The familiar inventory — noun, verb, adjective, adverb, preposition, article… — is not a natural classification of language; it is the Greek–Latin grammatical tradition exported (Dionysius Thrax's eight parts, 2nd c. BCE, carried worldwide by missionary grammar). Cross-linguistically the inventory is a gradient:
Near-universal: noun and verb — and only those
The strongest candidates for universal categories, with one famous battlefield: Salishan languages, Nuu-chah-nulth, Riau Indonesian have been argued to let any root predicate or refer, making the distinction nearly invisible. The typological consensus (Dixon, Croft): a noun/verb distinction can always be found, but it can be weak or covert. That contested floor is the hardest ground parts-of-speech have.
Famously non-universal: adjectives (Dixon)
- Igbo: a closed class of ~8 adjectives (big/small, black/white, new/old, good/bad) — every other property concept is a verb or noun.
- Mandarin: "adjectives" are stative verbs — 她高兴 takes no copula and patterns with 她跑.
- Korean: adjectives conjugate like verbs (descriptive verbs); Japanese splits the class in two — i-adjectives inflect verb-like, na-adjectives behave like nouns.
The rest of the European list
- Adverbs: a wastebasket category even at home; many languages replace them with case-marked adjectives, ideophones, or converbs.
- Adpositions: widely absent, with three substitutes — case suffixes (Finnish's ~15 cases), relational nouns (Mayan/Nahuatl "at the house's head" = on the house), verbs (serial-verb languages: Mandarin 在/给/到 as coverbs; West African take/give constructions).
- Articles: a minority feature — roughly a third of sampled languages have a definite article (WALS); Russian, Chinese, Japanese, Latin do without.
- Conjunctions: often replaced by clause-chaining, converbs, switch-reference.
The classes Europe never had
The traffic runs both ways — other languages have major classes the exported inventory lacks:
| Class | Example |
|---|---|
| Classifiers / measure words | 个 · 本 · 张 · 条 — a core class across Chinese, Japanese, Thai, Burmese |
| Ideophones | Japanese きらきら / ふわふわ; a major word class in many African languages |
| Evidentials | Turkish -mış, Quechua: seen / heard-of / inferred as an obligatory grammatical choice |
| Sentence-final particles | 吧 · 呢 · 吗; Cantonese has 30+, stackable |
| Converbs / serial verbs | Altaic and South Asian clause-chaining doing the work of conjunctions |
The theory in two sentences
What is universal is not the classes but the functions — reference, predication, modification — which every language must perform; languages differ in how morphosyntax packages them (Croft's prototypes: noun = object-reference, verb = action-predication, adjective = property-modification, with boundaries drawn differently everywhere). Haspelmath's sharper cut: parts of speech are language-particular descriptive categories; across languages there exist only comparative concepts — applying Latin's boxes to Xhosa is category smuggling.
The engineering answer
Universal Dependencies' 17 UPOS tags: a deliberately procrustean common denominator, chosen not because it is true but because every language can be force-annotated with it. NLP's parts of speech are that compromise, industrialized.