POS tagging 1. Introduction
Universal and language-specific POS tags
NLP pipelines that we practice in class provide two levels of POS information.
Universal POS tags
Universal POS (UPOS) tags provide a relatively small set of categories that can be applied across languages.
Examples include:
| UPOS tag | General category | Example |
|---|---|---|
NOUN |
common noun | student |
PROPN |
proper noun | Hawaiʻi |
VERB |
lexical verb | analyze |
AUX |
auxiliary | may |
ADJ |
adjective | complex |
ADV |
adverb | carefully |
PRON |
pronoun | they |
DET |
determiner | the |
ADP |
adposition | in |
CCONJ |
coordinating conjunction | and |
SCONJ |
subordinating conjunction | because |
PART |
particle | language-dependent |
PUNCT |
punctuation | . |
Language-specific part-of-speech tags
A language-specific tag (XPOS) provides a more detailed analysis based on a particular annotation scheme or treebank.
For example, English treebanks often use the Penn Treebank tag set, in which verbs may receive tags such as:
VB base-form verb
VBD past-tense verb
VBG gerund or present participle
VBN past participle
For Korean, a detailed analysis may distinguish categories such as common nouns, particles, connective endings, and sentence-final endings. Depending on the model and tokenization scheme, a single word may be assigned a combination of multiple Korean-specific tags.
In this tutorial:
- spaCy’s
token.pos_provides a universal part-of-speech category; - spaCy’s
token.tag_provides a more detailed English tag; - Stanza’s
word.uposprovides a universal part-of-speech category; - Stanza’s
word.xposprovides a treebank-specific Korean tag based on the Sejong tag set.