Prague Arabic Dependency Treebank

1
Prague Arabic DependencyTreebank
Development in Data and Tools
Jan Hajic Otakar SmržPetr ZemánekJan
ŠnaidaufEmanuel Beška
Faculty of Mathematics and Physics Faculty of
Philosophy and Arts Charles University in Prague
2
Project Release PADT 1.0

• December 2004, Linguistic Data Consortium
• 148 000 Morpho, 113 500 Syntax

AFP 13 000 N/A France Presse Penn ATB 1
UMH 38 500 N/A Ummah Press Penn ATB 2
XIN 13 500 N/A Xinhua News A Gigaword
ALH 10 000 73 500 Al-Hayat News A Gigaword
ANN 12 500 25 500 An-Nahar News A Gigaword
XIA 26 500 49 500 Xinhua News A Gigaword
3
Open-Source Tools

• TrEd Tree Editor
• Multi-purpose annotation environment
• Suite of programming utilities
• Netgraph Search Engine
• Server/Client system architecture
• Easy-to-learn query language
• EncodeArabic Perl Module
• Extension for processing of Arabic script
• ArabTeX, Buckwalter, Unicode,

4
PADT Functional Views

• Functional Generative Description
• Theory of linguistic meaning and its expression
• Prague Dependency Treebank for Czech
• Independence of representation levels
• Tectogrammatical linguistic meaning
• Analytical surface dependency syntax
• Morphological categories and lexical units
• Abstraction of the relations across levels
• Strict distinction between form and function
• Different units of description on each level

5
Functional Morphology

• Provides syntax levels with their abstract
  language, not just giving letters in tokens
• Revives multiple senses of categories
• Completeness of generation
• Strict modeling of grammatical control
• MorphoTrees human tagging
• Successful prototype feature-based tagger

6
Syntactic Levels of Description

• Analytical level
• Pragmatically motivated, close to surface syntax
• Every single token resulting frommorphological
  level forms one node
• Tree-like dependency structure for every sentence
• Tectogrammatical level
• Linguistic (literal) meaning, deep relations, TFA
• Initial structures transformed from AL
• Nodes for autosemantic words only
• Decisive role of valency frames

7
Logic of Analytical Trees

• Concepts of dependency and valency
• Reduction sentence must retain grammatical
  correctness if leaves(terminal nodes) are
  chopped off

• Trees clause components ? clauses ? sentences ?
  paragraphs etc.Subtrees of clauses exchangeable
  for non-clauses
• Nodes words, tokenized parts of words,
  punctuation marks marked by functions
• Edges syntactic relations governing node ?
  dependent node/subtree

8
Some Syntax Issues of Arabic

• Non-verbal predication of several types
• Subordinate non-verbal clauses / modification
• Verb-like behavior of many nominal forms
• Mostly VSO in verbal sentences, but
• vice-versa in non-verbal clauses
• different, depending on context boundness
• Compound verbs, fixed composite prepositions
• Grammatical co-reference, accusative ofinner
  object, complex referencing, etc.

9
Problem I Predication

• Head node of tree PREDICATE
• Why? Steady role in sentence, cannot be omitted
• Verbal predicate I-go to school
• Non-verbal predicate
• Nominal The-house a-big (the house is big)
• Existential There a-city (there is a city)
• Prepositional
• Possessive For him a-house (he has a house)
• Adverbial The-mosque in the-city (is)
• Conjunctional The-problem that (is that)

10
Predication Types in Trees
Verbal
Nominal
dAma Pred lasted
kabIrun Pnom a-big nom.
iqtirAHu Sb proposal
sAEatayni Adv two-hours acc.
Prepositional(possessive)
al-baytu Sb the-house nom.
Existential
-hu Atr his
al-EamalIyata Obj the-operation acc.
EalA AuxP on
vamata PredE there-is
zumalAi Obj colleagues
Prepositional(adverbial, locative)
la- PredP for
madInatun Sb a-city nom.
-hi Atr his
Verb-like behavior (object of noun?)
fI PredP in
-hu Obj him
baytun Sb a-house nom.
al-madInati Adv the-city gen.
al-jAmiEu Sb the-mosque nom.
11
Problem II Clauses Co-reference

• Recursiveness subordinate clause is con-tained
  as subtree in place of simple element
• Head-node of clause gets the same function
• Problem non-verbal structures clauses or not?
• Compound verbs (mA zAla etc.) treated equally
• Grammatical co-reference Personal pro- noun
  formally required by another element
• Pronoun must be marked to be treated as such
• Target of reference is unambiguously identifiable
• Often in subordinate clauses, mostly
  attributiveEx. He-wrote a-book number its-pages
  hundred

12
Clauses Co-reference in Trees
Compound verb, formed as main verb and its
complement
Attributive clause, prepositional predicate
(adverbial)
zAlat Pred she-stopped
kataba Pred he-wrote
kitAban Obj a-book
mA AuxM not
Objective clause, verbal predicate
tuHisu Atv she-feels
al-rajulu Sb the-man nom.
fI Atr_PredP in
zaybabu Sb Zaynab
Attributive clause, nominal predicate
miatu Sb hundred nom.
Referencing pronoun, as attribute in clause
anna AuxC that
-hi Adv_Ref it
tuEjibu Obj_Pred they-impress
SafHatin Atr pages gen.
jumalan Sb sentences acc.
Referencing pronoun, as adverbial in clause
wADiHun Atr_Pnom clear nom.
naHwu Sb grammar nom.
-hA Obj her
-hA Atr_Ref their
13
Future Prospects

• Implementation of Functional Morphology
• Tectogrammatical annotation
• Lexicons of valency frames
• Re-training the feature-based tagger on
  MorphoTrees
• Machine-learning on the treebank data for various
  purposes

14
Thank you

• Questions welcome!
• http//ckl.mff.cuni.cz/padt/