🌐 compound: compound
compound is the Universal Dependencies relation that links two or more words forming a single lexical unit, such as a noun–noun compound, a verb with its particle, or a serial verb, when that unit is written as separate words and none of the parts is a modifier, argument or function word of the others.
What it is
In Universal Dependencies (UD), the word is the basic unit of annotation, following the principle of lexical integrity (Marneffe et al. 2021). Many languages build complex lexical items from several words, and the spaces between them do not make the items any less word-like. The compound relation records this situation. It says that two syntactic words together form one lexeme-like unit, rather than one word governing a dependent phrase.
The head is the part that carries the category and the core meaning of the whole. The dependent is the part that narrows it. In English noun compounds the head is normally the rightmost noun, so in phone book the dependent phone attaches to book. The whole compound then takes its outer relation (nsubj, obj, obl and so on) from the head.
compound is one of the 37 universal relation types of UD v2. Languages may refine these types with subtypes but may not invent new basic categories (Nivre et al. 2020). It replaced the narrower nn label of earlier Stanford-style annotation. The cross-linguistic inventory itself came out of universal Stanford dependencies (Marneffe et al. 2014; Nivre et al. 2016).
Neighbouring relations
compound is one of three multiword relations in UD, and the three are kept apart:
| Relation | Internal structure | Typical cases |
|---|---|---|
| compound | headed; head is the semantic core | ice cream, pick up |
| flat | headless; first word is a formal head | names (Marie Curie), dates |
| fixed | grammaticalized function expression | because of, as well as |
compound is also distinct from amod and nmod. A noun that is a genuine modifying phrase, often marked by case or an adposition, takes nmod rather than compound. An adjective that remains a free modifier takes amod.
Common subtypes
- compound:prt links a verbal particle to its verb (English give up, German separable prefixes).
- compound:svc links verbs in a serial verb construction.
- compound:lvc links the nominal or other element of a light verb construction to the light verb.
- compound:redup marks reduplicated forms.
Examples
She bought a phone book. compound(book, phone)
They finally gave up the plan. compound:prt(gave, up)
The customs office opened late. compound(office, customs)
In the first and third sentences the dependent noun narrows the reference of the head noun without being a separate argument. In the second, up is not an adposition introducing a phrase: it belongs to the verb's lexical meaning, so it attaches as compound:prt. Compare They walked up the hill, where up introduces an oblique phrase and is marked case. Such phrasal verbs are a classic topic of multiword expressions.
Across languages
How often compound appears depends heavily on orthography and morphology.
In English, compounds are usually written with spaces, so the relation is frequent. Noun-on-noun attachments dominate its use. Proper-noun sequences and nouns attached to proper names come next. Verb-to-adposition links are also very common, and they reflect particles labelled compound:prt.
In French, the relation is rare. Most lexicalized combinations either take a preposition (machine à laver) and are analysed with nmod, or are hyphenated and treated as single tokens. What remains is almost entirely noun–noun.
German, Dutch and the Nordic languages write most nominal compounds as one orthographic word, which leaves little for the relation to mark. Their separable verb prefixes, however, use compound:prt. Serial verb languages of West Africa and Southeast Asia make heavy use of compound:svc. Because subtypes can be added per language, language-specific facts can be captured while the universal label stays comparable (Nivre et al. 2020).
In treebanks and research
The boundary between compound, flat, fixed and nmod is one of the less stable parts of UD annotation. Treebanks of the same language can label similar sequences differently. Work on annotation consistency has examined such variation across UD corpora (Marneffe et al. 2017). Tutorials on UD principles treat the multiword relations as a recurring source of annotator questions (Zeman 2021). Recent work on constructions and strategies in UD gives a framework for deciding which grammatical strategy a language uses for a given construction (Nivre 2025).
For dependency parsing, compound matters because long noun sequences are a classic source of attachment errors. UD treebanks underlay the CoNLL shared tasks of 2017 and 2018 on parsing raw text in many languages (Marneffe et al. 2021; Nivre et al. 2020). Applications that extract relations between content words often prefer graph representations. The enhanced English UD converters, for example, refine labels and treat certain multiword units as single nodes so that content words are linked directly (Schuster & Manning 2016). Recognizing a compound as one unit therefore helps downstream tasks such as information extraction and term detection in treebanks and raw text alike.
What compound joins: the most frequent head tag → dependent tag
The five most frequent in each language's treebanks, by tokens.
The numbers
| Pair | Tokens |
|---|---|
| English: NOUN → NOUN | 16,872 |
| English: PROPN → PROPN | 3,821 |
| English: VERB → ADP | 3,690 |
| English: NOUN → PROPN | 3,121 |
| English: VERB → ADV | 904 |
| French: NOUN → NOUN | 133 |
| French: ADP → NOUN | 3 |
| French: NOUN → ADP | 3 |
| French: PROPN → PROPN | 3 |
| French: NOUN → ADJ | 2 |
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-compound/