🌐 clf: classifier
clf (classifier) is the Universal Dependencies relation that attaches a classifier to the phrase it serves. A classifier is a word that picks out a class of nouns, by shape, animacy or kind, and is required with numerals or demonstratives in many East and Southeast Asian languages.
What it is
Many languages cannot put a number directly before a noun. Mandarin does not say "three book". It says "three běn book", where běn is a classifier for bound, flat objects. Universal Dependencies (UD) is a cross-linguistic scheme for annotating syntax. It labels the link between such a classifier and the rest of the noun phrase clf.
UD builds trees in which content words are heads and function words depend on them (Marneffe et al. 2021). Classifiers sit between the two. They carry some lexical meaning, yet they serve mainly to make counting or pointing grammatical.
The UD v2 guidelines describe how such a word attaches. The classifier depends on the numeral or demonstrative it combines with, and that word in turn attaches to the noun as nummod or det. Where no numeral is present, the classifier is usually attached to the noun itself. The relation belongs to the closed set of 37 universal relation types in v2. Languages may refine any of these types with subtypes but may not add new basic types (Nivre et al. 2020). No clf:subtype is in wide use across treebanks.
Neighbouring relations
clf has to be told apart from several relations nearby:
nummodlinks the numeral itself to the noun.detlinks a demonstrative or article to the noun.nmodcovers measure nouns in languages without classifiers, as in English "a cup of tea".compoundandflatcover multiword names.
English quantity phrases such as "a bunch of people" are a separate case. In the enhanced++ representation they are restructured, and the quantifying part attaches by det:qmod (Schuster & Manning 2016). That is a different solution from clf.
Examples
Each example marks the relation as clf(head, dependent).
| Language | Phrase | Gloss | Relation |
|---|---|---|---|
| Mandarin | 三 本 书 | three CLF book | clf(三, 本) |
| Thai | หมา สาม ตัว | dog three CLF | clf(สาม, ตัว) |
| Cantonese | 呢 架 車 | this CLF car | clf(呢, 架) |
In the Thai phrase, the numeral and classifier follow the noun. The tree still has the same shape: the classifier hangs on the numeral, and the numeral hangs on "dog". In the Cantonese phrase, the classifier is paired with the demonstrative rather than a number.
Across languages
The relation is concentrated in languages with grammaticalised classifier systems. These include Chinese varieties, Japanese, Korean, Thai, Vietnamese, Burmese and many languages of Mesoamerica and Oceania.
European languages lack the category. In the English and French UD treebanks the relation does not occur at all. Their measure expressions ("a piece of cake", "un morceau de gâteau") are analysed with ordinary nominal relations.
Even among classifier languages, practice diverges. In Vietnamese, a classifier such as con is often treated as the head noun of the phrase, so clf appears less often than one might expect. Japanese counters fuse with the numeral (san-satsu "three-CLF"). They are then handled through segmentation and morphology rather than a separate dependency.
Such divergence is what UD's subtypes and language-specific features are meant to absorb. The common relation set stays intact while each language records its own detail (Nivre et al. 2020). Nivre's later work recasts UD annotation in terms of constructions and the strategies languages use to express them (Nivre 2025). The classifier is a natural example of one such strategy for counting and reference.
In treebanks and research
UD grew out of universal Stanford dependencies, which proposed a cross-linguistic typology of relations (Marneffe et al. 2014). It also drew on Google's universal tags and on Interset (Nivre et al. 2016).
The v1 inventory had 40 dependency types (Nivre et al. 2016). The v2 revision reorganised the set, and clf took its place among the 37 universal types (Nivre et al. 2020).
For dependency parsing, clf is a short, local arc. It is usually easy to predict once part-of-speech tagging has identified the classifier. Errors come mostly from three places:
- word segmentation in scripts without spaces;
- confusing classifiers with nouns;
- the attachment choice between numeral and noun.
The CoNLL 2017 and 2018 shared tasks parsed raw text for many languages from UD data (Marneffe et al. 2021). They therefore scored clf alongside segmentation quality. For multilingual work, inconsistent classifier attachment across treebanks matters a great deal. Comparing treebanks only makes sense when they are annotated consistently.
Methods for detecting annotation inconsistencies in UD corpora are a practical tool for auditing such relations (Marneffe et al. 2017). Tutorials on UD principles and tools cover the same practical ground for annotators adding classifier languages to the treebanks (Zeman 2021).
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-clf/