The Universal Stanford Dependencies (USD) parsing scheme subordinates auxiliary verbs to content verbs and prepositions to their complements:


USD is contrary to most work in theoretical syntax in this regard. In effect, USD is denying the existence of nonfinite verb phrases and prepositional phrases. Apparently, USD is being widely adopted in many computational circles. Given my background in theoretical syntax, it is difficult to understand how USD can be taken seriously concerning its analysis of hierarchical structures.

I have the following questions in this regard:

  1. Is the USD approach to function words (e.g. auxiliaries and prepositions) a matter of debate in computational circles?

  2. What other parsing schemes that parse directly to dependencies (as opposed to converting from constituencies) produce structural analyses that are more in line with theoretical syntax (i.e. auxiliaries as heads over content verbs and prepositions as heads over their complements)?

  • Not most work, only the silly work. Head is not a well-defined phenomenon. And "denying in effect" is something you're doing for them, not something they're doing. The "in effect" only goes through if they share the same definitions of those entities that you do. And clearly they don't. The idea of markers being heads is simply a redefinition of what "head" means. Not a helpful move, really. Better to have meaningful heads, since they're what remains when the markers are sloughed off, as they so frequently are. – jlawler Jun 20 '14 at 15:29
  • 1
    @Jlawler, thanks for the comment. But I'm not clear about your point. You write "Not most work, only the silly work". Would you please write that out in fuller sentences, so I can see exactly what you mean. – Tim Osborne Jun 20 '14 at 15:35
  • Sorry, Tim, typing too fast. That was the phrase that popped up when I read your predicate contrary to most work in theoretical syntax. If I take you correctly, you think P and D and AuxV should be heads rather than the meaningful items they mark. As I said, that's a change of definition and presupposition. You may well be correct in computational terms -- I dunno; I don't keep up with that -- but I think it's simply silly in terms of descriptive syntax. What counts in syntax are the meaningful units; their phrasal apparatus is often in disarray. – jlawler Jun 20 '14 at 15:47
  • 1
    @Jlawler, I view more traditional approaches as correct: NPs (not DPs), PPs exist (prepositions as heads over their complement), and AuxPs (auxiliary verbs as heads over their content verbs). Most work in theoretical syntax acknowledges PPs, whereas USD really does not. Most work in theoretical syntax now acknowledges AuxPs, whereas USD does not. The NP vs. DP debate is an ongoing controversy, however. – Tim Osborne Jun 20 '14 at 15:53
  • @Jlawler, my question is concerned with the gulf between theoreticians and practitioners. There seems to be much willingness in computational circles to ignore theoretical syntax in general. My sense of this unfortunate situation is that it is due to the overinfluence of Chomskyan syntax. I think many computational people view Chomskyan syntax as impractical for their purposes, hence there is a willingness to ignore all of theoretical syntax. All other types of theoretical syntax are tossed into the basket with Chomskyan syntax. – Tim Osborne Jun 20 '14 at 15:58

I'm actually a grad student on the team at Stanford that's been developing Stanford Dependencies (on which Universal Dependencies is based), so I think I can clarify things for you.

You're assuming that SD and UD are theories, which they are not. They're frameworks for annotating strings of words, and they never claim to be predictive in any way. The goal of generative syntactic theory is to define formally the set of acceptable sentences in a language; that is, you feed the grammar a string, and if it can come up with an acceptable constituency structure for it then the grammar returns "true", and if not, it returns "false". That's it--it makes no claims about psychological plausibility or computational feasibility. SD (and UD) has a different goal--it doesn't try to make claims about psychological plausibility or about grammaticality, but it does try to be useful for computational research. Comparing SD to something like Minimalism is like comparing pens to pencils--it may look like they do the same thing, but there are fundamental differences between them that often render one totally unsuitable to the task at hand.

Our framework actually does draw very heavily from theoretical linguistics, just not from Chomskyan linguistics, looking to LFG for answers instead. In LFG, constituent structure involves relations between tree nodes, like "specifier", "head", and "complement". But these trees don't get mapped directly to semantics--they first get mapped to functional structure, an intermediate level that involves grammatical function relations like "subject", "object", and "clausal complement" (compare these to the UD relations "nsubj", "dobj", and "ccomp"). This bipartite system of syntax makes tons of things easier from a theoretical perspective (Warlpiri is the go-to example). You're mistakenly assuming our framework is a representation of c-structure--where you're certainly right that empirically, functional material tends to be heads--when in fact it aims to be closer to f-structure--where LFG prefers to analyze lexical material as heads.

I should also mention that we at Stanford have a parser that automatically produces UD-style dependency trees, and the way it does this is by first parsing the sentences into tree structures, and then running lots of complicated tregex expressions over those trees to pull out grammatical functions; this is highly reminiscent of how LFG operates. So it actually does make a lot of sense from a theoretical perspective, as long as you see it through the right theory.

SD and UD have really been designed for NLP. NLP people don't care about whether the tools they're using are theoretically well-motivated--they care about whether their model does better than everyone else's, and our dependency representations help them do that. Relation extraction is one example of a computational task that's easy with SD and difficult with a traditional tree grammar. At first glance it seems simple--all you have to do is pull out triples of "predicate(agent,patient)" from a collection of sentences. But, in a tree structure, this gets very messy, because the agent can occur in a wide variety of different tree structure positions. In SD, we abstract away from a lot of this extra noise, which makes the job more manageable for our users (and we even have special relations for passive sentences to make this task even easier).

So, to summarize: SD/USD doesn't make any claims about the existence of phrase structure categories (such as nonfinite VPs or PPs) because a) it's not meant to be theoretically predictive and b) it's meant to represent functional structure, not constituent structure. SD/USD is taken seriously in computational circles because it is easier to apply to practical applications, such as relation extraction, than theoretically motivated constituency structure trees. The USD approach to function words is not a matter of theoretical debate in computational circles, but it is sometimes noted when the choice of hierarchical vs flat structure affects performance. Our SD/USD parser actually does have the option to produce hierarchical structure if the user so desires, where copulas, auxiliaries, and prepositions are heads. But I don't know of any schemes where hierarchical structure is hard-coded in.

I hope that answers the questions you had.

| improve this answer | |
  • Thanks for your response. There are a couple of things you write here in your answer that contradict what you write in your paper linked to above. If you are interested, i would enjoy pointing out where the contradictions lie. In this regard, I encourage you to not reedit your answer here, so that we have something concrete on the record. If you want to add to your answer, fine, but please leave what is here now on the record. And further for the record, I am not a Chomskyan syntactician. I am a DG guy; I think much of Chomskyan syntax is nonsense. – Tim Osborne Feb 8 '15 at 7:21
  • To come to Thomas' comment, you have added some material that answers the first question above to an extent, but the second question above is not addressed. Let me restate that second question: Are there other parsing schemes in NLP circles that reject the USD choice to subordinate function words to content words? If they exist, what are these other parsing schemes? Why do they not choose to subordinate function words to content words? That would be directly relevant to my appearance at the upcoming Depling conference in Uppsala, where the status of function words will be debated at length. – Tim Osborne Feb 8 '15 at 7:29
  • If I understand correctly, you're a theoretical syntactician seeking to develop a dependency formalism that makes the same kind of empirical predictions that phrase structure grammars make. Your question suggested to me that you believed USD aimed to be such a formalism, and I wanted to correct what I saw as a misunderstanding. If our paper suggests something different, let me know and I'll attempt to explain it or raise the issue at our next meeting. However, I'm unable to answer your second question, as I don't know of any such dependency schemes. – Timothy Dozat Feb 8 '15 at 8:13
  • You are indicating a willingness to communicate about the issue. Good. I encourage you to send me an email, so that we can exchange some content without filling up the space here (they will cut us off if the exchange goes too long). My email address is tjo3ya@yahoo.com. Or provide your email address, and I will contact you. If anyone else reading this exchange wants to know where the contradictions are in the USD paper and Timothy's answer here, I will point them out. Contact me via email. The contradictions concern claims about surface syntax, linguistic validity, and f-structure. – Tim Osborne Feb 8 '15 at 8:36

Are you familiar with Sgall's Functional Generative Description (the theoretical framework used in the Prague Dependency Treebank)? He doesn't have synsemantic words in dependency trees and I think this is the right approach. He has the simple rule that in a phrase with an auxiliary or a preposition the autosemantic word is the head and whatever the auxiliary or preposition contributes to the syntactic structure is an attribute of the corresponding node in the tree. In other words, "I saw him" and "I have seen him" have the same dep. tree. Likewise, "Mary loves John" and "María ama a Juan" have the same dep. tree as far as structure is concerned. To sum up, only content words appear in syntax trees. If for some reason one would like to have function words in dependency trees they should be children of the nodes that represent content words (since they modify them, not vice versa).

| improve this answer | |
  • My knowledge of FGD is thin. But my firm stance is that from a syntactic point of view, there is loads and loads and loads of evidence that auxiliary verbs are heads over content verbs. If you are interested in this particularly issue, I suggest attending this conference linguistics.stackexchange.com/questions/8425/…. The issue will be debated extensively at the conference. – Tim Osborne Jan 25 '15 at 1:15
  • @TimOsborne I'm sorry, the link doesn't work. I'd really like to know which conference you meant. As for the original question, in the article you cited there are arguments given against the hypothesis that auxiliary verbs are heads. What's the evidence you mention? – Atamiri Jan 25 '15 at 1:33
  • The conference is Depling 2015 in Upsalla Sweden in August. You must know about this, since you are connected to the DG/FGD people in Prague. The evidence that auxiliary verbs are heads over prepositions is tremendous in my view. I (and at least one of my colleagues) will be in Uppsala presenting our evidence. But I'm not sure what your question is. Are you looking for evidence for or against the stance that the auxiliary verb is the head over the content verb? – Tim Osborne Jan 25 '15 at 8:36
  • If you are looking for evidence that the content verb is head over the auxiliary verb, I cannot produce any. But if you are looking for evidence that the auxiliary verb is head over the content verb, I can and will happily produce tons of that sort of evidence. I'll send you a paper. Just send me a quick email inquiry (tjo3ya@yahoo.com). – Tim Osborne Jan 25 '15 at 8:38
  • Depling is a nice conference, I'm planning to go to Uppsala. – Atamiri Jan 26 '15 at 7:12

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.