Dependency Parsing

That is the entire system. Pick any word; there is exactly one other word it depends on, called its head. The word doing the depending is the dependent. The link between them is an arc, and the arc carries a relation label naming the kind of dependency.

One word in each sentence has no head — the root, almost always the main verb. Everything else hangs off it, directly or through a chain.

The root has nothing above it, so the parser records it as being its own head. That is not a trick — it is how the data is genuinely stored, and it is why the root is drawn with a short stub rising out of the top of its box rather than an arc from somewhere else.

It matters in practice: in the Dependencies exercise, the root takes one click, not two. Every other relation asks you to click a head and then a dependent; for the root there is only one word to click.

Words stay in sentence order horizontally, which means arcs sometimes cross. That is accepted deliberately: rearranging words to avoid crossings would destroy the one thing that makes a tree easy to check against the sentence above it.

There are forty-six relation labels, which is far too many to give distinct colours that stay readable in every theme. They are grouped into six families instead, and the colour tells you the family, the label tells you the relation. Two arcs of the same colour are relatives, not the same thing.

The clearest way to understand either notation is to find the thing it is bad at.

A dependency parse cannot show a phrase. There is no arc meaning “these four words are one unit”. You can recover it — a word plus everything hanging below it is a phrase — but the picture never says so directly.

So a relation names the head of a phrase, while a Reed-Kellogg role names the whole of it: in the Dependencies exercise nsubj marks passengers alone, and in the Roles exercise the Subject is the passengers. The two labels can read alike — pobj and “Object of preposition” — and still point at different amounts of the sentence.

A Reed-Kellogg diagram cannot name a relationship. A slant says “modifier” and nothing more; whether it is a determiner, an adjective or a possessive is left to the reader. The dependency parse distinguishes all three.

The practical upshot: use the Reed-Kellogg diagram when you want to see structure, and the dependency tree when you want to see relationships. This app draws both for every sentence so you never have to choose.

A dependency tree cannot give a word two heads. Every word has exactly one incoming arc. That rule is what makes the picture a tree, and it is what the Dependencies exercise asks you to find: one relation and one head per word. Most sentences never put it under strain. This one does — She both knows and corresponds with the queen.

Queen is the direct object of knows and the object of the preposition with. The Reed-Kellogg diagram can say both at once: the two verbs fork, knows ends in the direct object’s upright separator, with hangs from corresponds, and the two lines run back together so that queen stands where they meet. The tree has to choose. The parser attaches queen to with as pobj, and knows is left with no object at all — the dobj relation is not hidden somewhere in the tree, it is absent from it.

So the Roles exercise accepts the queen as both Direct object and Object of preposition, while the Dependencies exercise asks for one arc, pobj, because that is all the tree contains. Richer representations exist — enhanced dependency graphs let a word carry a second, secondary head — but this app draws the basic tree deliberately: it is what the parser produces, and the point of showing both views is that each says something the other cannot. Try both exercises on this sentence in the Practice Lab.