The extended PTB POS tagset used in GUM: https://gucorpling.org/ptb_tags.html. All other POS tags in GUM are derived from these tags plus information from other gold standard annotations.
Tagging "don't" (or really two tokens, "do" and "n't"): the verb 'do' is not considered an auxiliary in the PTB scheme in the sense of having a special tag. If it’s a present form like "I don’t do X" then the first ‘do’ is VVP and the second 'do' is VV (a base form); if it’s an imperative like "Don't go!", it’s VV for both verbs (imperative is not considered present). The negation ‘not’ and also the form ‘n’t’ is considered adverbial (compare ‘very good’ vs. ‘not good’ – both modifiers are adverbs). As a result, it’s tagged RB.
Titles of books, films, etc.: tokens are considered NP or NPS if they are capitalized, but function words are tagged as normal. So for Starship Troopers, both words are considered ‘proper’ and tagged: Starship = NP and Troopers = NPS. But for "Beauty and the Beast" we get: NP, CC, DT, NP
In cases like "more interesting", we have two tokens - ‘more’ itself is tagged JJR, but 'interesting' is still just a normal JJ. If you're counting comparatives in the corpus, counting JJR still gets you
When a hyphen or dash appears in a number or date range, it means (and would be pronounced as) 'to', and is therefore tagged TO.
When an error (annotated with sic in markup) is a grammatically plausible construction, tag the word as it is found in the text, rather than what it "should" be:
Misspellings are tagged as if they were correctly spelled, even if the misspelling has the form of a different word. For example, if 'too' appears in a construction where only 'to' would be grammatically conceivable, it is considered a misspelling of 'to' and tagged accordingly:
A disfluent token is tagged based on what you think it would have been had it not been disfluent, as long as it contains at least a complete syllable. For an incomplete syllable we always tag UH:
If it's hard to know with reasonable certainty what a disfluent token was "supposed" to have been, consider using UH or SYM too.
The generic pronoun "one" is tagged PP, not CD:
However some uses of one in reference to people are still CD:
Use WDT for which, as well as that when it is used as a relative pronoun:
And use WP for what, who, and whom:
But use WDT and not WP when what is modifying a noun:
Cf. PTB guidelines for more details.
If an interrupted word can be obviously reconstructed and has at least one syllable, it is given it's normal POS tag and the equivalent lemma:
If the reconstruction is uncertain or the word has less than a complete syllable, the tag UH is used and the word form itself is used as the lemma:
URLs should be tagged as proper noun (NP) (effectively the name of a ‘place’). The same applies to hashtags: #success/NP
No verb should have -ing in its lemma. However, nouns ending with -ing should keep the -ing in their lemma. Some words can be both nouns and verbs; categorize them based on the specific instance.
Tokens tagged as propernouns with tags like NP(S) receive uppercase lemmas:
Proper name tokens which belong underlyingly to non-nominal but inflected categories receive the normal lemma of their category, but retain capitalization:
Some abbreviations receive expanded lemmas. Refer to the following list:
Note that the place name guideline does not apply to multiword acronyms. For example, the lemmas of N.Y., D.C., U.S. are identical to the tokens, and not "New York" (with a space) etc.