🏷️ PART: particle

Universal Dependencies · 7 key papers · all concepts

Part of: 🗨️ Natural language processing · 🌐 Universal Dependencies

PART (particle) is the Universal Dependencies part-of-speech tag for function words that must attach to another word or phrase to contribute meaning and that fit no other universal category. Examples include negators, infinitive markers, possessive markers and question particles.

Also: particle, PRT (Petrov et al. 2012 tagset), UPOS PART

What it is

In Universal Dependencies, every word receives one of 17 universal part-of-speech tags from a closed list (Nivre et al. 2020). PART is the residual class among function words. A word is tagged PART when it is grammatical rather than lexical, when it needs a host word or phrase to mean anything, and when it is not better described as an adposition, conjunction, determiner, auxiliary or adverb. The guidelines admit that the class is defined negatively and is therefore heterogeneous. They prefer to assign a word to a more specific tag whenever one applies.

The tag descends from PRT, one of the twelve coarse categories of the universal tagset proposed by Petrov, Das and McDonald (Petrov et al. 2012). That tagset was defined pragmatically from existing annotation guidelines rather than from theory, and it was mapped onto 25 treebanks covering 22 languages (Petrov et al. 2012). UD v1 extended this Google set to 17 tags and kept the particle category under the name PART (Nivre et al. 2016).

Borderline cases

• PART vs ADV: negators such as English not are PART. Similar words in other languages may be tagged ADV.
• PART vs ADP: English verb particles (give up) are usually ADP attached with compound:prt. The possessive 's is PART.
• PART vs SCONJ: infinitival to is PART, while clause-introducing that is SCONJ.
• PART vs AUX: aspect or mood markers that behave like verbal auxiliaries go to AUX.

Features and relations

PART words often carry few features. The most common are Polarity=Neg on negators, a language-specific PartType on some particles, and occasionally Poss=Yes or Mood. UD v2 renamed the older Negative feature to Polarity (Nivre et al. 2020). The features themselves derive from Interset, a universal inventory of morphosyntactic values that serves as an interlingua between tagsets (Zeman 2008; Nivre et al. 2016).

Because UD makes content words heads and attaches function words to them (Marneffe et al. 2021), particles appear as dependents. Typical relations are mark, advmod, case and discourse, and in some languages compound:prt.

Examples

The English clitic 's is split off as a separate syntactic word even though orthography joins it to its host (Marneffe et al. 2021).

Across languages

The weight of PART differs greatly between languages.

In English treebanks the tag is fairly frequent. Its distribution is dominated by three pairings: infinitival to as mark, negation as advmod, and possessive 's as case. Open clausal complements and conjuncts are only marginal. The tag therefore bundles three unrelated grammatical functions into one class.

In French treebanks PART is nearly absent, and its few tokens attach only as advmod. Negative ne and pas are normally tagged ADV, and there is no particle-marked infinitive or possessive.

Languages with rich sentence-final or discourse particles use the tag much more broadly. These include Mandarin, Japanese, Vietnamese and Thai, where particles mark questions, politeness and modality. Highly inflecting languages, by contrast, often express the same meanings morphologically and need PART only for a few negators or modal words. This variation illustrates how a residual category absorbs language-specific grammar (English, French).

In treebanks and research

PART matters for part-of-speech tagging and dependency parsing because cross-lingual transfer depends on comparable categories. Petrov et al. showed that coarse universal tags halved the cross-language spread of tagging accuracy. Their delexicalized parser transferred from English outperformed contemporary unsupervised methods (Petrov et al. 2012).

The residual nature of PART also makes it a frequent source of disagreement between treebanks of the same language. Annotation consistency in UD corpora has been audited systematically (Marneffe et al. 2017). Aggarwal and Zeman proposed a symmetric divergence over UPOS trigram distributions to detect inconsistent tagging between treebanks without relying on a trained tagger (Aggarwal & Zeman 2020).

θpos(A,B) = KLcpos3(A,B) + KLcpos3(B,A)

Here A and B are two treebanks, and KLcpos3 is the Kullback–Leibler divergence between their distributions of UPOS-tag trigrams. Summing both directions makes the measure symmetric. A threshold that depends on genre composition then separates consistent pairs from inconsistent ones (Aggarwal & Zeman 2020). Shifts of words between PART, ADV and ADP are exactly the kind of difference that such trigram statistics expose.

Practical overviews of the UD principles and tools also discuss these tag decisions for annotators (Zeman 2021).

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-part/