🏷️ ADJ: adjective

Universal Dependencies · 7 key papers · all concepts

Part of: 🗨️ Natural language processing · 🌐 Universal Dependencies

ADJ (adjective) is the Universal Dependencies part-of-speech tag for words that typically modify nouns and name a property or quality of what the noun refers to, such as *red*, *old* or *Swedish*. It covers adjectives used both attributively and predicatively.

Also: adjective, UPOS ADJ, adjectival word

What it is

In Part-of-speech tagging, ADJ is one of the universal tags of Universal Dependencies. The guidelines define the adjective by function. It is a word that typically modifies a noun or noun phrase by stating an attribute. It can also serve as the predicate of a copular clause. The tag is assigned in context, so the same written form may be ADJ in one sentence and another category in the next.

The tag goes back to the twelve-category universal tagset of Petrov, Das and McDonald. That set included ADJ next to NOUN, VERB, ADV and others, and its categories were fixed pragmatically from existing annotation guidelines (Petrov et al. 2012). One of their decisions was to map Korean stative verbs to adjectives (Petrov et al. 2012). UD v1 enlarged this inventory to 17 tags and kept ADJ (Nivre et al. 2016). The list remained closed at 17 in UD v2 (Nivre et al. 2020).

Borderline cases

• ADJ vs DET: words such as some, every, my and this are determiners, not adjectives, even when they precede a noun.
• ADJ vs NUM: cardinal numbers are NUM. Ordinals such as first or third are usually ADJ, marked with NumType=Ord.
• ADJ vs VERB: participles like broken are VERB when verbal properties dominate and ADJ when they behave as plain modifiers. The choice is partly language-specific.
• ADJ vs NOUN/PROPN: in Swedish law, the word Swedish is ADJ. The noun in apple pie stays NOUN, because its relation is compound and not amod.

Examples

• English attributive: a [steep] hill, where steep is ADJ attached by amod.
• English predicative: The hill is [steep]. Here steep is the head of the clause and is depends on it as cop.
• English secondary predicate: She seems [tired], where the adjective is an xcomp.
• French: une maison [blanche], a feminine singular form placed after the noun.
• German: ein [großes] Haus, which inflects for gender, number and case.

These attachments follow UD's design. Content words are heads, function words such as copulas depend on them, and modifier labels like amod encode the type of the dependent (Marneffe et al. 2021).

Across languages

Adjectives differ widely in how much morphology they carry. The FEATS commonly found on ADJ are listed below.

These features descend from Interset. Zeman built it as a universal inventory of morphosyntactic features and values, used as an interlingua between tagsets (Zeman 2008). In some languages, property words behave like verbs. Korean is the classic case, and the decision to treat its stative verbs as adjectives shows that the category boundary is a choice of annotation policy (Petrov et al. 2012).

In treebanks and research

In the English and French treebanks, adjectives most often attach as amod, by a wide margin. The next most common relations are root (predicative use), conj and xcomp. English adds ccomp among its frequent relations, while French adds advcl. The predicative pattern is more prominent in English than in French, where post-nominal modification is pervasive.

For dependency parsing, coarse tags such as ADJ cost little. An English parser using the 12 universal tags instead of the 45 Penn Treebank tags lost only about 0.6% accuracy (Petrov et al. 2012).

The fuzzy ADJ boundaries make the tag a frequent source of disagreement between treebanks. Consistency of UD annotation has been studied directly (Marneffe et al. 2017). One proposal compares UPOS trigram distributions between two treebanks A and B (Aggarwal & Zeman 2020):

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

In this formula, KLcpos3 is an asymmetric Kullback–Leibler divergence over trigrams of universal POS tags. Summing both directions gives a symmetric score θpos. For treebanks of the same genre, a score at or below 0.5 counts as consistent (Aggarwal & Zeman 2020). Because the measure works on tag sequences, a treebank that labels participles or ordinals differently will show up in such comparisons.

UD's tools and principles, including the handling of ADJ, are summarised in tutorial form by Zeman (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-adj/