🏷️ NUM: numeral

Universal Dependencies · 7 key papers · all concepts

Part of: 🗨️ Natural language processing · 🌐 Universal Dependencies

NUM is the Universal Dependencies part-of-speech tag for numerals. It covers words, digit strings and symbols that state a quantity or a number, typically cardinal numbers such as *three*, *17* or *XIV*. These usually modify a noun to say how many things there are.

Also: numeral, cardinal number, Num (as a lexical category in descriptive grammar)

What it is

In Universal Dependencies, NUM is one of seventeen universal part-of-speech tags. The inventory is closed: languages may refine relations with subtypes but may not add new basic tags (Nivre et al. 2020). In the guidelines' terms, a numeral expresses a number and answers the question "how many?". The tag is applied by function and form together. Cardinals written as words, as Arabic digits or as Roman numerals all receive NUM.

The category is deliberately narrow. Several kinds of number words go elsewhere:

• Ordinals such as first or 21st behave like adjectives and are tagged ADJ.
• Multiplicatives such as twice are ADV.
• Quantifiers that name no exact number (many, several, few) are usually DET.
• A numeral used as a noun phrase on its own, as in "one of them", may be tagged PRON in some treebanks, depending on language-specific practice.
• Nouns of quantity such as dozen or couple are usually NOUN.

The tag goes back to the compact universal tagset of Petrov, Das and McDonald (2012). NUM was one of its twelve coarse categories, defined pragmatically after the authors surveyed the guidelines of many existing corpora (Petrov et al. 2012). When UD v1 extended that set to seventeen tags, NUM was carried over unchanged (Nivre et al. 2016).

Examples

In English, a NUM word normally depends on the noun it counts through the nummod relation:

• She bought three apples — three (NUM) is attached to apples by nummod.
• The 1990s or in 2019 — the year is NUM. In a prepositional phrase it heads its own nominal and is attached as obl or nmod.
• two or three days — three is a conj of two.
• Chapter 4 — the digit modifies or follows the noun, and its attachment varies by treebank.

Typical morphological features (FEATS) for NUM words include:

These features descend from the Interset approach of Zeman (2008). Interset is a universal inventory of morphosyntactic features that serves as an interlingua when converting between tagsets (Zeman 2008). UD adopted such features as part of its morphological layer (Nivre et al. 2016).

Across languages

Numerals are a classic area of cross-linguistic variation, and UD handles it mostly through relations and features rather than through new tags.

• Slavic languages: from "five" upward, Russian and Czech numerals govern the genitive plural of the noun, as in пять книг 'five books'. UD marks this with the subtype nummod:gov, while agreeing numerals take plain nummod.
• Chinese and Japanese: a classifier stands between the numeral and the noun. In Chinese 三本书 'three books', 三 is NUM and the classifier 本 is linked by clf.
• French and other Romance languages: complex numerals such as vingt et un 'twenty-one' or deux mille 'two thousand' are split into several tokens. These are joined by flat or compound relations.
• Arabic and Hebrew: cardinals show gender polarity, so the numeral's gender is the opposite of the counted noun's.
• Spaces inside numbers: UD v2 allows a space inside a word only for numbers and abbreviations in languages that otherwise separate words with spaces. A form like French 10,000 can therefore remain a single NUM token (Nivre et al. 2020).

These choices follow UD's general design. Content words head the tree, and the relation label signals what kind of dependent each word is (Marneffe et al. 2021).

In treebanks and research

In both English and French treebanks, NUM accounts for well under two percent of tokens. nummod is by far its dominant relation. The next most frequent relations are nmod and obl, which cover dates, years, quantities and numbered references that head their own phrases. English shows NUM unusually often as root and conj, reflecting headings, list items and ranges. French makes more frequent use of flat for multiword numerals.

The tag matters for several kinds of work:

• Part-of-speech tagging: digit strings make NUM easy to recognise at the surface. Spelled-out numerals and borderline words such as one remain a source of disagreement.
• Cross-lingual transfer: Petrov et al. showed that coarse tags such as NUM support direct transfer of delexicalised parsers across languages (Petrov et al. 2012).
• Annotation consistency: inconsistencies in the tagging of numerals and quantifiers are among the issues that UD consistency checks aim to detect (Marneffe et al. 2017). Treebanks of the same language can also be compared through the divergence of their UPOS trigram distributions (Aggarwal & Zeman 2020).
• Downstream tasks: numerals in a dependency parsing output give information extraction and named entity recognition systems a direct route to quantities, dates and amounts.

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…
• 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-num/