🏷️ ADV: adverb
ADV (adverb) is the universal part-of-speech tag in Universal Dependencies for words that typically modify verbs, adjectives, other adverbs or whole clauses. They express time, place, manner, degree, direction or the speaker's stance, and they usually do not inflect for case or agreement.
What it is
In part-of-speech tagging, ADV marks words such as quickly, very, here, yesterday and why. These words add circumstances or degree to something else in the sentence. The guidelines of Universal Dependencies define the class mainly by function. An adverb is a modifier of a predicate, a property word or another modifier, and it does not head a nominal phrase.
UD sorts phrasal units into nominals, clauses and modifiers (Marneffe et al. 2021). Adverbs are the most typical members of the modifier type.
ADV was already one of the twelve coarse categories in the universal tagset of Petrov, Das and McDonald (Petrov et al. 2012). UD extended that set to 17 tags (Nivre et al. 2016), and the list is closed in v2 (Nivre et al. 2020). The adverb class itself carried over largely unchanged.
Borderline cases
The tag has several neighbours, and the boundary with each is drawn differently:
- ADV vs ADJ: fast in runs fast is ADV, but in a fast car it is ADJ.
- ADV vs ADP: English verb particles (give up) are usually ADP, attached as compound:prt. Upstairs, which takes no object, is ADV.
- ADV vs SCONJ: when and where are ADV when they modify the verb of their own clause. A pure subordinator such as because is SCONJ.
- ADV vs PART: English treebanks normally tag not as PART. Many other languages tag their negator as ADV.
- ADV vs PRON/DET: pronominal adverbs (there, somehow, nowhere) stay ADV but carry a PronType feature.
Features
Typical features in the morphology layer include:
- Degree=Cmp or Degree=Sup for faster and most.
- PronType=Int, Rel, Dem, Ind, Neg or Tot for the pronominal series.
- NumType=Mult or Ord for twice and firstly.
- Polarity=Neg where the negator is an adverb.
Many adverbs carry no features at all. The feature inventory derives from Interset (Nivre et al. 2016), Zeman's interlingua of morphosyntactic values (Zeman 2008).
Examples
- English: She almost (ADV) never (ADV) arrives late (ADV). Each adverb attaches by advmod. Almost modifies never, while never and late modify arrives.
- French: Il parle très bien. Here très depends on bien, which depends on parle.
- German: Wir fahren morgen dorthin. Both morgen and dorthin are ADV.
- Czech: rychleji "faster" is ADV with Degree=Cmp.
Across languages
Languages draw the line between adjective and adverb in different places. German and Dutch often use the bare adjective form adverbially. Treebanks then must decide by syntactic position whether to tag ADJ or ADV.
Romance and Slavic languages derive manner adverbs productively, for example with French -ment or Czech -e/-y. English does the same with -ly.
Languages with rich pronominal series, such as Slavic kde/tam/nikde "where/there/nowhere", lean heavily on PronType.
Multiword adverbials such as French tout à fait "entirely" or English of course are split into words. Their internal links are marked with fixed.
In treebanks and research
In English and French UD data, advmod is overwhelmingly the dominant relation for ADV.
- In English, the smaller shares go to root (adverbs heading fragments such as Now?), compound, obl and conj.
- In French, obl comes second, followed by mark and fixed. This reflects adverbs that introduce clauses and adverbs inside fixed multiword expressions.
Adverbs also make up a noticeably larger share of tokens in French than in English. See the English and French topic pages for more.
Moving from fine-grained tagsets to the coarse universal one costs little for downstream use. An English parser using 12 tags instead of the 45 Penn Treebank tags lost only about 0.6% accuracy (Petrov et al. 2012). Even so, the fuzzy ADV boundaries are a known source of disagreement between treebanks of one language (Marneffe et al. 2017).
Aggarwal and Zeman compare treebanks of the same language through their UPOS trigram distributions (Aggarwal & Zeman 2020):
θpos(A,B) = KLcpos3(A,B) + KLcpos3(B,A)
In this formula, A and B are two treebanks. KLcpos3 is the asymmetric Kullback–Leibler divergence between their distributions of UPOS trigrams, and summing both directions makes θpos symmetric. Values at or below 0.5 indicate consistent annotation, and values at or above 4.0 indicate inconsistent annotation (Aggarwal & Zeman 2020). A treebank that tags not or verb particles differently will shift such trigrams.
For dependency parsing, the near-deterministic pairing of ADV with advmod makes the tag a strong cue. Cross-lingual transfer exploits this cue, as in the delexicalized parsing experiments that first motivated the universal tagset (Petrov et al. 2012).
The relations ADV words take most often
The five most frequent in each language's treebanks, by tokens.
The numbers
| Pair | Tokens |
|---|---|
| English: advmod | 46,798 |
| English: root | 2,524 |
| English: compound | 972 |
| English: obl | 826 |
| English: conj | 749 |
| French: advmod | 53,184 |
| French: obl | 5,171 |
| French: mark | 2,043 |
| French: fixed | 1,378 |
| French: conj | 776 |
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…
- Part-of-speech tagging — Part-of-speech tagging is the task of assigning each token in a text a grammatical category, such as noun, verb, adjective or determiner…
- Morphology — Morphology is the branch of linguistics that studies the internal structure of words. It covers how words are built from meaningful units…
More papers
Topics: 🔤 Part-of-speech tagging · 🌐 Universal Dependencies
Shelves: 🗣️ Language and Linguistics
This page in Markdown → https://md.research.elamur.ai/free/en/concepts/upos-adv/