🌐 nummod: numeric modifier
nummod (numeric modifier) is the Universal Dependencies relation that attaches a numeral to the nominal whose quantity it specifies, as in *three books*, where *three* modifies *books*. It covers cardinal numbers written in words or digits.
What it is
In Universal Dependencies (UD), every word of a sentence hangs from a head through a typed relation. nummod is the label for a numeral that says how many of something there are. The head is a nominal, usually a noun, sometimes a proper noun, pronoun or symbol. The dependent is a cardinal number, normally tagged NUM in part-of-speech tagging.
UD sorts dependents by the kind of phrase they are: nominals, clauses and modifiers. The relation label itself records the dependent's type, so the phrase category can be read off the tree (Marneffe et al. 2021). nummod belongs to the modifiers of nominals. Its siblings there are amod for adjectives, det for determiners and nmod for nominal modifiers.
The UD v2 inventory has 37 universal relation types. Languages may refine them with subtypes but may not add new basic categories (Nivre et al. 2020). The relation goes back to the num label of the universal Stanford scheme (Marneffe et al. 2014).
What it is told apart from
- det: quantifying words such as some, many or every are determiners, not numerals.
- amod: ordinals such as third are usually adjectives and take amod.
- compound or flat: the internal parts of a complex number (two hundred) attach to each other, and only the number's head attaches to the noun via nummod.
- nmod or obl: a number standing alone as a nominal (two of them, at five) is not a modifier.
Subtypes
The best-known subtype is nummod:gov. It marks numerals that govern the case of their noun, as in Slavic languages where a higher numeral requires the genitive. Determiner-like quantifiers with the same behaviour take the parallel det:numgov.
Examples
Arrows run from head to dependent:
She bought three books. books → three (nummod)
The ticket costs $ 20. $ → 20 (nummod)
We met on May 5. May → 5 (nummod)
In the first sentence the noun is the head. In the second, UD treats the currency symbol as the nominal head and the amount as its modifier. The third shows a date, where the day number modifies the month name.
A contrast case is Two of the boys left. Here Two is the nominal subject, and boys attaches to it.
Across languages
Counts of UD 2.18 show the same picture in English and French. Noun heads with a numeral dependent make up the great majority of nummod edges. Symbol heads (currency, percent, units) come a distant second. Proper-noun heads, mostly dates, come next, and the relation's share of all tokens is well under one percent in both languages.
Typology changes the picture elsewhere:
- Classifier languages: many treebanks for Chinese attach the numeral to the classifier rather than directly to the noun.
- Case-governing numerals: Russian, Czech or Polish use nummod:gov to flag numerals that impose case on the noun.
- Word segmentation: UD v2 allows a space inside a single word as an exception for numbers, so a digit group like 10,000 can stay one token (Nivre et al. 2020).
This variation is the kind of thing UD separates into constructions and the language-particular strategies that express them (Nivre 2025).
In treebanks and research
UD began with release v1.2, which covered 37 treebanks in 33 languages in November 2015 (Nivre et al. 2016). By v2.7 in November 2020 it had reached 183 treebanks for 104 languages (Marneffe et al. 2021). These treebanks supplied the data for the CoNLL 2017 and 2018 shared tasks on raw-text dependency parsing across many languages (Marneffe et al. 2021; Nivre et al. 2020).
For parsers, nummod is a short, mostly local edge. Errors tend to come from the boundaries listed above:
- numeral versus determiner;
- nummod versus compound inside complex numbers;
- dates and measure phrases with symbol heads.
These boundaries are also where cross-treebank consistency checks are informative (Marneffe et al. 2017).
Enhanced representations reshape quantity expressions for downstream use. In English enhanced++, partitives and light-noun constructions take the content noun as head, and the quantifying part is attached with det:qmod (Schuster & Manning 2016).
The principles and tooling for annotating such relations are surveyed in tutorial form by (Zeman 2021). Further readings are listed under Universal Dependencies and syntax.
What nummod 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 → NUM | 5,680 |
| English: SYM → NUM | 598 |
| English: PROPN → NUM | 154 |
| English: ADV → NUM | 98 |
| English: NUM → NUM | 66 |
| French: NOUN → NUM | 6,299 |
| French: SYM → NUM | 590 |
| French: PROPN → NUM | 90 |
| French: VERB → NUM | 70 |
| French: NUM → NUM | 49 |
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-nummod/