🌐 conj: conjunct

Universal Dependencies · 6 key papers · all concepts

Part of: 🗨️ Natural language processing · 🌐 Universal Dependencies

conj (conjunct) is the Universal Dependencies relation that links two elements standing in a coordinate structure, such as two nouns joined by "and". The first conjunct is the head, and each later conjunct depends on it.

Also: conjunct, conj:and, conj:or (enhanced subtypes)

What it is

In Universal Dependencies, coordination joins two or more elements of equal grammatical rank. The relation conj records this tie. Its head is always the first conjunct in linear order. Every later conjunct attaches directly to that first one, not to the conjunct just before it. Any coordinating conjunction ("and", "or", "but") hangs from the conjunct that follows it, under the separate relation cc.

The relation belongs to the closed set of universal relation types, which numbered 40 in version 1 (Nivre et al. 2016). In version 2 the set was cut to 37. Languages may add subtypes to these but may not invent new basic categories (Nivre et al. 2020). The design goes back to the universal Stanford dependencies (Marneffe et al. 2014). The label for the conjunction word was changed along the way: the part-of-speech tag CONJ became CCONJ in v2 (Nivre et al. 2020).

Because only the first conjunct is linked to the rest of the sentence, the basic tree shows the clause's predicate or argument once. The other conjuncts do not receive that link. The enhanced representation repairs this by adding implicit relations for conjuncts, so shared material reaches every member of the coordination (Marneffe et al. 2021).

Neighbouring relations

Subtypes

Basic trees rarely subtype conj. In enhanced English graphs, the conjunction is copied into the label, giving conj:and or conj:or. Governors and dependents shared by the coordination are also spread to all conjuncts (Schuster & Manning 2016).

Examples

In each example, the head comes first and the dependent second.

• "Anna sings and dances."
  • conj(sings, dances)
  • cc(dances, and)
  • In the enhanced graph, "Anna" is also the nsubj of "dances".
• "We bought apples, pears or plums."
  • conj(apples, pears)
  • conj(apples, plums)
  • cc(plums, or)
  • Both later conjuncts hang from "apples".
• "The flat is small but cheap."
  • conj(small, cheap)
  • cc(cheap, but)
  • In the enhanced graph, the relation becomes conj:but.

Coordination without any conjunction, as in "came, saw, conquered", still uses conj. The commas attach as punctuation.

Across languages

The relation is defined the same way everywhere, but how often it appears and what it joins depend on the language. In the English treebanks of UD 2.18, conj makes up about three percent of all tokens. Two pairs dominate almost equally: noun with noun, and verb with verb. These are followed by proper-noun pairs, adjective pairs and numeral pairs.

French shows a slightly higher share. There, noun–noun coordination clearly outweighs verb–verb coordination. Mixed proper-noun-plus-noun pairs are also frequent enough to rank among the top pairs. Such contrasts reflect both syntax and genre, for example news versus fiction. They can be compared for English and French because the scheme is harmonized.

Typologically, languages differ in their coordination strategies:

• they may use a single conjunction, a conjunction on every member, or plain juxtaposition;
• some use comitative constructions ("I with you") where English uses "and";
• this raises the question of whether such constructions count as conj or as oblique modifiers.

UD's treatment of constructions versus the strategies languages use to express them has been examined directly (Nivre 2025). Its underlying functional-typological theory is described in (Marneffe et al. 2021).

In treebanks and research

Coordination is a classic weak spot of dependency parsing. The parser must decide both the scope of the coordination and which element is the first conjunct. In UD's design, the leftmost-head choice makes these decisions visible as a single attachment.

Because the guidelines are shared, conj annotations from many treebanks feed:

• multilingual parsers trained on the same data;
• the CoNLL shared tasks of 2017 and 2018 (Marneffe et al. 2021; Nivre et al. 2020);
• cross-lingual transfer.

Consistency across corpora is not automatic. Automatic checks of UD annotation have been proposed to find such disagreements (Marneffe et al. 2017). The guidelines and the tools for them are taught as a unit, as in a tutorial on UD principles and tools (Zeman 2021).

For information extraction and other understanding tasks, the enhanced layer matters most. Propagating subjects and objects across conjuncts gives a direct edge between content words that a surface tree lacks. The labels conj:and and conj:or also keep the type of coordination, which downstream systems can use (Schuster & Manning 2016). Converters that build these enhanced graphs from basic trees are distributed with Stanford CoreNLP (Schuster & Manning 2016).

Key papers

The works this page cites, the most cited first; citations as counted by OpenAlex plus the papers of this library that cite the work, read 2026-10-24.

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-conj/