What are the chief advantages & disadvantages of describing sentences with dependency vs. phrase structure (aka. constituency) trees?

From what I've read, dependency grammar trees lack phrase nodes and mark everything as dependent on the verb.

Phrase structure trees start with the highest constituent then analyzes it into phrases like NP, VP, etc., as I'm sure we all know.

Under what circumstances would it be preferable to use one type of tree diagram over the other?

  • Why force a choice? Some things you represent one way, others a different way, still others you use both. There isn't any universally correct way; this is simply representation, and that is always arbitrary and variable. – jlawler Mar 23 '13 at 22:01
  • 4
    My question doesn't imply that there is one universally correct way to represent grammatical structures. It (hopefully) implies that under some circumstances one would use one method and under other circumstances one would use the other. What kinds of grammatical structures are easiest to represent with PR trees and what kind are easiest to represent with dependency trees? – James Grossmann Mar 24 '13 at 6:16
  • Grammatical relations, for instance, are dependencies, while constituent structure is what PS trees were invented for. That's why one hears "NP VP", rather than "SVO"; the two terms NP and VP refer to constituent structure and the two terms S and O refer to grammatical relations. One doesn't mix them up; the only way O's are distinguishable in PS is because they're NP's that are inside the VP. – jlawler Mar 24 '13 at 14:45
  • @Pelliot See my edited response. Dick Hudson, Tim Osborne, and about 30 people giving papers in all were at this Depling along with a few who were just there to listen and ask questions, etc. – Dan Maxwell Apr 3 '14 at 2:53

Yes, there are concrete differences between dependency-based and constituency-based tree representations (D-trees vs. PS-trees).

D-trees have the great advantage that they are minimal compared to PS-trees. Thus one can produce a D-tree for a given sentence with much less effort than one needs to produce the corresponding PS-tree for the same sentence. The number of nodes and edges in D-trees tends to be about half as many as for the corresponding PS-tree.

The proponents of PS-trees will object, however, that minimalism and simplicity are not the entire story. Their objection is of course correct insofar as simplicity alone is of no value. A simpler model is only better if it succeeds at clarifying the phenomenon at hand as well as or better than a more complex model. (The principle is known as Occam's Razor)

My personal view is that the simpler D-trees can accomplish most anything that the more complex PS-trees can. In the scheme of things, each linguist has to form his/her own opinions about the strength/weaknesses of the competing principles for representing syntactic structure.

To provide a brief and direct answer to the question, D-trees are more suited for those situations where simplicity and transparency are the primary concern. Dependency is now the preferred means for representing syntactic structure in computational linguistics, where one is interested above all in automated parsing of natural language.

Phrase structure, in contrast, remains dominant in theoretical linguistics, where one tends to pay more attention to fine-grained syntactic and semantic distinctions.

| improve this answer | |

Seeing this question reminded me of a section in Peter Matthews book 'Syntax' (1981) (it's meant to be a textbook, but it's more like a monograph really). In Chapter 4, p84-93, there's an explicit comparison and evaluation of dependency grammars vs. constituency grammars.

Matthews shows that for any d-grammar, there is a ps-grammar which will generate the same set of sentences, and vice versa - that is to say that d-grammars and ps-grammars are weakly equivalent.

Matthews goes on to discuss a different and more interesting way in which the two can be compared: Can an analysis in one framework be shown, in each instance, to be isomorphic to an analysis in the opposing framework? In other words, can we say everything using d-grammar that we can using a ps-grammar, and vice versa? If the answer is yes, then the two are strongly equivalent (Chomsky '63, formal properties of grammar). As far as i understand it, strongly equivalent grammars are essentially notational variants.

In comparing d-grammars and ps-grammars, we shouldn't necessarily pay too much heed to the variable simplicity of the notation - we can only use this as an argument if it's already been established that the two are strongly equivalent.

Matthews pretty convincingly shows that there is no mechanical procedure from which d-trees can be derived from ps-trees, and no mechanical procedure from which ps-trees can be derived from d-trees across the board. There are things of descriptive value that can be said in d-grammars that can't be said in ps-grammars, and vice versa, and therefore they aren't strongly equivalent. You could entertain the idea of combining the two representations (which is something like what we have in, e.g. modern minimalism), although this is something that Matthews ultimately rejects in favour of his own d-grammar, for interesting and unrelated reasons.

Dick Hudson is a proponent of 'word grammar', a modern theory in the d-grammar school, and he has some powerful arguments in his various papers (available here: http://www.phon.ucl.ac.uk/home/dick/wg.htm#main_ideas) that a d-grammar is to be preferred to a ps-grammar. Given that ps-grammars and d-grammars aren't mere notational variants, it seems to me that one or the other will ultimately turn out to be a more accurate abstraction over speakers' mental representations. Ultimately, that's still an open question.

| improve this answer | |
  • Hello P Elliot, I agree with much of what you write here and I have some familiarity with the sources that you cite. Matthews (1981) overview of the dependency vs. constituency distinction is excellent. – Tim Osborne Aug 20 '13 at 9:56

I think there will be an increasing amount of debate about this in the next few years. The above answers from Tim Osborne and P. Elliot give a lot of general discussion about how to compare these two forms of grammar, but no comparison with respect to specific constructions. I provide a lot of this in my article in the proceedings from the second Depling conference held in Prague last August. Those interested in more details can see these proceedings at a website called Proceedings.com. It invites you to buy these proceedings for $95, but I think you can also read individual articles online for free. Mine is on pages 197-207.

My own answer is that DG is better, partly because it is simpler (many fewer nodes), but also because of the specific difficulties for constituent grammar discussed in that article. At the instigation of PElliot, I have found another source which is free. The link is: http://ufal.mff.cuni.cz/project/depling13/proceedings/pdf/W13-3722.pdf

More recent posts on this topic show that computational linguists have come to the same conclusion. Other things being equal, less is better!

Dan Maxwell

| improve this answer | |
  • Can you please provide us with a link to your article? – P Elliott Apr 1 '14 at 14:06
  • Yes, less is better if other things are equal (Occam's Razor). However, other things are not even equal. Dependendency makes a much more accurate prediction concerning the results of diagnostics for constituent structure. For instance, constituency takes each individual word to be a constituent, yet most constituency tests suggest that many individual words are not constituents. Dependency, in contrast, takes only those individual words as constituents that do not dominate any other words. Dependency's prediction in the area is therefore much more accurate. – Tim Osborne Apr 1 '14 at 15:33
  • Constituency has to introduce the distinction between minimal, intermediate, and maximal projections to deal with the problem. Dependency does not need that extra baggage. – Tim Osborne Apr 1 '14 at 15:34

PS has the advantage that it can name the structures (NP, VP, etc.) below each node. If you don't need those names then it is easy to shorten one edge of each branch. Shorten to zero and you've halved the number of edges. Sound familiar?

Do the differences between PS and D matter if they both have their heads on backwards regarding nouns?

(I apologize for sounding opinionated.)

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.