By David Schueler, Computational Linguist
The ‘ncomp’ dependency label
We at Codeq recently undertook a major revision to the corpus we use to train our dependency parser. We corrected some obvious annotator errors and enforced consistency of conventions of parsing for various patterns. As part of this process we decided to introduce a new dependency label to encode certain relations between syntactic heads and their dependents. The original labels we start from come from De Marneffe et al. 2008; when I mention modifications to the original set of dependency labels in this article, it should be understood as modifications to the system presented in De Marneffe et al. 2008.
The new label is
ncomp, which designates a noun phrase complement dependency type. The new label is parallel to the existing labels
acomp, which designates adjective phrase complements; and
xcomp, which designate different types of clausal complements. (It is also reminiscent of the label
pcomp, though that label is used a bit differently, to designate certain complements of prepositions, rather than complements which are prepositional phrases, the latter designated by the label
ncomp is used for cases where a noun phrase does not fit into existing dependency types that are used for noun phrases, such as
dobj (direct object),
iobj (indirect object), or
npadvmod (noun phrase adverbial).
A common instance of the use of
ncomp , though not the only one, is to indicate that a noun phrase is the second argument of the verb ‘be’. For instance, in (1), the noun phrase “a plaster saint” forms an
ncomp dependency of the form of the verb ‘be’, in this case the contracted form ‘’m’, as parsed in figure 1. (Note that the sentences are given in tokenized form.)
(1) I ’m not a plaster saint .
ncompis also used with the verb ‘become’, as in sentence (2), where “the new king” is an
ncomp dependent of ‘become’, as parsed in figure 2.
(2) His son Pekahiah became the new king after him .
We believe that the introduction of
ncomp has increased the consistency of the data used to train our parser. We believe that this further translates to a better application of the various dependency labels to the relations the parser will find in new data.
De Marneffe, M.-C. and Manning, C. D. (2008). Stanford typed dependencies manual. Technical report, Technical report, Stanford University.