1

Whenever we have a Prepositional Phrase as a subject, how should the dependency relations hold? Specifically, consider the following sentence:

Before Wednesday does not work for me.

One dependency parse could be:

ROOT(work)
NEG(work, not)
AUX(work, does)
NSUBJ(work, Wednesday)
PREP(Wednesday, Before)

PREP(work, for)
POBJ(for, me)

Another could be

ROOT(work)
NEG(work, not)
AUX(work, does)
PREP(work, Before)
POBJ(Before, Wednesday)

PREP(work, for)
POBJ(for, me)

Basically, should "Wednesday" be the subject for "work" and and "Before" the prepositional modifier for "Wednesday" Or "Before Wednesday" is the prepositional modifier for "work" and "Wednesday" is the prepositional object for "Before".

I've tried the Stanford Dependency Parser (http://nlp.stanford.edu:8080/parser/index.jsp), the Spacy Dependency Parser (https://explosion.ai/demos/displacy) and AllenNLP's dependency parser (https://demo.allennlp.org/dependency-parsing), but they all give inconsistent results.

Would be grateful for any help !

1

The first dependency grammar (DG) parse given in the question looks like this:

enter image description here

The second DG parse given in the question looks like this:

enter image description here

Neither of those two parses is linguistically well-motivated, although the second is somewhat better than the first. There is another parse that is much better than both of those. This other parse is next:

enter image description here

This parse sees auxiliary verbs dominating content verbs, and prepositions dominating their noun phrases. These assumptions about dependency structures are well-motivated linguistically. The matter is discussed at length in this recent article in Glossa.

The practice of positioning function words as dependents of content words, which is what you see in the first two parses to varying degrees, is something that arose at Stanford in the early 2000s. That trend continues today, most prominently in the form of the Universal Dependencies (UD) annotation scheme. However, a lengthy tradition in the DG community -- a tradition reaching back about four decades -- produces parses along the lines of the third tree, which views most function words as heads over the co-occurring content words. The history of this issue is discussed in these conference proceedings.

This issue currently splits the DG community. Theoretical DG people like myself argue vehemently in favor of the third parse above, which views most function words as heads over the co-occurring content words. Computational DG people are split into two camps concerning the matter. The one camp is represented by the annotation scheme of the UD project, which is already linked to above. The other camp of computational DG people is more in line with theoreticians insofar as they prefer parses along the lines of the third tree. See the Surface-Syntactic Universal Dependencies (SUD) webpage in this regard.

Turning to the example sentence from the question, it challenges theoretical analyses regardless of the annotation choices. The problem it poses is that it appears to lack a subject phrase, for there is no overt subject noun phrase present. What seems to be occurring in such cases, however, is that a nominal of some sort is elided. While the nature of the ellipsis mechanism involved is not immediately evident, one might posit the existence of an elided gerund, e.g.

enter image description here

The conditions that must be met for this type of ellipsis to occur require investigation. I cannot say anything more about the potential of such an account.

As a final comment, there are of course various annotation schemes out there that can be accessed. I think each linguist needs to be aware of what is at stake when choosing an annotation scheme and decide for him- or herself. If the annotation scheme should be linguistically solid, then parses along the lines of the third tree above are the best choice.

| improve this answer | |
  • huge answer with a downvote but no comment as to the downvote. i think it's fair i upvote without even reading. – vectory Apr 26 at 18:40
  • @vectory FYI: Comments are not mandatory for downvoting. Or upvoting. – Alenanno Apr 28 at 17:38
1

I am part of the team that developed the SUD syntactic format mentioned by Tim in his answer. The other participants are Sylvain Kahane, Kim Gerdès and Bruno Guillaume. The third analysis presented by Tim is exactly the syntactic annotation of the sentence in SUD. I think it is not necessary to introduce an ellipsis to be consistent with the usual syntactic category of a subject. This issue is similar to title parsing. Titles can have any syntactic category but they behave like proper nouns with respect to the outside. In SUD, we consider that title heads have an internal category in relation to the other elements of the title and an external category PROPN in relation to the outside. We also use this notion of external and internal category with multi-word expressions.

| 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.