UPOS    ...   REQUIRED    EXCLUSION
A       _        _          VERB    ...    1          _
B       _        sembrare   VERB    ...    1          CHILDREN:DEPREL=.+

SEM_FEATS

The semantic counterpart of FEATS: constraints on open slots not covered by morphosyntactic features. Since there’s no semantically annotated Italian corpus, these constraints are descriptive — they don’t currently drive the actual matching process. See Semantic features for the full tagset (OntoClass from Open Multilingual Wordnet Topics, Aktionsart from UniMorph). Formally identical to FEATS, including the pipe (|) for combining multiple features on one token:

ID    UD.FORM    LEMMA      UPOS    ...   SEM_FEATS
A       _        fare       VERB    ...    _
B       _        _          NOUN    ...    OntoClass=feeling

SEM_ROLES

Semantic roles of participants, mainly in argument structure constructions, annotated on both open and filled slots. Like SEM_FEATS, this doesn’t currently drive matching — it’s there to later annotate the roles of matched argument-structure constructions in corpora. Based on an adapted version of the UVI (Unified Verb Index) role hierarchy — see Semantic roles for the full tagset and its mapping to MoCCa comparative concepts.

ADJACENCY

Constrains linear adjacency: when no element can intervene between a token and its left-adjacent neighbor, the field holds the ID of that left-adjacent token. Also annotated in morphological constructions. E.g. in Num N in croce, neither the noun-in nor the in-croce boundary can have anything inserted:

ID    UD.FORM    LEMMA     UPOS     ...    ADJACENCY
A       _         _        NUM      ...    _
B       _         _        NOUN     ...    _
C      in         in       ADP      ...    B
D      croce      croce    NOUN     ...    C

IDENTITY

Annotates coindexation. Since coindexation can apply at different levels (a whole field, or one key inside FEATS), the value is COLUMN_NAME:token_ID — or FEATS.key:token_ID for a single morphosyntactic feature (see the agreement example in the main CoNLL-C format page). E.g. in the discontinuous reduplication construction Noun non Noun ‘not properly a Noun’, the second noun must match the first noun’s form:

ID    UD.FORM    LEMMA     UPOS     ...    IDENTITY
A       _         _        NOUN     ...    _
B      non       non       ADV      ...    _
C       _         _        NOUN     ...    FORM:A

Metadata

Construction-level information sits in #-prefixed comment lines above the token rows. Every construction needs at least cxn_id, name, and function; horizontal_links and vertical_links are optional.

cxn_id — matches the entry’s ID.
name — matches the entry’s name.
function — a concise natural-language paraphrase of the construction’s meaning (narrower in scope than the entry’s Definition, which can also cover pragmatics/information structure). Can reference specific tokens with ref:token_ID:
```
# cxn_id = 1
# name = fare una PROPN-ata
# function = ref:A does something typical of ref:D-1
```
horizontal_links / vertical_links — same information as the entry fields, space-separated IDs:
```
# horizontal_links = 1 34 39
# vertical_links = 10
```

Holistic properties

Sometimes a multiword expression behaves, as a whole, differently from what its parts would predict — e.g. giorno dopo giorno ‘day after day’ is two nouns and a preposition, but functions as an adverb (advmod) in the sentence. Since UD’s lexicalist approach can’t assign a UPOS/DEPREL to a multi-token span, this information is instead recorded as extra metadata lines:

# UPOS = ADV
# DEPREL = advmod

For morphological constructions this isn’t an issue, since the full word already has its own fields available for this purpose.

Formal variation: worked example

As mentioned in Formal variation in CoNLL-C, when formal variants can’t be captured with a regex or a disjunction, ItCon adds more than one graph to the same entry, suffixing the ID (a, b, …). For instance, the semi- + ADJ construction (entry 169) can appear bonded or hyphenated — since a single graph can’t represent both a morphological and a multiword form, the file gets two # cxn_id blocks:

# cxn_id = 169
# name = semiADJ
# function = not fully or not properly ref:A-2
# ...

ID   UD.FORM     LEMMA     UPOS   ...
A       _        semi.+     ADJ   ...
A-1     _        semi     BMORPH  ...
A-2     _         _         ADJ   ...

# cxn_id = 169b
# name = semiADJ
# function = not fully or not properly ref:C
# ...

ID   UD.FORM     LEMMA     UPOS   ...
A      semi      semi       ADJ   ...
B       _         -        PUNCT  ...
C       _         _         ADJ   ...