Domain Driven Disambiguation DDD

1

• Domain Driven Disambiguation (DDD)
• Alfio Gliozzo, Bernardo Magnini, Carlo
  Strapparava.
• gliozzo,magnini, strappa_at_itc.it
• ITC-irst
• http//tcc.itc.it/research/textec/topics/disambig
  uation/

2
Outline 1/2 DDD system architecture

• General Object Oriented Architecture to easily
  develop a great variety of WSD systems
  implementing different Categorization and Feature
  Extraction (FE) techniques.
• Domain Driven Disambiguator (DDD) as an instance
  of the general architecture.

3
Outline 2/2 DDD Methodology

• WordNet Domains
• Textual proprieties of Semantic Domains.
• Text categorization using WN-Domains
• DDD classifier

4
A general WSD System Architecture
Instance (Feat1,Value1) (FeatN,ValueN)
Text

• Corpus
• Reader
• get-text
• (text-id)

• Classifier
• Classify-
• Instance
• (instance)
• Import Model
• (model)
• Build Model
• (instances)

• Feature
• Extractor
• Get-token-
• features
• (text,tok-id)

sense
Sense
Model
Corpus
Supervised Learning
ARFF
5
DDD system

• Is a specific implementation of the general WSD
  system architecture.
• The modules feature extractor and classifier are
  specialized..

WN-D
6
Domain Driven Disambiguation

• Underlying assumption
• The polisemy of terms in domain specific corpora
  tends to disappear.
• Knowing in advance the relevant semantic
  domain(s) of a text makes Word Sense
  Disambiguation easier.
• Semantic domains play an important role in the
  disambiguation process

7
WordNet Domains 1/3(Magnini and Cavaglià,2000)

• Developed at ITC-Irst (Magnini and Cavaglià,
  2000)
• All synsets in WordNet have been annotated by a
  Domain Label (i.e SPORT, ART, ECONOMY, )
• FACTOTUM label is used for generic synsets (not
  belonging to any domain).

8
WordNet Domains 2/3 Polisemy Reduction
9
WordNet Domain 3/3Semantic Domains Organization

• 163 Domain labels collected from dictionaries
• Four level hierarchy (Based on Dewey Decimal
  Classification)
• Mapping between DDC and WND (completeness)
• 41 basic domains used for the experiments (2nd
  level)

10
Domains and Texts (Magnini et al. 2002)

• Many words in a document belongs to the domain of
  the text (statistics on Semcor).
• One Domain per Discourse is stronger than One
  Sense per Discourse (10 vs. 31 exceptions in
  Semcor)

11
TC using WND 1/2 (Magnini et al., 2001)

• Domain frequency is evaluated for each domain
  (excluding FACTOTUM) counting its occurrences
  inside the text.
• Domain relevance is the normalized domain
  frequency value (in the range 0,1) computed on
  windows of fixed size (20-100 tokens).
• Text vectors collects relevance values for each
  domain
• Example ((sport . 0.93) (economy . 0.70) (art
  . 0))

12
Text Categorization using WND2/2 example

From the plush Connolly hide leather sofa and
chairs in the living room to the Bang and Olufsen
stereo, and remote control television complete
with video, you're surrounded by the HIGHEST
QUALITY. The inlaid chequerboard top of the
coffee table houses all kind of games, including
backgammon, chess and Scrabble. You'll also find
a selection of books, from Queen Victoria's
Highland journals, to the very latest bestselling
thriller. The dinner table and chairs are
elegant yet comfortable, and you can be assured
of the finest tableware and crystal for meals at
home.
13
Text Categorization using WND2/2 example
From the plush Connolly hide leather sofa and
chairs in the living room to the Bang and Olufsen
stereo, and remote control television complete
with video, you're surrounded by the HIGHEST
QUALITY. The inlaid chequerboard top of the
coffee table houses all kind of games, including
backgammon, chess and Scrabble. You'll also find
a selection of books, from Queen Victoria's
Highland journals, to the very latest bestselling
thriller. The dinner table and chairs are
elegant yet comfortable, and you can be assured
of the finest tableware and crystal for meals at
home.
14
Text Categorization using WND2/2 example
From the plush Connolly hide leather sofa and
chairs in the living room to the Bang and Olufsen
stereo, and remote control television complete
with video, you're surrounded by the HIGHEST
QUALITY. The inlaid chequerboard top of the
coffee table houses all kind of games, including
backgammon, chess and Scrabble. You'll also find
a selection of books, from Queen Victoria's
Highland journals, to the very latest bestselling
thriller. The dinner table and chairs are
elegant yet comfortable, and you can be assured
of the finest tableware and crystal for meals at
home.
15
Text Categorization using WND2/2 example
From the plush Connolly hide leather sofa and
chairs in the living room to the Bang and Olufsen
stereo, and remote control television complete
with video, you're surrounded by the HIGHEST
QUALITY. The inlaid chequerboard top of the
coffee table houses all kind of games, including
backgammon, chess and Scrabble. You'll also find
a selection of books, from Queen Victoria's
Highland journals, to the very latest bestselling
thriller. The dinner table and chairs are
elegant yet comfortable, and you can be assured
of the finest tableware and crystal for meals at
home.
16
Sense Vector

• Represents the domain/s of the typical contexts
  in which the sense occurs
• DV for bank1 is ((economy . 1.75) (sport . 0.2)
  (law . 0) ))
• DV for a generic sense (FACTOTUM) is uniform
• Can be evaluated from
• Sense Tagged Corpora summing the Text Vectors of
  the contexts of the examples (supervised)
• WordNet-Domains (unsupervised)

17
DDD Classifier

• Two steps
• Context Categorization (text vector)
• Similarity between synsets and contexts vectors
• Example
• Bank1 depository financial institution ...
• Bank2 sloping land
• TEXT He cashed a check at the bank

Dot Product
1,731878
0,06185
18
IRST results at SENSEVAL-2
19
Domains and Texts (Magnini et al. 2002)

• Many words in a document belongs to the domain of
  the text (statistics on Semcor).

20
Conclusions and Future Works

• DDD architecture is an experimental framework to
  easily design and develop several WSD techniques.
• At the moment the Domain Oriented Methodology has
  been (partially) explored and implemented.
• We plan to improve the system using machine
  learning techniques (i.e Support Vector Machines,
  AdaBoost, )
• We plan to integrate the domain approach with
  other WSD techniques to take into account also
  syntactic information.

21
Other directions

• Acquisition of domain information from Text
  categorized corpora (Magnini et al., 2002a)
• Investigations on the connections polisemy,
  synonimy and domains
• Using domain information in order to improve and
  develop ML and memory based TC and IR techniques
• WDD

22
References 1/2

• (Magnini and Cavaglià, 2000) Bernardo Magnini
  and Gabriela Cavaglia'. Integrating Subject Field
  Codes into WordNet. In Gavrilidou M., Crayannis
  G., Markantonatu S., Piperidis S. and Stainhaouer
  G. (Eds.) Proceedings of LREC-2000, Second
  International Conference on Language Resources
  and Evaluation, Athens, Greece, 31 MAY- 2 JUNE
  2000, pp. 1413-1418.
• (Magnini and Strapparava, 2000) Bernardo Magnini
  and Carlo Strapparava. Experiments in Word Domain
  Disambiguation for Parallel Texts. Proceedings of
  the ACL workshop on Word Senses and
  Multilinguality, pag. 27-33, October 7, 2000,
  Hong Kong.
• (Magnini and Strapparava,2001) Bernardo Magnini
  and Carlo Strapparava. Using WordNet to Improve
  User Modelling in a Web Document Recommender
  System. Proceedings of the NAACL Workshop,
  "WordNet and Other Lexical Resources
  Applications, Extensions and Customizations",
  Pittsburgh, June 3-4, pp. 132-137, 2001. Also
  ITC-irst Technical Report Ref. No. 0106-02.

23
References 2/2

• (Magnini et al.,2001) Bernardo Magnini, Carlo
  Strapparava, Giovanni Pezzulo and Alfio Gliozzo.
  Using Domain Information for Word Sense
  Disambiguation. Proceedings of Senseval-2, Second
  international Workshop on Evaluationg Word Sense
  Disambiguation Systems, pag. 111-114, 5-6 July,
  2001, Toulose, France.
• (Magnini et al., 2002a) Bernardo Magnini, Carlo
  Strapparava, Giovanni Pezzulo and Alfio Gliozzo.
  Comparing ontology-Based and Corpus-Based Domain
  Annotations in WordNet in Proceedings of the
  First Global WordNet Conference, Mysore(India),
  2002.
• (Magnini et al.,2002b) B. Magnini, C.
  Strapparava, G. Pezzulo, A. Gliozzo (2002) "The
  Role of Domain Information in Word Sense
  Disambiguation". to appear in Special Issue of
  Journal of Natural Language Engineering on
  "Evaluating Word Sense Disambiguation Systems".
• (Gliozzo, 2001) Alfio Gliozzo, Ruolo dei campi
  semantici nella struttura di un lessico
  computazionale utilizzo per la disambiguazione
  automatica di senso Thesis, University of
  Bologna, 2001

24

• Domain Driven Disambiguation (DDD)
• Alfio Gliozzo, Bernardo Magnini, Carlo
  Strapparava.
• gliozzo,magnini, strappa_at_itc.it
• ITC-irst
• http//tcc.itc.it/research/textec/topics/disambig
  uation/