Multilingual Lexical Resources application requirements

1
Multilingual Lexical Resourcesapplication
requirements

• Gr. Thurmair
• 3-6-2000

2
Outline

• 1 Multilingual Applications
• 2 Lexicon Requirements
• 3 Paradigms of ML resources
• 4 Synthesis proposal
• 5 Challenges and actions

3
ML applications (1)

• 1. Cross-Lingual Information Retrieval
• multilingual Concept Net
• words with representation of senses
• words with links to concepts
• ontology relations between concepts
• multilingual equivalents between concepts
• gt ML Concept Net, EuroWordNet

4
ML Applications (2)

• 2. ML Information Extraction
• Domain objects with special semantics
• person - river - company designator - drug
• Application specific semantic frames
• (immigrate -gt agent ltpersongt)
• Ontology (domain model) for objects
• multilingual Links to concept nodes
• (concepticon)

5
ML Applications (3)

• 3. Multilingual Key Topic identificationand
  classification
• key topic identifier (key terms / key objects)
• key topic translation
• standard terminology lookup tool

6
ML Applications (4)

• 4. Machine Translation
• full morphosyntactic coding of mono entries
• usually entry-based, not reading-based
• transfer lexicons
• bilingual, directed
• linguistic tests and actions
• sense disambiguation as part of
  transferselection (subject area codes)

7
Requirements for ML lexicons

• ML lexicons must support
• some monolingual descriptions
• morphosyntax, word senses, pragmatics
  (application-specific issues)
• monolingual hierarchies
• conceptual hierarchies / ontologies
• subject area classifications
• multilingual links between the monos

8
Paradigms of ML resources

• 2 Paradigms are implemented
• word-to-word links in translation technology
• concept-to-concept links in ML concept nets

9
Paradigms of ML resources

• Concept-to-concept links
• are more general, support recall
• have broader language coverage
• are not sufficient for translation
• Word-to-work links
• are more specific, cover meaning nuances
• support precision
• cost more effort to maintain
• no empty transfers must occur

10
Synthesis Proposal (1)

• Combine word and concept links
• Base the resource on word-to-word links
• Promote some words to concepts
• for concepts, word-to-word links are identical
  to concept-to-concept links
• for synset members, use
• word-to-word transfer if appropriate and
  available
• conceptual transfer otherwise
• give attributes to links

11
Synthesis Proposal (2)

• Keep synset relationships for IR applications
• Keep transfer relations for MT applications

12
Synthesis Proposal (3)

• Consequences
• we save all 11 concept - synset links
• (this is the majority) (kill -gt ltkillgt)
• all concept nodes must also be words
• (can of course be multiwords)
• we translate on both general and specific level
• we can offer default translations
• in cases no word-to-word links are available

13
Consequences (1)

• Design a database to support this
• monolingual entries with annotations
• transfer links for these entries
• multilingual, directed
• cross-reference links between entries
• on word level abb_for, forbidden_for
• between words and concepts (synset)
• between concepts (hierarchical)
• attributes for the links

14
Consequences (1)

• Entry
• canonical_form char 40language intpart_of_sp
  eech charreading_number charsubject_area char
  concept_type word - term - concept
• TransferLink
• EQ-type full - partial - none
• CrossrefLinkLinkstatus t-t, t-c,
  c-cLinktype hyponym, ....

15
Consequences (2)

• Convert existing resources
• MT lexicons are entry-based
• entries can have several readings
• entries have several translations
• kill -gt töten (gv)
• -gt abbrechen (dp)
• Concept nets / term bases are concept-based
• kill_1 -gt töten
• kill_2 -gt abbrechen_1
• interrupt_3 lt- abbrechen_2

16
Consequences (2)
CN -gt MT merge readings MT -gt CN create
readings (SL and TL!)
17
Consequences for MT systems

• Lexical resources
• same mono lexicons used in many language
  pairs(impact on other languages?)
• Combinatorics
• ambiguous words cause combinatoric problemsin
  parsing (impact on performance?)
• Disambiguation Strategies
• disambiguaiton in analysis, not transfer(impact
  on output?)

18
Thank you for your attention