Olga Pustylnikov, Alexander Mehler

1
A Unified Database of Dependency
TreebanksIntegrating, Quantifying
EvaluatingDependency Data

• Olga Pustylnikov, Alexander Mehler
• Bielefeld University

2
Motivation

• Exploring similarities among languages by means
  of syntactic treebanks
• We collected a database covering 11 languages
• Treebanks have been developed separately by
  different research projects
• quantitative investigations on these treebanks -gt
  the need for unification

3
Motivation
corpus
structure
annotation
(loves v ( (John n) (Mary n) )
loves
John loves Mary
1 John n 2 2 loves v 0 3 Mary n 2
Mary
John
ltS ID"1"gt ltW DOM"1" ID"2"gt John lt/Wgt ltW
DOM"_root" ID"2"gt loves lt/Wgt ltW DOM"2"
ID"3 Mary lt/Wgt lt/Sgt
4
Motivation
Demands on the unified format of treebanks

• () generic allowing to represent as many
  treebanks as possible
• () extensible to new treebanks
• () complete preserving all corpus specific
  information
• () transferable to other kinds of corpora
• () complex exhibiting the minimal
• complexity
• -gt graph representations

5
Motivation
GXL (Holt et al., 2006)

• Graph eXtensible Language is a graph model
  representig corpora in terms of graphs

XML
Multimodal Data
GXL
TOOLS
eGXL
WIKI
Treebanks
Treebanks

• GXL can be applied to any kinds of corpora. (See
  e.g. Mehler and Gleim (2005), Ferrer i Cancho et
  al. (2007), Pustylnikov and Mehler (2008))

6
Agenda
7
eGXL
2-level data model
Types
ltgraph idTypesgt ltnode idPOS /gt ltnode
idt245 nameVERB /gt lt/graphgt
IDREF
ltgraph id"Sentences"gt ltgraph id"g8"gt ltnode
id"s8_1" form"Detta" pos"t151" /gt ltnode
id"s8_2" form"vill" pos"t245" /gt ...
ltrelgt ltrelend direction"in" target"s8_2"
/gt ltrelend direction"out" target"s8_1" /gt
lt/relgt ... lt/graphgt
Sentences
8
eGXL
2-level data model
Types
ltgraph idTypesgt ltnode idPOS /gt ltnode
idt245 nameVERB /gt lt/graphgt
IDREF
ltgraph id"Sentences"gt ltgraph id"g8"gt ltnode
id"s8_1" form"Detta" pos"t151" /gt ltnode
id"s8_2" form"vill" pos"t245" /gt ...
ltrelgt ltrelend direction"in" target"s8_2"
/gt ltrelend direction"out" target"s8_1" /gt
lt/relgt ... lt/graphgt
Sentences
9
The eGXL Types-graph

• The Types-graph contains treebank specific
  attributes (e.g.POS, morphological attribute
  etc.) -gt nodes
• Each instance of an attribute is given a unique
  identifier

ltgraph idTypesgt ltnode idPOS /gt ltnode
idt245 nameVERB /gt lt/graphgt
a unique identifier
a unique identifier
the value of the attribute
the value of the attribute
10
The eGXL Sentences-graph
vill
.
Detta
bestämt
jag
bemöta
each token of a treebank
each token of a treebank
an IDREF to the POS-node of the Types-graph
an IDREF to the POS-node of the Types-graph
ltgraph id"Sentences"gt ltgraph id"g8"gt ltnode
id"s8_1" form"Detta" pos"t151" /gt ltnode
id"s8_2" form"vill" pos"t245" /gt ...
ltrelgt ltrelend direction"in" target"s8_2"
/gt ltrelend direction"out" target"s8_1" /gt
lt/relgt ... lt/graphgt
word form
word form
a (syntactic) relation
a (syntactic) relation
from (e.g. a head verb)
from (e.g. a head verb)
to (e.g. a dependent argument)
to (e.g. a dependent argument)
11
The eGXL Sentences-graph
vill
.
Detta
bestämt
jag
bemöta
node each token of a treebank
id a unique identifier
form word form
pos an IDREF to the POS-node of the Types-graph
rel a (syntactic) relation
relend a relation anchor
in from (e.g. a head verb)
out to (e.g. a dependent argument)
ltgraph id"Sentences"gt ltgraph id"g8"gt ltnode
id"s8_1" form"Detta" pos"t151" /gt ltnode
id"s8_2" form"vill" pos"t245" /gt ...
ltrelgt ltrelend direction"in" target"s8_2"
/gt ltrelend direction"out" target"s8_1" /gt
lt/relgt ... lt/graphgt
12
eGXL
13
Agenda
14
11 Dependency Treebanks
7 different formats
15
Input vs. Output Formats

• Examples from Dutch, Swedish, Italian treebanks

16
Unification is possible

• due to the separation of the core from the
  secondary parts

ltgraph idTypesgt ltnode idPOS /gt ltnode
idt245 nameVERB /gt lt/graphgt
diversity
ltgraph id"Sentences"gt ltgraph id"g8"gt ltnode
id"s8_1" form"Detta" pos"t151" /gt ltnode
id"s8_2" form"vill" pos"t245" /gt ...
ltrelgt ltrelend direction"in" target"s8_2"
/gt ltrelend direction"out" target"s8_1" /gt
lt/relgt ... lt/graphgt
commonality
17
The TreebankWiki

• http//ariadne.coli.uni-bielefeld.de/wikis/treeban
  kwiki/

18
Agenda
19
Complexity of eGXL

• Logical Scalling Factor (LSF) number of logical
  elements (e.g. XML-element) required to represent
  a treebank unit (e.g. a word form, POS etc.)

node
rel
other
eGXL
other
eGXL
20
Agenda
21
DTDB
22
Agenda
23
Conclusions

• a database covering 11 languages
• eGXL a generic XML graph model adopted to
  syntactic treebanks
• use of treebanks within a single application
  (Ariadne)
• olga.pustylnikov_at_uni-bielefeld.de
• alexander.mehler_at_uni-bielefeld.de
• ruediger.gleim_at_uni-bielefeld.de
• SFB 673
• Thank you for your attention!