Numbers, Symbols & the Rest

A cardinal number. The single fine tag under NUM, so the two scales agree completely here. Its dependency label is usually nummod, and in a Reed-Kellogg diagram it slants under the noun it counts like any other modifier.

How spaCy decides: the same lookup as the category. Words and digits get one tag — three and 42 in the same sentence are both CD.

Nine^CDcandles^NNSburned^VBDfor^INtwo^CDhours^NNS.

Chapter^NNthree^CDends^VBZon^INpage^NN42^CD.

The interjection tag. The single fine tag under INTJ. There is none in the corpus at the time of writing — the sentences there are statements and questions, and interjections belong to speech — so the first one the Tags exercise offers will be from a newly added sentence.

How spaCy decides: the same lookup. Note that yes and no as answers are UH, while no before a noun (no bus) is a determiner.

Oh^UH,I^PRPforgot^VBDmy^PRP$umbrella^NNagain^RB.

Yes^UH,the^DTtickets^NNSarrived^VBDthis^DTmorning^NN.

Ouch^UH,that^DTpan^NNis^VBZhot^JJ!

The symbol tag proper: mathematical operators and the like. The Penn set separates it from the dollar sign, which has a tag of its own below.

How spaCy decides: a symbol character between numbers is the reliable case — 3 x 4, 4 / 5. Most other symbols land elsewhere.

The^DTanswer^NNis^VBZ3^CDx^SYM4^CD.

Rate^VBthe^DTfilm^NN4^CD/^SYM5^CD.

The dollar sign — and, in the Penn convention, other currency signs too. It gets its own tag because it behaves unlike other symbols: it precedes its number rather than sitting between two. In the dependency tree the number is the head and the sign hangs off it — and the label it hangs by is nmod, the one construction in this app that produces that relation at all (see Nominal Dependents).

How spaCy decides: the character. The tokenizer splits $40 into $ and 40, so the sign is its own token with the tag $ and the number is a separate CD.

Tickets^NNScost^VBP$^$12^CDfor^INadults^NNS.

The^DTrepair^NNcosts^VBZ$^$40^CDat^RBmost^JJS.

Sentence-final punctuation: the full stop, question mark and exclamation mark. In the dependency tree it attaches to the root verb with the label punct.

The^DTlamp^NNflickered^VBDtwice^RB.^.

Did^VBDthe^DTguests^NNSleave^VBearly^RB?^.

The^DTkettle^NNis^VBZboiling^VBG!^.

The comma. The one punctuation mark that matters to the diagrams indirectly: a comma after a sentence-initial adverb is what the parser reads as the signal for a sentence disjunct, and a comma pair is what sets off an appositive or a vocative. It is still never drawn.

After^INlunch^NN,^,the^DTguests^NNSdozed^VBD.^.

Two commas doing two different jobs — one after a fronted phrase, one inside a list:

After^INlunch^NN,^,we^PRPpacked^VBDthe^DTcar^NN,^,the^DTtent^NNand^CCthe^DTdog^NN.^.

Colon or semicolon — and, by the Penn convention, the dash and the ellipsis as well. Any mark that separates two parts of a sentence without ending it.

Bring^VBtwo^CDthings^NNS:^:a^DTcoat^NNand^CCa^DTmap^NN.^.

The^DTtide^NNturned^VBD;^:the^DTboats^NNSreturned^VBD.^.

The^DTverdict^NNwas^VBDclear^JJ—^:nobody^NNwas^VBDhurt^VBN.^.

The opening quotation mark. Written as two backticks because the original Penn Treebank predates curly quotes and used `` and '' to tell opening from closing. The parser assigns it from position: the first of a pair opens.

The^DTsign^NNread^VBD"^``closed^VBN"^''in^INred^JJletters^NNS.^.

A quoted word inside a sentence — the opening mark gets this tag:

She^PRPwhispered^VBD"^``wait^VB"^''and^CCleft^VBD.^.

The closing quotation mark — the second of the pair in the example above. The corpus has exactly one, a stray closing quote at the end of a sentence that was pasted in with it.

The same sentence, with the closing mark picked out:

She^PRPwhispered^VBD"^``wait^VB"^''and^CCleft^VBD.^.

A left bracket — an opening parenthesis, square bracket or brace. The name is spelled out in letters because the original treebank format could not hold a literal bracket character.

The^DTrecipe^NN(^-LRB-from^INmy^PRP$aunt^NN)^-RRB-needs^VBZsaffron^NN.^.

A parenthetical aside in the middle of a clause:

The^DTferry^NN(^-LRB-the^DTlast^JJone^CD)^-RRB-left^VBDat^INnoon^NN.^.

A right bracket — the closing half of the pair above. Also the closing parenthesis after a list marker such as 1), which is how it most often appears in the probes for this page.

The same aside, with the closing bracket picked out:

The^DTferry^NN(^-LRB-the^DTlast^JJone^CD)^-RRB-left^VBDat^INnoon^NN.^.

A hyphen inside a word. The parser splits a hyphenated compound into its parts — well, -, known — so the hyphen is its own token. The pieces on either side keep their own tags (RB and VBN here), and the display joins the three back into one visual word because none of them has a space before it.

This is the one punctuation tag with a job in the app: co-captains in the corpus is three tokens, and the drag surface and the diagrams draw them as one word precisely because the middle one is HYPH with no spacing on either side.

The^DTwell^RB-^HYPHknown^VBNauthor^NNsigned^VBDcopies^NNS.^.

The^DTtwenty^CD-^HYPHyear^NN-^HYPHold^JJbridge^NNfinally^RBreopened^VBD.^.

Superfluous punctuation — marks that punctuate nothing: emoticons, a row of asterisks, a tilde, decorative dashes. The parser reaches for it whenever a punctuation-like character is doing no grammatical work. Absent from the corpus, and likely to stay so.

Good^JJluck^NN:-)^NFPand^CCsee^VByou^PRPsoon^RB.^.

Meet^VBme^PRPat^INnoon^NN~^NFPbring^VBlunch^NN.^.

A foreign word — a word from another language used untranslated: de facto, à la carte. The tag is for the word’s status, not its meaning, so a borrowed phrase gets it whatever the phrase does in the sentence.

How spaCy decides: poorly. The model recognises a handful of Latin and French phrases and treats most foreign words as names (NNP) or plain nouns — bon appétit came back half FW, half NN; merci beaucoup was two proper nouns; status quo was two English nouns. Expect it to be inconsistent.

Their^PRP$plan^NNwas^VBD,de^FWfacto^FW,a^DTretreat^NN.

The^DTdish^NNis^VBZserved^VBNà^FWla^NNPcarte^NN.

An affix — a prefix or suffix that the tokenizer has split from its stem, as in co-captains or pre-war. The tag exists so that the detached piece is not mistaken for a word in its own right.

How spaCy decides: it does not, in practice. Every hyphenated prefix probed for this page — pre-, co-, ex-, anti-, non-, post-, semi-, sub-, re- — came back as a noun or an adjective, never AFX.

The two instances in the corpus were both entered as NN by the parser and corrected to AFX/X by an administrator, which is the only way the tag reaches an exercise. The hyphen beside it had been called a plural noun, and was corrected to HYPH in the same pass — the raw parse below shows both.

The parser’s own output for the corpus sentence:

Two^CDstudents^NNSwere^VBDmade^VBNco^NN-^NNScaptains^NNS.

And the same sentence as it is stored, after correction:

Two^CDstudents^NNSwere^VBDmade^VBNco^AFX-^HYPHcaptains^NNS.^.

A list item marker — the 1) or b) that numbers an item in a list. Not a numeral: it labels the line rather than counting anything.

How spaCy decides: a bare digit or letter at the very start, followed by a closing bracket. A marker written 1. is not recognised — the period is split off as sentence-final punctuation and the digit is still tagged LS, but then the model reads the next word as a name. A list marker will not survive this app’s sentence normalisation as anything meaningful, so treat the tag as documentation.

1^LS)Bring^VBa^DTcoat^NN.

b^LS)Bring^VBa^DTcoat^NNand^CCa^DTtorch^NN.

An email or web address. The tokenizer recognises URLs and email addresses as single tokens, and this tag marks the token as an address rather than a word.

How spaCy decides: the token shape, and only for the well-formed cases — hello@example.com and https://example.com both work. A bare domain without a scheme (www.example.com) does not, and comes back as a proper noun or an adjective. No sentence in the corpus contains an address.

Email^VBme^PRPat^INhello@example.com^ADD.

Visit^VBhttps://example.com^ADDfor^INdetails^NNS.

Unknown. The tagger’s way of saying it cannot classify the token at all — stray characters, fragments of markup, damaged text. It is also the palette’s fallback colour: a tag with no colour of its own renders in the XX hue, which is why an undefined token in the app looks like this rather than vanishing.

How spaCy decides: hard to provoke on purpose. Nonsense words are confidently tagged as nouns or names; what actually triggers it is a stray symbol like the > of a text arrow. The corpus has none.

See^VBthe^DTnote^NN-->^XXbelow^RB.

It is hard to provoke on purpose. Lorem ipsum dolor sit amet. comes back as five proper nouns, and The password was xq7zt. makes the password a proper noun too — an unknown word is guessed to be a name long before it is written off as XX.