Sweetening WordNet with DOLCE and OntoClean

1
Sweetening WordNet with DOLCE(and OntoClean)
Nicola GuarinoLaboratory for Applied
Ontology Institute for Cognitive Science and
Technologies (ISTC-CNR) National Research
Council Rome - Trento, Italy
2
Acknowledgements

• Stefano Borgo
• Bob Colomb
• Aldo Gangemi
• Claudio Masolo
• Alessandro Oltramari
• Luc Schneider
• Chris Welty

3
Summary

• Some problems of WordNetstaxonomy of nouns
• DOLCEa cognitively-biased upper level ontology
• Cleaning up WordNetusing DOLCE and OntoClean

4
Some problems when WordNet is used as an ontology
Unclear semantic interpretation of
hyperonimy Instantiation vs. subsumption Object-le
vel vs. meta-level Hyperonymy used to account for
polysemy (law both a document and a rule) Unclear
taxonomic structure - Glosses not consistent
with taxonomic structure Heterogeneous leves of
generality Formal constraints violations
(especially concerning roles) Polysemous use of
antonymy (child/parent vs. daughter/son) Poor
ontology of adjectives and qualities Shallow
taxonomy of verbs
5
Instantiation vs. subsumption

• Fall, Social_Event, _at_? Event
• Macao, Trust_Territory, _at_? Territorial_Dominio
  n
• Bach, Songwriter, _at_? Composer
• Red_Cross, Company, _at_? Organization
• Prime Minister, Appointment, _at_? Job
• Problem Expressivity lack
• Solution We need an instance-of relation
• (marking individuals doesnt work!)

6
Object-level vs. Meta-level

• The synset Abstraction (a general concept formed
  by extracting common features from specific
  examples) includes both abstract entities, such
  as Set, Time, and Space, and abstractions such as
  Attribute, Relation, Quantity.
• Problem confusion due to different domains.
• Solution Attribute, Relation, and Quantity are
  meta-level concepts, while Set, Time, and Space
  are object-level concepts.

7
Roles vs Types

• Formal constraint A Role cannot subsume a Type
• Person (that we consider as a type) is subsumed
  by two different concepts, Organism and
  Causal_Agent.
• Organism can be conceived as a type as well,
  while we have two options for Causal_Agent
• A role link to Person violates constraint
• A type hyponymys of Causal Agent must be
  understood as essentially causal agents. This may
  be not so intuitive in some cases (Catalyst).
• Solution Enforce OntoCleans constraints

8
Heterogeneous levels of generality

• Under Animal (subsumed by Life_Form) we find out
  specific concepts, such as Work_Animal,
  Domestic_Animal, Captive together with general
  classes such as Chordate, Fictional_Animal, etc.
• Under Phenomenon there are general partitioning
  concepts such as Process and Natural_Phenomenon
  together with a more specific concept,
  Consequence
• Solution distinguishing between types and roles
  helps to pinpoint this different generality

9
Developing foundational ontologies

• List the basic options
• Explore most relevant mutual dependencies
• Propose one preliminary upper level which is
  carefully justified and positioned with respect
  to the space of possible choices
• Add some minimal ontologies specifically relevant
  for selected domains
• Explore alternative upper levels

10
The WonderWeb Library of Foundational Ontologies

• No single upper level
• Rather, a (small) set of foundational ontologies
  carefully justified and positioned with respect
  to the space of possible choices
• Basic options clearly documented
• Clear branching points to allow for easy
  comparison of ontological options)

11
The WFO architecture
Choose Vision
4D
3D
Top
Choose Subject
Bank
Formal Links Between Visions Modules
Law
Single Module
Single Vision
12
DOLCE a Descriptive Ontology for Linguistic and
Cognitive Engineering

• A first reference module for the Foundational
  Ontology Library
• Strong cognitive bias influenced by
• Perception
• Culture
• Social conventions
• Rich axiomatization
• Categories as conceptual containers no deep
  metaphysical implications

13
DOLCEs basic taxonomy
Quality Physical Spatial location Temporal
Temporal location Abstract Abstract Quali
ty region Time region Space region Color
region
Endurant Physical Amount of matter Physical
object Feature Non-Physical Mental
object Social object Perdurant Static Stat
e Process Dynamic Achievement Accomplishmen
t
14
Abstract vs. Concrete Entities

• Concrete located in space-time (regions of
  space-time are located in themselves)
• Abstract - two meanings
• - Result of an abstraction process (something
  common to multiple exemplifications)
• ? Not located in space-time
• Mereological sums (of concrete entities) are
  concrete, the corresponding sets are abstract...

15
Endurance vs. Perdurance

• Endurants
• All proper parts are present whenever they are
  present (wholly presence, no temporal parts)
• Exist in time
• Can genuinely change in time
• May have non-essential parts
• Need a time-indexed parthood relation
• Perdurants
• Only some proper parts are present whenever they
  are present (partial presence,temporal parts )
• Happen in time
• Do not change in time
• All parts are essential
• Do not need a time-indexed parthood relation

16
Qualities and qualia

• Linguistic evidence
• This rose is red
• Red is a color
• This rose has a color
• The color of this rose turned to brown in one
  week
• The rooms temperature is increasing
• Red is opposite to green and close to brown
• Every entity comes with certain qualities that
  permanently inhere to it and are unique of it
• Qualities are perceptually mapped into qualia,
  which are regions of quality spaces.
• Properties hold because qualities have certain
  locations in their quality spaces.
• Each quality type has its own quality space

17
Qualities

• The rose and the chair have the same color
• different color qualities inhere to the two
  objects
• they are located in the same quality region
• Therefore,the same color attribute (red) is
  ascribed to the two objects

18
Qualities
Quality
Quality attribution
Quality space
Color-space
Red-obj
Rose
Has-part
Color
Red-region
q-location
Has-part
Color of rose1
Red421
Rose1
Inheres
Has-quale
19
Aggregate vs. Object
Both are enduring entities An object has a
unity criterion, while an aggregate does not.
20
Physical vs. Non-physical Object
FIAT SpA

• Physical objects
• inherent spatial localization
• not dependent on other objects (physical objects,
  like cars) or no inherent localization and be
  dependent on agents (non-physical objects, like
  laws and institutions).
• Non-physical objects can also be divided into
  mental (depending on singular agents) and social
  (depending on communities of agents).

21
Features
Features are parasitic entities, that exist
insofar their host exists. Features may be
relevant parts of their host, or places (which
are not parts of their hosts). All features are
essential wholes, but no common unity criterion
may exist for all of them (U).
22
Abstracts

• Abstracts are entities that have no inherent
  spatial or temporal localization. Examples of
  Abstract are propositions, sets, symbols,
  regions, etc.
• Quality regions and quality spaces are relevant
  examples of abstract entities

23
Figure Conventions

• Given an entity x to be characterised as D(x),
  its properties are written with the following
  compact syntax (in the attribute slot of the
  next UML class diagrams)
• C ?x D(x) ? C(x)
• NOT(C) ?x D(x) ? ?C(x)
• RC ?x D(x) ? ?y R(x,y) ? C(y)
• SOMEgtlt ?x D(x) ? ?(ngtnltn)(y) R(x,y) ?
  C(y)
• ALLRC ?x,y D(x) ? R(x,y) ? C(y)
• NOT(RC) ?x D(x) ? ??y R(x,y) ? C(y)

24
(No Transcript)
25
Perdurants
26
Qualities
27
Abstracts
28
Basic Relations

• Parthood
• Between quality regions (immediate)
• Between arbitrary objects (temporary)
• Dependence
• Specific/generic constant dependence
• Constitution
• Inherence (between a quality and its host)
• Quale
• Between a quality and its region (immediate, for
  unchanging ent)
• Between a quality and its region (temporary, for
  changing ent)
• Participation
• Representation

29
Part, Constitution, and Identity

• Structure may change identity

• Mereological extensionality is lost

• Constitution links the two entities
• Constitution is asymmetric (implies dependence)

Two blocks
30
The case of organizations

• Is my finger part of CNR?

31
Quality relations
32
Primitive relations and basic categories
33
Dependence relations
34
Participation relations

• Hold between a perdurant and its involved
  endurants
• Extremely relevant for domain modelling
• Current axiomatization covers
• constant vs. temporary
• complete vs. partial
• Further distinctions are currently primitive
  (thematic roles)
• Agent, Theme, Substrate, Instrument, Product
• More is needed on event structure,
  intentionality, and artifacts to produce analytic
  definitions

35
Representation relations

• Ongoing axiomatization (semiotics ontology)
• Extremely relevant for domain modelling
• Concepts
• PhysicalRepresentation vs. Expression
• Expression vs. Content
• Content vs. Reference
• Relations
• Realization, Interpretant, Reference, Description
• Non-trivial dependences between use and
  instantiation of expressions and contents

36
Axiomatizing basic relations

• Ground axioms (mainly algebraic)
• Links to other relations
• Dependence on time
• FO Modal Theory (S5Barcan)
• WonderWeb D17 v.2 for details

37
Alignining WordNet to DOLCE (1)

• Quality
• positionplace
• time_intervalinterval
• chromatic_color

38
Alignining WordNet to DOLCE (2)

• Quality Region
• Space_1
• time_1
• chromatic_color

39
Aligning WordNet to DOLCE (3)

• Aggregate
• Amount of matter
• body_substance
• chemical_element
• mixture
• compoundchemical_compound
• mass_5
• fluid_1
• Arbitrary collection

40
Aligning WordNet to DOLCE (4)

• Physical Object
• Physical_Body
• blackbodyfull_radiator
• body_5
• universeexistence
• Ordinary _Object
• collectionaggregation
• biological_group
• Body_Of_WaterWater
• LandDry_LandEarth
• artifactartefact
• Life_formorganismbeing

41
Aligning WordNet to DOLCE (5)

• Non-Physical Object
• Mental object
• Cognitionknowledge
• Structure
• statement_1
• Proposition
• Social object
• Possession_2
• Law
• Abstract
• symbol
• set_5

42
Aligning WordNet to DOLCE (6)

• Feature
• Relevant_Part
• edge_3
• skin_4
• paringparings
• Dependent_Region
• opening_3
• excavationhole_in_the_ground

43
Aligning WordNet to DOLCE (7)

• Occurrence
• State
• Non-relational
• Cognitve_state
• Emotional_state
• death
• Relational
• Relationship
• conflict

44
Aligning WordNet to DOLCE (8)

• Process
• Non-relational
• Shaping
• Increment
• decrement
• Relational
• Execution
• activity
• Accomplishmentachievement

45
Further problems

• Possession is a role, and it includes both roles
  and types
• Some hyponyms of Physical_Object are mapped to
  Feature
• Abstraction is the most heterogeneous Unique
  Beginner. It contains abstracts (Set), quality
  spaces (Chromatic_Color), qualities (mostly from
  the Attribute) and a hybrid concept (Relation)
  that contains abstracts, other entities, and even
  meta-level categories
• Psychological_Feature contains both mental
  objects (Cognition) and Events (emotional_state)

46
Some applications

• Ontology merging and building (e.g. fishery, bank
  norms)
• Catalogues creation/maintenance/integration (e.g.
  portals)
• DB design and requirement analysis
• Behaviour description and detection
• Quality/anomaly assessment from legacy DBs
  (money-laundering procedures)
• Quality/anomaly assessment of runtime operations
  (service level agreement)

47
Conclusions

• Our analysis of WordNet supplies
• A formal semantics for lexical relations
• A formal check of taxonomic links
• Rigorous bases to intuitive distinctions
• On-going work
• Further analysis of WordNet database (nouns,
  verbs, adjectives)
• Feedbacks from the applications side
• Report on DOLCE and the Foundational Ontologies
  Library downloadable from the WonderWeb site
  (Deliverable 17)
• Wonderweb.semanticweb.org

48
Developing foundational ontologies

• List the basic options
• Explore most relevant mutual dependencies
• Propose one preliminary upper level which is
  carefully justified and positioned with respect
  to the space of possible choices
• Add some minimal ontologies specifically relevant
  for the semantic web
• Explore alternative upper levels