The CrossLingual Reuse and Extension of Knowledge Resources in Ontological Semantics

1
The Cross-Lingual Reuse and Extension of
Knowledge Resources in Ontological Semantics

• Marjorie McShane, Sergei Nirenburg, Stephen
  Beale, Margalit Zabludowski
• The Institute for Language and Information
  Technologies (ILIT)
• University of Maryland Baltimore County

2
Plan of the Talk

• Background
• the OntoSem environment
• the OntoSem processors and their parameterization
  potential
• language-independent and language-dependent
  knowledge resources
• Focus of the talk the benefits and challenges of
  using our current, English, lexicon to seed the
  building of lexicons for other languages
• For comparison (one of many possible) the
  European Unions SIMPLE lexicon-building project

3
What is OntoSem?

• Ontological Semantic (OntoSem) text processing
  takes as input open text and returns text-meaning
  representations (TMRs), which are structured
  representations of its meaning (Nirenburg and
  Raskin, Ontological Semantics, MIT Press,
  forthcoming).
• TMRs are the basis for all other applications
  MT, QA, Knowledge Extraction
• OntoSem is a practical, non-toy system that
  concentrates on encoding and interpreting text
  meaning (in contrast with most stochastic methods
  that focus on surface strings)

4
OntoSem Processors

• Processors include
• pre-processor (tokenization, POS tagging,
  morphological analysis, etc.)
• syntactic analyzer
• semantic analyzer
• All processors are parametrizable i.e., basic
  functionality can apply to any language, but
  certain language-specific information must be
  recorded
• Recent experiments port analyzers to Arabic and
  Persian (CRL) successful for small experiment

5
Language Independent Knowledge Sources

• the TMR language
• an ontology, currently containing 5500 concepts
  designed to support many-to-one lexical mappings
• a fact repository, which is a database of
  real-world facts that represent instances of
  ontological concepts
• Brief descriptions of each follow.

6
Static Resources 1 The TMR Language

• He asked the UN to authorize the war.
• REQUEST-ACTION-69   AGENT HUMAN-72
  THEME ACCEPT-70   BENEFICIARY
  ORGANIZATION-71   SOURCE-ROOT-WORD ask
  TIME (lt (FIND-ANCHOR-TIME)) ACCEPT-70  
  THEME WAR-73   THEME-OF REQUEST-ACTION-69
    SOURCE-ROOT-WORD authorizeORGANIZATION-71
    HAS-NAME United-Nations  BENEFICIARY-OF
  REQUEST-ACTION-69   SOURCE-ROOT-WORD
  UNHUMAN-72   HAS-NAME Colin Powell 
  AGENT-OF REQUEST-ACTION-69 SOURCE-ROOT-WORD
  he reference resolution has been carried
  outWAR-73   THEME-OF ACCEPT-70
    SOURCE-ROOT-WORD war

7
Static Resources 2 The Ontology
8
Ontology (cont) Why our ontology is not just a
word net it contains many properties and their
values(an average of 16 per concept), scripts,
etc.

• BIOLOGICAL-WEAPON
• MADE-OF MICRO-ORGANISM
• INSTRUMENT-OF DESTROY, ANIMAL-DISEASE,
• PLANT-DISEASE
• GUN
• INSTRUMENT-OF DISCHARGE, HUNTING-EVENT
• PRODUCED-BY GUNSMITH
• MADE-OF METAL (PLASTIC, WOOD)
• COLON-CANCER
• LOCATION COLON
• ESTABLISHED-BY COLONOSCOPY, SIGMOIDOSCOPY
• HAS-SYMPTOM CONSTIPATION
• REMEDIED-BY DRUG, PERFORM-SURGERY
• EXPERIENCER ANIMAL

9
Static Resources 3 Fact Repository
10
Language Dependent Knowledge Sources

• Lexicons and onomastica (lexicons of proper
  names) are language dependent
• For lexicons, however, the semantic
  representation is largely transferable across
  languages, and the syntactic patterns that
  realize a given semantic meaning often are as
  well this is the source of the cross-linguistic
  portability well explore here

11
Example of a Basic Lexicon Entry

• watch
• watch-v1
• synonyms observe
• anno
• definition to observe, look at
• example Hes watching the demolition
  team.
• syn-struc
• subject var1
• v var0
• directobject var2
• sem-struc
• VOLUNTARY-VISUAL-EVENT
• agent var1
• theme var2

12
Methods of Expressing Word / Phrase Meaning in
OntoSem

• mapping directly to an ontological concept (dog
  maps to DOG)
• mapping to an ontological concept with
  modification by properties e.g.,
• Zionist maps to POLITICAL-ROLE
• AGENT-OF SUPPORT
• THEME Israel
• asphalt (v.) maps to COVER
• INSTRUMENT ASPHALT
• recall (v., as in They recalled the high chairs)
  maps to
• RETURN-OBJECT
• THEME ARTIFACT,
  INGESTIBLE, MATERIAL
• CAUSED-BY
  FOR-PROFIT-CORPORATION

13
Methods of Expressing Meaning (Cont.)

• using modality or aspect
• (might-aux1
• (def "expresses the possibility of something
  happening - epistemic .5")
• (ex "he might come over")
• (syn-struc
• ((subject ((root var1) (cat n)))
• (root var0) (cat v)
• (inf-cl ((root var2) (cat v)))))
• (sem-struc
• (var2
• (epistemic .5)
• (agent (value var1)))))
• (meaning-procedure
• (fix-case-role (value var1) (value
  var2))))

14
Methods of Expressing Meaning (Cont.)

• 4. using our non-ontological methods of
  expressing time, sets, etc
• (yesterday-adv1
• (def the day before the speech time")
• (ex "he admitted that yesterday")
• (syn-struc
• ((root var1) (cat v)
• (mods ((root var0) (cat adv) (type
  pre-verb-post-clause)))))
• (sem-struc
• (var1
• (time (combine-time (find-anchor-time)
  (day 1) before)))))

15
Methods of Expressing Meaning (Cont.)

• 5. calling a meaning procedure
• (she-pro1
• (def "the pronoun 'she'") (ex "she kicked
  the can.")
• (syn-struc
• ((root var0) (cat n) (type pro)))
• (sem-struc
• (animal))
• (meaning-procedure
• (trigger-reference
• (person third) (number sing) (gender
  female)
• (same-clause .1) (preceding-clause
  .7) (pre-preceding-clause .5)
  (preceding-sent .5) (sentence-minus-2 .2)
  (sentence-minus-3 .1) (para-break .5)
  (repeat-collocation .7) (synonym-collocation
  .6)
• (agent-theme .8) (pp-embedded .2)
  (function-match .7) (coord .7))))

16
Why Port Semantic Representations?

• Once a semantic representation of a word sense
  has been created along with the concurrent
  extension, if necessary, of other resources it
  not only can but should be used to represent the
  same sense in any language. Why?
• Time (we want to save it!)
• Paraphrase (weapons of mass destruction a)
  weapons that can kill more than people b)
  nuclear and/or bio weapons)
• Options for Resource Development (e.g., one can
  ontologize a fine-grained notion or describe its
  properties in the lexicon either is fine)

17
A Driving Principle The Principle of Practical
Effability

• What can be expressed in one language can be
  expressed in every language so the sem-strucs,
  by definition, must be portable (apart from their
  variables)

18
What Can Be Involved in Editing a Word/Phrase
Sense for a New L (on the example of Polish)

• No modification required, just a new translation
  dog gt pies. This may rely on global syntactic
  rules for parameterization e.g., subject in
  English maps to Nominative case in Polish as a
  default so a basic transitive frame in English
  can generally map to a basic transitive frame in
  Polish
• Manual syntactic modification required e.g., an
  object in Polish can have quirky case-marking an
  xcomp in English might be realized as a comp in
  Polish a category that is optional in English
  might be required in Polish or vice versa, etc.
• Linking of variables might be different in
  different languages
• Semantic distinctions in one language that are
  missing in another (e.g., the English hand/arm
  distinction is missing in many languages)

19
From Porting Senses to Porting a Whole Lexicon

• Porting senses is fairly straight forward one
  can provide as many translations of a given sense
  as needed using the synonyms field or a new
  entry, if there are any syntactic differences or
  semantic nuances to be capture translations can
  be words or phrases of any complexity
• However, porting a whole lexicon introduces
  difficulties of a more organizational nature, to
  which we turn now

20
Organizational Issues

• Leave the base lexicon as is or attempt to
  improve its quality while building L2 (e.g., add
  more distinguishing properties and values)?
• Expand lexicons simultaneously (e.g., add more
  senses to English words and their corresponding
  L2 equivalents add more words in general)?
• Be driven by correspondences in head words or
  simply by sem-struc meanings? (e.g., all English
  senses of table will be in one head entry should
  all senses of all L2 translations of table be
  handled at once during L2 acquisition?)
• To what extent should the regular acquisition
  process including ontology supplementation be
  carried out on L2?
• Automate? If so, how and how much?

21
Insight from an Experiment

• We attempted to do a fast port of part of the
  English lexicon to Polish to determine time
  savings, problems, automation potential
• The experiment was by carried out by one
  bilingual English and Polish speaker working for
  about a week.
• A portion of the lexicon was ported,
  problems/successes were reported.

22
As Regards Automation

• If on-line lexicon bilingual English-L2 lexicon
  has 1 sense for a given word form, it is
  reasonable to assume identity (big time savings
  for technical terms, real-world objects (frying
  pan), etc.). Presenting results in quickly
  inspectable format would help
• If there is a many-to-one correspondence, would
  need an interface to really exploit time savings
  otherwise, somewhat difficult to keep track of
  senses.

23
As Regards Content of The English and L2 Lexicons

• How the lexicons would best develop
  simultaneously depends in large part upon the
  capabilities of acquirers, their training, etc.
  Driving English acquisition from the L2 side is
  perfectly fine, since the sem-struc is language
  independent.
• Need to divide tasks according to their
  difficulty a relatively untrained informant
  could do simple nouns, whereas a trained
  informant is necessary for polysemous verbs,
  phrasals, etc.
• Time savings clearly depends upon organization of
  efforts getting bogged down in simultaneous
  development of multiple lexicons and the ontology
  is a real risk

24
For Comparison The SIMPLE Project

• Goal develop 10K-sense semantic lexicons for 12
  European Union languages (the earlier PAROLE
  project developed 20K-sense morphological and
  syntactic ones)
• Each lexicon is built separately the word list
  for each L is based on corpus evidence for that L
  
• Each L must cover a given inventory of high-level
  concepts in EuroWordNet to ensure some overlap
• Each L uses the same inventory of template
  types, which indicates which types of properties
  should be described for different types of words

25
SIMPLE Aims

• Apparently, translation is the main aim, since
  semantic description is shallow (wouldnt support
  reasoning)
• Semantic description is limited to
• mapping to EuroWordNet concepts (which are
  iconic, not descriptive few properties are used)
• using a slightly expanded version of qualia,
  which are properties that support reasoning about
  generative properties of words (Pustejovsky, The
  Generative Lexicon)
• qualia represent just the generative corner of
  lexical description they do not provide breadth
  of description

26
Generalized vs. Application-Specific Resources

• A dichotomy at stake here is the one between
  generality of a LR lexical resource vs.
  usefulness for applications. In principle, only
  when we know the actual specific use we intend to
  do of a LR can we build the very best LR for
  that use, but this has proved to be too expensive
  and not realistic. In practice, however, there
  exists a large core of information that can be
  shared by many applicative uses, and this leads
  to the concept of generic LR, which is at the
  basis for the EAGLES initiative and of the
  PAROLE/SIMPLE projects, to be then enhanced and
  tuned with other means (Syntactic/semantic
  lexicons for the European languages towards a
  standardised infrastructure, Calzolari 1999 42)
• However, rebuilding resources for every
  application is not cost effective either, which
  brings us back to the approach were taking in
  OntoSem