Static and Dynamic Resource Management in Ontological Semantics

1
Static and Dynamic Resource Management in
Ontological Semantics

• Max Petrenko
• RiverGlass, Inc.

2
Presentation plan

• 1) OnSe architecture
• 2) Static resources management
• Ontology management
• Lexicon management
• 3) Dynamic resources management
• Semantic Text Analyzer (STAn)
• Modules
• Pre-processing options
• Processing options
• 4) STAn assessment

3
OnSe architecture

• Ontology 1873 concepts
• Lexicon 10500 entries
• Ontology browser/acquisition tool (OAT)
• Lexicon browser/acquisition tool (LAT)
• STAn 14 modules

4
OnSe architecture
SEMANTIC TEXT ANALYZER Lexicon reader Format
checker Parsing modules
LOAD
Ontology acquisition tool
Lexicon acquisition tool
SUBMIT
SUBMIT
Lexicon acquisition
Ontology acquisition
Sync
5
Static resource management

• Ontology
• - language-independent
• - culture-independent
• - parsimonious
• Lexicon
• - language-specific
• - carries syntactic and semantic information
• - guided by ontology

6
Ontology management

• Core features
• - hierarchically structured
• - hundreds of properties
• - case-roles
• - modality values
• - literal and relative attributes
• - relations between objects and events
• - triplet (property(facet(filler))) format
• - filler inheritance
• - inheritance blocking mechanism
• - inverse processing
• - metonymic relation HUMAN(has-social-rol
  e(SOCIAL-ROLE))

7
Ontology management

• Ontological facets
• ARREST
• agent default POLICE-OFFICER
• sem HUMAN
• beneficiary default CRIMINAL
• sem HUMAN
• DEFAULT indicates a more typical filler
  (important to TMR weight assignment)
• DEFAULT should be in the class of SEM

8
Ontology management
9
Ontology management
10
Ontology management
11
Ontology management

• Concept structure BOX
• (box
• (definition (value("a rectangular
  container")))
• (is-a (hier(container)))
• (made-of
• (sem(ceramic metal paper plastic
  wood))) (shape(value(rectangular square)))
• )

12
Ontology management
13
Ontology management

• Concept structure INJURY
• (injury
• (definition (value("to damage a body part")))
• (is-a(hier(damage physiological-event)))
• (subclasses
• (hier(suffocate contusion pull-ligament
  break-bone burn-injury bruise concussion
  electric-shock)))
• (theme(sem(animal-part)))
• (beneficiary(sem(animal))))

14
Ontology management
15
Ontology management

• Concept structure SHAPE
• (shape
• (definition (value("the shape of objects")))
• (is-a (hier(physical-object-attribute)))
• (domain (sem(physical-object)))
• (range (value(curved cylindrical egglike hooked
  linear oval rectangular round square
  triangular)))
• )

16
Ontology management

• shape(domain(value(physical-object)))
• BOX
• IS-A hier CONTAINER
• IS-A hier ARTIFACT
• IS-A hier PHYSICAL-OBJECT
• shape(domain(value(box))) true
• Question How about the shape of gold?

17
Ontology management
18
Ontology management

• SHAPE
• (domain(value(PHYSICAL-OBJECT )
• MATERIAL
• METAL
• METALLIC-ELEMENT
• GOLD
• (shape(domain(value(gold)))) true

IS-A
IS-A
IS-A
IS-A
19
Lexicon management

• Lexicon
• Core features
• - cannot transgress ontological restrictions
• - can only further restrict ontological fillers
• - carries syntactic and semantic info
• - covers closed classes (prepositions, pronouns,
  etc)
• - represents synonymy
• - covers derivatives
• - covers alternations

20
Lexicon management

• Lexical entry template
• (head-entry
• (sense-1, 2, 3
• (cat(n/v/adj/pro/prep))
• (synonyms "")
• (anno(def "")
• (comments "Acquired by ltacquirer namegt on
  ltdategt at lttimegt. ")
• (ex ""))
• (syn-struc((root(var0))(cat(n/v/adj/pro/prep)))
  )
• (sem-struc(CONCEPT))
• ))

21
Lexicon management
22
Lexicon management

• (box
• (box-n1
• (cat(n))
• (synonyms "")
• (anno(def "a rectangular container")
• (comments "Acquired by mpetrenko on 05-25-2009
  at 064959 CDT. ")
• (ex "the man put the book in the box"))
• (syn-struc((root(var0))(cat(n))))
• (sem-struc(box))
• )

23
Lexicon management

• (box-v1
• (anno(def "to put in a box")(ex "he boxed the
  glassware")(comments ""))(synonyms "")(cat(v))
• (syn-struc(
• (subject((root(var1))(cat(np))))(root(var0))(ca
  t(v)) (directobject((root(var2))(opt())(cat(np))
  ))))
• (sem-struc
• (put(end-location(sem(box)))
• (agent(value(var1(should-be-a(sem(human))))))
• (theme(value(var2(should-be-a(sem(artifact
  material plant))))))))
• )

24
Lexicon management

• Variables function
• syn-struc var1 var0 var2 varx
• sem-struc var1 concept var2 varx
• var0 stands for the ontological concept
• Syntactic variables
• specify position in the sentence
• specify syntactic category
• Semantic variables
• provide semantic information
• constrain ontological properties
• specify ontological fillers

25
Lexicon management

• Constraining ontological properties
• box-v1
• (sem-struc
• (put(end-location(sem(box)))
• (agent(value(var1(should-be-a(sem(human))))))
• (theme(value(var2(should-be-a(sem(artifact
  material plant))))))))

26
Lexicon management

• (agent(value(var1(should-be-a(sem(human))))))
• Ontology
• (agent
• (definition (value("the doer of an action")))
• (is-a (hier(case-role)))
• (inverse (value(agent-of)))
• (domain (sem(event)))
• (range (sem(animate group))))
• HUMAN ? HOMINIDAE ? ltgt ANIMAL ? ANIMATE
• var1 tells STAn to search for children of HUMAN

27
Lexicon management

• (theme
• (value(var2(should-be-a(sem(artifact material
  plant))))))))
• (theme
• (definition (value("the object of an
  event")))
• (is-a (hier(case-role)))
• (inverse (value(theme-of)))
• (domain (sem(event)))
• (range (sem(event object property))))
• ARTIFACT ? PHYSICAL-OBJECT
• MATERIAL ? PHYSICAL-OBJECT
• PLANT ? ANIMATE ? PHYSICAL-OBJECT

28
Lexicon management

• Why are the restrictions necessary?
• (box-v1
• (sem-struc
• (put(end-location(sem(box)))
• (agent(value(var1)))
• (theme(value(var2))))))
• Sentences now allowed
• The flyAg boxed the bananaTh
• The coralAg boxed the bridgeTh
• The horseAg boxed the starTh

29
Lexicon management

• PROBLEM acquiring verbs with Subject-Instrument
  alternation (Beth Levins index)
• (1) The man Agent broke the window
• (2) The hammer Instrument broke the window
• (3) The asteroid Instrument broke the window
• (4) The hurricane Precondition broke the window
  
• Solution expand the sem-struc to include more
  case roles mapping on the same variable

30
Lexicon management

• (break-v1
• (cat(v))(anno(def "to cause to break")
• (ex "")(comments "")(senseprim(1)))(synonyms "")
• (syn-struc ((subject((root(var1))(cat(np))))(roo
  t(var0))(cat(v))
• (directobject((root(var2))(cat(np))))))
• (sem-struc (damage(agent(value(var1)))
• (instrument(value(var1(should-be-a(sem(artifact
  animate-part material celestial-object)))
• (precondition(value(var1))))))
• (theme(value(var2(should-be-a(sem(artifact)))))
  ))))

31
Lexicon management

• Problem how to avoid long enumerations in the
  restrictions for variables
• break-v1
• (instrument
• (value(var1(should-be-a
• (sem (artifact
• animate-part
• material
• celestial-object)))
• same for damage-v1, destroy-v1, devastate-v1,
  etc.

32
Lexicon management
33
Lexicon management

• (physical-object
• (definition (value("objects that physically
  exist")))
• (is-a (hier(object)))
• (subclasses
• (hier (surface-feature
• landscape-object
• animate agent of break-v1
• animate-part
• animal-artifact instrument of break-v
• material
• artifact
• celestial-object))))

34
Lexicon management

• Why do we need this enumeration?
• Take a sentence, The man broke the window
• Remove the restriction from (instrument(value(va
  r1)))
• Resulting TMR is bad
• TMR 1Weight(TMR) 2.22 Event break-v1, damage1
• agent(value (man-n1, human1(gender(value(male)))))
  
• instrument(value(man-n1, human(gender(value(male
• )))))
• theme(value (window-n1, window1 ))

35
Lexicon management

• Alternative solution introduce the division of
  concepts INANIMATE vs. ANIMATE as children of
  PHYSICAL-OBJECT
• PHYSICAL-OBJECT
• subclasses value ANIMATE
• INANIMATE
• ANIMATE
• subclasses value ANIMAL FUNGUS
• PLANT VIRUS
• INANIMATE
• Subclasses value ANIMAL-ARTIFACT
• ARTIFACT
• MATERIAL
• CELESTIAL-OBJECT, etc.

36
Lexicon management

• Problematic entries closed classes
  (prepositions, parametric verbs, adjectives,
  pronouns, conjunctions, etc.)
• Prepositions specific format parsing module
• Parametric verbs specific format
• Adjectives specific format
• Pronouns specific format parsing module

37
Lexicon management

• Prepositions
• (with-prep1
• (anno(def "by the means of")(comments "")(ex "he
  shot the man with the gun"))(cat(prep))(synonyms
  "")
• (syn-struc((vp((root(var1))(cat(vp))))
  (root(var0))(cat(prep))(obj((root(var2))(cat(np)
  )))))
• (sem-struc(var1(instrument(value(var2(should-b
  e-a(sem(artifact animal-part material
  plant-part)))))))))

38
Lexicon acquisition

• Problem distinguishing strong and weak
  prepositions
• - The criminal killed the victim with a knife.
• Preposition is strong indicates instrumentality
• The FBI are opening an Investigation of into
  over regarding a fight.
• Prepositions are weak they are
  interchangeable, carry indirect meaning.
• How do we acquire them?

39
Lexicon management

• Solution acquire weak prepositions as part of
  lexical entries as a (pp-adjunct(root)(object)))
• (investigation-n1
• (cat(n))(anno(def "a legal process")
• (syn-struc((root(var0))(cat(n))
• (pp-adjunct((opt())(root(of into over
  regarding)) (cat(prep))(obj((root(var2))(cat(np))
  ))))
• (sem-struc(investigate
• (agent(value(var1))) (theme(value(var2))))))

40
Dynamic resource management

• Semantic Text Analyzer (STAn) modules
• Event identification module
• Property filler module Basic
• TMR building module processing stage
• Passives module
• Multiword Entry Resolution module
• Preposition module Extended
• Possessives resolution module processing stage
• Event embedment module
• Pronoun resolution module Advanced
• Unattested Personal Names module processing stage

41
Semantic Text Analyzer (STAn)

• STAn
• Input Natural Language text
• Output Text Meaning Representation (TMR)
• Basic functionality
• - performs parsing, event identification, and
  case-role filling within and across clauses
• - performs pronoun, possessives, multiword
  resolution
• - co-reference under development
• - multi-clausal parsing under development

42
Semantic Text Analyzer
43
Semantic Text Analyzer

• Module management
• Preposition processing module
• 1. A PP phrase is broken into Noun Prep Noun
• 2. Look for pp-adjuncts in the noun.
• 3. Look for preposition senses in the file.
• 4. Try to resolve the PP by filling the nouns
  based on the property

44
Semantic Text Analyzer

• Input The man closed the door of the car
• 1. Look for pp-adjunct in the noun car-n1? No.
• 2. Look for a sense of of-prep in the file?
  Found.
• of-prep1 (sem-struc(var1(origin(value(var2(s
  hould-be-a(sem(territory))))))))
• of-prep2 (sem-struc(var1(made-of(value(var2(s
  hould-be-a(sem(material))))))))
• of-prep3(sem-struc(var1(should-be-a(sem(physica
  l-object)))(part-of-object(value(var2(should-be-
  a(sem(physical-object)))))))) -- SELECTED

45
Semantic Text Analyzer
46
Resource management cycle

• For large corpora use gap detector
• For small corpora run STAn
• Detect the gap
• Acquire the entry
• Test the entry by STAn
• Locate inconsistencies (extra TMRs wrong
  case-roles)
• Verify ontological restrictions
• Research documentation
• Revise the entry
• Test the entry by STAn