품사 태깅

A word can belong to different grammatical categories depending on its use. Book is a noun in read a book and a verb in book a flight. A tagger therefore uses context rather than assumes that a dictionary entry determines one fixed tag. The relevant unit is a token, and the chosen annotation convention affects what counts as a category. English work often uses the Penn Treebank tags. Its core set contains 36 part-of-speech tags, with additional punctuation tags also used in annotations. NN denotes a singular or mass noun, JJ an adjective, and VBD a past-tense verb. Detailed tags distinguish linguistic properties that a coarser inventory may represent separately. Universal Dependencies version 2 defines 17 universal POS categories, including NOUN, VERB, ADJ, and PUNCT, for a shared annotation framework. Its additional features can express properties beyond the core category. POS categories should also be distinguished from dependency relations. NOUN describes a word category, whereas nsubj describes a nominal subject relation to a syntactic head. A noun is not automatically the subject of every sentence in which it occurs. Classic hidden Markov model taggers treat tags as hidden states and words as observations. They score tag sequences using transitions between tags and emissions of words given tags. Neural taggers can instead learn contextual representations and predict token categories. Both approaches require evaluation against an identified tagset and dataset; an accuracy claim for unspecified clean text is not a universal result.