Week5-7Dependency parsing

Dependency structure

• blue
  
  • modifier, dependent, child, subordinate
• house
  
  • head, governor, parent, regent

• Phrase structure

• Dependency structure
  

Dependency grammar

• Characteristics
  
  • lexical/syntactic dependencies between words
  • top-level predicate of the sentence is the root
  • simpler to parse than CFG
  • particularly useful for free word order languages(Latin)

How to identify the head

• H=head, M=modifier
  
  • H determines the syntactic category of the construct
  • H determines the semantic category of the construct
  • H is required, M can be skipped( house is a must, but blue can be skipped)
  • Fixed linear position of M with respect to H

Head rules from Collins

Techniques

• Dynamic Programming
  
  • CKY - similar to lexicalized PCFG, $O(n^3)$
    
• Constraint based methods
  
  • Maruyama 1990, Karlsson 1990
  • NP complete problem; heuristics need
• Constraint Graph
• Deterministic Parsing
  
  • Covington 2001
  • Maltparser by Nivre
    
    • Similar to shift-reduce parser
• Graph based methods
  
  • Maximum spanning trees(MCDonald et al.)

Non-projectivity

If you allow your parse tree to have crossings as shown at the bottom of the sentence, the parse tree is non-projective parse tree.

Dependency parsing

• Background
  
  • McDonald et al. 2005
• Projectivity
  
  • English dependency trees are mostly projective (can be drawn without crossing dependencies)
• Idea
  
  • Dependency parsing is equivalent to search for a maximum spanning tree in a directed graph.

MaltParser

• It includes following components
  
  • a stack
  • a buffer
  • set of dependencies (arcs)
• There is also arc-eager parser having different actions

Example

• There is no search, the next action is chosen using a classifier.
• The final list of arcs is returned as the dependency tree.

Evaluation metric

• Labelled dependency accuracy
  
  • # correct deps / # deps

Complexity

• Projective (CKY) $O(n^5)$
• Projective (Eisner) $O(n^3)$
• Non-projective (MST- Chu-Liu-Edmonds) $O(n^2)$
• Projective (Malt) $O(n)$

Application

Use in Information Extraction

Dependency kernels

Decides how similar 2 sentences are based on how similar their dependency structures are.