🌐 advmod: adverbial modifier

Universal Dependencies · 6 key papers · all concepts

Part of: 🗨️ Natural language processing · 🌐 Universal Dependencies

advmod (adverbial modifier) is the Universal Dependencies relation linking a predicate or modifier word to an adverb or adverbial phrase that modifies it, adding manner, degree, time, place, frequency or negation without being a clause or a nominal.

Also: adverbial modifier, adverb modifier, advmod:emph, advmod:tmod, advmod:neg

What it is

In Universal Dependencies, every word except the root attaches to one head by a typed relation. advmod is the label for one common case. The dependent is an adverb, or a small phrase headed by one, that refines the meaning of its head. In She runs quickly, the word quickly depends on runs as advmod.

The head is usually a verb, but it can also be an adjective (very tall), another adverb (quite slowly) or, less often, a noun or numeral (almost everyone, about ten).

UD describes syntax in terms of three kinds of units: nominals, clauses and modifiers. The relation label tells which kind of unit the dependent is (Marneffe et al. 2021). advmod is the modifier relation for adverbial words. It sits beside amod for adjectival modifiers and beside nmod and obl for nominal dependents.

The first version of the guidelines already separated phrasal from clausal dependents in pairs such as advmod and advcl (Nivre et al. 2016). Version 2 keeps a fixed inventory of 37 universal relations. Languages may refine these with subtypes but may not invent new basic ones (Nivre et al. 2020). The label goes back to the cross-linguistic typology proposed as Universal Stanford dependencies (Marneffe et al. 2014).

Neighbouring relations

Negation is a special case. In v2, words like not and never are attached as advmod rather than by a separate negation relation. Some treebanks mark them as advmod:neg.

Other common subtypes are:

• advmod:emph, for emphasizing or intensifying words
• advmod:tmod, for temporal adverbials
• advmod:lmod, for locative adverbials

Because advmod is never a core argument, it falls on the non-core side of the argument split discussed for UD by Zeman (Zeman 2017).

Examples

Arrows run from head to dependent:

She answered (→ politely)         advmod(answered, politely)
a (→ remarkably) clear proof      advmod(clear, remarkably)
We did (→ not) see it             advmod(see, not)

In the first line, politely gives the manner of the verb. In the second, remarkably sets the degree of the adjective. In the third, the negator attaches directly to the content verb see, not to the auxiliary did. This follows the content-word-as-head principle of UD (Marneffe et al. 2021).

A degree adverb may in turn modify another adverb, as in very → slowly. In that case advmod links adverb to adverb.

Across languages

The relation is defined by function, so its dependents need not carry the part-of-speech tag ADV.

In English treebanks, the dominant pattern is a verb heading an adverb. The second most frequent pattern is a verb heading a particle, because English not is tagged PART. Adjective, noun and adverb heads follow at a distance.

French treebanks are similar but even more verb-centred. A small residue of adpositions is used adverbially, without an object. In both languages, roughly one token in twenty carries the label.

Languages with rich case systems often express adverbial meanings through case-marked nouns, which UD labels obl rather than advmod. As a result, the share of advmod in a treebank partly reflects how a language divides labour between adverbs and nominals. Nivre frames such contrasts as different strategies serving one construction (Nivre 2025). Clauses whose main content is not a verb bring their own questions of where adverbials attach (Nivre et al. 2026).

In treebanks and research

advmod has belonged to the inventory since release v1.2 (Nivre et al. 2016). It is used in every UD treebank, which made it part of the shared benchmark behind the CoNLL 2017 and 2018 multilingual parsing tasks (Marneffe et al. 2021). In dependency parsing evaluation, it is one of the labels behind labelled attachment scores.

The relation is hard to get right in two ways. Choosing the head is tricky: the same adverb may plausibly modify a verb, an adjective or a whole clause. Choosing the label is also tricky: the boundaries with obl, discourse and advcl are not always sharp.

Such borderline choices are among the issues addressed by work assessing annotation consistency across UD corpora (Marneffe et al. 2017). Tutorials on UD principles and tools use relations like this one to show how the guidelines are applied in practice (Zeman 2021).

In the enhanced representation for English, the basic tree is kept and extra edges are added. Shared dependents spread to all conjuncts, so an adverb modifying coordinated verbs can be linked to each of them (Schuster & Manning 2016). For downstream tasks such as information extraction and semantic role labeling, advmod edges help separate modifiers of manner, time and negation from core arguments.

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-advmod/