Arabic Syntactic Trees

1
Arabic Syntactic Trees
from Constituency to Dependency

• Zdenek abokrtský
• Otakar Smr
• Center for Computational Linguistics
• Faculty of Mathematics and Physics
• Charles University in Prague

2
Motivation Background

• Linguistic Data Consortium Arabic Treebank
• Constituent-syntax bracketing 100k words
  published
• Modification from English to Arabic
• Prague Arabic Dependency Treebank
• Dependency approach to syntax 50k words in
  progress
• Pre-step to tectogrammatical description
• Motivation co-operation and resource exchange
• Our goal transform the data from one annotation
  scheme to the other

3
Constituency X Dependency

• Non-terminal nodes Text tokens
• Constituent labeling on non-terminals
• Slots and traces
• Linguistic Data Consortium, University of
  Pennsylvania

• Sentence root node Text tokens
• Analytical function for every tree node
• Government and roles
• CCL IFAL ICL, Charles University in Prague

4
Model Arabic Phrase I

• Trace of the antecedent subject

• Compound function of the head of the clause
  outer and inner perspectives
• Free word-order compliant

5
Outline of the Transformation

• 1. Build temporary dependency tree
• Contraction of the input phrase-structure tree
• Uniquely determined by head selection function
• Implementation simple recursive procedure
• 2. Create analytical tree topology
• Post-processing (corrections) of the temporary
  dep. tree, e.g., substituting traces with trace
  coindexed fillers
• Re-arrangement of special complex constructs
• 3. Assign analytical functions

6
Head Selection Function

• For each constituent, select the head constituent
  among its children
• Based on (ordered) handcrafted rules
• Examples
• If there is a node with tagPREP among the
  children, then it is the head
• If there is a node with phrase_labelVP among the
  children, then it is the head
• ... etc ...
• If nothing was selected by the rules, then the
  rightmost child is selected

7
Analytical Function Assignment

• Based on (ordered) handcrafted rules and lexical
  lists
• Completes the process, does not override previous
  assignments
• Examples
• phrase_labelNP-SBJ ? afunSb
• lemmawa- ? afunCoord
• pos_tagCONJ ? afunAuxC
• ... etc ...

8
Model Arabic Phrase II

• Sister-like co-ordination

• Conjunction of co-ordination
• Status constructus

9
Model Arabic Phrase III

• Non-expressed subject (?)
• Complex modality constructs
• Principal discrepancies between descriptions
  both in topology and labeling

10
Model Arabic Sentence

• Wa lam yakun mina s-sahli alay hi muwagahatu
  kamirati t-tilfizyuni wa adasati l-musawwirina
  wa huwa yasadu l-basa.
• It was not easy for him to face the television
  cameras and the lenses of photographers as he was
  getting on the bus.

11
Constituency Annotation
12
Dependency Annotation
13
Evaluation Conclusion

• Implementation still in progress, fine-tuning
  needed
• 10,000 words manually annotated in both styles
• 60 of correctly aimed dependencies
• 2nd Prague Penn Arabic Treebanking Workshop, May
  2003 in Prague
• Transfer from dependency to constituency?

14
Related Work

• New tool for assignment of analytical functions
• Based on machine learning (C5-trained decision
  trees)
• Error rate 17 (supposing the topology of the
  tree is correct)
• First experiments with Arabic dependency parser
• Incorporated into the process of annotation of
  Prague Arabic Dependency Treebank