Used to determine which sorts of contexts in a dependency parse tree are meaningful and which are not.
It maintains a state variable, patterns, each of can be matched against a different sort of meaningful path in a
dependency parse tree. The patterns are composed of predicate functions which select for either the nodes or edges
in a graphical representation of a dependency parse.
The model used for context is based on the DepDM model described in Baroni, et. al. 2010
This is the description of DepDM from the DM paper:
Our first DM model relies on the classic intuition that dependency paths are a good approximation to semantic
relations between words (Grefenstette 1994; Curran and Moens 2002; Pado and Lapata 2007; Rothenhausler and
Schutze 2009). DepDM is also the model with the least degree of link lexicalization among the three DM instances
we have built (its only lexicalized links are prepositions). DepDM includes the following noun-verb, noun-noun,
and adjective-noun links (in order to select more reliable dependencies and filter out possible parsing errors,
dependencies between words with more than five intervening items were discarded):
- sbj-intr
- sbj-tr
- sbj-tr
- obj
- iobj
- nmod
- coord
- prd
- verb
- preposition