🌐 cc: coordinating conjunction
cc (coordinating conjunction) is the Universal Dependencies relation that attaches a coordinating word, such as English and, or, but, to a conjunct of the coordination it introduces. In v2 that conjunct is normally the one immediately following the conjunction.
What it is
In Universal Dependencies, the words of a sentence form a tree of typed relations. Content words are heads, and function words hang from them (Marneffe et al. 2021). Coordination uses two relations.
- conj links the first conjunct to each later conjunct.
- cc links the linking word to a conjunct.
The dependent of cc is the coordinator itself. It is usually tagged CCONJ, a tag that version 2 of the guidelines adopted in place of the older CONJ (Nivre et al. 2020). The head is the conjunct that the coordinator introduces. In UD v2 this is the following conjunct, not the first one. Version 1 attached the coordinator to the first conjunct, and the change keeps the conjunction inside the phrase it actually belongs to.
The relation descends from the cross-linguistic typology of universal Stanford dependencies (Marneffe et al. 2014). That typology was one of the strands merged into UD v1 (Nivre et al. 2016). cc is one of the 37 universal relation types. Like all of them, it may take language-specific subtypes but cannot be replaced by new basic categories (Nivre et al. 2020).
Neighbouring relations
| Relation | What it marks | Typical dependent |
|---|---|---|
| cc | coordinator introducing a conjunct | and, or, but |
| cc:preconj | first half of a paired coordinator | both, either, neither |
| conj | a later conjunct, attached to the first | the conjunct's head word |
| mark | subordinator introducing a clause | that, because, if |
| advmod / discourse | connective adverbs and particles | however, so, well |
The main boundary is with mark. A subordinator makes one clause dependent on another, while a coordinator joins equals. A second boundary separates cc from adverbs such as however, which also link sentences but are not coordinators. The one common subtype is cc:preconj, used for the first word of correlative pairs.
Examples
Arrows run from head to dependent.
- Tea and coffee were served.
- conj(Tea → coffee)
- cc(coffee → and)
- She sang, but nobody listened.
- conj(sang → listened)
- cc(listened → but)
- Either Anna or Ben will drive.
- cc:preconj(Anna → either)
- conj(Anna → Ben)
- cc(Ben → or)
With three or more conjuncts, as in red, white and blue, every conjunct attaches to red by conj. Only blue carries the cc dependent. A coordinator that opens a sentence, as in And then it rained, attaches to the main predicate with no visible first conjunct.
Across languages
In the English and French treebanks, cc most often depends on a verb. Noun heads come second, then adjectives and proper nouns. The relation makes up a somewhat larger share of tokens in French than in English. The fifth most frequent head category also differs: adverbs in English, pronouns in French.
Further differences follow from the word-based design of UD. Lexical integrity makes the syntactic word the unit of annotation, and clitics are split off even where spelling joins them (Marneffe et al. 2021; Nivre et al. 2016). Enclitic coordinators therefore become separate cc tokens, as with Latin -que or the prefixed w- of Hebrew and Arabic.
Other patterns leave no cc at all. Asyndetic coordination simply chains conj edges. Languages that coordinate nouns with a comitative word must decide whether that word is cc or a case marker.
Framing UD analysis in terms of constructions and the strategies that express them gives a principled way to make such decisions (Nivre 2025). Checking whether treebanks actually apply these labels consistently is a known concern of the project (Marneffe et al. 2017; Zeman 2021).
In treebanks and research
Parsing
For dependency parsing, cc is usually an easy edge to predict. The difficult part is coordination as a whole: finding where each conjunct begins and which word heads it. UD treebanks supplied the data for the CoNLL shared tasks of 2017 and 2018, in which systems parsed raw text in many languages (Marneffe et al. 2021; Nivre et al. 2020). Coordination was evaluated there like any other construction.
Enhanced representations
The basic tree encodes coordination asymmetrically. Only the first conjunct is linked directly to the shared governor, so a subject or object shared by all conjuncts appears only once. The enhanced English representation addresses this in two ways (Schuster & Manning 2016):
- It folds the coordinator into the conjunct label, giving conj:and or conj:or.
- It copies shared governors and dependents to every conjunct.
These steps give a direct path between content words, which applications in natural language understanding can use. The enhanced graph's handling of conjuncts is one of the main additions UD makes on top of the basic tree (Marneffe et al. 2021).
Meaning of the coordinator
The conjunction a cc edge points to matters for meaning, not just structure. Coordination is common: about 38.8% of Penn Treebank sentences contain and, or or but (Saha et al. 2020).
The ConjNLI stress test builds inference pairs by deleting, inserting or swapping conjuncts. Models fine-tuned on MNLI score poorly on it (Saha et al. 2020):
- Or is the hardest conjunction to reason about, and but the easiest.
- Non-boolean readings, such as collective sums, cost every model around ten accuracy points compared with boolean ones.
Adding predicate structure from semantic role labeling helped only a little (Saha et al. 2020). A correctly parsed cc edge is therefore necessary for handling coordination, but it is not enough.
What cc joins: the most frequent head tag → dependent tag
The five most frequent in each language's treebanks, by tokens.
The numbers
| Pair | Tokens |
|---|---|
| English: VERB → CCONJ | 13,371 |
| English: NOUN → CCONJ | 9,090 |
| English: ADJ → CCONJ | 2,868 |
| English: PROPN → CCONJ | 2,496 |
| English: ADV → CCONJ | 728 |
| French: VERB → CCONJ | 20,912 |
| French: NOUN → CCONJ | 11,984 |
| French: ADJ → CCONJ | 2,857 |
| French: PROPN → CCONJ | 2,348 |
| French: PRON → CCONJ | 1,012 |
Key papers
Related concepts
- Universal Dependencies — Universal Dependencies is a framework and an open collection of treebanks that annotates the grammar of many languages with one shared…
- Dependency parsing — Dependency parsing is the task of recovering a sentence's syntactic structure as directed, labelled links between its words. Each word is…
More papers
Topics: 🌐 Universal Dependencies · 🌳 Dependency parsing
Shelves: 🗣️ Language and Linguistics
This page in Markdown → https://md.research.elamur.ai/free/en/concepts/ud-cc/