Institute of Formal and Applied Linguistics Wiki

Marie-Catherine de Marneffe, Christopher D. Manning: The Stanford typed dependencies representation, Coling 2008.

1. What do the authors give as a good example of a frequently used linguistic data resource? What reason(s) do they give for its frequent use?
   WordNet. Its design, which is straightforward enough even for non-linguists to understand how to use it.
   (Link Parser and MiniPar are not data resources.)
2. What is more important for SDs: semantic relevance or linguistic accuracy?
   Semantic relevance.
3. The version of SDs described in the paper collapses prepositions. If they were represented as individual nodes instead (which they are in the “basic” version of SDs, not described in the paper), would you expect them to be head nodes or dependent nodes of the prepositional phrase? I.e., which representation of “go to school” seems more appropriate according to the Design Principles?
   • a) prep(go, to); pobj(to, school)
   • b) pobj(go, school); prep(school, to)
     I would expect b), i.e. the preposition as a leaf node, according especially to DP5. (However, it actually is a), although this might change in future versions.)
4. In the variant of SDs described in the paper, each word of the sentence appears as a dependent in a relation:
   • a) exactly once
   • b) once or not at all
   • c) once or more times
   • d) zero or more times
     Does not appear if it is collapsed (e.g. prepositions) or left out (punctuation).
     Appears multiple times if it is a shared modifier of a coordination. (See examples in the paper, or the answer to question 7 where you can see all this happen.)
5. In the variant of SDs described in the paper, each word of the sentence appears as a head in a relation:
   • a) exactly once
   • b) once or not at all
   • c) once or more times
   • d) zero or more times
     Does not appear as a head if it is a leaf or does not appear at all.
     Appears multiple times if it has more than one dependent.
     (Note that this is true for probably all dependency representations…)
6. Based on the paper, what properties do you think that a representation of a sentence in SD has (if you think about it as a graph)?
   (This is actually discussed in detail in Stanford typed dependencies manual, but I wanted you to try to guess the properties based on what you had read.)
   • a) Is it connected?
     Yes, although this might not have been explicitly mentioned in the paper.
   • b) Is it a DAG (directed acyclic graph)?
     The paper probably suggests that yes, although in practice this is not true – SDs capture correfence-like relations, which may form a directed cycle. See page 17 of Stanford typed dependencies manual, which shows this on “the woman who introduced you”.
   • c) Is it rooted?
     Yes, although this might not have been explicitly mentioned in the paper.
   • d) Is it a tree?
     No – e.g. coordinations form undirected cycles. (See examples in the paper, or the answer to question 7 where you can see this happen.)
7. Try to devise the SD representation of the following sentence: Small boys and girls must go to school.

Hint: you should probably use the following SD types: amod, aux, conj_*, nsubj, prep_*, root.
The 'root' relation was probably not mentioned in the paper (so you don't have to use it) – it is used to denote the dependent of the technical ROOT-0 node.
If you want, try to also draw that as a graph representation.

amod(boys-2, Small-1)
amod(girls-4, Small-1)
nsubj(go-6, boys-2)
conj_and(boys-2, girls-4)
nsubj(go-6, girls-4)
aux(go-6, must-5)
root(ROOT-0, go-6)
prep_to(go-6, school-8)

If you try it out (e.g. here http://nlp.stanford.edu:8080/parser/index.jsp),
you may notice that the second relation is missing. However, “Small” is
definitely a shared modifier of the whole coordination “boys and girls”, not a
private modifier of “boys”. Unfortunately, this kind of disambiguation is
usually hard to make for the automated tools (and often even for people).

The 'root' relation was probably not mentioned in the paper.