Root & Leftovers

The one word in a sentence with no head. Everything else hangs off it, directly or through a chain, which makes it the first thing to find when you read a tree.

How spaCy decides: the main verb, almost always. It is stored as being its own head — not a trick, but genuinely how the data is written — which is why it is drawn with a stub rising out of the top of its box rather than an arc from elsewhere.

The root is not always a verb. A verbless exclamation still needs one, and the parser promotes whatever is left.

An interjection — oh, hey, wow. It stands outside the grammar of the sentence and simply attaches to the main verb.

How spaCy decides: the word is tagged as an interjection. It is the most reachable of the three here; ordinary written prose just rarely contains one.

Two clauses placed side by side with no conjunction and no subordination — an interrupting comment, or clauses joined only by a semicolon.

How spaCy decides: a second clause attached to the first with nothing grammatical joining them. It is hard to elicit deliberately; most sentences that look like candidates are parsed as ccomp or advcl instead.

The parser’s admission of defeat. dep means “there is a dependency here and I cannot name it”, and it appears when the input is degenerate enough that no ordinary analysis fits.

How spaCy decides: it does not, really — this is the fallback when nothing else applies. Seeing it on well-formed English is a strong signal that the parse has gone wrong somewhere earlier.

Even deliberately broken input rarely produces it. Um the thing uh broke. comes back as two interjections around an ordinary clause, and Yes, well, no. is three interjections — the parser would rather guess a real relation than admit defeat.

Direct address — naming the person you are speaking to. Children in Children, we must hurry is not the subject and not a modifier; it is outside the clause entirely.

How spaCy decides: it does not. This model never produced a vocative once across the whole corpus. Fed that sentence, it labels Children as a noun phrase acting adverbially:

An administrator corrected that token by hand, and the corrected value is what the corpus stores and what the exercise grades against:

Reserved for markup and metadata that is in the text but not in the sentence — editorial marks, tracking artefacts, stray formatting.

How spaCy decides: in practice, it does not. meta appears nowhere in the corpus and could not be elicited by any test input. It exists for cleaning up scraped text rather than for analysing curated sentences, and every sentence in this app is curated.

There is no “what it emits instead” figure for this one, because there is no construction it competes for. If you ever see it on a Grammify sentence, it was set by hand.

The forty-sixth label, and the only one that is both extremely common and completely invisible. Punctuation is the single most frequent relation in the corpus — and it is excluded from every surface in the app.

Where it is excluded: punctuation is dropped from the dependency tree, so it is never a node and never an arc. It is absent from the POS, Tags and Dependencies palettes. It is not clickable in any exercise, and no syntactic span in a Reed-Kellogg diagram ever includes it.

The one place a reader can see the tags themselves is a chips figure that labels them:

Well^UH,^,the^DTferry^NNleft^VBD!^.

The one place it is visible is the Workbench token table, which shows every row including punctuation, so an administrator can see the complete tokenisation. Note that a mid-word hyphen is a punctuation token too — co-captains is three tokens, not one.