Part-of-Speech Tagging and Universal Dependencies: How NLP Labels Every Word

Before a program can work out who did what to whom in a sentence, it usually has to decide what kind of word each word is. "Book" is a noun in "a good book" and a verb in "book a table"; "that" can be a determiner, a pronoun or a conjunction. Part-of-speech (POS) tagging is the task of assigning each token its grammatical category. It is one of the oldest problems in natural language processing, and the way it is now annotated across languages, through the Universal Dependencies project, says a lot about how multilingual NLP is built.
What a tagger does
The input is a sequence of tokens; the output is one label per token. The difficulty is ambiguity. In English, many frequent word forms can belong to several categories: "round" can be a noun, verb, adjective, adverb or preposition, and "back" is nearly as versatile. Context resolves most cases ("the round table", "round the corner"), and a tagger is a model of that context. Rare and unseen words are the other half of the problem: a tagger has to guess the category of a word it never saw in training from its spelling (a suffix such as -ness or -ly), its capitalisation and its neighbours.
Tagging depends on an earlier decision, the segmentation of text into tokens, which is covered in our article on tokenization. It is not the same as the subword segmentation used by large language models: a POS tagset is defined over syntactic words, so a form such as French au ("to the") has to be split into the two words à and le before each can receive its own tag.
From one tagset per corpus to universal tags
Early annotated corpora each defined their own inventory. The Penn Treebank, the corpus of American English that dominated tagging research from the 1990s, uses 36 tags for words, plus tags for punctuation and symbols. It distinguishes, for example, singular and plural nouns (NN, NNS), proper nouns and six verb forms. Corpora of German, Czech or Chinese used inventories built around their own grammars, some with hundreds or thousands of fine-grained tags that encode case, gender or tense.
That variety made sense for linguists and was a problem for everyone else. A tagger's accuracy on one corpus could not be compared with another's, and a model trained on one language could not be evaluated on another with the same categories. In 2012, Slav Petrov, Dipanjan Das and Ryan McDonald proposed "A Universal Part-of-Speech Tagset": twelve coarse categories, with mappings from the tagsets of 25 existing treebanks. The idea was not to replace detailed tags but to add a shared layer above them.
Universal Dependencies
Universal Dependencies grew out of that tagset, the Stanford dependency scheme and other harmonisation efforts. It is an open community project that annotates parts of speech, morphological features and syntactic dependencies with the same guidelines across languages. Its home page describes more than 600 contributors producing more than 200 treebanks in more than 150 languages; release 2.19 is dated 15 November 2026. Treebanks are released on a regular schedule, each with its own licence, which matters if you plan to reuse them.
The 17 universal tags
UD version 2 uses 17 universal POS tags, grouped as open-class words, closed-class words and others:
| Group | Tags |
|---|---|
| Open class | ADJ adjective, ADV adverb, INTJ interjection, NOUN noun, PROPN proper noun, VERB verb |
| Closed class | ADP adposition, AUX auxiliary, CCONJ coordinating conjunction, DET determiner, NUM numeral, PART particle, PRON pronoun, SCONJ subordinating conjunction |
| Other | PUNCT punctuation, SYM symbol, X other |
Some choices are deliberate departures from school grammar. "Adposition" covers prepositions, postpositions and circumpositions, so that English in, Japanese de and Turkish için share a tag. Auxiliaries are separated from main verbs. Everything finer, such as number, tense, case or politeness, goes into a separate column of morphological features (Number=Plur, Tense=Past), and each treebank may keep its original language-specific tag alongside the universal one.
Dependencies and the CoNLL-U format
Beyond tags, each word is attached to a head word by one of 37 universal relations, such as nsubj (nominal subject), obj (object), obl (oblique nominal) or case (case marker), with optional language-specific subtypes. Treebanks are distributed in the CoNLL-U format: plain text, one word per line, ten tab-separated columns, a blank line between sentences. Here are six of those columns for the French sentence "Il va au marché" ("He goes to the market"), slightly simplified:
# text = Il va au marché.
ID FORM LEMMA UPOS HEAD DEPREL
1 Il il PRON 2 nsubj
2 va aller VERB 0 root
3-4 au _ _ _ _
3 à à ADP 5 case
4 le le DET 5 det
5 marché marché NOUN 2 obl
6 . . PUNCT 2 punct
Line 3-4 is a multiword token: the surface form au appears once in the text but stands for two syntactic words, each tagged separately. The verb has head 0 and the relation root. The format is easy to read by eye and to parse.
How taggers work
The methods follow the general history of NLP. Rule-based taggers came first, followed by statistical sequence models: hidden Markov models that combine the probability of a tag given the previous tags with the probability of a word given its tag, then discriminative models such as maximum-entropy taggers and conditional random fields, which can use many overlapping features of a word and its context. Bidirectional LSTMs over word and character representations raised accuracy again, and current systems typically fine-tune a pretrained transformer encoder, often multilingual. Open toolkits such as UDPipe, Stanza and spaCy offer models trained on UD treebanks for dozens of languages.
Tagging is rarely the end product. Tags feed lemmatisers, dependency parsers, the feature sets of older named entity recognisers, terminology extraction and corpus linguistics queries ("find every adjective before this noun"). They also help with lexical ambiguity: knowing that "bass" is a noun does not tell you whether it is a fish or an instrument, which is the separate problem of word sense disambiguation, but it rules out readings of other words.
What "97% accurate" means
For English newswire, per-token tagging accuracy has hovered around 97% for years. Christopher Manning's 2011 paper "Part-of-Speech Tagging from 97% to 100%: Is It Time for Some Linguistics?" examined the remaining errors and found that many came from inconsistencies and mistakes in the gold-standard annotation itself, and from genuinely hard distinctions, rather than from obvious model failures. Two cautions follow.
- Per token is not per sentence. If errors were independent, a 97% tagger would get a 25-token sentence entirely right with probability 0.9725, about 0.47. Downstream components see whole sentences, and one wrong tag on a verb can derail a parse.
- Benchmarks flatter the easy cases. Most tokens are frequent, unambiguous words. Accuracy on unknown words, on social-media text, on transcribed speech or on a language with few annotated sentences is much lower, which is the general lesson of our article on evaluating NLP models.
Comparing scores across languages needs the same care. A tagger can score lower on a morphologically rich language because the universal tags are harder to assign there, because the treebank is small, or because its text comes from a different genre. UD makes the comparison possible; it does not make it automatically fair.
Practical advice
- Use universal tags when you need comparability across languages, and the language-specific tags or features when you need detail within one.
- Check which treebank, release and genre a pretrained model was trained on before trusting it on your text.
- Report accuracy on unknown words and, where possible, sentence-level accuracy alongside the headline figure.
Sources
- Universal Dependencies, "Universal POS tags" and "CoNLL-U Format" (UD version 2 guidelines); home page and release 2.19 information consulted on 19 November 2026.
- Petrov, Das and McDonald, "A Universal Part-of-Speech Tagset", LREC 2012 (arXiv:1104.2086).
- Marcus, Santorini and Marcinkiewicz, "Building a Large Annotated Corpus of English: The Penn Treebank", Computational Linguistics, 1993; Manning, "Part-of-Speech Tagging from 97% to 100%: Is It Time for Some Linguistics?", CICLing 2011; de Marneffe and colleagues, "Universal Dependencies", Computational Linguistics, 2021.
Methods & evaluationHow tokenization works
Why models read text in pieces, how BPE, WordPiece and SentencePiece split words, and why the same sentence costs more tokens in some languages.
Methods & evaluationNamed entity recognition
Spotting the names in a text sounds trivial until you have to decide where a name ends, what kind of thing it refers to, and how to score a system that half gets it right.
Methods & evaluationWord sense disambiguation
Bank, bass, bat: how NLP systems choose the right meaning of an ambiguous word, and how the field measures whether they succeed.