|
My go-to reference to nlp, Jurafsky and Martin's book, fails to mention what exactly is a projective dependency/phrase structure grammar, or what is a projective parser, and why is it different (worse?) than a non-projective one. Where can I see this clearly explained? |
|
Non-projective dependency trees are ones with crossing edges. They allow for more constructions to represented in your parse than the projectivity constraint, and is prevalent in languages with free word order. Look at Figure 2 of McDonald et al (2005) "Non-projective dependency parsing using spanning tree algorithms". (Actually, this is a seminal work, and it started the whole MST-based approach to dependency parsing. I encourage you to read it.) From their work:
1
Just a quick historic note: the non-projective dependency parsing as graph search was described already at 1998 for parsing Czech. The parameter estimation was poor (count-and-divide, like most work at that time period), and the decoding was a greedy-heuristic instead of the exact CLE MST search (no good excuse for that one..), but the principle components are there.
(Oct 10 '10 at 21:05)
yoavg
A small, loosely related question: what sort of feature is commonly used in a dependency parser? I didn't find even a rough list in the paper, and it doesn't seem obvious to me which way to go.
(Oct 11 '10 at 14:26)
Alexandre Passos ♦
For a graph-based parser, you can find a good list in Ryan's thesis (page 51). People are doing more than that these days, but this is the core feature set and it is quite hard to beat. For a transition based parser, there is a good list in Huang et.al. 2009. You can also check out the feature list of my easy-first parser.
(Oct 11 '10 at 16:42)
yoavg
|
|
To add to the discussion, non-projectivity is dependent mostly on the way it is represented. As of now, the present theory of dependency parsing might poke you towards several cases of non-projectivity - partially non-projective, strictly non-projective, etc.. If you are looking into the concept of non-projectivity, then the above definition is sufficient. But if you are deeply interested in it, I would suggest you go through an indepth analysis here: "Marco Kuhlmann, Giorgio Satta: Treebank Grammar Techniques for Non-Projective Dependency Parsing. EACL 2009: 478-486" |
